High-flow vertical centrifugal pump water medium test method and system

Abstract

The invention relates to the technical field of centrifugal pumps, and discloses a high-flow vertical centrifugal pump water medium test method and system, and the method comprises the following steps: carrying out the preparation before water cut-off; setting a time window and a sampling frequency, and dividing the time window into a front section and a rear section; after water is cut off, the bearing temperature, the mechanical seal temperature and the motor winding temperature are collected according to the sampling frequency, and the water temperature, the environment temperature, the vibration intensity and noise data are synchronously recorded; fusing data obtained from the front section of a time window into an input feature based on window sliding; and inputting the input features into a pre-trained LSTM model to obtain a predicted temperature rise trend, and calculating a deviation between the actual temperature rise and the predicted temperature rise, if the deviation is greater than a preset threshold, indicating that the temperature rise trend is abnormal, and if the deviation is less than the preset threshold, indicating that the temperature rise trend is normal. The method has the advantages that historical data are fully utilized, the deep time sequence features are extracted, and the accuracy of temperature rise prediction is improved.

Description

Technical Field

[0001] This application relates to the technical field of centrifugal pumps, and in particular to a water medium test method and system for a large-flow vertical centrifugal pump. Background Art

[0002] In application scenarios such as the nuclear industry, petrochemical industry, ship power, and large water-cooling systems, large-flow vertical centrifugal pumps are one of the key equipment, undertaking the task of high-flow and long-term stable coolant transportation. Its core structure includes an impeller, a volute, a pump shaft, a mechanical seal, bearings, etc. The liquid is sucked in from the inlet through the high-speed rotation of the impeller and discharged tangentially under the action of centrifugal force, thus forming a fluid transportation with high flow and stable pressure. Compared with horizontal centrifugal pumps, the structural design of vertical centrifugal pumps is more conducive to reducing the floor area, improving the equipment compactness, and facilitating integration with large cooling systems. Especially in high-temperature and high-pressure or closed systems, such as the secondary loop cooling system of a nuclear reactor, the reliability of large-flow vertical centrifugal pumps directly affects the safety and operation stability of the entire system. Therefore, in practical applications, the cooling performance, seal reliability, and bearing durability of the pump become key evaluation indicators, and the water cut-off condition is an important means to conduct limit tests on these indicators.

[0003] The water cut-off test is an extreme condition test specifically for centrifugal pumps, aiming to simulate the sudden failure of the cooling system to evaluate the tolerance ability and temperature rise characteristics of the pump when the cooling water supply is interrupted. During normal operation, components such as the bearings and mechanical seals of large-flow vertical centrifugal pumps rely on cooling water for heat exchange to ensure that the temperature is maintained within a safe range. However, in the water cut-off state, the friction, shear, and high-speed rotation inside the pump will cause the temperature to rise sharply. If the heat dissipation is poor, it may lead to overheating of the bearings, failure of the lubrication system, deterioration of the seal material, and even damage to the pump body. Through the water cut-off test, the temperature rise rate of these components under extreme conditions can be quantified, and their failure modes can be analyzed, providing important references for optimizing the pump's structural design, improving the cooling system, and enhancing the overall reliability.

[0004] Existing water cutoff test methods still have many deficiencies in testing and data analysis, mainly reflected in the following aspects. First, traditional water cutoff tests usually adopt the method of sampling at fixed times, recording data through temperature sensors, vibration sensors, and noise measurement devices. However, the analysis of the temperature rise trend mainly relies on manual experience and it is difficult to predict the temperature change trend in real time. Since the bearing temperature rise is affected by multiple factors, such as pump speed, ambient temperature, lubrication status, etc., the manual analysis method is prone to misjudgment or delayed response, and it is difficult to accurately predict when the pump reaches the dangerous temperature. Second, the existing tests lack a systematic study on the influence of microbubbles. In the pump cavity, the presence of microbubbles may cause local lubricating oil film rupture, thereby increasing friction and resulting in abnormal temperature rise. However, most of the current test methods only focus on the overall temperature rise change and fail to analyze the influence mechanism of microbubbles by combining high-frequency noise signals, making it difficult to accurately judge the cause of abnormalities under complex working conditions. Summary of the Invention

[0005] In order to make full use of historical data, extract deep time series features, and improve the accuracy of temperature rise prediction, this application provides a water medium test method and system for large-flow vertical centrifugal pumps.

[0006] In the first aspect, this application provides a water medium test method for large-flow vertical centrifugal pumps, adopting the following technical solutions:

[0007] A water medium test method for large-flow vertical centrifugal pumps includes the following steps:

[0008] S1. Prepare before water cutoff for the large-flow vertical centrifugal pump;

[0009] S2. Set the time window and sampling frequency, and divide the time window into two segments, where the first segment of the time window is used to obtain data to calculate the predicted temperature rise trend, and the second segment of the time window is used to obtain data to get the actual temperature rise trend corresponding to the first segment;

[0010] S3. After water cutoff, collect the bearing temperature, mechanical seal temperature, and motor winding temperature at the sampling frequency, and synchronously record the water temperature, ambient temperature, vibration intensity, and noise data;

[0011] S4. Based on window sliding, fuse the bearing temperature, mechanical seal temperature, motor winding temperature, water temperature, ambient temperature, vibration intensity, and noise data obtained in the first segment of a time window as input features;

[0012] S5. Input the input features into the pre-trained LSTM model to obtain the predicted temperature rise trend, and calculate the deviation between the actual temperature rise and the predicted temperature rise. If the deviation is greater than the preset threshold, it indicates that the temperature rise trend is abnormal; if the deviation is less than the preset threshold, it indicates that the temperature rise trend is normal.

[0013] Optionally, S1 includes the following steps:

[0014] S101. Keep the water temperature of the test loop at 85 ± 10 °C;

[0015] S102. Keep the liquid level in the pump chamber 990 ± 200 mm below the installation plane of the large flange of the outer casing;

[0016] S103. Keep the air pressure covering the pump chamber at 0.33 ± 0.05 MPa;

[0017] S104. Adjust the frequency converter to 20 Hz and start the pump;

[0018] S105. Adjust the outlet valve until the main circuit flowmeter reaches 4888 ± 200 m³ / h, and then stop adjusting the valve;

[0019] S106. After the pump unit runs stably, gradually adjust the frequency converter to 50 Hz to make the pump reach the rated flow of 12220 ± 367 m³ / h;

[0020] S107. After stable operation, close the main valve of the cooling water system.

[0021] Optionally, S106 includes the following steps:

[0022] After the pump unit runs stably, successively adjust the frequency converter to 25 Hz, 30 Hz, 35 Hz, 40 Hz, 45 Hz, and 50 Hz to increase the speed;

[0023] After the pump unit runs stably, observe whether the main circuit flow is 12220 ± 367 m³ / h. If it exceeds, slightly adjust the outlet valve to make the main circuit flow reach the range of 12220 ± 367 m³ / h.

[0024] Optionally, S1 includes the following steps:

[0025] S111. Keep the water temperature of the test loop at 85 ± 10 °C;

[0026] S112. Keep the liquid level in the pump chamber 990 ± 200 mm below the installation plane of the large flange of the outer casing;

[0027] S113. Keep the air pressure covering the pump chamber at 0.33 ± 0.05 MPa;

[0028] S114. Adjust the frequency converter to 20 Hz and start the pump;

[0029] S115. Adjust the outlet valve until the main circuit flowmeter reaches 4888 ± 200 m³ / h, and then stop adjusting the valve;

[0030] After the pump unit operates stably, gradually adjust the frequency of the frequency converter to 12.5 Hz so that the pump reaches the rated flow rate of 3055 ± 92 m³ / h;

[0031] After stable operation, close the main valve of the cooling water system.

[0032] Optionally, S116 includes the following steps:

[0033] After the pump unit operates stably, adjust the frequency of the frequency converter to 12.5 Hz;

[0034] After the pump unit operates stably, observe whether the flow rate of the main circuit is 3055 ± 92 m³ / h. If it exceeds, slightly adjust the outlet valve to make the flow rate of the main circuit reach the range of 3055 ± 92 m³ / h.

[0035] Optionally, the pre-training steps of the LSTM model include:

[0036] S501. Perform data collection and data preprocessing;

[0037] S502. Set 70 seconds as the time window, and generate a training set, a validation set, and a test set; among them, the training set uses the data of the previous 60 seconds as the input and the temperature rise trend of the next 10 seconds as the output;

[0038] S503. Construct an LSTM network;

[0039] S504. Input the training set into the LSTM network and calculate the loss function;

[0040] S505. Calculate the gradient of the loss with respect to the weights of the LSTM layer, and use the Adam optimizer to perform gradient descent to update the weights of the LSTM layer;

[0041] S506. After the model training is completed, use the test set and calculate the test set error, and adjust the model complexity based on the test set error.

[0042] Optionally, the structure of the LSTM network in S503 includes:

[0043] An input layer for receiving input features corresponding to 60 time steps;

[0044] The first layer of LSTM, containing 64 units, is used to extract time series features and output the information of all time steps;

[0045] The second layer of LSTM, containing 32 units, is used to extract deep time series information and only output the information of the last time step;

[0046] A fully connected layer for mapping the features extracted by the LSTM to a numerical value to predict the temperature rise trend in the next 10 seconds.

[0047] Optionally, S501 includes the following steps:

[0048] S5011. Obtain training data, where the training data comes from normal water cut-off experiment data without microbubbles, abnormal water cut-off experiment data with microbubbles, and test data with partially artificially introduced microbubbles under different environmental temperatures and different test conditions;

[0049] S5012. Detect outliers and remove the outlier points;

[0050] S5013. Interpolate and complete the missing values;

[0051] S5014. Normalize the data.

[0052] In a second aspect, the present application provides a water medium test system for a large-flow vertical centrifugal pump, adopting the following technical solution:

[0053] A water medium test method for a large-flow vertical centrifugal pump includes a processor, and a program of the water medium test method for a large-flow vertical centrifugal pump described in any one of the above is run in the processor.

[0054] In a third aspect, the present application provides a storage medium, adopting the following technical solution:

[0055] A storage medium stores a program of the water medium test method for a large-flow vertical centrifugal pump described in any one of the above.

[0056] In summary, the present application includes at least one of the following beneficial technical effects: This solution proposes an intelligent diagnosis method based on deep learning (LSTM). By constructing a time series prediction model, it can accurately predict the temperature rise trend, and combine with an anomaly detection algorithm to determine whether the temperature rise exceeds the safety threshold. The advantage of this method is to make full use of historical data, extract deep time series features, improve the accuracy of temperature rise prediction, and at the same time analyze the influence of microbubbles on temperature rise by combining high-frequency noise, greatly improving the intelligent level of the water cut-off test. Description of the Drawings

[0057] Figure 1 is a flowchart of a program of a water medium test method for a large-flow vertical centrifugal pump in an embodiment of the present application. [[ID=I37]]Detailed Embodiments

[0058] The following details the embodiments of the present application, and the examples of the embodiments are shown in the drawings.

[0059] In the description of this specification, the description with reference to the terms "certain embodiments", "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0060] The embodiments of the present application disclose a test method for water medium of a large-flow vertical centrifugal pump, with reference to Figure 1 , including the following steps S1 - S5.

[0061] S1. Prepare before water cut-off of the large-flow vertical centrifugal pump.

[0062] A large-flow vertical centrifugal pump is a type of pump specifically used for high-flow liquid transportation. Its impeller rotates in the vertical direction, relying on centrifugal force to lift the liquid to a high-pressure state, and is widely used in the nuclear industry, petrochemical industry, and large hydraulic systems. In these applications, the stability and reliability of the pump are crucial. Especially in the case of interruption of the cooling water supply, whether the pump can withstand short-term extreme working conditions is a key factor affecting the safety of the system. For this reason, a water cut-off test is usually carried out, that is, the cooling system is artificially shut down, and the temperature rise trend, vibration intensity, and noise change of the pump are observed under different working conditions to verify its design rationality and operation reliability.

[0063] The purpose of the water cut-off test is to simulate possible cooling failure situations and evaluate the tolerance of the pump. If this test is not carried out, it may lead to the inability to predict the ultimate bearing capacity of the equipment when sudden cooling failure occurs during actual use, resulting in serious problems such as overheating of bearings, failure of mechanical seals, and even damage to the pump body. In addition, cooling failure is often sudden. If the reaction characteristics of the pump are not mastered in advance, a reasonable coping strategy cannot be formulated, which may lead to uncontrollable failures of the entire system.

[0064] Generally, the results of the water cut-off test mainly include the temperature rise of the pump, vibration changes, and noise level. Among them, the temperature rise is affected by various factors, including the initial water temperature, ambient temperature, lubrication state of the bearings, and the sealing of the test circuit. The influence of microbubbles is particularly significant. The bubbles remaining in the test circuit will reduce the heat transfer efficiency of the liquid, causing the bearing temperature rise rate to be low at the beginning and then rise steeply. In addition, when the microbubbles burst, local pressure fluctuations will occur, resulting in a short-term increase in vibration and the generation of short-term high-frequency noise. If the influence of microbubbles is not fully considered, the test data may show abnormal deviations, leading to misjudgment of the equipment health status. Therefore, thorough exhaust treatment is required before the test, and intelligent data analysis methods are combined to detect the influence of microbubbles on the measurement data.

[0065] The implementation process of this test starts from the preparations before water cut-off, that is, ensuring the standardization of test conditions, ensuring that the water temperature, pump chamber liquid level, and air pressure are within the set range to eliminate the influence of environmental variables. Subsequently, start the pump and gradually adjust the rotational speed to reach the rated flow rate, and after stable operation, turn off the cooling system to officially enter the water cut-off test stage.

[0066] Optionally, in an embodiment, S1 includes the following steps S101 - S107.

[0067] S101. Keep the water temperature in the test circuit at 85 ± 10 °C.

[0068] Before the test starts, the water temperature in the test circuit needs to be kept at 85 ± 10 °C, and this setting is based on the physical properties of water. The temperature of water affects its viscosity and heat transfer coefficient, and thus affects the flow resistance and heat dissipation capacity of the pump. If the temperature is too low, the viscosity of water increases, which may lead to insufficient bearing lubrication and increased frictional losses; if the temperature is too high, the heat dissipation capacity of water decreases, which may cause the pump temperature rise rate to be too high. Therefore, maintaining an appropriate water temperature can ensure that the heat conduction characteristics of the pump meet the engineering design expectations and reduce the error of abnormal temperature rise. In addition, the control of water temperature involves a heat exchange system, and usually a PID controller is used to adjust the heating element to keep the water temperature within the target range:

[0069]

[0070] where e is the deviation between the current water temperature and the target water temperature, are PID control parameters to ensure stable operation within the allowable error range of temperature.

[0071] S102. Keep the pump chamber liquid level 990 ± 200 mm below the installation plane of the large flange of the outer housing.

[0072] The setting ensures that the liquid in the pump is under sufficient static pressure to prevent cavitation problems caused by too low liquid level. Cavitation refers to the generation of bubbles in the low-pressure area of the liquid. When the bubbles burst in the high-pressure area, they release huge energy, which may cause damage to the impeller surface and even affect the vibration and efficiency of the pump. Too high liquid level may lead to abnormal inlet pressure of the pump and affect the hydrodynamic characteristics. Therefore, it is necessary to install a liquid level sensor to monitor the liquid level height in real time and combine it with a solenoid valve to control the water level replenishment system to ensure that the liquid level is within the set range and reduce the test errors caused by liquid level fluctuations.

[0073] S103. The air pressure in the pump chamber is maintained at 0.33 ± 0.05 Mpa.

[0074] In addition to liquid level control, the air pressure in the pump chamber is set at 0.33 ± 0.05 MPa. This is to ensure that the internal environment of the pump operates under controlled conditions and prevent external gas from entering the pump chamber and affecting the fluid characteristics. Usually, a pressure sensor is used to detect the air pressure in the pump chamber, and a hermetic system is used for dynamic adjustment to ensure that the air pressure is maintained within the set range. Appropriate pump chamber pressure can optimize the sealing performance of the pump, reduce leakage, and improve operation stability. For example, in a high-pressure environment, the mechanical seal may bear greater pressure, which will affect its friction performance and increase the risk of seal failure.

[0075] S104. Adjust the frequency converter to 20 Hz and start the pump.

[0076] After setting the parameters of the test environment, it is necessary to start the pump and adjust its operating state. First, adjust the frequency converter to 20 Hz and start the pump. This is to make the pump gradually enter a stable operating state at a lower speed and reduce the mechanical impact caused by starting with instantaneous high load. Under the control of the frequency converter, the acceleration process of the pump can be smoothly transitioned, thus avoiding affecting the measurement data due to the transient changes during startup.

[0077] S105. Adjust the outlet valve until the main circuit flowmeter reaches 4888 ± 200 m³ / h, and then stop adjusting the valve.

[0078] Gradually adjust the outlet valve to make the main circuit flow reach 4888 ± 200 m³ / h. The purpose of this step is to establish a controllable liquid flow state so that the pump can operate according to the set flow rate. The adjustment process of the valve is usually completed by an automatic flow control system, in which a flow sensor monitors the flow rate in real time, and the system dynamically adjusts the valve opening according to the feedback signal to ensure that the flow rate is stable within the set range.

[0079] S106. After the pump unit operates stably, gradually adjust the frequency converter to 50 Hz to make the pump reach the rated flow rate of 12220 ± 367 m³ / h.

[0080] Specifically, in this embodiment, S106 includes the following steps a - b.

[0081] a. After the pump group runs stably, adjust the inverter frequency to 25Hz, 30Hz, 35Hz, 40Hz, 45Hz, and 50Hz in sequence to increase the speed.

[0082] b. After the pump unit runs stably, observe whether the main circuit flow is 12220±367m³ / h. If it exceeds, fine-tune the outlet valve to make the main circuit flow within the range of 12220±367m³ / h.

[0083] After completing low-speed operation and flow adjustment, the pump unit needs to enter rated flow operation. This involves gradually adjusting the inverter frequency to 50Hz to achieve a rated flow of 12,220 ± 367 m³ / h. This process isn't completed all at once, but rather proceeds incrementally to minimize flow field disturbances caused by sudden speed changes. During frequency adjustment, the frequency can be increased in the order of 25Hz → 30Hz → 35Hz → 40Hz → 45Hz → 50Hz. After each increase, observe the pump flow stability to ensure the pump remains stable throughout each transition.

[0084] S107. After stable operation, close the main valve of the cooling water system.

[0085] After the pump stabilizes, the cooling water system's main valve is closed to artificially create an extreme cooling failure condition. This allows the pump to assess its ability to withstand the absence of external cooling and its temperature rise trend. This step officially begins the water shut-off test. All test parameters (bearing temperature, mechanical seal temperature, motor winding temperature, water temperature, ambient temperature, vibration intensity, noise, etc.) will now enter a free-range phase. The pump's operating state will rely entirely on its own heat transfer and dissipation capabilities, no longer relying on the external cooling medium.

[0086] Optionally, in another embodiment, S1 includes the following steps S111-S117.

[0087] S111. Maintain the test loop water temperature at 85±10℃.

[0088] S112. The liquid level in the pump chamber should be maintained at 990±200mm below the mounting plane of the large flange of the outer casing.

[0089] S113. The pump chamber covering air pressure is maintained at 0.33±0.05Mpa.

[0090] S114. Adjust the inverter frequency to 20Hz and start the pump.

[0091] S115. Adjust the outlet valve until the main circuit flow meter reaches 4888±200m³ / h, and stop adjusting the valve.

[0092] After the pump unit runs stably, gradually adjust the frequency of the frequency converter to 12.5 Hz so that the pump reaches the rated flow rate of 3055 ± 92 m³ / h.

[0093] Specifically, in this embodiment, S116 includes the following steps a - b.

[0094] a. After the pump unit runs stably, adjust the frequency of the frequency converter to 12.5 Hz.

[0095] b. After the pump unit runs stably, observe whether the main circuit flow rate is 3055 ± 92 m³ / h. If it exceeds, slightly adjust the outlet valve to make the main circuit flow rate reach the range of 3055 ± 92 m³ / h.

[0096] After the pump unit runs stably, gradually adjust the frequency of the frequency converter to 12.5 Hz so that the pump reaches the rated flow rate of 3055 ± 92 m³ / h, ensuring that the pump runs stably at low speed and providing a reliable test environment for the long - term water cut - off test (72 hours).

[0097] S117. After stable operation, close the main valve of the cooling water system.

[0098] After the pump runs stably, close the main valve of the cooling water system to artificially create an extreme working condition of cooling failure, so as to evaluate the tolerance ability of the pump when lacking external cooling and its temperature rise trend. This step marks the official start of the water cut - off test, and all test parameters (bearing temperature, mechanical seal temperature, motor winding temperature, water temperature, ambient temperature, vibration intensity, noise, etc.) will enter the free - change stage at this moment. The working state of the pump will completely depend on its own heat transfer and heat dissipation capabilities, rather than relying on the action of external cooling media.

[0099] S2. Set the time window and sampling frequency, and divide the time window into two segments. Among them, the first segment of the time window is used to obtain data to calculate the predicted temperature rise trend, and the second segment of the time window is used to obtain data to get the actual temperature rise trend corresponding to the first segment.

[0100] In the water cut - off test, the temperature change of the pump is a time - series problem, that is, the current temperature not only depends on the instantaneous state of the pump, but also is affected by factors such as the temperature, vibration, and ambient temperature in the past period. Therefore, it is necessary to construct samples through a sliding time window so that the model can predict the future temperature rise trend based on historical data. After setting the time window W, the data will be split into two parts:

[0101]

[0102] Among them: (The first segment of the time window): Represents the data of the past 60 seconds and is used as the model input.

[0103] (The latter part of the time window): Represents the temperature rise trend in the next 10 seconds, which is used as the true value for calculating the prediction error.

[0104] The selection of the time window needs to balance the sufficiency of historical information and the timeliness of prediction. If the window is too short, the model cannot capture the long-term trend; if the window is too long, the computational complexity increases, and the contribution of remote historical information may decay. Usually, the window size is optimized through experiments, such as predicting the 10-second future trend with 60 seconds of historical data.

[0105] In this solution, the acquisition frequency of data such as temperature rise and vibration is set to 1Hz, that is, recorded once per second, ensuring that the data is dense enough to capture the subtle changes in temperature rise:

[0106]

[0107] Among them, each data point contains multiple measured values, such as bearing temperature, mechanical seal temperature, motor winding temperature, vibration, noise, ambient temperature, etc. To reduce noise interference, the data needs to be preprocessed before entering the time window, such as moving average filtering.

[0108] Optionally, in this solution, a dynamic window mechanism can also be adopted. This is because within one minute before the water cut-off starts, the temperature rise trend is usually relatively gentle, while the temperature rise rate rises sharply within 10 - 20 seconds after the water cut-off. Therefore, .

[0109] S3. After the water cut-off, collect the bearing temperature, mechanical seal temperature, and motor winding temperature at the sampling frequency, and synchronously record the water temperature, ambient temperature, vibration intensity, and noise data.

[0110] After the water cut-off occurs, the temperature change of the pump is no longer controlled stably by the cooling water, but is completely determined by internal heat conduction, lubrication state, and surrounding environmental conditions. To capture this dynamic change, data must be sampled at a high frequency (usually 1Hz) to make the data points dense enough to ensure that no subtle temperature rise fluctuations are missed. The data acquisition system usually uses high-precision temperature sensors, such as thermocouples or platinum resistance sensors RTD, to transmit the measured data to the central control system in real time.

[0111] For the acquisition of vibration data, accelerometers are usually used, and the output signal of which can analyze the spectral characteristics through Fourier transform to detect whether there are abnormal high-frequency components.

[0112] The recording of noise data is highly relevant for detecting microbubbles and lubrication anomalies. For example, after water interruption, if there are microbubbles or lubricant deterioration, the noise level will briefly increase. If a high-frequency noise peak appears within 1 - 2 minutes after water interruption, it may be a precursor to cavitation or bearing lubrication failure.

[0113] In addition, the data of all sensors must be sampled and stored strictly according to the same time step. If the data is not synchronized, it may lead to misalignment during the extraction of time series features, affecting the learning effect of the LSTM model. Therefore, a timestamp synchronization mechanism is usually adopted in the data acquisition system.

[0114] S4. Based on window sliding, fuse the bearing temperature, mechanical seal temperature, motor winding temperature, water temperature, ambient temperature, vibration intensity, and noise data obtained from the first part of a time window as input features.

[0115] In the water interruption test, the temperature rise of the pump is a dynamic evolution process, which is not only affected by the current state but also depends on the temperature change trend, environmental conditions, and mechanical operating conditions in the past period. Therefore, it is not enough to analyze the data at a certain moment alone. It is necessary to construct a time series input through a sliding time window so that the model can learn the long-term dependence relationship of the temperature rise. The setting method of the time window is as follows:

[0116]

[0117] Among them, each time window contains the data of the past 60 seconds, which is used as input features to predict the temperature rise trend in the next 10 seconds.

[0118] The key to feature fusion lies in data synchronization and feature transformation, ensuring that all variables are aligned in the time dimension and at the same time eliminating the interference of noise data on the model. During the implementation process, data standardization is required first to avoid the influence of the order of magnitude differences of different physical quantities on model training. Min-Max normalization is used to map all features to the [0,1] interval:

[0119]

[0120] where X is the original data, are the minimum and maximum values of the data set respectively. This can ensure that features such as temperature, vibration, and noise are on the same numerical scale, avoiding the dominant influence of certain features on model training.

[0121] In addition, in order to improve the robustness of the prediction model, statistical features need to be constructed, such as calculating the temperature mean, variance, maximum and minimum values in the past 60 seconds, etc., to enhance the feature expression ability of the time series:

[0122]

[0123] The ultimate goal of data fusion is to construct a structured input matrix for LSTM training and prediction:

[0124]

[0125] Among them, each row represents the measurement data at a certain moment, and each column represents a certain feature variable. This input format meets the requirements of LSTM, enabling it to process time series data and learn the dependencies across time steps.

[0126] S5. Input the input features into the pre-trained LSTM model to obtain the predicted temperature rise trend, and calculate the deviation between the actual temperature rise and the predicted temperature rise. If the deviation is greater than the preset threshold, it indicates that the temperature rise trend is abnormal; if the deviation is less than the preset threshold, it indicates that the temperature rise trend is normal.

[0127] This step inputs the fused input features into the pre-trained LSTM model to obtain the predicted temperature rise trend, and calculates the deviation between the actual temperature rise and the predicted temperature rise to determine whether the temperature rise trend is abnormal. The data matrix generated in the previous step is input into the input layer (Input Layer) of the LSTM model. After two layers of LSTM calculations, the fully connected layer (DenseLayer) outputs the predicted temperature rise values for the next 10 seconds:

[0128]

[0129] The predicted temperature rise trend is compared with the actual temperature rise data to calculate the error:

[0130]

[0131] The mean squared error is used for error calculation to measure the deviation between the predicted value and the true value:

[0132]

[0133] If exceeds the preset threshold, it is considered that the temperature rise trend is abnormal, which may mean lubrication failure of the pump, bearing overheating or microbubble interference. For example, under normal circumstances, the temperature rise deviation should be less than 3°C. If , an alarm may need to be issued.

[0134] Specifically, in a certain embodiment, the pre-training steps of the LSTM model include S501 - S506.

[0135] S501. Perform data collection and data preprocessing.

[0136] Specifically, in a certain embodiment, S501 includes the following steps S5011 - S5014.

[0137] S5011. Obtain training data, where the training data comes from normal water cut-off experiment data without microbubbles, abnormal water cut-off experiment data with microbubbles, and some test data with artificially introduced microbubbles under different environmental temperatures and different test conditions.

[0138] S5011 is responsible for obtaining training data, which comes from water cut-off tests under multiple different working conditions, including normal water cut-off experiment data (without microbubbles), abnormal water cut-off experiment data (with microbubbles), and some test data with artificially introduced microbubbles. The purpose of these data is to cover the operating states of the pump under various possible water cut-off conditions, ensuring that the LSTM model can learn the temperature rise change patterns in different situations. The construction of the training data depends on the sensor data acquisition system. The variables collected include bearing temperature, mechanical seal temperature, motor winding temperature, water temperature, environmental temperature, vibration intensity, and noise data, and are recorded at a fixed frequency (such as 1Hz) to form a continuous time series. For the acquisition of microbubble data, a high-sensitivity noise sensor is used because microbubbles may burst at the initial stage of water cut-off, generating high-frequency noise, and the timing characteristics of these noises can reflect the impact of microbubbles on the pump operating state. After data acquisition, all data will be aligned according to the time stamp to ensure the consistency of input features.

[0139] S5012. Detect outliers and remove the outlier points.

[0140] S5012 is responsible for detecting outliers and removing the outlier points to ensure the stability of the training data. Outliers may come from sensor failures, environmental interference, or abnormal fluctuations during the test process.

[0141] S5013. Interpolate to fill in the missing values.

[0142] During the sensor data acquisition process, due to signal interference or equipment failures, some data may be lost. Common methods for filling in missing values include linear interpolation and time series interpolation.

[0143] S5014. Normalize the data.

[0144] This step keeps the numerical ranges of all variables consistent, preventing the scale differences of different physical quantities from affecting model training. Common methods for normalization include Min-Max normalization and Z-score standardization.

[0145] S502. Set 70 seconds as the time window, and generate a training set, a validation set, and a test set; where the training set uses the data of the previous 60 seconds as input and the temperature rise trend of the next 10 seconds as output.

[0146] The division of the dataset follows the standard process of training set, validation set, and test set, usually divided in a ratio of 8:1:1:

[0147] Training set (80%): Used for LSTM training to enable the model to learn the basic patterns of temperature rise trends.

[0148] Validation set (10%): Used during training to monitor the generalization ability of the model and prevent overfitting.

[0149] Test set (10%): Used for the final evaluation of the model's prediction ability to ensure it can still maintain high accuracy on unseen data.

[0150] The organization of training data usually adopts the sliding window method, enabling new training samples to be generated at each time step.

[0151] S503. Construct an LSTM network.

[0152] Specifically, in a certain embodiment, the structure of the LSTM network in S503 includes: an input layer, a first-layer LSTM, a second-layer LSTM, and a fully connected layer.

[0153] The input layer is used to receive the input features corresponding to sixty time steps; the first-layer LSTM contains sixty-four units, which is used to extract time series features and output the information of all time steps; the second-layer LSTM contains thirty-two units, which is used to extract deep time series information and only output the information of the last time step; the fully connected layer is used to map the features extracted by the LSTM to a numerical value to predict the temperature rise trend in the next ten seconds.

[0154] The input layer is responsible for receiving the sensor data of the past 60 seconds, and the data format is:

[0155]

[0156] Among them, each row represents the sensor data of a certain time step, and each column represents a feature variable, such as bearing temperature, mechanical seal temperature, motor winding temperature, etc. The role of this input layer is to format the data into a time series structure that can be processed by the LSTM, enabling the model to establish data associations between multiple time steps.

[0157] The input data enters the first-layer LSTM, which contains 64 LSTM units and can extract complex time series features in the time dimension. The LSTM controls the information flow through memory gates and gating mechanisms, overcoming the problem of gradient disappearance that traditional recurrent neural networks are prone to when dealing with long time series. The first-layer LSTM uses return_sequences=True, which means it will output a 64-dimensional feature vector at each time step, forming a complete time series feature matrix:

[0158]

[0159] Among them, represents the eigenvalue of the i-th dimension calculated by the LSTM at time step t. The role of this layer is to establish patterns within the time series, such as periodic changes in temperature rise, mutation points, etc.

[0160] Subsequently, the data enters the second layer of LSTM, which contains 32 LSTM units and is used to further extract deep features of the time series. Different from the first layer of LSTM, the return_sequences of this layer is False, which means it only outputs the information of the last time step, that is:

[0161]

[0162] This indicates that the second layer of LSTM has highly compressed the entire time series, only retaining the most representative eigenvalue as the final output. The main role of this layer is to further screen and refine the time series features with the most predictive value.

[0163] Finally, the time series features extracted by the LSTM enter the fully connected layer, and the role of this layer is to map the 32-dimensional features calculated by the LSTM to the final predicted value of the temperature rise:

[0164]

[0165] The number of neurons in the fully connected layer determines the shape of the predicted output. For this task, only one value needs to be output finally, that is, the predicted result of the temperature rise trend in the next 10 seconds. Therefore, the number of neurons in the fully connected layer is 1. The calculation method of this layer can be expressed as:

[0166]

[0167] Among them, W is the weight matrix of the fully connected layer, b is the bias term, and the finally obtained is the predicted temperature rise trend.

[0168] S504. Input the training set into the LSTM network and calculate the loss function.

[0169] The preprocessed time series data is input into the LSTM model. Each sample contains the sensor data of the past 60 seconds as the input and the temperature rise trend in the next 10 seconds as the target output. The forward propagation process of the LSTM involves a series of matrix operations, and the calculation of each time step includes input gate, forget gate, output gate, and cell state update.

[0170] In this step, the loss function is calculated to measure the difference between the LSTM predicted value and the true value, and based on this error, the weight parameters of the LSTM layer are adjusted to minimize the prediction error. In this embodiment, the mean square error is used as the loss function.

[0171] S505. Calculate the gradient of the loss with respect to the LSTM layer weights, and use the Adam optimizer for gradient descent to update the LSTM layer weights.

[0172] After the loss is calculated, LSTM uses backpropagation to calculate the gradient of the loss with respect to the network parameters, and updates the weights through the Adam optimizer.

[0173] S506. After the model training is completed, use the test set and calculate the test set error, and adjust the model complexity based on the test set error.

[0174] This step is responsible for model testing and optimization, that is, using unseen test data to evaluate the prediction performance of LSTM, and adjusting the model complexity according to the test error to ensure that it can maintain high prediction accuracy under different water cut-off test conditions. During the test, calculate the MSE on the test set and compare it with the training error:

[0175]

[0176] If the test error is much larger than the training error, it indicates that the model may be overfitting, and the LSTM structure needs to be adjusted, such as reducing the number of LSTM layers, increasing regularization, or adopting an early stopping strategy.

[0177] After the test is completed, the final model is saved and used for real-time prediction of the temperature rise trend in the water cut-off test. In practical applications, the model receives new sensor data and generates temperature rise prediction values for the next 10 seconds. This mechanism enables the LSTM model to detect the abnormal temperature rise trend in real time during the water cut-off test and issue an alarm in advance to avoid equipment damage.

[0178] The embodiment of the present application also discloses a large-flow vertical centrifugal pump water medium test system, including a processor, and a program of the large-flow vertical centrifugal pump water medium test method described in any one of the above is run in the processor.

[0179] The embodiment of the present application also discloses a storage medium, storing a program of the large-flow vertical centrifugal pump water medium test method described in any one of the above.

[0180] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A test method for water medium of a large-flow vertical centrifugal pump, characterized in that It includes the following steps: S1. Prepare before the water cut-off of the high-flow vertical centrifugal pump; S2. Set the time window and sampling frequency, and divide the time window into two segments, the front segment of the time window is used to obtain data to calculate the predicted temperature rise trend, and the rear segment of the time window is used to obtain data to get the actual temperature rise trend corresponding to the front segment; S3. After the water cut-off, collect the bearing temperature, mechanical seal temperature and motor winding temperature at the sampling frequency, and synchronously record the water temperature, ambient temperature, vibration intensity and noise data; S4. Based on window sliding, fuse the bearing temperature, mechanical seal temperature, motor winding temperature, water temperature, ambient temperature, vibration intensity and noise data obtained in the front segment of a time window into input features; S5. Input the input features into the pre-trained LSTM model to obtain the predicted temperature rise trend, and calculate the deviation between the actual temperature rise and the predicted temperature rise. If the deviation is greater than the preset threshold, it indicates that the temperature rise trend is abnormal. If the deviation is less than the preset threshold, it indicates that the temperature rise trend is normal.

2. The large-flow vertical centrifugal pump water medium test method according to claim 1, characterized in that, The S1 includes the following steps: S101. Keep the water temperature of the test loop at 85±10°C; S102. Keep the pump chamber liquid level 990±200mm below the installation plane of the large flange of the outer casing; S103. Keep the air pressure covering the pump chamber at 0.33±0.05MPa; S104. Adjust the frequency converter frequency to 20Hz and start the pump; S105. Adjust the outlet valve until the main circuit flowmeter reaches 4888±200m³ / h, and then stop adjusting the valve; S106. After the pump unit runs stably, gradually adjust the frequency converter frequency to 50Hz to make the pump reach the rated flow of 12220±367m³ / h; S107. After stable operation, close the main valve of the cooling water system.

3. The large-flow vertical centrifugal pump water medium test method according to claim 2, characterized in that The S106 includes the following steps: After the pump unit runs stably, adjust the frequency converter frequency to 25Hz, 30Hz, 35Hz, 40Hz, 45Hz, 50Hz in sequence to increase the speed; After the pump unit runs stably, observe whether the main circuit flow is 12220±367m³ / h. If it exceeds, slightly adjust the outlet valve to make the main circuit flow reach within the range of 12220±367m³ / h.

4. The large-flow vertical centrifugal pump water medium test method according to claim 1, characterized in that The S1 includes the following steps: S111. Keep the water temperature of the test loop at 85±10°C; S112. Keep the pump chamber liquid level 990±200mm below the installation plane of the large flange of the outer casing; S113. Keep the air pressure covering the pump chamber at 0.33±0.05MPa; S114. Adjust the frequency converter frequency to 20Hz and start the pump; S115. Adjust the outlet valve until the main circuit flowmeter reaches 4888±200m³ / h, and then stop adjusting the valve; S116. After the pump unit runs stably, gradually adjust the frequency converter frequency to 12.5Hz to make the pump reach the rated flow of 3055±92m³ / h; S117. After stable operation, close the main valve of the cooling water system.

5. The large-flow vertical centrifugal pump water medium test method according to claim 4, wherein The S116 includes the following steps: After the pump unit runs stably, adjust the frequency converter frequency to 12.5Hz; After the pump unit runs stably, observe whether the main circuit flow is 3055±92m³ / h. If it exceeds, fine-tune the outlet valve to make the main circuit flow within the range of 3055±92m³ / h.

6. The large-flow vertical centrifugal pump water medium test method according to any one of claims 1-5, characterized in that, The pre-training steps of the LSTM model include: S501. Perform data collection and data preprocessing; S502. Set a time window of 70 seconds and generate a training set, a validation set, and a test set; wherein the training set uses the first 60 seconds of data as input and the temperature rise trend after 10 seconds as output; S503. Build LSTM network; S504. Input the training set into the LSTM network and calculate the loss function; S505. Calculate the gradient of the loss with respect to the LSTM layer weights, and use the Adam optimizer to perform gradient descent to update the LSTM layer weights. S506. After the model training is completed, use the test set and calculate the test set error, and adjust the model complexity based on the test set error.

7. The large-flow vertical centrifugal pump water medium test method according to claim 6, characterized in that The structure of the LSTM network in S503 includes: The input layer is used to receive input features corresponding to sixty time steps; The first LSTM layer contains 64 units, which is used to extract time series features and output information of all time steps; The second LSTM layer contains thirty-two units, which is used to extract deep time series information and only outputs the information of the last time step; The fully connected layer is used to map the features extracted by LSTM to a numerical value to predict the temperature rise trend in the next ten seconds.

8. The large-flow vertical centrifugal pump water medium test method according to claim 7, characterized in that The S501 includes the following steps: S5011. Obtain training data, wherein the training data comes from normal water-off test data without microbubbles under different ambient temperatures and different test conditions, abnormal water-off test data with microbubbles, and some artificially introduced microbubbles test data; S5012. Detect outliers and remove outliers; S5013. Interpolation to fill missing values; S5014. Normalize the data.

9. A water medium test system for a large-flow vertical centrifugal pump, characterized in that, The invention comprises a processor in which a program of the water medium test method for a large flow vertical centrifugal pump according to any one of claims 1 to 8 is run.

10. A storage medium, characterized in that, A program for a water medium test method for a large flow vertical centrifugal pump according to any one of claims 1 to 8 is stored.

Images