Centrifugal impeller pump life prediction method based on LSTM model

Abstract

The invention provides a centrifugal impeller pump service life prediction method based on an LSTM model. The centrifugal impeller pump service life prediction method comprises the steps that centrifugal impeller pump operation data are collected, and a sample data set is constructed; preprocessing the sample data set; constructing a long and short term memory network model; training and testing the long-short-term memory network model to obtain a trained long-short-term memory network model; and predicting the service life of the centrifugal impeller pump based on the trained long-short term memory network model. According to the method, model training and centrifugal impeller pump service life prediction are carried out by constructing the long and short term memory network model, features can be automatically extracted through data driving, complex nonlinear time series data are processed, dependence on a complex physical modeling process is not needed, and the method has higher accuracy, flexibility and robustness and higher practicability.

Description

Technical Field

[0001] The invention belongs to the technical field of centrifugal impeller pump life prediction, and in particular to a centrifugal impeller pump life prediction method based on an LSTM model. Background Art

[0002] Centrifugal impeller pumps are widely used in petrochemical, water treatment, pharmaceutical and other industries. During operation, they convert mechanical energy into kinetic energy of fluid by rotating impellers, which is used to transport liquid or gas. Its working flow rate, that is, the volume of liquid transported per unit time, is closely related to the performance of the pump, and changes in flow rate directly affect the operating status of the pump. Under different working conditions, the flow rate, pressure, power and efficiency of the pump will change. Fluctuations or abnormalities in flow rate may be early signs of pump failure or damage. Changes in flow rate will affect the stress distribution and fatigue accumulation of the pump body and key components. Therefore, by monitoring flow changes, a predictive basis can be provided for the accumulation of fatigue damage, which is crucial to improving the reliability and safety of centrifugal impeller pumps.

[0003] However, the existing traditional centrifugal impeller pump life prediction methods usually rely on physical models or expert experience, and have many shortcomings in the prediction of centrifugal impeller pump life, including limited modeling capabilities for complex nonlinear behaviors, poor real-time performance, reliance on manual intervention, and difficulty in dealing with incomplete or noisy data. These shortcomings make traditional methods have great limitations in practical applications and it is difficult to provide accurate and timely predictions in dynamic and complex environments. Therefore, in order to process complex nonlinear time series data, it is urgent and necessary to seek a centrifugal impeller pump life prediction method based on the LSTM model, which can automatically extract features through data-driven, without complex physical modeling or frequent updates, to achieve complex and dynamically changing centrifugal impeller pump life prediction. Summary of the invention

[0004] In view of the defects in the above-mentioned prior art, the present invention proposes a centrifugal impeller pump life prediction method based on the LSTM model. The method includes collecting centrifugal impeller pump operation data and constructing a sample data set; preprocessing the sample data set; constructing a long short-term memory network model; training the long short-term memory network model and testing it to obtain a trained long short-term memory network model; and predicting the life of the centrifugal impeller pump based on the trained long short-term memory network model. The present invention constructs a long short-term memory network model for model training and centrifugal impeller pump life prediction, and can automatically extract features through data-driven, process complex nonlinear time series data, and does not need to rely on complex physical modeling processes. It has higher accuracy, flexibility and robustness, and is more practical.

[0005] The present invention provides a centrifugal impeller pump life prediction method based on an LSTM model, which comprises the following steps:

[0006] S1, collect the operation data of the centrifugal impeller pump and construct a sample data set;

[0007] S11, using a sensor to collect centrifugal impeller pump operation data, wherein the centrifugal impeller pump operation data includes centrifugal impeller pump operation time and flow rate;

[0008] S12. Based on the operation data of the centrifugal impeller pump, a sample data set required for training the model is constructed;

[0009] S2. Preprocessing the sample data set: the preprocessing includes removing abnormal data, cleaning outlier data and smoothing data;

[0010] S3, dividing the sample data set: dividing the sample data set into a first data set and a second data set, the first data set is used for model training, and the second data set is used for model testing;

[0011] S4. Constructing a long short-term memory network model: the long short-term memory network model includes a forget gate, an input gate and an output gate;

[0012] S41. Establish a forget gate: Use the Sigmoid activation function to convert the memory cell h at the previous moment t-1 t-1 The running time x of the centrifugal impeller pump input at the current time t t Converted to forget value f t ;

[0013] S42. Establish input gate: use Sigmoid activation function and tanh function to obtain new memory cell state;

[0014] S421, use the Sigmoid activation function to activate the memory cell h at the previous time t-1 t-1 The running time x of the centrifugal impeller pump input at the current time t t Convert to input value I t , expressed as:

[0015] I t =σ(W1[h t-1 , x t ]+b1) (4)

[0016] Wherein, W1, b1 represent the first weight value and the first compensation value of the input gate respectively; σ(·) represents the Sigmoid activation function operation;

[0017] S422, using the tanh function to compare the memory cell at the previous moment with the centrifugal impeller pump running time x input at the current moment t t Transformation into new candidate memory cell states It is expressed as:

[0018]

[0019] Among them, W c , b c They are respectively represented as the second weight value and the second compensation value of the input gate;

[0020] S423, based on the forget gate acting on the memory cell state C at the previous time t-1 t-1 , will I t Acting on new candidate memory cell states Get the new memory cell state C at the current time t t :

[0021]

[0022] S43. Establish output gate: obtain new memory cells with the help of Sigmoid activation function and tanh function;

[0023] S431, use the Sigmoid activation function to associate the memory cell at the previous time t-1 with the centrifugal impeller pump running time x input at the current time t t Converted to output value O t , expressed as:

[0024] O t =σ(W0[h t-1 , x t ]+b0) (7)

[0025] Among them, W0 and b0 represent the weight value and compensation value of the output gate respectively;

[0026] S432, based on output value O t , apply the tanh function to the new memory cell state C t , get the new memory cell h at the current time t t :

[0027] h t =O t tanhC t (8);

[0028] S5. Training and testing the long short-term memory network model to obtain a trained long short-term memory network model: training the long short-term memory network model based on the first data set, and testing the long short-term memory network model based on the second data set;

[0029] S6. Predict the life of the centrifugal impeller pump based on the trained long short-term memory network model: Based on the trained long short-term memory network model, continue to increase the running time of the centrifugal impeller pump and predict the flow of the centrifugal impeller pump. When the flow of the centrifugal impeller pump is lower than the set threshold, the centrifugal impeller pump fails, thereby obtaining the life of the centrifugal impeller pump.

[0030] Further, the step S2 specifically includes the following steps:

[0031] S21. Elimination of abnormal traffic data: According to the normal distribution law of random variables, the traffic data falling outside the mean by 3 times the standard deviation is discarded;

[0032] S22. Cleaning traffic outlier data: using median absolute deviation (MAD) to determine traffic outlier data, and using linear interpolation to replace traffic outlier data;

[0033] S221. Calculate the median absolute deviation MAD:

[0034] MAD=median(|y i -M|) (1)

[0035] Among them, y i represents the i-th flow value in the sample data set; M represents the median of all flow values ​​in the sample data set; median represents the median operation; || represents the absolute value operation;

[0036] S222, determine the outlier data in the sample data set: set a judgment formula, if the i-th flow value y in the sample data set i If the judgment formula is not satisfied, then y i Determined as traffic outlier data;

[0037] S223, based on the traffic data before and after the traffic outlier data, using a linear interpolation method to obtain traffic data interpolation and replace the traffic outlier data;

[0038] S23, smoothing the flow data: using Gaussian filtering to smooth the flow data in the sample data set.

[0039] Preferably, the forgotten value f in step S41 is t The calculation is:

[0040] f t =σ(W f [h t-1 , x t ]+b f ) (3)

[0041] Among them, W f , b fThey are respectively represented as the weight value and compensation value of the forget gate.

[0042] Preferably, the judgment formula in step S222 is set as:

[0043] M-3×MAD <y i <M+3×MAD (2)。

[0044] Preferably, the loss function J during the training in step S5 is:

[0045]

[0046] Among them, Y, y represent the traffic prediction value and the actual traffic value respectively; || || 2 represents the 2-norm.

[0047] Preferably, in step S5, the device operation data flow observed at the next moment is used as the true value of the previous sampled data flow. When the deviation between it and the true value accumulates to an error threshold δ, that is, J ≥ δ, the weight matrix and bias vector in the model are updated by back propagation.

[0048] Preferably, the forgotten value f in step S41 is t The value range is [0,1], where 0 means discarding all information and 1 means keeping all information.

[0049] Preferably, the input and output of the sample data set in step S1 are respectively the running time and flow rate of the centrifugal impeller pump.

[0050] Compared with the prior art, the technical effects of the present invention are:

[0051] 1. The centrifugal impeller pump life prediction method based on the LSTM model designed by the present invention collects the operating status data of the centrifugal impeller pump, constructs a long short-term memory network model for model training and centrifugal impeller pump life prediction. Compared with the prior art, the proposed method can predict the life of the centrifugal impeller pump more accurately and reliably, and is suitable for the life prediction of the centrifugal impeller pump in complex and dynamically changing working environments.

[0052] 2. The centrifugal impeller pump life prediction method based on the LSTM model designed by the present invention can automatically extract features through data-driven, process complex nonlinear time series data, effectively deal with complex nonlinear time series data, and can carry out prediction work in real time without relying on complex physical modeling processes or frequent updates. It has higher accuracy, flexibility and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Other features, objects and advantages of the present application will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings.

[0054] Figure 1 It is a flow chart of the centrifugal impeller pump life prediction method based on the LSTM model of the present invention;

[0055] Figure 2a is a data distribution diagram of a first group of centrifugal impeller pump test parts in a specific embodiment of the present invention;

[0056] Figure 2b is a data distribution diagram of a second group of centrifugal impeller pump test parts in a specific embodiment of the present invention;

[0057] Figure 2c is a data distribution diagram of a third group of centrifugal impeller pump test parts in a specific embodiment of the present invention;

[0058] Figure 3 is a schematic diagram of the training process of a model in a specific embodiment of the present invention;

[0059] Figure 4 is a comparison result diagram of the first data set prediction result and the real data of the present invention;

[0060] Figure 5 This is a comparison diagram of the predicted results of the second data set of the present invention and the actual data. DETAILED DESCRIPTION

[0061] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It will be appreciated that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the inventions. It should also be noted that, for ease of description, only portions related to the relevant inventions are shown in the accompanying drawings. It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0062] Figure 1 The invention shows a centrifugal impeller pump life prediction method based on the LSTM model, which comprises the following steps:

[0063] S1. Collect the operation data of the centrifugal impeller pump and construct a sample data set; the input and output of the sample data set are the operation time and flow rate of the centrifugal impeller pump respectively.

[0064] S11. Collect centrifugal impeller pump operation data using sensors, where the centrifugal impeller pump operation data includes centrifugal impeller pump operation time and flow rate.

[0065] S12. Based on the centrifugal impeller pump operation data, construct a sample data set required for training the model.

[0066] S2. Preprocess the sample data set: Preprocessing includes removing abnormal data, cleaning outlier data and smoothing data.

[0067] S21. Elimination of abnormal flow data: According to the normal distribution law of random variables, in multiple tests, the probability of flow measurement values ​​falling within the range of the mean plus or minus 3 times the standard deviation is 99.73%, and the probability of appearing outside this range is only 0.27%. Therefore, in actual tests, once it appears, the flow measurement data is considered unreliable and should be discarded. Therefore, the flow data falling outside the mean 3 times the standard deviation is discarded.

[0068] S22. Cleaning flow outlier data: Affected by working conditions, the collected flow data is usually missing in time. The present invention uses the median absolute deviation to determine and linear interpolation to supplement the missing flow data.

[0069] The median absolute deviation (MAD) was discovered and promoted by Hampel in 1974. Like the mean, the median M is a measure of central tendency, but it has the advantage of being very insensitive to the presence of outliers. When the breakdown point is 0.5, the median is the outlier estimation model with the highest breakdown point. As an outlier detection model, MAD takes advantage of this property of the median. In addition, MAD is completely unaffected by sample size.

[0070] The median absolute deviation (MAD) is used to identify outlier data and linear interpolation is used to replace outlier data.

[0071] S221. Calculate the median absolute deviation MAD:

[0072] MAD=median(|y i -M|) (1)

[0073] Among them, y i represents the i-th flow value in the sample data set; M represents the median of all flow values ​​in the sample data set; median represents the median operation; | | represents the absolute value operation.

[0074] S222, determine the outlier data in the sample data set: set a judgment formula, if the i-th flow value y in the sample data set i If the judgment formula is not satisfied, then y i Identified as outlier data.

[0075] Generally speaking, fluctuations within the range of 3 times the MAD value above and below the median are considered normal data, and values ​​exceeding 3 times the MAD value are considered outliers. The judgment formula is set as:

[0076] M i -3×MAD <y i <M i +3×MAD (2).

[0077] S223. Based on the flow data before and after the outlier flow data, a linear interpolation method is used to obtain flow interpolation and replace the outlier flow data. Linear interpolation refers to an interpolation method in which the interpolation function is a first-order polynomial, and the interpolation error at the interpolation node is zero.

[0078] S23. Smoothing flow data: Gaussian filtering is used to smooth the flow data in the sample data set. The basic principle of Gaussian filtering is to use the Gaussian function to perform convolution operations on the image to achieve image smoothing and noise reduction. The Gaussian function is a bell-shaped curve with central symmetry and unimodality. Its mathematical expression is:

[0079]

[0080] Among them, (x, y) represents the spatial coordinates, ε represents the standard deviation, controls the width of the Gaussian function, and determines the smoothness of the filter. A smaller ε will result in less blurring effect, while a larger ε will produce a stronger blurring effect.

[0081] Gaussian filtering uses a convolution kernel (also called a template or mask) to scan each pixel in the image, and replaces the value of the central pixel of the template with the weighted average grayscale value of the pixels in the neighborhood determined by the template. The value of the convolution kernel is no longer 1, but is set according to the value of the Gaussian function, so that the weight value of the center point is larger and the weight value far from the center point is smaller, thereby achieving image smoothing and noise reduction effects.

[0082] S3. Divide the sample data set: divide the sample data set into a first data set and a second data set, the first data set is used for model training, and the second data set is used for model testing.

[0083] Set the first length to λ s The sequence is used as the model input, and the second length is λ t The sequence is used as the label when training the model. The formula for calculating the sequence length is:

[0084]

[0085] Among them, α represents the partition factor; λ n Represents all sequences.

[0086] S4. Construct a long short-term memory network model: The long short-term memory network model includes a forget gate, an input gate, and an output gate.

[0087] The long short-term memory network model is a variant of the recurrent neural network RNN, which is particularly suitable for processing sequence data. It can handle variable-length sequence inputs and use the hidden state of the previous moment to affect the output of the current moment. However, traditional RNNs encounter the problem of gradient vanishing or gradient exploding when dealing with long-term dependency tasks, making it difficult to learn long-term memory. The LSTM network is a special implementation of RNN, which mainly improves the hidden layer in the RNN and adds 3 gate control units in the hidden layer to control the self-loop, solving the problem that RNN cannot capture long-term dependencies.

[0088] S41. Establish a forget gate: Use the Sigmoid activation function to convert the memory cell h at the previous moment t-1 t-1 The running time x of the centrifugal impeller pump input at the current time t t Converted to forget value f t ; Forgetting value f t The calculation is:

[0089] f t =σ(W f [h t-1 , x t ]+b f ) (3)

[0090] Among them, W f , b f They represent the weight value and compensation value of the forget gate respectively; σ(·) represents the Sigmoid activation function operation.

[0091] Forgetting value f t The value range is [0,1], where 0 means discarding all information and 1 means keeping all information.

[0092] S42. Establish input gate: Use Sigmoid activation function and tanh function to obtain new memory cell state.

[0093] S421, use the Sigmoid activation function to activate the memory cell h at the previous time t-1 t-1 The running time x of the centrifugal impeller pump input at the current time t t Convert to input value I t , expressed as:

[0094] I t =σ(W1[h t-1 , x t ]+b1) (4)

[0095] Wherein, W1 and b1 represent the first weight value and the first compensation value of the input gate respectively.

[0096] S422, using the tanh function to compare the memory cell at the previous moment with the centrifugal impeller pump running time x input at the current moment t t Transformation into new candidate memory cell states It is expressed as:

[0097]

[0098] Among them, W c , b c They are represented as the second weight value and the second compensation value of the input gate respectively.

[0099] S423, based on the forget gate acting on the memory cell state C at the previous time t-1 t-1 , will I t Acting on new candidate memory cell states Get the new memory cell state C at the current time t t :

[0100]

[0101] S43. Establish output gate: Use Sigmoid activation function and tanh function to obtain new memory cells.

[0102] S431, use the Sigmoid activation function to associate the memory cell at the previous time t-1 with the centrifugal impeller pump running time x input at the current time t t Converted into output value O t , expressed as:

[0103] O t =σ(W0[h t-1 , x t ]+b0) (7)

[0104] Among them, W0 and b0 represent the weight value and compensation value of the output gate respectively.

[0105] S432, based on output value O t , apply the tanh function to the new memory cell state C t , get the new memory cell h at the current time t t :

[0106] h t =O t tanhC t (8).

[0107] S5. Train and test the long short-term memory network model to obtain a trained long short-term memory network model: train the long short-term memory network model based on the first data set, and test the long short-term memory network model based on the second data set.

[0108] The loss function J during training is:

[0109]

[0110] Among them, Y, y represent the traffic prediction value and the actual traffic value respectively; || || 2 represents the 2-norm.

[0111] The device operation data flow observed at the next moment is taken as the true value of the previous sampled data flow. When the deviation between it and the true value accumulates to the error threshold δ, that is, J ≥ δ, the weight matrix and bias vector in the model are updated through back propagation. As the model running time increases, the model accuracy will continue to improve. δ is customized according to the actual situation.

[0112] S6. Predict the life of the centrifugal impeller pump based on the trained long short-term memory network model: Based on the trained long short-term memory network model, continue to increase the running time of the centrifugal impeller pump and predict the flow of the centrifugal impeller pump. When the flow of the centrifugal impeller pump is lower than the set threshold, the centrifugal impeller pump fails, that is, the life of the centrifugal impeller pump is obtained.

[0113] In a specific embodiment, taking a company's life survey test for a centrifugal impeller pump as an example, a simulation is performed based on the LSTM model to predict the life survey behavior of the centrifugal impeller pump.

[0114] Continue the test with the test piece that has been subjected to the durability test. Power on the test bench motor, and then start the centrifugal impeller pump according to the steps in the equipment manual. After the centrifugal impeller pump is started, adjust the pressure gauge and flow valve to control the flow rate output by the centrifugal impeller pump. Use the pressure gauge and flow meter to record the head and flow value of the water pump under 100% working conditions in real time.

[0115] The operating status data of the centrifugal impeller pump, including pressure and flow, are collected and preprocessed. The values ​​that do not change over time are deleted, and the data are normalized so that each group of values ​​is close to 1. The operating status data are used to obtain random variable-length sample data by randomly taking values ​​to enhance the diversity of the sample data.

[0116] The distribution of some data of centrifugal impeller pump test is as follows Figure 2a to 2c As shown, the horizontal axis is the data serial number, and the vertical axis is the pump flow m 3 / h, Figure 2aThe flow rate increases due to the large opening of the medium valve (i.e., the head change), the outlier data are determined based on the median absolute deviation MAD, and the outlier data are replaced by linear interpolation.

[0117] The operating status data is divided into a first data set and a second data set, 60% of the operating status data is used as the first data set for model training, and the remaining 40% of the operating status data is used as the second data set for life prediction.

[0118] A long short-term memory network model was established, and the operating status data was used for model training. The overall degradation characteristics of the centrifugal pump life cycle were obtained through weighted smoothing.

[0119] After obtaining the trained model, the second data set is input into the model for iterative calculation to obtain the flow prediction result of the centrifugal impeller pump. When the flow of the centrifugal impeller pump is lower than the set threshold, the centrifugal impeller pump fails, that is, the life prediction of the centrifugal impeller pump is obtained.

[0120] In this example, the training LSTM model sets the learning rate to 0.0001 and completes 200 iterations. Determine the loss function, compare the prediction results with the actual data, determine the prediction accuracy, and plot the training process. Figure 3 The comparison between the prediction results of the first data set and the real data is shown in Figure 4 As shown, the comparison results between the prediction results of the second data set and the real data are as follows Figure 5 shown.

[0121] The present invention designs a centrifugal impeller pump life prediction method based on the LSTM model. By collecting the operating status data of the centrifugal impeller pump, a long short-term memory network model is constructed to perform model training and centrifugal impeller pump life prediction. Compared with the prior art, the proposed method can more accurately and reliably predict the life of the centrifugal impeller pump, and is suitable for centrifugal impeller pump life prediction in complex and dynamically changing working environments; it can automatically extract features through data-driven, process complex nonlinear time series data, effectively cope with complex nonlinear time series data, and can carry out prediction work in real time without relying on complex physical modeling processes or frequent updates, and has higher accuracy, flexibility and robustness.

[0122] Finally, it should be noted that the above embodiments are only intended to illustrate rather than limit the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the present invention can still be modified or replaced by equivalents. Any modification or partial replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A centrifugal impeller pump life prediction method based on LSTM model, characterized in that: It includes the following steps: S1, collect the operation data of the centrifugal impeller pump and construct a sample data set; S11, using a sensor to collect centrifugal impeller pump operation data, wherein the centrifugal impeller pump operation data includes centrifugal impeller pump operation time and flow rate; S12. Based on the operation data of the centrifugal impeller pump, a sample data set required for training the model is constructed; S2. Preprocessing the sample data set: the preprocessing includes removing abnormal data, cleaning outlier data and smoothing data; S3, dividing the sample data set: dividing the sample data set into a first data set and a second data set, the first data set is used for model training, and the second data set is used for model testing; S4. Constructing a long short-term memory network model: the long short-term memory network model includes a forget gate, an input gate and an output gate; S41. Establish a forget gate: Use the Sigmoid activation function to convert the memory cell h at the previous moment t-1 t-1 The running time x of the centrifugal impeller pump input at the current time t t Converted to forget value f t ; S42. Establish input gate: use Sigmoid activation function and tanh function to obtain new memory cell state; S421, use the Sigmoid activation function to activate the memory cell h at the previous time t-1 t-1 The running time x of the centrifugal impeller pump input at the current time t t Convert to input value I t , expressed as: AND t =σ(W1[h t-1 ,x t ]+b1) (4) Wherein, W1, b1 represent the first weight value and the first compensation value of the input gate respectively; σ(·) represents the Sigmoid activation function operation; S422, using the tanh function to compare the memory cell at the previous moment with the centrifugal impeller pump running time x input at the current moment t t Transformation into new candidate memory cell states It is expressed as: Among them, W c , b c They are respectively represented as the second weight value and the second compensation value of the input gate; S423, based on the forget gate acting on the memory cell state C at the previous time t-1 t-1 , will I t Acting on new candidate memory cell states Get the new memory cell state C at the current time t t : S43. Establish output gate: obtain new memory cells with the help of Sigmoid activation function and tanh function; S431, use the Sigmoid activation function to associate the memory cell at the previous time t-1 with the centrifugal impeller pump running time x input at the current time t t Converted into output value O t , expressed as: ABOUT t =σ(W0[h t-1 ,x t ]+b0) (7) Among them, W0 and b0 represent the weight value and compensation value of the output gate respectively; S432, based on output value O t , apply the tanh function to the new memory cell state C t , get the new memory cell h at the current time t t : h t =Oh t underC t (8); S5. Training and testing the long short-term memory network model to obtain a trained long short-term memory network model: training the long short-term memory network model based on the first data set, and testing the long short-term memory network model based on the second data set; S6. Predict the life of the centrifugal impeller pump based on the trained long short-term memory network model: Based on the trained long short-term memory network model, continue to increase the running time of the centrifugal impeller pump and predict the flow of the centrifugal impeller pump. When the flow of the centrifugal impeller pump is lower than the set threshold, the centrifugal impeller pump fails, thereby obtaining the life of the centrifugal impeller pump.

2. The centrifugal impeller pump life prediction method based on the LSTM model according to claim 1 is characterized in that: The step S2 specifically includes the following steps: S21. Elimination of abnormal traffic data: According to the normal distribution law of random variables, the traffic data falling outside the mean by 3 times the standard deviation is discarded; S22. Cleaning traffic outlier data: using median absolute deviation (MAD) to determine traffic outlier data, and using linear interpolation to replace traffic outlier data; S221. Calculate the median absolute deviation MAD: MAD=median(|y i -M|) (1) Among them, y i represents the i-th flow value in the sample data set; M represents the median of all flow values ​​in the sample data set; median represents the median operation; || represents the absolute value operation; S222, determine the outlier data in the sample data set: set a judgment formula, if the i-th flow value y in the sample data set i If the judgment formula is not satisfied, then y i Determined as traffic outlier data; S223, based on the traffic data before and after the traffic outlier data, using a linear interpolation method to obtain traffic data interpolation and replace the traffic outlier data; S23, smoothing the flow data: using Gaussian filtering to smooth the flow data in the sample data set.

3. The centrifugal impeller pump life prediction method based on the LSTM model according to claim 1 is characterized in that: The forgotten value f in step S41 t The calculation is: f t =σ(W f [h t-1 ,x t ]+b f ) (3) Among them, W f , b f They are respectively represented as the weight value and compensation value of the forget gate.

4. The centrifugal impeller pump life prediction method based on the LSTM model according to claim 2 is characterized in that: The judgment formula in step S222 is set as: M-3×MAD<y i <M+3×MAD (2)。 5. The centrifugal impeller pump life prediction method based on the LSTM model according to claim 2 is characterized in that: The loss function J during the training in step S5 is: Among them, Y, y represent the traffic prediction value and the actual traffic value respectively; |||| 2 represents the 2-norm.

6. The centrifugal impeller pump life prediction method based on the LSTM model according to claim 1 is characterized in that: In step S5, the device operation data flow observed at the next moment is used as the true value of the previous sampled data flow. When the deviation between it and the true value accumulates to the error threshold δ, that is, J ≥ δ, the weight matrix and bias vector in the model are updated through back propagation.

7. The centrifugal impeller pump life prediction method based on the LSTM model according to claim 1 is characterized in that: The forgotten value f in step S41 t The value range is [0,1], where 0 means discarding all information and 1 means keeping all information.

8. The centrifugal impeller pump life prediction method based on the LSTM model according to claim 1 is characterized in that: The input and output of the sample data set in step S1 are respectively the running time and flow rate of the centrifugal impeller pump.

Images