Hydro-junction pump station fault diagnosis method based on LSTM deep learning algorithm

Abstract

The invention relates to the field of hydro-junction pump station fault diagnosis, in particular to a hydro-junction pump station fault diagnosis method based on an LSTM deep learning algorithm, which comprises a hardware system and a software system, and is characterized in that the hardware system comprises a flow sensor, a temperature sensor, a vibration sensor, a noise sensor and an edge computing power device; the software system comprises a pump station fault diagnosis system and a regional data center covering an adjacent water system pump station, the software system analyzes multi-source heterogeneous data from a sensor through a long short-term memory (LSTM) network and performs fault diagnosis, and a generative adversarial network (GAN) is assisted to perform data enhancement on the station so as to enlarge the network training sample size; newly born data of the station are converged into a regional data center to be accumulated and fed back to all pump stations in the region for fault diagnosis. The method has the advantages that the hydro-junction pump station fault diagnosis system is endowed with a self-evolution function of adapting to regional characteristics, so that the diagnosis efficiency is continuously improved along with the running time of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent fault diagnosis and predictive maintenance of water conservancy hub pump stations, and in particular relates to a water conservancy hub pump station fault diagnosis method based on an LSTM deep learning algorithm, which is used for the automatic fault diagnosis of key units (such as water pumps and motors) of water conservancy hub pump stations. BACKGROUND

[0002] Water conservancy hub pump stations are core infrastructure for guaranteeing water resource allocation, flood control and drainage, and agricultural irrigation, and the sustained safe and stable operation of the main pump units thereof is of vital importance. Once a fault occurs, it will result in huge economic losses and social risks. Traditional pump station fault diagnosis mainly relies on periodic inspection, threshold alarm, and expert analysis based on vibration spectrum. These methods have obvious shortcomings: threshold alarm is lagging and prone to false positives; spectrum analysis is heavily dependent on expert experience, and knowledge is difficult to pass on and standardize, and the ability to identify early weak faults, intermittent faults, and concurrent multiple faults is limited.

[0003] In recent years, deep learning-based fault diagnosis methods have begun to be applied, but the accuracy of deep learning fault diagnosis is highly dependent on the experience (data samples) of existing diagnosis cases. The number and suitability of existing cases have an important impact on current and future fault diagnosis work. In the initial stage of the current application of intelligent fault diagnosis of water conservancy hub pump stations, the lack of suitable data samples seriously weakens the generalization ability of the model, thus restricting the improvement of diagnosis efficiency.

[0004] The long short-term memory network (LSTM) is very suitable for fault diagnosis of rotating machinery such as water pumps and motors due to its excellent time series modeling capability, but the current LSTM network is only based on a single signal source (mainly vibration signals) in use, and fails to fully utilize the complementary information contained in the multi-source heterogeneous data of pump stations, and also ignores the completely different time series patterns and fault representation forms of different physical signals. At the same time, the output form of the diagnosis result remains at the qualitative level, making it difficult to provide quantitative guidance for predictive maintenance. SUMMARY

[0005] The purpose of the present application is to provide a water conservancy hub pump station fault diagnosis method based on an LSTM deep learning algorithm to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides the following technical solutions: A water conservancy hub pump station fault diagnosis method based on an LSTM deep learning algorithm, comprising a hardware system and a software system, the hardware system comprising a flow sensor, a temperature sensor, a vibration sensor, a noise sensor, and an edge computing device, and the software system comprising a pump station fault diagnosis system and a regional data center covering all pump stations in the adjacent water system. The fault diagnosis method comprises: S1: The flow sensor transmits the flow signals of the key nodes upstream and downstream of the water pump, the temperature sensor transmits the temperature signals of the heat-emitting devices including the water pump driving motor, the vibration sensor transmits the vibration signals of the pump body and shaft system, and the noise sensor transmits the noise signals of the pump body, shaft system, pipeline and electrical equipment to the data collector in the edge computing device; S2: The data collector pre-processes the multi-source heterogeneous data from different sensors, first performs wavelet threshold denoising processing on all collected signals, and then performs normalization processing on all recorded data based on the minimum-maximum value range to eliminate dimension differences and form a unified scale; S3: The processed data is input into a deep learning neural network, including a main network, i.e., an LSTM network, and an auxiliary network, i.e., a GAN network. The LSTM network is a double-branch time series feature extraction network used to extract long-term dependence features and time series evolution rules of multi-dimensional operating parameter time series data. Meanwhile, data about pump station failures are retrieved from an industrial time series database. The GAN network generates synthetic time series data simulating fault features based on the new data of the current pump station and is used to expand the training set. The high-level features output by the two branches of the LSTM network are spliced and input into a Transformer network to adaptively weight the importance of different sensor signal features. S4: The industrial time series database and the synthetic data generated by the GAN network are divided into a training set and a test set according to a certain proportion. The LSTM network is trained and diagnosed to output. The output layer is connected to two sub-networks. One sub-network outputs the fault type, and the other sub-network estimates the fault severity by regression. S5: The new data of the current pump station fault diagnosis result is imported into the regional data center for storage and is called by all pump stations in the adjacent water system. As the amount of pump station fault data covered by the regional data center increases, the amount of synthetic fault data generated from the industrial time series database and the GAN network is correspondingly reduced, gradually achieving efficient diagnosis in the region.

[0007] In S1, the flow sensor, temperature sensor, vibration sensor and noise sensor are connected to the software system through a 5G module serial port. In S2, the wavelet threshold denoising processing function is: , ω is the wavelet variation coefficient of the noisy signal, th is the set threshold value, and the size of the wavelet variation coefficient ω of the noisy signal is compared with the preset value th. If it is less than th, the output is zero, otherwise, the threshold function processing is performed on it. The normalization processing function is: , F max , F minrespectively the maximum and minimum values of the measured physical quantity within the signal acquisition period.

[0008] The mathematical model of the LSTM network in S3 is: , where h t-1 is the output value of the LSTM neural network at time t-1, x t is the input value of the LSTM neural network at time t, w i , w f , w o , w c and b i , b f , b o , b c represent the weights and biases of the input gate, the forget gate, the output gate, and the candidate memory, respectively, i t , f t , o t are the states of the input gate, the forget gate, and the output gate at time t, c t is the state of the memory unit at time t, and sigma is the Sigmoid activation function, and tanh() is the hyperbolic tangent function.

[0009] The LSTM network contains two independent branches. Branch one receives pre-processed multi-dimensional operating parameter time series data, including flow, temperature, and noise parameter time series values. Branch two is a time series branch in the frequency domain. First, the pre-processed vibration signal is subjected to short-time Fourier transform to generate a time-frequency map sequence, and the transform function is: , Then the LSTM network of this branch extracts the time series evolution rule of the frequency domain features. The GAN network trains a generator and a discriminator to synthesize multi-dimensional operating parameter time series data based on new data from the pump station. The mathematical model of the GAN network is: , After splicing, the Transformer network applies a multi-head attention mechanism to adaptively weight the importance of different sensor signal features.

[0010] In S4, the data distribution ratio of the training set and the test set is 7:3~9:1. The data from the industrial time series database and the synthetic data from the GAN network are equally distributed in the training set and the test set.

[0011] In S5, with the increase of the number of samples called from the regional data center to cover the pump station fault data, the decrease rate of the number of samples called from the industrial time series database is greater than the decrease rate of the number of samples called from the GAN network, and when the number of samples from the regional data center reaches the initial sum of the number of samples from the industrial time series database and the GAN network, the local pump station fault diagnosis is entirely based on the samples called from the regional data center.

[0012] Compared with the prior art, the beneficial effects of the present application are:

[0013] The method includes a regional data center covering all pump stations in the adjacent water system, and realizes sharing and rapid accumulation of fault diagnosis data samples. Since the pump stations in the adjacent water system are affected by similar hydrological and environmental characteristics, the fault generation mode has high comparability, and the suitability of new data can enhance the generalization ability of the deep learning model. Initially, the industrial time series database and the GAN network are called to implement local data enhancement, which can quickly start the fault diagnosis work and accelerate the self-evolution of the diagnosis capability of the pump stations in the region. The double-branch LSTM network can assign excellent time series modeling capability to the processing of multi-source heterogeneous signals, fully adapt to the time series mode and fault representation form of different sensor signals, avoid missed diagnosis, and the output layer is connected to the subnetwork, which can simultaneously reflect the fault type and severity, and provide fine quantitative guidance for predictive maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 It is the software and hardware framework of the data calling, collecting, training and diagnosis of the deep learning system for the fault diagnosis of the water conservancy hub pump station.

[0015] Figure 2 It is the neural network working logic architecture of the present application.

[0016] Figure 3 It is the principle diagram of the LSTM neural network structure of the present application.

[0017] Figure 4 It is the principle diagram of the multi-head attention mechanism structure of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0019] Embodiment: The shaft flow pump unit of the water conservancy hub pump station is subjected to fault diagnosis, and there are n pump stations in the adjacent water system, sharing a regional data center.

[0020] S1: The flow sensor transmits the flow signals of the key nodes upstream and downstream of the water pump, the temperature sensor transmits the temperature signals of the heat-emitting devices including the water pump drive motor, the vibration sensor transmits the vibration signals of the pump body and shaft system, and the noise sensor transmits the noise signals emitted by the pump body, shaft system, pipeline and electrical equipment to the data collector in the edge computing device. Optionally, a turbine flowmeter or an electromagnetic flowmeter is arranged at the inlet and outlet pipelines of the water pump, a patch type thermocouple or an infrared thermal imager is arranged at the motor bearing and winding, a three-coordinate acceleration sensor is arranged at the pump body, shaft system and motor, and a near-field capacitive microphone is arranged around the main devices such as the pump body, shaft system, pipeline and electrical equipment. All the collected signals are transmitted to the multiplexer in the edge computing device.

[0021] S2: The data collector pre-processes the multi-source heterogeneous data from different sensors. Firstly, all the collected signals are subjected to wavelet threshold denoising processing, and then all the recorded data are subjected to normalization processing based on the minimum-maximum value range to eliminate the dimension difference and form a unified scale. Optionally, the data collector collects the temperature, flow and noise at a low frequency, collects the vibration signals from the acceleration sensor at a high frequency, and subjects the collected data to normalization processing with a time window of 5-10 seconds to relieve the influence of data baseline drift under different working conditions.

[0022] S3: The processed data are input into a deep learning neural network including a main network, i.e., an LSTM network, and an auxiliary network, i.e., a GAN network. The LSTM network is a double-branch time sequence feature extraction network for extracting long-term dependence features and time sequence evolution rules of multi-dimensional running parameter time sequence data. Meanwhile, data about pump station faults are called from an industrial time sequence database. The GAN network generates synthetic time sequence data simulating fault features based on the new data of the current pump station and is used to expand the training set. The high-level features output by the two branches of the LSTM network are spliced and input into a Transformer network for self-adaptive weighting of the importance of different sensor signal features. Optionally, the LSTM network includes a forget gate, an input gate, an output gate and a cell state for storing and transmitting information. Each gate generates an output between 0 and 1 by a Sigmoid activation function, indicating the opening degree of the gate. A tanh activation function is used to generate a candidate value. The generator of the GAN network receives the pre-processed data from the data collector as input to synthesize samples similar to the local real data. The discriminator of the GAN network is composed of multiple convolutional layers for extracting features from the input information of the data collector and outputting a discrimination result through a fully connected layer. The industrial time sequence database adopts TDengine, and the Transformer network uses a multi-head attention mechanism.

[0023] S4: The synthetic data generated by the industrial time series database and the GAN network is divided into a training set and a test set according to a certain proportion, and the LSTM network is trained and diagnosed to output. The output layer is connected to two sub-networks, one of which outputs the fault type, and the other of which estimates the fault severity. Optionally, the sub-network outputting the fault type is connected to a softmax classification layer to output "no fault, rotor imbalance, shaft wear, cavitation, electrical short circuit, component loosening, and multiple faults coexisting" seven types, and the sub-network outputting the fault severity is connected to a linear regression layer followed by a sigmoid function to output a severity index I between 0 and 1. The closer I is to 1, the more serious the fault.

[0024] S5: The current pump station fault diagnosis result new data is imported into the regional data center for storage, and is called by n pump stations in the adjacent water system. As the amount of pump station fault data covered by the regional data center increases, the amount of synthetic fault data generated from the industrial time series database and the GAN network is correspondingly reduced, gradually and eventually achieving efficient diagnosis in the region. Optionally, the regional data center is located in the central control platform of the pump station fault diagnosis system supplier or in the central control room of one of the pump stations, and the regional data center covers n pump stations in the jurisdictional water system.

[0025] In S1, the flow sensor, temperature sensor, vibration sensor and noise sensor are connected to the software system through the 5G module serial port, and the wavelet threshold denoising function in S2 is: , ω is the wavelet variation coefficient of the noisy signal, th is the set threshold value, and the size of the wavelet variation coefficient ω of the noisy signal is compared with the preset value th. If it is less than th, the output is zero, otherwise, it is processed by the threshold function.

[0026] The normalization processing function is: , F max 、F min are the maximum and minimum values of the measured physical quantity during the signal acquisition period. Preferably, the original F value of the temperature, flow and noise signal respectively uses the Kelvin temperature scale, hourly volume flow and decibel value as the dimension, and the original F value of the vibration signal uses the vibration frequency and amplitude as the dimension.

[0027] The mathematical model of the LSTM network in S3 is: , In the formula, h t-1 is the output value of the LSTM neural network at t-1, x t is the input value of the LSTM neural network at t, w i , w f , wo , w c , and b i , b f , b o , b c respectively represent the input gate, the forget gate, the output gate, the weight and the bias of the candidate memory, i t , f t , o t are the states of the input gate, the forget gate and the output gate at time t, c t is the state of the memory cell at time t, and σ is the Sigmoid activation function, tanh() is the hyperbolic tangent function.

[0028] The LSTM network contains two independent branches. Branch one receives the pre-processed multi-dimensional operating parameter time series data, including flow, temperature and noise parameter time series values. Branch two is the time series branch of the vibration frequency domain. First, the pre-processed vibration signal is subjected to short-time Fourier transform to generate a time-frequency diagram sequence. The transform function is: , Then, the LSTM network of this branch extracts the time series evolution rule of the frequency domain features. The GAN network trains a generator and a discriminator to synthesize multi-dimensional operating parameter time series data through new data of the pump station. The mathematical model of the GAN network is: , After splicing, the Transformer network applies a multi-head attention mechanism to adaptively weight the importance of different sensor signal features.

[0029] Optionally, for branch one, input the data of the parameters corresponding to the dimension of the signal acquisition point in a continuous number of time steps to form a "time step-parameter dimension" matrix, and output high-level time sequence features after passing through two layers of stacked LSTM units; for branch two, divide and window the continuous vibration signal with a fixed length window, perform short-time Fourier transform on each window data, arrange a series of time-frequency diagrams in chronological order to form a three-dimensional tensor as the input of this branch, take the vibration signal in the corresponding time length (one data segment per second), obtain the amplitude spectrum (take the first few frequency points), arrange the continuous amplitude spectrum to form a "amplitude spectrum number-frequency point number" time-frequency sequence matrix, input it into another independent two-layer LSTM network, and output time sequence features. Using normal samples and known fault samples in historical data, a one-dimensional convolutional GAN network is trained, the generator learns to generate simulated fault feature flow, temperature, noise, and vibration time sequence segments, and after training, a specified number of synthetic fault data is generated, mixed with real data for subsequent model training. After splicing the output features of branch one and branch two, input them into a four-head self-attention layer, each head will learn the importance of different aspects of the features, and finally the outputs of each head are spliced and linearly transformed to obtain a weighted fusion feature vector.

[0030] In S4, the training set and test set data allocation ratio is 7:3~9:1, wherein the data from the industrial time series database and the synthetic data from the GAN network are allocated equally in the training set and the test set. Preferably, if the training set data amount is 90% and the test set data amount is 10%, then the initial training data amount and test data amount from the industrial time series database are 45% and 5% respectively, and the synthetic fault training data amount and test data amount from the GAN network are 45% and 5% respectively.

[0031] In S5, as the increase of the pump station fault data amount covered by the regional data center, the reduction rate of the sample amount called from the industrial time series database is greater than that of the sample amount called from the GAN network, and when the sample amount from the regional data center reaches the sum of the initial sample amounts from the industrial time series database and the GAN network, the local pump station fault diagnosis uses only the samples called from the regional data center. Preferably, the fault data increment imported into the regional data center is counted in units of quarters, 20% of the fault data increment is allocated as the reduction amount of the sample amount called from the GAN network, and 80% of the fault data increment is allocated as the reduction amount of the sample amount called from the industrial time series database, which has the advantage of making the basis samples of local diagnosis come from the regional data as much as possible, so as to continuously improve the accuracy of diagnosis.

Claims

1. A fault diagnosis method for hydraulic engineering pumping stations based on LSTM deep learning algorithm, characterized in that: The system includes hardware and software systems. The hardware system includes flow sensors, temperature sensors, vibration sensors, noise sensors, and edge computing devices. The software system includes a pump station fault diagnosis system and a regional data center covering all pump stations in the adjacent water system. The fault diagnosis method includes: S1: The flow sensor transmits the flow signals from key upstream and downstream nodes of the water pump, the temperature sensor transmits the temperature signals from heat-generating equipment including the water pump drive motor, the vibration sensor transmits the vibration signals from the pump body and shaft system, and the noise sensor transmits the noise signals from the pump body, shaft system, pipelines, and electrical equipment to the data acquisition unit in the edge computing device. S2: The data acquisition unit preprocesses the multi-source heterogeneous data from different sensors. First, it performs wavelet threshold noise reduction on all acquired signals. Then, it normalizes all recorded data based on the minimum-maximum range to eliminate dimensional differences and form a scale uniformity. S3: Input the processed data into a deep learning neural network, including a main network, namely an LSTM network, and an auxiliary network, namely a GAN network. The LSTM network is a two-branch time series feature extraction network, which is used to extract the long-term dependency features and time series evolution rules of multi-dimensional operating parameter time series data. At the same time, it retrieves data on pump station failures from the industrial time series database. The GAN network generates synthetic time series data of simulated failure features based on the current new data of the pump station and uses it to expand the training set. The high-level features output by the two branches of the LSTM network are concatenated and input into a Transformer network to adaptively weight the importance of different sensor signal features. S4: Divide the industrial time series database and the synthetic data generated by the GAN network into training set and test set according to a certain ratio, train and diagnose the LSTM network, and connect two sub-networks in parallel to the output layer. One sub-network outputs the fault type and the other sub-network performs regression estimation of the fault severity. S5: The newly generated data of the current pump station fault diagnosis results are imported into the regional data center for storage, which is then used by all pump stations in the adjacent water system. As the amount of pump station fault data covered by the regional data center increases, the amount of synthetic fault data generated from the industrial time series database and the GAN network is reduced accordingly, gradually and eventually achieving efficient diagnosis within the region.

2. The method for fault diagnosis of hydraulic hub pumping stations based on LSTM deep learning algorithm according to claim 1, characterized in that: The flow sensor, temperature sensor, vibration sensor, and noise sensor mentioned in S1 are connected to the software system via a 5G module serial port. The wavelet threshold noise reduction processing function mentioned in S2 is: ， ω is the wavelet transform coefficient of the noisy signal, and th is the preset threshold. The wavelet transform coefficient ω of the noisy signal is compared with the preset value th. If it is less than th, the output is zero; otherwise, the threshold function is applied to it. The normalization function is: ， F max F min These represent the maximum and minimum values ​​of the physical quantities measured during the signal acquisition period, respectively.

3. The method for fault diagnosis of hydraulic hub pumping stations based on LSTM deep learning algorithm according to claim 2, characterized in that: The mathematical model of the LSTM network described in S3 is: ， In the formula, h t-1 Let x be the output value of the LSTM neural network at time t-1. t Let w be the input value of the LSTM neural network at time t. i w f w o w c and b i b f b o b c These represent the weights and biases of the input gate, forget gate, output gate, and candidate memory, respectively. t f t o t Let c represent the states of the input gate, forget gate, and output gate at time t. t Let σ be the state of the memory unit at time t, σ be the Sigmoid activation function, and tanh() be the hyperbolic tangent function. The LSTM network comprises two independent branches. Branch one receives preprocessed multi-dimensional operating parameter time-series data, including time-series values ​​of flow, temperature, and noise parameters. Branch two is a time-series branch in the vibration frequency domain. It first performs a short-time Fourier transform on the preprocessed vibration signal to generate a time-frequency sequence. The transform function is: ， Subsequently, the LSTM network of this branch extracts the temporal evolution patterns of frequency domain features. The GAN network trains a generator and a discriminator using newly generated data from this pumping station to synthesize time-series data of multi-dimensional operating parameters. The mathematical model of the GAN network is: ， The concatenated Transformer network uses a multi-head attention mechanism to adaptively weight the importance of different sensor signal features.

4. The method for fault diagnosis of hydraulic hub pumping stations based on LSTM deep learning algorithm according to claim 3, characterized in that: In S4, the data allocation ratio between the training set and the test set is 7:3 to 9:1, wherein the data from the industrial time series database and the synthetic data from the GAN network are distributed equally in the training set and the test set.

5. The method for fault diagnosis of hydraulic hub pumping stations based on LSTM deep learning algorithm according to claim 4, characterized in that: In S5, as the amount of pump station fault data covered by the regional data center increases, the rate of decrease in the number of samples called from the industrial time-series database is greater than the rate of decrease in the number of samples called from the GAN network. When the number of samples from the regional data center reaches the sum of the number of samples from the industrial time-series database and the GAN network at the initial stage, local pump station fault diagnosis uses all samples called from the regional data center.

Images