A method and system for determining a water pump unit failure

Abstract

The application provides a kind of water pump unit fault determination method and system, the method is: acquisition water pump unit's multiple groups of vibration signals to be handled;Multiple groups of vibration signals to be handled are preprocessed, and feature matrix to be handled is obtained;The feature matrix to be handled is input into the preset fault diagnosis model to carry out fault identification and fault diagnosis, and the fault diagnosis result of water pump unit is obtained, and the fault diagnosis model is based on sample vibration signal Training improved long short-term memory neural network model is obtained.In the scheme, the characteristics of the vibration signal to be processed are fully extracted by preprocessing to obtain the feature matrix to be processed.The improved long short-term memory neural network model enhances the stability and prediction accuracy of the model, and the fault diagnosis model trained by the improved long short-term memory neural network model is used to identify and diagnose the fault of the feature matrix to be processed, and the fault diagnosis result of the water pump unit is obtained, which can improve the accuracy of the fault diagnosis of the water pump unit.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a water pump unit fault determination method and system. BACKGROUND

[0002] With the accumulation of the running time of the water pump unit, the internal parts of the water pump unit will be damaged, thereby causing a series of abnormal failures.

[0003] The current way to diagnose the fault of the water pump unit is to collect the fault vibration signals of the water pump unit and use the fault vibration signals for fault diagnosis. However, the working conditions of the water pump unit are relatively harsh, which causes the collected fault vibration signals to often have strong noise, nonlinearity and non-stationarity, so that the fault vibration signals cannot be used to accurately diagnose the fault, and the accuracy of the fault diagnosis is poor. SUMMARY

[0004] Therefore, the embodiments of the present application provide a water pump unit fault determination method and system to solve the problem of poor accuracy of the current way to diagnose the fault of the water pump unit.

[0005] To achieve the above object, the embodiments of the present application provide the following technical solutions:

[0006] The first aspect of the embodiments of the present application discloses a water pump unit fault determination method, which comprises:

[0007] Collecting a plurality of groups of vibration signals to be processed of the water pump unit;

[0008] Pretreating the plurality of groups of vibration signals to be processed to obtain a feature matrix to be processed;

[0009] Inputting the feature matrix to be processed into a preset fault diagnosis model for fault recognition and fault diagnosis to obtain a fault diagnosis result of the water pump unit, wherein the fault diagnosis model is obtained based on a sample vibration signal training an improved long short-term memory neural network model, and the improved long short-term memory neural network model at least contains a layer normalization layer.

[0010] Preferably, the pretreating the plurality of groups of vibration signals to be processed to obtain the feature matrix to be processed comprises:

[0011] For each group of the vibration signals to be processed, performing sliding window slicing processing on the vibration signals to be processed to obtain corresponding vibration data to be processed;

[0012] Extracting specified features of the vibration data to be processed, wherein the specified features at least include time domain features, frequency domain features and time-frequency domain features;

[0013] construct a feature matrix using all the specified features of the vibration data to be processed;

[0014] normalize the feature matrix to obtain a processed feature matrix.

[0015] Preferably, the process of training the improved long short-term memory neural network model based on the sample vibration signal to obtain the fault diagnosis model comprises:

[0016] The pre-acquired sample vibration signal is pre-processed to obtain a sample feature matrix, wherein the sample vibration signal carries a pre-added fault label.

[0017] The training set and the test set are divided from the sample feature matrix.

[0018] Based on the preset classification loss function, the adaptive matrix estimation optimizer, the training set and the test set, the improved long short-term memory neural network model is trained until the improved long short-term memory neural network model converges to obtain a fault diagnosis model.

[0019] Preferably, it further comprises:

[0020] Output the fault diagnosis result, wherein the fault diagnosis result comprises a first mark indicating that the water pump unit is fault-free, or a second mark indicating that the water pump unit has a fault and the fault type of the water pump unit.

[0021] Preferably, the plurality of groups of vibration signals to be processed of the water pump unit are collected, comprising:

[0022] The plurality of groups of vibration signals to be processed of the water pump unit are collected by using vibration sensors arranged at specified positions of the water pump unit.

[0023] The second aspect of the embodiment of the application discloses a system for determining the fault of a water pump unit, comprising:

[0024] The acquisition unit is configured to acquire a plurality of groups of vibration signals to be processed of a water pump unit;

[0025] The preprocessing unit is configured to pre-process the plurality of groups of vibration signals to be processed to obtain a processed feature matrix.

[0026] The processing unit is configured to input the processed feature matrix into a preset fault diagnosis model for fault recognition and fault diagnosis to obtain a fault diagnosis result of the water pump unit, wherein the fault diagnosis model is obtained by training an improved long short-term memory neural network model based on a sample vibration signal, and the improved long short-term memory neural network model comprises at least a layer normalization layer.

[0027] Preferably, the preprocessing unit comprises:

[0028] a processing module, configured to perform sliding window slicing processing on the vibration signals to be processed to obtain vibration data to be processed for each group of the vibration signals to be processed;

[0029] an extraction module, configured to extract specified features of the vibration data to be processed, the specified features including at least time domain features, frequency domain features and time-frequency domain features;

[0030] a construction module, configured to construct a feature matrix by using the specified features of all the vibration data to be processed;

[0031] a normalization module, configured to normalize the feature matrix to obtain a processed feature matrix.

[0032] Preferably, the processing unit comprises:

[0033] a preprocessing module, configured to perform preprocessing on sample vibration signals collected in advance to obtain a sample feature matrix, wherein the sample vibration signals carry pre-added fault labels;

[0034] a division module, configured to divide the sample feature matrix to obtain a training set and a test set;

[0035] a training module, configured to train an improved long short-term memory neural network model based on a preset classification loss function, an adaptive matrix estimator optimizer, the training set and the test set until the improved long short-term memory neural network model converges, so as to obtain a fault diagnosis model.

[0036] Preferably, the processing unit further comprises:

[0037] an output unit, configured to output the fault diagnosis result, wherein the fault diagnosis result includes a first mark indicating that the water pump unit is fault-free, or a second mark indicating that the water pump unit has a fault and a fault type of the water pump unit.

[0038] Preferably, the collection unit is specifically configured to collect the vibration signals to be processed of the water pump unit by using vibration sensors arranged at specified positions of the water pump unit.

[0039] Based on the water pump unit fault determination method and system provided by the above embodiment of the application, the method is: collecting multiple groups of vibration signals to be processed of the water pump unit; pre-processing the multiple groups of vibration signals to be processed to obtain a feature matrix to be processed; inputting the feature matrix to be processed into a preset fault diagnosis model for fault recognition and fault diagnosis to obtain a fault diagnosis result of the water pump unit, the fault diagnosis model being obtained based on a sample vibration signal training an improved long short-term memory neural network model, and the improved long short-term memory neural network model at least comprising a layer normalization layer. In the scheme, the features of the vibration signals to be processed are fully extracted through pre-processing to obtain the feature matrix to be processed. The improved long short-term memory neural network model enhances the stability and prediction accuracy of the model. The fault diagnosis model trained by the improved long short-term memory neural network model is used for fault recognition and fault diagnosis on the feature matrix to be processed, and the fault diagnosis result of the water pump unit is obtained, which can improve the accuracy of the fault diagnosis of the water pump unit. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0041] Figure 1 The flowchart of the water pump unit fault determination method provided by the embodiment of the present application;

[0042] Figure 2 The flowchart of training the fault diagnosis model provided by the embodiment of the present application;

[0043] Figure 3 The structural block diagram of the water pump unit fault determination system provided by the embodiment of the present application. DETAILED DESCRIPTION

[0044] 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 some embodiments of the present application, not all 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.

[0045] In this application, the terms "comprise", "contain", or any other variant thereof, are intended to cover non-exclusive inclusions, so that a process, method, article, or device that includes a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such a process, method, article, or device. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of additional identical elements in the process, method, article, or device that includes the element.

[0046] As can be known from the background art, in diagnosing the fault of a water pump unit, the fault vibration signal of the water pump unit is currently collected, and the fault vibration signal is used for fault diagnosis. However, the working conditions of the water pump unit are relatively harsh, which causes the collected fault vibration signal to often have strong noise, nonlinearity, and non-stationarity, and the fault vibration signal cannot be used to accurately diagnose the fault, and the accuracy of fault diagnosis is poor.

[0047] Therefore, the embodiments of the present application provide a water pump unit fault determination method and system, which fully extracts the features of the to-be-processed vibration signal through preprocessing to obtain a to-be-processed feature matrix. The improved long short-term memory neural network model enhances the stability and prediction accuracy of the model, and a fault diagnosis model trained by the improved long short-term memory neural network model is used to perform fault recognition and fault diagnosis on the to-be-processed feature matrix to obtain a fault diagnosis result of the water pump unit, so as to improve the accuracy of fault diagnosis of the water pump unit.

[0048] Referring to Figure 1 , a flowchart of a water pump unit fault determination method provided by the embodiments of the present application is shown, and the determination method comprises:

[0049] Step S101: Collecting a plurality of groups of to-be-processed vibration signals of a water pump unit.

[0050] In the process of specifically implementing step S101, a vibration sensor arranged at a specified position on the water pump unit is used to collect a plurality of groups of to-be-processed vibration signals of the water pump unit; specifically, a continuous vibration signal is collected by using the vibration sensor to obtain a plurality of groups of to-be-processed vibration signals.

[0051] In some embodiments, the vibration sensor is arranged at a specified position in the vertical direction of the water pump unit.

[0052] It should be noted that in the process of collecting the continuous vibration signal by using the vibration sensor, the sampling point number of the vibration signal is related to the frequency of the vibration sensor; for example, when the frequency of the vibration sensor is 50kHZ, the sampling point number is 50000 data collected in 1 second.

[0053] Each group of vibration signals to be processed can be denoted as "x(i), i = 1, 2, 3..., N"; wherein, N is the number of sampling points.

[0054] Step S102: Preprocessing a plurality of groups of vibration signals to be processed to obtain a feature matrix to be processed.

[0055] In the process of implementing step S102, for each group of vibration signals to be processed, the sliding window slicing processing is performed on the group of vibration signals to be processed to obtain the corresponding vibration data to be processed. Specifically, the sliding window slicing processing is performed on each group of vibration signals x(i), and the sliding overlap degree is a, to obtain the vibration data F(t) corresponding to each group of vibration signals x(i). The specific content of F(t) is shown in formula (1).

[0056] F(t) = X[t:t+n] (1)

[0057] In formula (1), t = an x j; j = 0, 1, 2..., m; Let z = N / an, represents the maximum integer not exceeding z; n is the sliding window slicing width, and an is the sliding window step length; after the sliding window slicing processing is performed on x(i), X[t:t+n] can be obtained, which is used to represent a group of data such as x(t) to x(t+n).

[0058] In the above manner, the vibration data to be processed corresponding to each group of vibration signals to be processed can be determined. For the vibration data to be processed corresponding to each group of vibration signals to be processed, the specified features of the vibration data to be processed are extracted, and the specified features at least include time domain features, frequency domain features and time-frequency domain features.

[0059] Specifically, the specified features of the vibration data to be processed are extracted by using signal processing methods such as Hilbert transform, wavelet transform and empirical mode decomposition; the specified features of the vibration data to be processed corresponding to each group of vibration signals to be processed can be extracted.

[0060] The specified features of all the vibration data to be processed are used to construct a feature matrix; that is, the specified features of the vibration data to be processed corresponding to each group of vibration signals to be processed are used to construct a feature matrix. Specifically, the specified features of all the vibration data to be processed are arranged in sequence to form a feature matrix.

[0061] The feature matrix composed of the specified features of the vibration data to be processed is normalized to obtain a feature matrix to be processed. Specifically, according to formula (2), the min-max normalization method is used to normalize the feature matrix to obtain a feature matrix to be processed.

[0062]

[0063] In formula (2), x i_trans is the normalized i-th data in the to-be-processed vibration data, x i is the i-th data in the to-be-processed vibration data, min(x j ) is the minimum data in the to-be-processed vibration data, and max(x j ) is the maximum data in the to-be-processed vibration data.

[0064] Step S103: inputting the to-be-processed feature matrix into a preset fault diagnosis model to perform fault recognition and fault diagnosis, so as to obtain a fault diagnosis result of the water pump unit.

[0065] It should be noted that the fault diagnosis model is obtained by training an improved Long Short-Term Memory (LSTM) model based on sample vibration signals, and the improved LSTM model at least comprises a Layer Normalization (LN) layer, so the improved LSTM model can be referred to as an LN-LSTM model.

[0066] It should be further noted that the fault diagnosis model can determine whether the water pump unit has a fault based on the to-be-processed feature matrix; when it is determined that the water pump unit has a fault, the fault diagnosis model can determine a probability that the fault belongs to different fault types, and the fault type with the maximum probability is the fault type of the water pump unit.

[0067] In the process of specifically implementing step S103, the to-be-processed feature matrix is input into the preset fault diagnosis model to perform fault recognition and fault diagnosis, so as to obtain a fault diagnosis result.

[0068] In some embodiments, after the fault diagnosis model performs fault recognition and fault diagnosis, if the water pump unit does not have a fault, the fault diagnosis result determined by the fault diagnosis model comprises a first mark indicating that the water pump unit has no fault. If the water pump unit has a fault, the fault diagnosis result determined by the fault diagnosis model comprises a second mark indicating that the water pump unit has a fault and a fault type of the water pump unit.

[0069] In some embodiments, the fault diagnosis result is output, for example, the fault diagnosis result is displayed on a display; wherein the fault diagnosis result comprises the first mark indicating that the water pump unit has no fault, or the second mark indicating that the water pump unit has a fault and the fault type of the water pump unit.

[0070] The user can know whether the water pump unit has a fault through the fault diagnosis result, and when the water pump unit has a fault, the user can further know the fault type of the water pump unit.

[0071] In some embodiments, the fault diagnosis model is obtained by training an improved long short-term memory neural network model based on sample vibration signals; to better explain how to train the fault diagnosis model, the training process of the fault diagnosis model is explained as follows.

[0072] Explanation of the process of training the fault diagnosis model:

[0073] Obtain vibration signals collected under different fault types of the water pump unit to obtain multiple groups of sample vibration signals; pre-process the multiple groups of sample vibration signals to obtain a sample feature matrix, wherein the sample vibration signals carry pre-added fault labels; the fault label of each group of sample vibration signals is used to indicate whether the group of sample vibration signals is a normal sample or a fault sample (the fault sample is marked with the corresponding fault type).

[0074] Specifically, each group of sample vibration signals is subjected to sliding window slicing processing to obtain data samples corresponding to each group of sample vibration signals; the fault labels of the data samples are subjected to one-hot encoding.

[0075] It can be understood that in the process of one-hot encoding the fault labels of the data samples, N-bit state registers are used to one-hot encode the fault labels of N types of data samples, each type of data sample has its own independent register bit, and only one bit is valid at any time (or each type of data sample corresponds to only one fault label).

[0076] Extract specified features (time domain features, frequency domain features, and time-frequency domain features) of each data sample, and group the specified features of each data sample in sequence to form a feature matrix; according to formula (2), normalize the feature matrix formed based on the specified features of the data samples to obtain a sample feature matrix.

[0077] According to a preset proportion, divide the training set and the test set from the sample feature matrix; specifically, for the same type of data sample in the sample feature matrix, divide the same type of data sample into the training set and the test set according to the preset proportion to obtain a sample feature matrix corresponding to the training set and a sample feature matrix corresponding to the test set.

[0078] For example: for the same type of data sample in the sample feature matrix, divide 70% of the data samples in the same type of data sample into the training set, and divide 30% of the data samples in the same type of data sample into the test set.

[0079] Based on the preset classification loss function, the adaptive matrix estimation optimizer, the training set and the test set, the improved long short-term memory neural network model is trained until the improved long short-term memory neural network model converges, so as to obtain the fault diagnosis model.

[0080] Specifically, the LN-LSTM model is constructed; the sample feature matrix corresponding to the training set and the sample feature matrix corresponding to the test set are input into the constructed LN-LSTM model for training, and the LN-LSTM model is trained to convergence to obtain the fault diagnosis model.

[0081] In the process of training the LN-LSTM model by using the sample feature matrix, the LN layer transforms the input value of each neuron of the hidden layer by the following formula (3) using the sample sliding mean E(X) and the sample sliding variance Var(X) of each layer.

[0082]

[0083] In formula (3), X is a specified feature used to compose the sample feature matrix, β is a learnable parameter, γ is a linear operation on the input feature, and ε is used to prevent the denominator from being 0 in the calculation process.

[0084] It should be noted that, compared with LSTM, LN-LSTM also includes input gate, forget gate and output gate, so the LN process shown in the above formula (3) can be simplified as LN(·; γ).

[0085] In the training process, the output value of the update gate at time t is as formula (4).

[0086]

[0087] In formula (4), W u is the weight parameter of the input value X <t>< / t> and the historical data stream h <t-1>< / t-1> input to the update gate at the current time, b u is the bias value of the input value X <t>< / t> and the historical data stream h <t-1>< / t-1> input to the update gate at the current time, σ(·) is a sigmoid function, and γ h is γ in the above LN(·; γ).

[0088] In the training process, the output value of the forget gate at time t is as formula (5).

[0089]

[0090] In formula (5), W f is the weight parameter of the input value X <t>< / t>The historical data stream h <t-1>< / t-1> The weight parameter b input into the process of the forget gate f The input value X <t〉 The historical data stream h <t-1〉 The bias value input into the process of the forget gate

[0091] In the training process, the memory cell state value at time t is as formula (6).

[0092]

[0093] In formula (6), And The determination method is shown in formula (4) and formula (5) above, The input data at time t after the hyperbolic tangent activation, The determination method is shown in formula (7).

[0094]

[0095] In formula (7), W c And b c Respectively represent the weight parameter value and the bias value in the hyperbolic tangent layer, and tanh(·) is the hyperbolic tangent activation function.

[0096] In the training process, the memory cell output value is as formula (8), and the output value of the output gate is as formula (9).

[0097]

[0098]

[0099] In formula (8) and formula (9), W o The input value X <t〉 The historical data stream h <t-1〉 The weight parameter b input into the process of the output gate o The input value X <t>< / t> The historical data stream h <t-1>< / t-1> The bias value input into the process of the output gate

[0100] It can be understood that, for the problem of training the model for the data sample of the unbalanced fault type, an improved classification loss function (Focal loss) as shown in formula (10) is introduced in the training process.

[0101] FL(p t )=-a t (1-p t ) γlog(p t ) (10)

[0102] In the formula (10), a can be used to control the sharing weight of positive and negative sample pairs on the loss value; γ is a modulation coefficient, which can reduce the weight of easy-to-classify samples, so that the model training is more focused on difficult-to-classify samples; p t The specific value of y is as formula (11).

[0103]

[0104] In the formula (11), y is the accuracy of the training set on classification, y = 1 is used to represent correct classification, and p is a preset value.

[0105] In the process of training the LN-LSTM model, in order to realize more efficient optimization of the network structure, an adaptive moment estimation optimizer with momentum term and weight decay (referred to as Adam optimizer) can be used in the training process, and the learning rate lr gradually decreases with the iteration steps; The principle of adaptive moment estimation optimizer is as follows formula (12) to formula (14).

[0106] m t = beta1 * m t-1 + (1 - beta1) * g (12)

[0107] v t = beta2 * v t-1 + (1 - beta2) * g * g (13)

[0108]

[0109] In formula (12) to formula (14), m t is the first moment of gradient g (that is, the expectation of gradient g), beta1 is the first moment decay coefficient, g is the gradient obtained by deriving the optimization objective function with respect to X, v t is the second moment of gradient g (that is, the expectation of g * g), beta2 is the second moment decay coefficient, and ε is used to prevent the denominator from being 0 in the calculation process.

[0110] It should be noted that formula (12) calculates the first exponential smoothing value of the historical gradient, which is used to obtain the gradient value with momentum; Formula (13) calculates the first exponential smoothing value of the square of the historical gradient, which is used to obtain the learning rate weight parameter of each weight parameter; Formula (14) calculates the update value of the variable, which is proportional to the first exponential smoothing value of the historical gradient and inversely proportional to the first exponential smoothing value of the square of the historical gradient.

[0111] During the training process, the output of the LN-LSTM network layer is input into a softmax classification layer, and the probability values of the data sample belonging to each fault type are calculated through a softmax function; the softmax function is as shown in formula (15).

[0112] FD j =soft max(wx+b) j (15)

[0113] In formula (15), FD j is the predicted probability distribution of different fault types, w is a weight, b is a bias, and x is the output of the LN-LSTM network layer.

[0114] The above is the related description of training LN-LSTM; in order to effectively avoid overfitting and underfitting during the training process, K-fold cross-validation is used to verify the training process, so as to obtain a better fault diagnosis model.

[0115] In the embodiment of the application, the characteristics of the vibration signal to be processed are fully extracted through preprocessing to obtain a feature matrix to be processed. The improved long short-term memory neural network model enhances the stability and prediction accuracy of the model. The fault diagnosis model trained by the improved long short-term memory neural network model is used for fault recognition and fault diagnosis of the feature matrix to be processed, and the fault diagnosis result of the water pump unit is obtained, which can improve the accuracy of the fault diagnosis of the water pump unit.

[0116] The above embodiment of the application Figure 1 The process of training the LN-LSTM model in step S103 is better explained by Figure 2 The flowchart for training the fault diagnosis model is shown in the figure; Figure 2 comprises the following steps:

[0117] Step S201: Collecting multiple groups of sample vibration signals.

[0118] In the process of specifically implementing step S201, the vibration sensor arranged in the vertical direction of the water pump unit is used to collect vibration signals of different fault types from the pump end to obtain multiple groups of sample vibration signals.

[0119] It should be noted that the sample vibration signals of different fault types are vibration signals collected under different states of the water pump unit; the states of the water pump unit include but are not limited to: normal state, rotor imbalance, rotor misalignment, rotor rubbing, rolling bearing cage fault, rolling bearing inner ring fault, rolling bearing outer ring fault, rolling bearing rolling element fault, rotor fault, mechanical looseness fault, etc.

[0120] That is, the fault type is divided into normal, rotor imbalance, rotor misalignment, rotor rubbing, rolling bearing cage failure, rolling bearing inner ring failure, rolling bearing outer ring failure, rolling bearing rolling element failure, rotor failure, mechanical looseness failure.

[0121] Step S202: performing sliding window slicing processing on each group of sample vibration signals to obtain data samples corresponding to each group of sample vibration signals.

[0122] It should be noted that in the process of performing sliding window slicing processing, the sliding window size is set to 4096, and the overlap degree is 50%;

[0123] In some embodiments, the fault labels of each data sample are one-hot encoded.

[0124] Step S203: extracting specified features of each data sample to form a feature matrix, and normalizing the feature matrix to obtain a sample feature matrix.

[0125] In the process of implementing step S203, the signal processing methods such as Hilbert transform, wavelet transform and empirical mode decomposition are used to obtain the expression of the data sample in different spaces, and then the statistical knowledge is used to calculate the statistical parameters such as signal peak value, mean value, variance, margin, kurtosis and skewness, to obtain the specified features (time domain features, frequency domain features and time-frequency domain features) of the data sample.

[0126] The specified features of each data sample are sequentially combined to form a feature matrix, and the feature matrix composed of the specified features based on the data samples is normalized based on formula (2) to obtain a sample feature matrix.

[0127] In some embodiments, for the same type of data sample in the sample feature matrix, 70% of the data samples in the same type of data sample are divided into the training set, and 30% of the data samples in the same type of data sample are divided into the test set.

[0128] Step S204: constructing an LN-LSTM model.

[0129] Step S205: initializing model parameters.

[0130] Step S206: training the LN-LSTM model using the training set and the test set to obtain a fault diagnosis model.

[0131] In the process of implementing step S206, the sample feature matrix corresponding to the training set and the sample feature matrix corresponding to the test set are input into the constructed LN-LSTM model for training, wherein the probability values of the data samples belonging to each fault type can be obtained after being processed by a softmax classifier.

[0132] It can be understood that in the training process of the LN-LSTM model, the model training parameters can be set according to the content shown in Table 1; it should be noted that Table 1 is only used for illustration, and the specific model training parameters can be set according to the actual situation.

[0133] Table 1:

[0134] Loss function Focal Loss Optimizer Adam with weight decay Learning rate Initial value 0.001, gradually decreased with the number of iterations Number of iterations 150 Batch Size 32 Number of units in the output layer 10 k-fold cross-validation k＝10

[0135] To effectively avoid overfitting and underfitting in the training process, K-fold cross-validation is used to verify the training process; specifically, the validation set is used to verify the fault recognition accuracy of the trained fault diagnosis model.

[0136] For example: in the training set, 10 (K=10) fold cross-validation is used to verify the fault recognition accuracy of the trained fault diagnosis model, and the fault recognition accuracy of the fault diagnosis model is shown in Table 2, which is only used for illustration.

[0137] Table 2:

[0138]

[0139] As can be seen from the above Table 2, the fault diagnosis model trained by the embodiment of the present application can accurately identify whether the water pump unit has failed and identify the fault type.

[0140] Corresponding to the above-mentioned method for determining the fault of the water pump unit, referring to Figure 3 The embodiment of the present application also provides a structural diagram of a system for determining the fault of the water pump unit, which comprises: an acquisition unit 301, a preprocessing unit 302 and a processing unit 303.

[0141] The acquisition unit 301 is used to acquire a plurality of groups of vibration signals to be processed of the water pump unit.

[0142] In a specific implementation, the acquisition unit 301 is specifically configured to: acquire the plurality of groups of vibration signals to be processed of the water pump unit by using the vibration sensor arranged at the specified position of the water pump unit.

[0143] The preprocessing unit 302 is configured to preprocess the plurality of groups of vibration signals to be processed to obtain a feature matrix to be processed.

[0144] The processing unit 303 is configured to input the feature matrix to be processed into a preset fault diagnosis model for fault recognition and fault diagnosis to obtain a fault diagnosis result of the water pump unit, and the fault diagnosis model is obtained based on a sample vibration signal and an improved long short-term memory neural network model, and the improved long short-term memory neural network model at least comprises a layer normalization layer.

[0145] In the embodiment of the present application, the characteristics of the vibration signals to be processed are fully extracted through preprocessing to obtain a feature matrix to be processed. The improved long short-term memory neural network model enhances the stability and prediction accuracy of the model. The fault diagnosis model trained by the improved long short-term memory neural network model is used to perform fault recognition and fault diagnosis on the feature matrix to be processed, and the fault diagnosis result of the water pump unit is obtained, which can improve the accuracy of fault diagnosis of the water pump unit.

[0146] Preferably, in combination with the content shown, the preprocessing unit 302 includes a processing module, an extraction module, a construction module and a normalization module, and the execution principles of each module are as follows: Figure 3 As shown, the preprocessing unit 302 includes a processing module, an extraction module, a construction module and a normalization module, and the execution principles of each module are as follows:

[0147] The processing module is configured to perform sliding window slicing processing on the vibration signals to be processed for each group of vibration signals to be processed to obtain corresponding vibration data to be processed.

[0148] The extraction module is configured to extract specified features of the vibration data to be processed, and the specified features at least include time domain features, frequency domain features and time-frequency domain features.

[0149] The construction module is configured to construct a feature matrix using the specified features of all vibration data to be processed.

[0150] The normalization module is configured to normalize the feature matrix to obtain a feature matrix to be processed.

[0151] Preferably, in combination with the content shown, the processing unit 303 includes a preprocessing module, a division module and a training module; the execution principles of each module are as follows: Figure 3 As shown, the processing unit 303 includes a preprocessing module, a division module and a training module; the execution principles of each module are as follows:

[0152] The preprocessing module is configured to preprocess the sample vibration signals collected in advance to obtain a sample feature matrix, wherein the sample vibration signals carry a pre-added fault label.

[0153] The division module is configured to divide the sample feature matrix to obtain a training set and a test set.

[0154] The training module is configured to train the improved long short-term memory neural network model based on a preset classification loss function, an adaptive matrix estimator optimizer, the training set and the test set until the improved long short-term memory neural network model converges to obtain a fault diagnosis model.

[0155] Preferably, in combination with the content shown, the determination system includes: Figure 3 As shown, the determination system includes:

[0156] The output unit is configured to output a fault diagnosis result, wherein the fault diagnosis result comprises a first mark indicating that the water pump unit is fault-free, or a second mark indicating that the water pump unit has a fault and a fault type of the water pump unit.

[0157] To sum up, the embodiment of the present application provides a water pump unit fault determination method and system, the features of the vibration signal to be processed are fully extracted through preprocessing to obtain a feature matrix to be processed. The improved long short-term memory neural network model enhances the stability and prediction accuracy of the model. The fault diagnosis model trained by the improved long short-term memory neural network model is used to perform fault recognition and fault diagnosis on the feature matrix to be processed, and the fault diagnosis result of the water pump unit is obtained, which can improve the accuracy of the fault diagnosis of the water pump unit.

[0158] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, it is described more simply, and the related parts can be referred to the part of the method embodiment. The above-described system and system embodiment are only illustrative, and the units described as separate components can be or can not be physically separated, and the components displayed as units can be or can not be physical units, that is, they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to the actual needs. Those skilled in the art can understand and implement without creative labor.

[0159] The skilled person can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0160] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and, while certain modifications are discussed, it is desired to be protected in accordance with the scope of this application, which is defined in the following claims.

Claims

1. A method of determining a failure of a water pump assembly, the method comprising: The method comprises: Collecting a plurality of groups of vibration signals to be processed of a water pump unit; Pretreating the plurality of groups of vibration signals to be processed to obtain a feature matrix to be processed; The feature matrix to be processed is input into a preset fault diagnosis model for fault recognition and fault diagnosis to obtain a fault diagnosis result of the water pump unit, the fault diagnosis model is obtained based on a sample vibration signal training an improved long short-term memory neural network model, the improved long short-term memory neural network model at least contains a layer normalization layer, wherein the layer normalization layer utilizes a sample sliding mean E(X) and a sample sliding variance Var(X) of each layer to transform an input value of each neuron of a hidden layer according to a formula , wherein X is a specified feature used to compose the feature matrix, is a learnable parameter, is a linear operation on the input feature, is used to prevent the denominator from being 0 in the calculation process, and LN represents the layer normalization layer.

2. The method of claim 1, wherein, Pretreating the plurality of groups of vibration signals to be processed to obtain a feature matrix to be processed, comprising: For each group of the vibration signals to be processed, performing sliding window slicing processing on the vibration signals to be processed to obtain corresponding vibration data to be processed; Extracting specified features of the vibration data to be processed, the specified features at least including time domain features, frequency domain features and time-frequency domain features; Using the specified features of all the vibration data to be processed to construct a feature matrix; Normalizing the feature matrix to obtain a feature matrix to be processed.

3. The method of claim 1, wherein, The process of training the improved long short-term memory neural network model based on sample vibration signals to obtain the fault diagnosis model comprises: Pretreating pre-collected sample vibration signals to obtain a sample feature matrix, wherein the sample vibration signals carry pre-added fault labels; Dividing the sample feature matrix to obtain a training set and a test set; Training the improved long short-term memory neural network model based on a preset classification loss function, an adaptive matrix estimator optimizer, the training set and the test set until the improved long short-term memory neural network model converges to obtain a fault diagnosis model.

4. The method of claim 1, wherein, Further comprising: Outputting the fault diagnosis result, wherein the fault diagnosis result includes a first mark indicating that the water pump unit is fault-free, or a second mark indicating that the water pump unit has a fault and a fault type of the water pump unit.

5. The method of claim 1, wherein, Collecting a plurality of groups of vibration signals to be processed of a water pump unit comprises: Using vibration sensors arranged at specified positions of the water pump unit to collect a plurality of groups of vibration signals to be processed of the water pump unit.

6. A system for determining a failure of a water pump assembly, comprising: The system comprises: A collection unit configured to collect a plurality of groups of vibration signals to be processed of a water pump unit; A pretreatment unit configured to pretreat the plurality of groups of vibration signals to be processed to obtain a feature matrix to be processed; The processing unit is configured to input the to-be-processed feature matrix into a preset fault diagnosis model for fault recognition and fault diagnosis to obtain a fault diagnosis result of the water pump unit. The fault diagnosis model is obtained based on a sample vibration signal and an improved long short-term memory neural network model. The improved long short-term memory neural network model at least comprises a layer normalization layer. The layer normalization layer utilizes a sample sliding mean E(X) and a sample sliding variance Var(X) of each layer to transform an input value of each neuron of a hidden layer according to a formula , wherein X is a specified feature used to compose the feature matrix, is a learnable parameter, is a linear operation on the input feature, is used to prevent a phenomenon that a denominator is 0 from occurring in a calculation process, and LN represents the layer normalization layer.

7. The system of claim 6, wherein, The pretreatment unit comprises: A processing module configured to, for each group of the vibration signals to be processed, perform sliding window slicing processing on the vibration signals to be processed to obtain corresponding vibration data to be processed; An extraction module configured to extract specified features of the vibration data to be processed, the specified features at least including time domain features, frequency domain features and time-frequency domain features; A construction module configured to use the specified features of all the vibration data to be processed to construct a feature matrix; A normalization module configured to normalize the feature matrix to obtain a feature matrix to be processed.

8. The system of claim 6, wherein, The processing unit comprises: A pretreatment module configured to pretreat pre-collected sample vibration signals to obtain a sample feature matrix, wherein the sample vibration signals carry pre-added fault labels; A division module configured to divide the sample feature matrix to obtain a training set and a test set; The training module is configured to train the improved long short-term memory neural network model based on a preset classification loss function, an adaptive matrix estimation optimizer, the training set and the test set until the improved long short-term memory neural network model converges, so as to obtain the fault diagnosis model.

9. The system of claim 6, wherein, Further comprising: The output unit is configured to output the fault diagnosis result, wherein the fault diagnosis result comprises a first mark indicating that the water pump unit is fault-free, or a second mark indicating that the water pump unit has a fault and a fault type of the water pump unit.

10. The system of claim 6, wherein, The collection unit is specifically configured to collect a plurality of groups of vibration signals to be processed of the water pump unit by using vibration sensors arranged at designated positions of the water pump unit.

Images