Impeller imbalance fault detection method based on vibration signal analysis

Abstract

The invention relates to the technical field of engine detection, and discloses an impeller imbalance fault detection method based on vibration signal analysis, which comprises the following steps: S1, collecting impeller vibration signals by arranging a plurality of three-axis acceleration sensors at key parts such as an impeller and a bearing; s2, preprocessing the collected vibration signals, including removing noise by using wavelet transform and standardizing the signals; and S3, decomposing the vibration signal into a plurality of intrinsic mode functions through empirical mode decomposition. According to the method, technologies such as multi-channel vibration signal acquisition, empirical mode decomposition and wavelet packet transformation feature extraction, multi-task learning classification and identification, generalization ability enhancement of a generative adversarial network and fault trend prediction of a long and short-term memory network are combined; the accuracy, the mode distinguishing capability and the dynamic trend prediction capability of impeller imbalance fault detection are remarkably improved, and meanwhile, the real-time performance and the intelligent early warning function of fault monitoring are achieved through edge calculation.

Description

Technical Field

[0001] The invention relates to the technical field of engine detection, and in particular to an impeller imbalance fault detection method based on vibration signal analysis. Background Art

[0002] Impeller imbalance failure is the most common fault in rotating machinery and has the most significant impact on the stability and safety of equipment operation. This fault is usually caused by manufacturing errors, uneven materials, wear or installation deviations, and manifests as an abnormal increase in vibration amplitude during equipment operation, which may cause noise, fatigue damage, and chain damage to components such as bearings and gears. In industrial sites, early detection and accurate identification of impeller imbalance failures are of great significance to ensuring equipment safety, extending service life, and reducing maintenance costs.

[0003] Traditional impeller imbalance fault detection methods mainly rely on frequency domain analysis of vibration signals (such as Fourier transform) or model matching based on empirical rules. However, these methods face significant limitations under complex working conditions:

[0004] Insufficient adaptability to complex working conditions: When multiple fault modes coexist, the characteristics of impeller imbalance faults are easily confused with other faults (such as bearing faults or gear wear), leading to misdiagnosis or missed diagnosis.

[0005] Limited feature extraction capability: Vibration signals are usually non-stationary signals. Traditional frequency domain analysis is difficult to capture their time-varying characteristics, and the signal features are not comprehensive and accurate enough.

[0006] Data scarcity and limited generalization ability: In actual operation, fault samples are scarce and the operating conditions are complex. Traditional methods rely on fixed feature extraction patterns and limited training samples, which are difficult to cope with unseen data and dynamically changing operating conditions.

[0007] Lack of real-time performance: Most existing methods rely on complex data transmission and centralized processing, which cannot meet the needs of industrial sites for real-time monitoring and rapid fault response.

[0008] In recent years, with the development of sensor technology, signal processing technology and artificial intelligence technology, fault detection methods based on vibration signal analysis have gradually become a research hotspot. By collecting vibration data through multi-channel sensors and combining signal decomposition, feature extraction and deep learning technology, accurate identification and trend prediction of fault modes under complex working conditions can be achieved. However, existing methods still have room for improvement in terms of feature extraction, classification accuracy, real-time response and model generalization ability.

[0009] Based on this, a method for detecting impeller imbalance fault based on vibration signal analysis is proposed. Summary of the invention

[0010] In view of the shortcomings of the prior art, the present invention provides an impeller imbalance fault detection method based on vibration signal analysis, which solves the problems of low detection accuracy, insufficient pattern distinction capability, weak real-time performance and poor generalization ability of impeller imbalance faults under complex working conditions.

[0011] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for detecting impeller imbalance fault based on vibration signal analysis, comprising the following steps:

[0012] S1. Multiple three-axis acceleration sensors are placed at key locations such as impellers and bearings to collect impeller vibration signals.

[0013] S2. Preprocessing the collected vibration signal, including removing noise using wavelet transform and performing signal standardization;

[0014] S3, decomposing the vibration signal into multiple intrinsic mode functions by empirical mode decomposition and performing wavelet packet transform on each intrinsic mode function component;

[0015] S4. Use convolutional neural network to extract local features in vibration signals, and classify and identify impeller unbalance faults and other fault modes through a multi-task learning framework;

[0016] S5, using a generative adversarial network to generate a vibration signal similar to the actual fault;

[0017] S6. Use long short-term memory network to analyze the vibration signal in time series and predict the development trend of impeller imbalance fault;

[0018] S7. Monitor faults in real time and issue warning signals when a fault occurs or is predicted to occur.

[0019] Preferably, the signal standardization in step S2 includes processing the signal as follows:

[0020]

[0021] Among them, x(t) is the original signal, μ is the signal mean, σ is the standard deviation, and x norm (t) is the normalized signal.

[0022] Preferably, the formula for decomposing the vibration signal in step S3 is as follows:

[0023]

[0024] Among them, x(t) is the vibration signal, r N (t) is the residual signal, IMF i (t) is the i-th intrinsic mode function.

[0025] Preferably, the multi-task learning framework in step S4 includes:

[0026] Extract common features from signals through shared convolutional layers;

[0027] Set up an independent task-specific layer for each task to accurately distinguish different failure modes;

[0028] Through the loss function optimization framework, combined with the loss function of each task:

[0029]

[0030] Among them, L taski is the loss of the i-th task, λ i is the task weight coefficient, L gen is the loss term of the generated adversarial network, and α is its weight coefficient.

[0031] Preferably, the generation of the adversarial network in step S5 includes:

[0032] The generator generates a vibration signal similar to the actual fault signal through a convolutional neural network. The generator input is random noise z, and the fault signal G(z) is generated through convolutional layer and deconvolution layer operations.

[0033] The discriminator determines the authenticity of the input signal and determines whether the signal x is a real signal by outputting the probability D(x). The loss function of the discriminator is:

[0034]

[0035] Among them, D(x) is the output of the discriminator for the real signal, G(z) is the false signal generated by the generator, and L dis is the discriminator loss;

[0036] The loss function of the generator is:

[0037]

[0038] Among them, D(G(z)) is the judgment of the discriminator on the generated signal, which indicates the probability that the generated signal is a true signal, and L gen is the generator loss;

[0039] During the training process, the generator and the discriminator are trained by alternating optimization, and the goal of the discriminator is to minimize the loss function L dis , and the goal of the generator is to minimize the loss function L gen , so that the generated signal is as close to the real signal as possible.

[0040] Preferably, the update formula of the long short-term memory network in step S6 includes:

[0041] Forget Gate t The control formula is:

[0042] f t =σ(W f [h t-1 ,x t ]+b f )

[0043] Among them, f t is the output of the forget gate, indicating the forgetting ratio, σ is the Sigmoid activation function; W f is the weight matrix of the forget gate, b f is the bias term, h t-1 is the hidden state at the previous moment, x t is the input at the current moment;

[0044] Input Gate i t The control formula is:

[0045] i t =σ(W i [h t-1 ,x t ]+b i )

[0046] Among them, i t is the output of the input gate, indicating the proportion of new information written into the cell state, W i is the weight matrix of the input gate, b i is the bias term;

[0047] Cell State C t The update formula is:

[0048] C t =f t ·C t-1 +i t tanh(W C [h t-1 ,x t ]+b C )

[0049] Among them, C t is the cell state at the current moment, C t-1 is the cell state at the previous moment, tanh is the hyperbolic tangent activation function, W C is the weight matrix updated for the cell state, b C is the bias term;

[0050] Output gate o t The control formula is:

[0051] ot =σ(W o [h t-1 ,x t ]+b o )

[0052] Among them, t is the activation value of the output gate, indicating the output ratio of the hidden state at the current moment; W o is the weight matrix of the output gate, b o is the bias term;

[0053] The final hidden state h t The calculation formula is:

[0054] h t =o t tanh(C t )

[0055] Among them, h t is the hidden state at the current moment, as the output of the long short-term memory network, C t is the cell state at the current moment.

[0056] Preferably, the step S7 specifically includes the following steps:

[0057] S7.1. Real-time data collection and analysis through edge computing devices;

[0058] S7.2. Rapidly process the real-time vibration signal to generate real-time diagnostic results and predict the fault mode using the following formula:

[0059]

[0060] in, is the fault prediction value, x t is the input feature, h t-1 is the state at the previous moment;

[0061] S7.3. When an impeller imbalance failure occurs, the system will prompt the equipment operator through an early warning signal to perform maintenance or shut down for inspection in advance.

[0062] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the above method is implemented.

[0063] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the method described above is implemented.

[0064] The present invention provides a method for detecting impeller imbalance fault based on vibration signal analysis. It has the following beneficial effects:

[0065] 1. The present invention uses a convolutional neural network to extract local features in vibration signals, and extracts common features of vibration signals through shared convolutional layers, such as the periodic vibration mode of impeller imbalance fault and the impact characteristics of other fault modes (such as bearing fault and gear wear). On this basis, the classification of different fault modes is realized through a multi-task learning framework. The multi-task learning framework classifies the specific features of each fault mode by setting a task-specific layer, and optimizes the model in combination with a joint loss function. In the joint loss function, the training focus of multi-task learning is dynamically adjusted through the task weight coefficient, and the learning ability of different tasks is further balanced, which significantly improves the ability to distinguish multiple fault modes, solves the problem of fault mode confusion when multiple faults coexist, and makes the classification results more accurate and robust.

[0066] 2. The present invention significantly improves the generalization ability of the model by introducing a generative adversarial network to generate vibration data similar to the actual fault signal. The generator of GAN maps random noise z through a convolutional neural network to generate a false signal G(z) consistent with the characteristics of the real fault signal. The discriminator determines the authenticity of the input signal through the convolution layer and optimizes the capabilities of the generator and the discriminator through the loss function. In adversarial training, the generator continuously improves the quality of its generated signal, so that the generated false signal is highly similar to the real signal in feature distribution. By generating rich and diverse fault samples, GAN overcomes the problem of scarcity of fault samples in practical applications, and at the same time enhances the performance of the model in complex working conditions and unseen data, so that the model can better adapt to dynamic working condition changes and improve the robustness and adaptability of fault detection.

[0067] 3. The present invention adopts a long short-term memory network to perform time series modeling on the vibration signal, and captures the long-term dependency in the vibration signal through the gating mechanism of the forget gate, input gate and output gate. LSTM can effectively learn the development trend of the impeller imbalance fault from the historical data of the vibration signal by dynamically updating the cell state and hidden state. In practical applications, LSTM combines the input features with the state of the previous moment to predict the fault development trend at the current and future moments, and provide reliable trend analysis results for the equipment operation status. Through accurate trend prediction, the present invention can provide scientific maintenance suggestions for equipment operators, and take measures in advance to avoid further deterioration of the fault. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is a flow chart of the method of the present invention;

[0069] Figure 2 It is a schematic diagram of the computer device structure of the present invention.

[0070] Among them, 40, computer equipment; 41, processor; 42, memory; 43, storage medium. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] Please refer to the attached Figure 1 The embodiment of the present invention provides a method for detecting an impeller imbalance fault based on vibration signal analysis, comprising the following steps:

[0073] S1. Multiple three-axis acceleration sensors are placed at key locations such as impellers and bearings to collect impeller vibration signals.

[0074] Specifically, in response to the detection needs of impeller imbalance faults, accurately obtaining vibration data of key parts during the vibration signal analysis process is a prerequisite for fault diagnosis. By deploying multiple three-axis acceleration sensors in key parts such as impellers and bearings, the vibration signals of the impeller and its accessories can be fully collected. These signals contain rich fault information, which is especially important for early warning and trend analysis of dynamic imbalance faults. Through reasonable deployment and precise collection, the sensor can obtain vibration signals in all directions and effectively prevent the failure of fault identification due to missing data in a certain direction.

[0075] In this embodiment, a plurality of three-axis acceleration sensors are first arranged at key locations such as the impeller and bearings, and the vibration signals of the impeller and its key components are collected in real time through these sensors. The three-axis acceleration sensor can simultaneously monitor and record the vibration information of the impeller in three orthogonal directions (X, Y, and Z axes), thereby fully capturing the motion state of the equipment. These vibration signals, as raw data, can provide important information such as the rotation characteristics and dynamic balance state of the impeller.

[0076] Specifically, the vibration signal acquisition module includes multiple triaxial acceleration sensors installed in key locations such as the impeller, bearings, and housing. These sensors transmit vibration signals to the signal processing module for analysis in real time through a high-speed data acquisition system. Each triaxial acceleration sensor collects acceleration signals in three directions at its location, and the vibration data obtained contains multidimensional information generated by different failure modes that occur during the operation of the equipment. These signals will serve as input data for subsequent fault detection and diagnosis algorithms.

[0077] In the process of processing vibration signals, by jointly analyzing the three-axis vibration data, the complementarity of signals in different directions can be used to further enhance the reliability and accuracy of fault detection. For example, for an impeller imbalance fault, the vibration signal will show different characteristics in different axial directions. By collecting vibration data in multiple directions, the dynamic effects and frequency characteristics caused by imbalance can be more comprehensively identified, thereby improving the sensitivity and accuracy of fault detection.

[0078] In addition, in this embodiment, the multi-scale characteristics of the vibration signal are taken into account, and signal processing methods such as wavelet packet transform (WPT) or empirical mode decomposition (EMD) are used to extract features of the vibration signal collected by the three-axis acceleration sensor, so as to better capture and analyze the vibration spectrum changes caused by the impeller imbalance fault.

[0079] S2. Preprocessing the collected vibration signal, including removing noise using wavelet transform and performing signal standardization;

[0080] Signal normalization in step S2 includes processing the signal as follows:

[0081]

[0082] Among them, x(t) is the original signal, μ is the signal mean, σ is the standard deviation, and x norm (t) is the normalized signal.

[0083] Specifically, in order to ensure the quality of vibration signals and provide an effective data basis for subsequent fault detection and analysis, preprocessing the collected vibration signals is a key step. The present invention uses wavelet transform to remove noise from the signal, and combines signal standardization technology to unify the dimension and range of the signal, thereby eliminating the amplitude difference between data collected by different sensors and the influence of environmental interference. This processing process can significantly improve the analysis accuracy and robustness of the signal.

[0084] In this embodiment, the collected vibration signal is preprocessed, and the specific steps include:

[0085] Use wavelet transform to remove noise:

[0086] Wavelet transform is a multi-resolution analysis technique that can analyze signals in both the time domain and the frequency domain. The collected vibration signal usually contains high-frequency noise (such as mechanical environmental noise, sensor interference) and useful low-frequency features. Using wavelet transform, the signal can be decomposed into different frequency bands, and the coefficients of each frequency band can be threshold filtered to effectively remove noise and retain the main signal components.

[0087] Signal decomposition: The vibration signal x(t) is decomposed into sub-signals of different frequency bands by wavelet transform, and the formula is:

[0088]

[0089] Among them, W(a,b) is the coefficient after wavelet transformation, ψ a,b (t) is the wavelet basis function, and the parameters a and bbb represent the scaling and translation factors respectively.

[0090] Denoising: Apply soft threshold filtering to the decomposed high-frequency coefficients to remove high-frequency noise. The filtered signal is:

[0091] x denoised (t) = WT -1 (W(a,b)*H(a,b))

[0092] Among them, WT -1 is the inverse wavelet transform, and H(a,b) is a filter used to retain the signal in the useful frequency band.

[0093] Signal standardization: The purpose of signal standardization is to adjust the vibration signal to a uniform dimension range to eliminate the amplitude difference between data collected by different sensors. Standardization is of great significance for subsequent feature extraction and deep learning model training. Specifically, standardization normalizes the signal by calculating the mean μ\muμ and standard deviation σ of the signal. The formula is:

[0094]

[0095] x is the original signal;

[0096] μ is the mean of the signal:

[0097]

[0098] Where N is the number of sample points of the signal.

[0099] σ is the standard deviation of the signal:

[0100]

[0101] Among them, the standard deviation reflects the fluctuation range of the signal.

[0102] x norm (t) is the standardized signal with a mean of 0 and a standard deviation of 1.

[0103] In this embodiment, wavelet transform denoising is used to effectively remove high-frequency noise in the vibration signal while retaining the low-frequency characteristic signal of the impeller imbalance fault; signal standardization solves the problem of inconsistent signal amplitude caused by different sensor sensitivities or installation position differences, providing unified and standardized data input for subsequent feature extraction and model training.

[0104] S3, decomposing the vibration signal into multiple intrinsic mode functions by empirical mode decomposition and performing wavelet packet transform on each intrinsic mode function component;

[0105] The formula for decomposing the vibration signal in step S3 is as follows:

[0106]

[0107] Among them, x(t) is the vibration signal, r N (t) is the residual signal, IMF i (t) is the i-th intrinsic mode function.

[0108] Specifically, in order to extract the key features of the impeller vibration signal, the present invention decomposes the vibration signal into multiple intrinsic mode functions (IMF) by empirical mode decomposition (EMD), and further performs time-frequency analysis on each IMF component in combination with wavelet packet transform (WPT). EMD can adaptively decompose non-stationary signals into a series of intrinsic mode functions, while wavelet packet transform further enhances the resolution of frequency decomposition, thereby providing multi-scale feature information for fault detection.

[0109] In this embodiment, in order to fully capture the dynamic characteristics of the impeller vibration signal, the empirical mode decomposition method is used to decompose the signal at multiple scales. EMD decomposes the original signal into multiple intrinsic mode functions (IMFs) and a residual signal by fitting the envelope of the local extreme points of the signal. Each IMF component represents the vibration characteristics of the signal within a specific frequency range.

[0110] The process of empirical mode decomposition is:

[0111] Empirical mode decomposition decomposes the original vibration signal x(t) into several IMF components and residual signal rN(t). The decomposition formula is:

[0112]

[0113] x(t): original vibration signal;

[0114] IMF i (t): the i-th intrinsic mode function, each IMF is a vibration component of a single frequency band;

[0115] rN(t): Residual signal, representing the low-frequency trend term that cannot be further decomposed.

[0116] Breakdown steps:

[0117] Extract extreme points: Extract all local maximum and minimum points from the original signal x(t).

[0118] Construct envelope: Connect the maximum and minimum points respectively through interpolation to form upper and lower envelopes.

[0119] Calculate the envelope mean: Calculate the mean of the upper and lower envelopes m1(t) and subtract it from the original signal to obtain the first modal component candidate:

[0120] h1(t)=x(t)-m1(t)

[0121] Determine the IMF conditions: If h1(t) satisfies the two conditions of IMF:

[0122] The difference between the number of extreme points and zero crossing points of the signal does not exceed 1;

[0123] The local mean of the signal is zero. Then h1(t) is the first IMF component IMF1(t). Otherwise, repeat the above steps until the condition is met.

[0124] Remove the extracted IMF components: Remove the first IMF from the original signal to obtain the remaining signal:

[0125] r1(t)=x(t)-IMF1(t)

[0126] Repeat the above steps for the remaining signal r1(t) until the remaining signal cannot be further decomposed.

[0127] Application of wavelet packet transform: For each IMF component, wavelet packet transform (WPT) is further used for frequency refinement analysis. Wavelet packet transform further decomposes the local characteristics of the signal in different frequency ranges into high-frequency and low-frequency components through multi-scale decomposition.

[0128] Wavelet packet transform formula:

[0129]

[0130] IMF i (t) is the i-th IMF component; c j is the wavelet packet decomposition coefficient; ψ j (t) is the wavelet packet basis function, which represents the decomposition scale and position of the signal.

[0131] Effect of combining EMD and WPT: In this embodiment, the IMF components obtained by EMD decomposition can capture the vibration characteristics of the signal at different time scales, and WPT further decomposes these characteristics into finer-grained frequency band characteristics, ensuring that the characteristic information of the impeller imbalance fault is fully extracted. For example, in the case of impeller imbalance, the vibration signal may appear as periodic vibration of a specific frequency, and these characteristics can be clearly identified and extracted through the joint analysis of EMD and WPT.

[0132] S4. Use convolutional neural network to extract local features in vibration signals, and classify and identify impeller unbalance faults and other fault modes through a multi-task learning framework;

[0133] The multi-task learning framework in step S4 includes:

[0134] Extract common features from signals through shared convolutional layers;

[0135] Set up an independent task-specific layer for each task to accurately distinguish different failure modes;

[0136] Through the loss function optimization framework, combined with the loss function of each task:

[0137]

[0138] in, is the loss of the i-th task, λ i is the task weight coefficient, L gen is the loss term of the generated adversarial network, and α is its weight coefficient.

[0139] Specifically, in order to achieve classification and identification of impeller imbalance faults and other fault modes, the present invention extracts local features in vibration signals by using convolutional neural networks (CNN) and adopts a multi-task learning framework (MTL) for classification and identification. Convolutional neural networks can efficiently extract spatial features in vibration signals, and the multi-task learning framework not only improves the training efficiency of the model by combining shared feature layers and task-specific layers, but also enhances the ability to distinguish between multiple fault modes. Combined with the optimization method of the loss function, the model can learn the features of different tasks at the same time, improving the accuracy and robustness of the overall classification.

[0140] In this embodiment, a convolutional neural network (CNN) is used to extract local features from the preprocessed vibration signal, and the impeller imbalance fault and other fault modes (such as bearing failure, gear wear, etc.) are classified and identified through a multi-task learning framework.

[0141] Convolutional neural network extracts local features:

[0142] Convolutional neural network is an important network structure in deep learning, and is particularly suitable for extracting local features from time series signals. In the present invention, the convolution layer performs local scanning on the vibration signal through a sliding convolution kernel, and can extract periodic features, impact features, etc. in the signal.

[0143] Convolutional layer operation formula:

[0144]

[0145] Among them, x is the input signal; w is the convolution kernel (weight matrix), which is used to extract specific patterns; y(t) is the output feature after the convolution operation.

[0146] Through multi-layer convolutional networks, the low-order features of vibration signals (such as frequency and amplitude) are gradually mapped to high-order features (such as pattern classification information).

[0147] Multi-task learning framework: In the multi-task learning framework (MTL), the model extracts common features through shared feature extraction layers and performs classification of different tasks in task-specific layers.

[0148] a) Shared convolutional layers extract common features:

[0149] The shared convolutional layer consists of multiple convolutional layers and pooling layers to extract common features in vibration signals that are relevant to all tasks. For example, impeller imbalance and bearing faults may share similar periodic vibration patterns, and these features are extracted in the shared layer to improve feature utilization.

[0150] b) Task classification at the task-specific layer:

[0151] The task-specific layer sets up an independent classification network for each task (such as impeller imbalance fault, bearing fault, etc.) based on the shared features to identify different fault modes.

[0152] c) Loss function optimization:

[0153] The multi-task learning framework uses a joint loss function to optimize each task to ensure the accuracy and robustness of each task during classification. The loss function formula is as follows:

[0154]

[0155] in, is the loss of the i-th task, λ i is the task weight coefficient, L gen is the loss term of the generated adversarial network, and α is its weight coefficient.

[0156] In this embodiment, the multi-task learning framework implements fault classification and optimization through the following steps:

[0157] Universal feature extraction: The universal features of vibration signals are extracted through shared convolutional layers to build basic feature representations for all tasks.

[0158] Task-specific classification: For each fault mode, a separate task-specific layer is set up for classification. For example, the classifier for impeller imbalance fault focuses on specific frequency features in the vibration signal, while the classifier for bearing fault focuses on impact features.

[0159] Loss function optimization: All tasks are jointly optimized through the joint loss function, using the weight coefficient λ i Balance the impact of different tasks to ensure the overall performance of the model in classification tasks.

[0160] S5, using a generative adversarial network to generate a vibration signal similar to the actual fault;

[0161] The generative adversarial network in step S5 includes:

[0162] The generator generates a vibration signal similar to the actual fault signal through a convolutional neural network. The generator input is random noise z, and the fault signal G(z) is generated through convolutional layer and deconvolution layer operations.

[0163] The discriminator determines the authenticity of the input signal and determines whether the signal x is a real signal by outputting the probability D(x). The loss function of the discriminator is:

[0164]

[0165] Among them, D(x) is the output of the discriminator for the real signal, G(z) is the false signal generated by the generator, and L dis is the discriminator loss;

[0166] The loss function of the generator is:

[0167]

[0168] Among them, D(G(z)) is the judgment of the discriminator on the generated signal, which indicates the probability that the generated signal is a true signal, and L gen is the generator loss;

[0169] During the training process, the generator and the discriminator are trained by alternating optimization, and the goal of the discriminator is to minimize the loss function L dis , and the goal of the generator is to minimize the loss function L gen , so that the generated signal is as close to the real signal as possible.

[0170] Specifically, in order to enhance the generalization ability and adaptability of the impeller imbalance fault detection model to unseen data, the present invention uses a generative adversarial network (GAN) to generate a vibration signal similar to the actual fault signal. GAN uses adversarial training between the generator and the discriminator. The generator generates a false signal to "cheat" the discriminator, while the discriminator optimizes the performance of the generator by identifying the authenticity of the input signal. Ultimately, the generator is able to generate vibration data close to the real fault signal, thereby improving the performance of the model on diversified data.

[0171] In this embodiment, the generative adversarial network (GAN) is composed of a generator and a discriminator. The generator is used to generate a vibration signal similar to a real fault signal, while the discriminator is used to determine whether the input signal is a real signal or a generated signal. Through the alternating optimization training of the generator and the discriminator, the generated false signal gradually approaches the real signal.

[0172] Implementation of the generator: The generator generates a vibration signal similar to the actual fault signal through a convolutional neural network (CNN). The input of the generator is random noise z, which is mapped into a false signal similar to the real fault signal through a series of convolutional layers, activation functions, and deconvolution layer operations.

[0173] Input Random Noise: The generator input z is a random noise vector sampled from a uniform or normal distribution.

[0174] Convolutional layer and deconvolution layer operations: The generator contains multiple layers of convolution operations. Each layer extracts and transforms the input signal, and the final output is a false signal G(z). The formula for generating the signal is:

[0175] G(z)=f(W g ·z+b g )

[0176] Where: G(z) is the false signal output by the generator; W g and b g is the weight matrix and bias term of the generator; f is the activation function (such as ReLU or tanh function).

[0177] Implementation of the discriminator: The discriminator is used to determine whether the input signal x is a real signal. The features of the input signal are extracted through the convolution layer, and the discriminator outputs a probability value D(x), which indicates the possibility that the signal is a real signal.

[0178] Loss function of the discriminator: The goal of the discriminator is to correctly distinguish between real signals and generated signals, and its loss function is:

[0179]

[0180] Where: D(x) is the output of the discriminator for the real signal, G(z) is the false signal generated by the generator, L dis is the discriminator loss.

[0181] Generator loss function: The goal of the generator is to generate a false signal that can "fool" the discriminator, that is, to make the discriminator judge the generated signal as a real signal. Therefore, the generator's loss function is:

[0182]

[0183] Where: L gen is the generator loss; D(G(z)) is the output probability of the discriminator for the generated signal G(z), indicating the possibility that the generated signal is a real signal.

[0184] Training process: In this embodiment, the training of GAN adopts the alternating optimization strategy of the generator and the discriminator:

[0185] Optimize the discriminator: When the generator is fixed, optimize the loss function L of the discriminator dis , the goal is to improve the discriminator's ability to distinguish between real signals and generated signals.

[0186] Optimize the generator: When the discriminator is fixed, optimize the loss function Lgen of the generator, the goal is to generate a fake signal that is difficult to distinguish from the real signal.

[0187] Alternating training: Repeat the above steps until the signal generated by the generator is very close to the real signal, that is, the discriminator cannot distinguish between real and fake signals.

[0188] S6. Use long short-term memory network to analyze the vibration signal in time series and predict the development trend of impeller imbalance fault;

[0189] The update formula of the long short-term memory network in step S6 includes:

[0190] Forget Gate t The control formula is:

[0191] f t =σ(W f [h t-1 , x t ]+b f )

[0192] Among them, f t is the output of the forget gate, indicating the forgetting ratio, σ is the Sigmoid activation function; W f is the weight matrix of the forget gate, b f is the bias term, h t-1 is the hidden state at the previous moment, x t is the input at the current moment;

[0193] Input Gate i t The control formula is:

[0194] i t =σ(W i [h t-1 , x t ]+b i )

[0195] Among them, i t is the output of the input gate, indicating the proportion of new information written into the cell state, W i is the weight matrix of the input gate, b i is the bias term;

[0196] Cell State C t The update formula is:

[0197] C t =f t ·C t-1 +i t tanh(W C [h t-1 , x t ]+b C )

[0198] Among them, C t is the cell state at the current moment, C t-1 is the cell state at the previous moment, tanh is the hyperbolic tangent activation function, W C is the weight matrix updated for the cell state, b C is the bias term;

[0199] Output gate o t The control formula is:

[0200] o t =σ(W o [h t-1 , x t ]+b o )

[0201] Among them, t is the activation value of the output gate, indicating the output ratio of the hidden state at the current moment; W o is the weight matrix of the output gate, b o is the bias term;

[0202] The final hidden state h t The calculation formula is:

[0203] h t =o t tanh(C t )

[0204] Among them, h t is the hidden state at the current moment, as the output of the long short-term memory network, C t is the cell state at the current moment.

[0205] Specifically, in order to predict the development trend of impeller imbalance fault, the present invention uses long short-term memory network to perform time series analysis on vibration signals. The long short-term memory network can effectively capture the long-term dependencies in the time series signals through its unique gating mechanism, including forgetting gate, input gate, output gate and cell state, and avoid the gradient vanishing problem in traditional recurrent neural network (RNN). By learning the temporal dynamic characteristics of vibration signals, the long short-term memory network can accurately predict the development trend of faults and provide a reliable basis for equipment maintenance.

[0206] In this embodiment, a long short-term memory network (LSTM) is used to perform time series analysis on the vibration signal. The LSTM selectively memorizes and forgets information through a gating mechanism, thereby capturing the time dependency in the vibration signal, especially the dynamic changes of the impeller unbalance fault over time.

[0207] 1. Gating mechanism of long short-term memory network

[0208] (1) Forget gate: The function of the forget gate is to determine how much information from the previous moment needs to be "forgotten" from the cell state at the current moment. The output of the forget gate is f t Represents the forgetting ratio, and its calculation formula is:

[0209] f t =σ(W f [h t-1 , x t ]+b f )

[0210] f t : The output of the forget gate, ranging from [0, 1]. Close to 0 means complete forgetting, and close to 1 means complete retention. σ is the activation function, which maps the output to [0, 1][0, 1][0, 1]. W f is the weight matrix of the forget gate; b f is the bias term of the forget gate; h t-1 is the hidden state of the previous moment; x t is the input signal at the current moment.

[0211] (2) Input gate: The input gate is used to determine the proportion of new information that needs to be written into the cell state at the current moment. The output i of the input gate t Indicates the current information writing ratio, and its calculation formula is:

[0212] it =σ(W i [h t-1 , x t ]+b i )

[0213] i t is the output of the input gate, ranging from [0, 1]; W i is the weight matrix of the input gate; b i is the bias term of the input gate; other symbols have the same meaning as above.

[0214] (3) Updating of cell status

[0215] Cell State C t It is a long-term memory unit of the long short-term memory network, which is used to store the historical information of the network. The update of the cell state is combined with the forget gate f t The forgotten information and input gate i t The new information is written. The update formula is:

[0216] C t =f t ·C t-1 +i t tanh(W C [h t-1 , x t ]+b C )

[0217] C t is the cell state at the current moment; C t-1 is the cell state at the previous moment; tanh is the hyperbolic tangent activation function, which is used to limit the value range; W C is the weight matrix updated for the cell state; b C Bias term for cell state update.

[0218] (4) Output gate: The output gate controls the hidden state h at the current moment t The output ratio of the output gate is calculated as:

[0219] o t =σ(W o [h t-1 , x t ]+b o )

[0220] o t is the activation value of the output gate, ranging from [0, 1]; W o is the weight matrix of the output gate; b o is the bias term of the output gate.

[0221] (5) Calculation of hidden state

[0222] Hidden state h t is the output of the long short-term memory network and the feature representation of the current moment. The hidden state is represented by the output gate o t and cell state C t The combined calculation yields:

[0223] h t =o t tanh(C t )

[0224] h t is the hidden state at the current moment, which is the output of the long short-term memory network; C t is the cell state at the current moment.

[0225] 2. Timing Analysis Process of Long Short-Term Memory Network

[0226] In this embodiment, the timing analysis of the vibration signal is completed by the following steps:

[0227] Signal input: The preprocessed vibration signal x t Input the long short-term memory network and combine it with the hidden state h at the previous moment t-1 , calculate the activation value of each gate.

[0228] State update: Update the current cell state C according to the control formula of the forget gate, input gate and cell state t .

[0229] Hidden state output: Calculate the hidden state h at the current moment through the output gate t , the hidden state will be passed to the next moment and used for the output at the current moment.

[0230] Fault trend prediction: The long short-term memory network captures the long-term dependencies in the vibration signal through multi-layer time series modeling, and finally outputs the fault development trend.

[0231] Through the above formula and process, the long short-term memory network can learn the long-term dependencies in the vibration signal and accurately predict the development trend of the impeller imbalance fault. Combined with its gating mechanism, the long short-term memory network effectively avoids the gradient vanishing problem and performs well in capturing the temporal dynamic characteristics of the vibration signal, providing important technical support for fault diagnosis and prediction.

[0232] S7. Monitor faults in real time and issue warning signals when a fault occurs or is predicted to occur.

[0233] Step S7 specifically includes the following steps:

[0234] S7.1. Real-time data collection and analysis through edge computing devices;

[0235] S7.2. Rapidly process the real-time vibration signal to generate real-time diagnostic results and predict the fault mode using the following formula:

[0236]

[0237] in, is the fault prediction value, x t is the input feature, h t-1 is the state at the previous moment;

[0238] S7.3. When an impeller imbalance failure occurs, the system will prompt the equipment operator through an early warning signal to perform maintenance or shut down for inspection in advance.

[0239] Specifically, in order to achieve real-time monitoring and early warning of impeller imbalance faults, the present invention designs a fault monitoring and prediction system based on edge computing and vibration signal analysis. In the system, the vibration signals collected in real time are quickly processed by edge computing devices, and the prediction model is used to generate real-time diagnostic results and fault development trends. When an impeller imbalance fault is detected or an impending fault is predicted, the system will issue a warning signal to prompt the equipment operator to take necessary maintenance measures. The entire process integrates real-time, high efficiency and predictiveness, providing a guarantee for the safe operation of the equipment.

[0240] In this embodiment, real-time data collection, rapid processing and fault prediction are achieved through edge computing devices, and the early warning signal function is combined to promptly notify operators to take measures. The specific steps are as follows:

[0241] Edge computing devices are installed near key parts such as impellers and bearings, and are responsible for collecting sensor data and processing it in real time. Compared with traditional centralized processing, edge computing can effectively reduce data transmission delays and improve system response speed.

[0242] Data collection: Multiple three-axis acceleration sensors are used to collect the impeller vibration signals in real time. These data contain the dynamic characteristics of the equipment during operation and potential fault information.

[0243] Preliminary processing: The collected vibration signals are preprocessed in the edge computing device, including noise removal (such as wavelet transform denoising) and signal standardization to ensure signal quality.

[0244] In this embodiment, the core of real-time processing is to quickly analyze the vibration signal through the deep learning model, generate real-time diagnosis results and predict the development trend of the fault mode. The process includes the following steps:

[0245] Signal processing: The vibration signal is decomposed into feature representations (e.g., extracting local features through convolutional neural networks) and passed to the prediction model.

[0246] Fault prediction model: Use the long short-term memory network (LSTM) to predict the fault mode. The formula is as follows:

[0247]

[0248] is the fault prediction value, indicating whether there is a fault or a fault development trend at the current moment; x t is the input feature, generated by the vibration signal processed in real time; h t-1 It is the state of the previous moment, which contains the network’s memory of historical information.

[0249] Model update and iteration: By inputting new data in real time, the model continuously adjusts the prediction value to ensure the accuracy of fault prediction results.

[0250] In this embodiment, when the system detects an impeller imbalance fault or predicts that a fault is about to occur, the system will automatically trigger an early warning signal. The early warning signal includes:

[0251] Alarm notification: The system issues a warning to the equipment operator through visual, auditory or remote communication methods. For example:

[0252] Flashing warning lights or buzzer reminders;

[0253] Send alarm information to the monitoring center via wireless network.

[0254] Maintenance recommendations: The system provides maintenance recommendations based on the prediction results, such as scheduling maintenance or emergency shutdown inspections in advance.

[0255] Please refer to the attached Figure 2 The present invention further provides a computer device 40, comprising: a processor 41 and a memory 42, wherein the memory 42 stores a computer program executable by the processor, and when the computer program is executed by the processor, the above method is executed.

[0256] The present invention further provides a storage medium 43 on which a computer program is stored. When the computer program is run by the processor 41, the above method is executed.

[0257] Among them, the storage medium 43 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, disk or optical disk.

[0258] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting impeller imbalance fault based on vibration signal analysis, characterized in that: The following steps are involved: S1. Multiple three-axis acceleration sensors are placed at key locations such as impellers and bearings to collect impeller vibration signals. S2. Preprocessing the collected vibration signal, including removing noise using wavelet transform and performing signal standardization; S3, decomposing the vibration signal into multiple intrinsic mode functions by empirical mode decomposition and performing wavelet packet transform on each intrinsic mode function component; S4. Use convolutional neural network to extract local features in vibration signals, and classify and identify impeller unbalance faults and other fault modes through a multi-task learning framework; S5, using a generative adversarial network to generate a vibration signal similar to the actual fault; S6. Use long short-term memory network to analyze the vibration signal in time series and predict the development trend of impeller imbalance fault; S7. Monitor faults in real time and issue warning signals when a fault occurs or is predicted to occur.

2. The impeller imbalance fault detection method based on vibration signal analysis according to claim 1 is characterized in that: The signal normalization in step S2 includes processing the signal as follows: Among them, x(t) is the original signal, μ is the signal mean, σ is the standard deviation, and x norm (t) is the normalized signal.

3. The impeller imbalance fault detection method based on vibration signal analysis according to claim 1 is characterized in that: The formula for decomposing the vibration signal in step S3 is as follows: Among them, x(t) is the vibration signal, r N (t) is the residual signal, IMF i (t) is the i-th intrinsic mode function.

4. The impeller imbalance fault detection method based on vibration signal analysis according to claim 1 is characterized in that: The multi-task learning framework in step S4 includes: Extract common features from signals through shared convolutional layers; Set up an independent task-specific layer for each task to accurately distinguish different failure modes; Through the loss function optimization framework, combined with the loss function of each task: in, is the loss of the i-th task, λ i is the task weight coefficient, L gen is the loss term of the generated adversarial network, and α is its weight coefficient.

5. The impeller imbalance fault detection method based on vibration signal analysis according to claim 1 is characterized in that: The generation of the adversarial network in step S5 includes: The generator generates a vibration signal similar to the actual fault signal through a convolutional neural network. The generator input is random noise z, and the fault signal G(z) is generated through convolutional layer and deconvolution layer operations. The discriminator determines the authenticity of the input signal and determines whether the signal x is a real signal by outputting the probability D(x). The loss function of the discriminator is: Among them, D(x) is the output of the discriminator for the real signal, G(z) is the false signal generated by the generator, and L dis is the discriminator loss; The loss function of the generator is: Among them, D(G(z)) is the judgment of the discriminator on the generated signal, which indicates the probability that the generated signal is a true signal, and L gen is the generator loss; During the training process, the generator and the discriminator are trained by alternating optimization, and the goal of the discriminator is to minimize the loss function L dis , and the goal of the generator is to minimize the loss function L gen , so that the generated signal is as close to the real signal as possible.

6. The impeller imbalance fault detection method based on vibration signal analysis according to claim 1 is characterized in that: The update formula of the long short-term memory network in step S6 includes: Forget Gate t The control formula is: f t =σ(W f [h t-1 ,x t ]+b f ) Among them, f t is the output of the forget gate, indicating the forgetting ratio, σ is the Sigmoid activation function; W f is the weight matrix of the forget gate, b f is the bias term, h t-1 is the hidden state at the previous moment, x t is the input at the current moment; Input Gate i t The control formula is: i t =σ(W i [h t-1 ,x t ]+b i ) Among them, i t is the output of the input gate, indicating the proportion of new information written into the cell state, W i is the weight matrix of the input gate, b i is the bias term; Cell State C t The update formula is: C t =f t ·C t-1 +i t ·tanh(W C [h t-1 ,x t ]+b C ) Among them, C t is the cell state at the current moment, C t-1 is the cell state at the previous moment, tanh is the hyperbolic tangent activation function, W C is the weight matrix updated for the cell state, b C is the bias term; Output gate o t The control formula is: the t =σ(W o [h t-1 ,x t ]+b o ) Among them, t is the activation value of the output gate, indicating the output ratio of the hidden state at the current moment; W o is the weight matrix of the output gate, b o is the bias term; The final hidden state h t The calculation formula is: h t =o t ·tanh(C t ) Among them, h t is the hidden state at the current moment, as the output of the long short-term memory network, C t is the cell state at the current moment.

7. The impeller imbalance fault detection method based on vibration signal analysis according to claim 1 is characterized in that: The step S7 specifically includes the following steps: S7.

1. Real-time data collection and analysis through edge computing devices; S7.

2. Rapidly process the real-time vibration signal to generate real-time diagnostic results and predict the fault mode using the following formula: in, is the fault prediction value, x t is the input feature, h t-1 is the state at the previous moment; S7.

3. When an impeller imbalance failure occurs, the system will prompt the equipment operator through an early warning signal to perform maintenance or shut down for inspection in advance.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images