A rolling bearing fault diagnosis method and system based on WPD-LNN

Abstract

The application discloses a WPD-LNN-based rolling bearing fault diagnosis method and system, and belongs to the technical field of engine fault diagnosis. The method comprises the following steps: synchronously collecting original vibration signals and instantaneous speed signals of a rolling bearing; adopting a sliding time window to divide the vibration signals, adaptively selecting an optimal wavelet base and a decomposition layer number in each time window based on a frequency band energy entropy minimization criterion to perform wavelet packet decomposition, and extracting energy features of each subband; calculating weighted mutual information between the energy of each subband and a fault type through a speed weight, screening the weighted mutual information, and splicing the screened weighted mutual information with time domain and frequency domain statistical features and a speed normalization factor into a working condition adaptive feature vector; and inputting the feature vector, the instantaneous speed value and a speed change rate into a working condition linkage type liquid neural network, dynamically adjusting a liquid time constant according to the speed change rate, controlling network states by a variable coefficient ordinary differential equation, and outputting a fault recognition result. The application realizes high-precision rolling bearing fault diagnosis under a non-steady working condition.

Description

Technical Field

[0001] This invention relates to the field of engine fault diagnosis, and specifically to a rolling bearing fault diagnosis method and system based on WPD-LNN. Background Technology

[0002] Rolling bearings are a key component of engines, playing a crucial role in supporting rotating parts, reducing friction, transmitting loads, and maintaining rotational accuracy. Their operating condition directly impacts the safety and reliability of mechanical equipment. Failures can lead to performance degradation or even shutdown. Therefore, fault diagnosis of rolling bearings is of great significance. In actual operation, the vibration signals of rolling bearings often exhibit characteristics such as non-stationarity, strong noise, and weak fault characteristics. Especially under unsteady conditions such as variable speed and load, the fault characteristic signals are modulated by changes in operating conditions, further increasing the difficulty of diagnosis.

[0003] Wavelet packet decomposition (WPD) can simultaneously decompose the high-frequency and low-frequency components of a signal and has been widely used in bearing fault diagnosis. Liquid neural networks (LNNs) use ordinary differential equations to describe the continuous-time dynamic characteristics of neuron states, adjusting the information flow speed through a liquid time constant, giving them advantages in processing time-series dynamic signals. However, existing WPD-based bearing fault diagnosis methods use fixed wavelet bases and fixed decomposition layers, failing to adaptively adjust decomposition parameters according to the signal's time-frequency characteristics. When changes in operating conditions alter the signal's frequency band structure, fixed parameters struggle to match the current signal's frequency distribution, easily leading to missed or distorted fault-related frequency band energy, affecting diagnostic accuracy. Existing LNN-based bearing fault diagnosis methods employ a CNN+LNN structure, where the CNN convolution kernel parameters are fixed after training, unable to dynamically adjust feature extraction strategies based on real-time operating conditions. Furthermore, while the LNN's time constant is dynamic, it cannot actively adjust network response characteristics when operating conditions change, resulting in insufficient adaptability to time-varying signals and decreased fault mode recognition accuracy. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a rolling bearing fault diagnosis method and system based on WPD-LNN. By using adaptive wavelet packet decomposition and a working condition-linked liquid neural network, it achieves high-precision fault diagnosis of rolling bearings under unsteady working conditions such as variable speed and variable load, effectively improving the working condition adaptability and recognition accuracy of the diagnostic model.

[0005] In a first aspect, the technical solution of the present invention provides a rolling bearing fault diagnosis method based on WPD-LNN, comprising the following steps: S1, synchronously acquire the original vibration signal and instantaneous speed signal of the rolling bearing in operation, and perform noise reduction and normalization preprocessing respectively; S2, the preprocessed vibration signal is divided into several analysis segments by using a sliding time window. Within each time window, based on the frequency band energy entropy minimization criterion, the optimal wavelet basis and optimal decomposition level are dynamically selected from the candidate wavelet basis set and candidate decomposition level. Adaptive wavelet packet decomposition is performed on the vibration signal segment of the current time window to extract the energy features of each sub-band. S3, calculate the average speed based on the instantaneous speed information within the current time window, and calculate the weighted mutual information between the energy of each sub-band and the fault type based on the average speed and the speed weight. Select the sub-band energy with the weighted mutual information greater than the dynamic threshold as the filtered feature subset, and concatenate the feature subset, time domain statistical features, frequency domain statistical features and speed normalization factor into the working condition adaptive feature vector. S4, input the working condition adaptive feature vector, the instantaneous speed value of the current time window and the speed change rate into the working condition linkage liquid neural network, and output the fault type identification result; the liquid time constant of each neuron in the liquid neural network is dynamically adjusted in real time according to the working condition parameters, and the network state is controlled by variable coefficient ordinary differential equations for iterative updates.

[0006] Secondly, the technical solution of the present invention provides a rolling bearing fault diagnosis system based on WPD-LNN, comprising: The signal acquisition module is used to synchronously acquire the original vibration signal and instantaneous speed signal of the rolling bearing in operation, and perform noise reduction and normalization preprocessing respectively. The adaptive decomposition module is used to divide the preprocessed vibration signal into several analysis segments using a sliding time window. Within each time window, based on the frequency band energy entropy minimization criterion, the optimal wavelet basis and optimal decomposition level are dynamically selected from the candidate wavelet basis set and candidate decomposition level. Adaptive wavelet packet decomposition is performed on the vibration signal segment of the current time window to extract the energy features of each sub-band. The feature construction module is used to calculate the average speed based on the instantaneous speed information within the current time window, and to calculate the weighted mutual information between the energy of each sub-band and the fault type based on the average speed by introducing speed weights. The sub-band energy with weighted mutual information greater than the dynamic threshold is selected as the filtered feature subset. The feature subset, time-domain statistical features, frequency-domain statistical features and speed normalization factor are concatenated into an operating condition adaptive feature vector. The fault diagnosis module is used to input the adaptive feature vector of the working condition, the instantaneous speed value of the current time window and the speed change rate into the working condition linkage liquid neural network, and output the fault type identification result; the liquid time constant of each neuron in the liquid neural network is dynamically adjusted in real time according to the working condition parameters, and the network state is controlled by variable coefficient ordinary differential equations for iterative updates.

[0007] As can be seen from the above technical solutions, this application has the following advantages: Within each time window, based on the criterion of minimizing the frequency band energy entropy, the optimal decomposition parameters are dynamically selected from the candidate wavelet basis set and candidate decomposition levels, ensuring that the decomposition scheme always matches the time-frequency characteristics of the current signal. This avoids the feature extraction distortion problem caused by fixed parameters, providing high-quality discrimination information for subsequent fault identification, thereby improving the accuracy of fault feature extraction under varying operating conditions. By calculating weighted mutual information using rotational speed weights, the modulation effect of rotational speed changes on frequency band energy is incorporated into the screening process. The selected feature subset maintains a strong correlation with the fault type across the entire operating condition range. Simultaneously, the rotational speed normalization factor is added as an auxiliary feature to the feature vector, eliminating the influence of rotational speed changes on the feature amplitude. This makes the feature vector comparable under different operating conditions, avoiding misclassification by the classifier due to feature distribution shifts caused by operating condition changes, and improving the diagnostic confidence of the model in varying operating condition scenarios. The liquid time constant is designed as a linkage function that adjusts in real time with the rate of change of rotational speed. Under steady-state conditions, the time constant is larger, allowing the network to retain more historical information to capture long-term fault trends. Under transient conditions, the time constant decreases, allowing the network to focus more on the current input and respond quickly to signal abrupt changes, thus maintaining high recognition accuracy across the entire operating range. Adaptive WPD ensures the operating condition matching of the feature extraction stage, speed-weighted mutual information filtering ensures the operating condition relevance of the feature selection stage, and operating condition-linked LNN ensures the operating condition adaptability of the classifier stage, enabling accurate identification of rolling bearing fault types under non-steady-state conditions such as variable speed and variable load. Attached Figure Description

[0008] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of a rolling bearing fault diagnosis method based on WPD-LNN provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic block diagram of a rolling bearing fault diagnosis system based on WPD-LNN provided in an embodiment of the present invention.

[0011] Figure 3 This is a schematic diagram of the structure of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The various terms used in this specification are merely for describing specific embodiments and do not constitute a limitation thereof.

[0014] Figure 1 This is a schematic flowchart of a rolling bearing fault diagnosis method based on WPD-LNN provided in an embodiment of the present invention. Figure 1 The executing entity can be a rolling bearing fault diagnosis system based on WPD-LNN. The rolling bearing fault diagnosis method based on WPD-LNN provided in this embodiment of the invention is executed by a computer device; correspondingly, the rolling bearing fault diagnosis system based on WPD-LNN runs on the computer device. Depending on different requirements, the order of the steps in this flowchart can be changed, and some steps can be omitted.

[0015] like Figure 1 As shown, the method includes the following steps.

[0016] S1 synchronously acquires the original vibration signal and instantaneous speed signal of the rolling bearing in operation, and performs noise reduction and normalization preprocessing respectively.

[0017] In this embodiment, the original vibration signal The accelerometer sensor mounted on the bearing housing collects data, and the sampling frequency should satisfy the Nyquist sampling theorem to ensure coverage of the bearing fault characteristic frequencies and their harmonics. Instantaneous speed signal. The vibration signal is synchronously acquired by a speed sensor to characterize the current operating condition. Both signals are sampled synchronously through the same data acquisition card to ensure strict time alignment between the vibration signal and the speed signal.

[0018] The acquired raw vibration signals and instantaneous speed signal Denoising and normalization preprocessing were performed separately.

[0019] Denoising is used to eliminate high-frequency noise components such as environmental noise and electromagnetic interference from contaminating signal quality. For vibration signals, wavelet thresholding or bandpass filtering can be used: wavelet thresholding involves thresholding the detail coefficients after wavelet decomposition to suppress noise components before reconstructing the signal; bandpass filtering sets the passband based on the distribution range of bearing fault characteristic frequencies to filter out interference components outside the passband. For speed signals, median filtering or low-pass filtering can be used to eliminate pulse interference and quantization noise introduced during speed measurement.

[0020] Normalization is used to eliminate the influence of differences in signal amplitude dimensions under different sensors and operating conditions on subsequent feature extraction. The normalization of vibration signals and rotational speed signals both adopt the minimum-maximum normalization method, which linearly maps the signal amplitude to the [0,1] interval, making signals with different dimensions numerically comparable.

[0021] The preprocessed vibration signal is denoted as The rotational speed signal is denoted as This serves as the input for step S2.

[0022] S2. The preprocessed vibration signal is divided into several analysis segments using a sliding time window. Within each time window, based on the criterion of minimizing the frequency band energy entropy, the optimal wavelet basis and the optimal decomposition level are dynamically selected from the candidate wavelet basis set and the candidate decomposition level. Adaptive wavelet packet decomposition is then performed on the vibration signal segment of the current time window to extract the energy features of each sub-band.

[0023] During actual operation, the engine undergoes various operating conditions such as starting, acceleration, cruising, and deceleration, resulting in vibration signals exhibiting significant non-stationary and time-varying modulation characteristics. A sliding time window strategy is employed to divide the continuous vibration signal into several short-time analysis segments. Within each time window, the optimal decomposition parameters are independently selected, ensuring that the decomposition scheme always adapts to the time-frequency characteristics of the current signal.

[0024] S2.1, set the window length and overlap of the sliding time window to divide the preprocessed vibration signal into several analysis segments.

[0025] Preferably, the window length of the sliding time window is... It corresponds to a time span of 2 to 3 revolutions of the engine rotor, and can achieve a good balance between frequency domain resolution and time domain tracking capability.

[0026] As a preferred option, the overlap rate is 50% to ensure the continuity of information between adjacent segments and to prevent key features from being missed because they fall at the window boundary.

[0027] The preprocessed vibration signal is After being segmented by a sliding time window, K analysis segments were obtained. .

[0028] S2.2 Select a wavelet basis from the candidate wavelet basis set and select a number of layers from the candidate decomposition layer range. Use the wavelet basis and the number of layers to perform wavelet packet decomposition on the vibration signal segment of the current time window to obtain each sub-band under the decomposition layer. Calculate the energy of each sub-band and calculate the frequency band energy entropy based on the energy of each sub-band.

[0029] The candidate wavelet basis set is pre-set to include the db family of wavelets (such as db4, db8, db11, etc.), the sym family of wavelets (such as sym4, sym8, etc.), and the coif family of wavelets (such as coif4, etc.). Different wavelet bases have different regularity, vanishing moments, and tight support characteristics, making them suitable for signal decomposition of different waveforms. The number of candidate decomposition levels is pre-set to 3, 4, 5, and 6 levels. The range of 3 to 6 levels covers different resolution requirements from coarse to fine.

[0030] The calculation of the energy of each sub-band specifically includes: performing wavelet packet decomposition on the vibration signal segment of the current time window based on the wavelet basis and the layer number to obtain the wavelet packet coefficients of each sub-band; reconstructing the signal from the wavelet packet coefficients of each sub-band to obtain the time-domain reconstructed signal of each sub-band; calculating the integral value of the square of the amplitude of the time-domain reconstructed signal of each sub-band as the energy of that sub-band; and summing the energy of each sub-band to obtain the sub-band energy feature vector of the current time window.

[0031] Specifically, let the current decomposition level be J, the wavelet packet decomposition at this level yields... Each subband has its own energy. The specific steps for calculating the energy of each subband are as follows.

[0032] The first step is to perform wavelet packet decomposition on the vibration signal segment of the current time window using the currently selected wavelet basis and decomposition level to obtain the wavelet packet coefficients of each sub-band.

[0033] Wavelet packet decomposition is a generalization of wavelet transform, decomposing not only the low-frequency components but also the high-frequency components layer by layer. Let the vibration signal segment of the current time window be... , with the selected wavelet basis Wavelet packet decomposition is performed on the decomposition level J to obtain the wavelet packet coefficients of each sub-band. ,in It is the wavelet packet coefficient sequence of the i-th sub-band within the k-th time window.

[0034] The second step is to reconstruct the wavelet packet coefficients of each sub-band to obtain the time-domain reconstructed signal of each sub-band.

[0035] The wavelet packet coefficients of each sub-band are reconstructed individually, and the wavelet packet coefficients of each sub-band are reconstructed into a time-domain signal separately. Let the time-domain signal of the i-th sub-band after reconstruction be... ,but:

[0036] In the formula, Let be the reconstructed wavelet basis function corresponding to the i-th subband of the J-th layer. Through single-branch reconstruction, the frequency domain subband coefficients are restored to the time domain signal, which facilitates subsequent energy calculation.

[0037] The third step is to calculate the integral of the square of the time-domain reconstructed signal amplitude for each sub-band, which is then used as the energy of that sub-band.

[0038] Sub-band energy reflects the signal strength within that frequency band. The energy of the i-th sub-band within the k-th time window. Calculate using the following formula:

[0039] For discrete signals, the integral can be approximated as the sum of the squares of the amplitudes at each sampling point.

[0040] The fourth step is to aggregate the energy of each sub-band to obtain the sub-band energy feature vector for the current time window.

[0041] Arrange the energy of all sub-bands within the current time window in frequency band order to form the sub-band energy feature vector. .

[0042] The definition of band energy entropy is as follows: Let the number of decomposition layers be J, then we get... For each sub-band, calculate the proportion of energy in each sub-band i to the total energy of all sub-bands as the energy distribution probability. :

[0043] The frequency band energy entropy is:

[0044] The smaller the frequency band energy entropy, the higher the energy concentration, the more significant the energy difference between the sub-bands after decomposition, and the better the decomposition effect; conversely, the larger the entropy value, the more dispersed the energy distribution between the sub-bands, and the worse the decomposition effect.

[0045] S2.3, traverse all wavelet bases in the candidate wavelet base set and all layers in the candidate decomposition layers, repeat step S2.2, and obtain the frequency band energy entropy corresponding to each candidate combination.

[0046] Let the set of candidate wavelet bases be The set of candidate decomposition levels is For the vibration signal segment of the k-th time window Iterate through all candidate combinations Calculate the corresponding frequency band energy entropy respectively. .

[0047] S2.4 Compare the frequency band energy entropy of all candidate combinations, and select the candidate combination with the smallest frequency band energy entropy as the optimal wavelet basis and the optimal number of decomposition layers for the current time window.

[0048] Optimal decomposition parameters Determined by the following formula:

[0049] in For the set of candidate wavelet bases, Let be the vibration signal segment in the k-th time window. Minimizing energy entropy means that the signal energy after decomposition is concentrated in a few sub-bands, indicating that the combination of this wavelet basis and the number of decomposition layers can most effectively match the time-frequency structure of the current signal, making the frequency band where the fault characteristics are located more prominent.

[0050] S2.5, the energy of each sub-band corresponding to the optimal wavelet basis and the optimal decomposition level is used as the extracted sub-band energy features to form the sub-band energy feature vector of the current time window.

[0051] The optimal wavelet basis determined in step S2.4 and the optimal number of decomposition levels Wavelet packet decomposition is performed on the vibration signal segment within the current time window to extract the energy features of each sub-band, thus constructing the sub-band energy feature vector for the current time window. .

[0052] It should be noted that, since the optimal decomposition level may differ for each time window, the dimensionality of the sub-band energy feature vectors may also differ for different time windows. In subsequent steps, a unified selection of sub-bands is performed using rotational speed-weighted mutual information filtering, mapping the feature vectors of different time windows to a unified dimension.

[0053] Through adaptive wavelet packet decomposition, the vibration signal of each time window can be analyzed in time and frequency with optimal decomposition parameters. This avoids the problem of decomposition mismatch caused by changes in frequency band structure when the working conditions change due to fixed wavelet basis and fixed decomposition level. This provides high-quality time and frequency features for subsequent feature selection and fault identification.

[0054] S3. Calculate the average speed based on the instantaneous speed information within the current time window. Based on the average speed, introduce speed weights to calculate the weighted mutual information between the energy of each sub-band and the fault type. Select the sub-band energy with weighted mutual information greater than the dynamic threshold as the filtered feature subset. Concatenate the feature subset, time-domain statistical features, frequency-domain statistical features, and speed normalization factor into an adaptive feature vector for the working condition.

[0055] In this embodiment, step S2 obtains the sub-band energy feature vector for each time window. However, the optimal decomposition level may differ for different time windows, resulting in inconsistent dimensions of the sub-band energy feature vectors. Furthermore, the original sub-band energy features may contain frequency band components with weak correlation to the fault type. Step S3 uses speed-weighted mutual information filtering to select sensitive sub-bands with strong correlation to the fault type from the full-band sub-bands, and constructs a feature vector with unified dimensions and adaptive operating conditions using operating condition information.

[0056] S3.1 Calculate the average rotational speed of the current time window based on the instantaneous rotational speed signal within the current time window. And obtain the minimum rotational speed during the entire sampling process. and maximum speed .

[0057] For the k-th time window, based on the instantaneous rotational speed signal within that time window... Calculate the average rotational speed within this time window. :

[0058] In the formula, This is the window length of the sliding time window. For discrete sampled signals, the average rotational speed is obtained by taking the arithmetic mean of the rotational speed values ​​at each sampled point within this time window.

[0059] S3.2, normalize the average speed, and calculate the speed weight of the current time window based on the normalized average speed and the preset speed influence factor. .

[0060] The rotational speed weight for the current time window is calculated using the following formula. :

[0061] In the formula, α is the preset rotational speed influence factor.

[0062] when hour, That is, the speed weight takes the minimum value, indicating that the subband energy is less affected by speed modulation at low speeds, and its mutual information value accounts for a relatively low proportion in the weighted average; when hour, The maximum value for the rotational speed weight indicates that the subband energy is significantly affected by rotational speed modulation at high speeds, and its mutual information value accounts for a relatively high proportion in the weighted average. By using rotational speed weights, the mutual information contribution of each subband at different speeds can be reasonably adjusted.

[0063] The sub-band energy of vibration signals is affected by rotational speed; the higher the rotational speed, the greater the vibration energy in each frequency band. If mutual information is calculated directly without considering rotational speed, the contribution of samples collected under high-speed conditions to the mutual information will be amplified, and the selected features may be biased towards characterizing rotational speed changes rather than fault conditions. By compressing the contribution of high-speed samples through rotational speed weighting, the selected features truly reflect fault information rather than rotational speed information.

[0064] S3.3, calculate the mutual information value between the energy of each sub-band and the corresponding fault tag within each time window. Then, the weighted mutual information of the energy of each subband is calculated. In the formula, K represents the total number of time windows. The energy characteristic of the i-th sub-band within the k-th time window. This corresponds to the fault label.

[0065] Mutual information (MI) is a measure of the degree of interdependence between two random variables. It can capture nonlinear correlations and is suitable for feature selection in fault diagnosis.

[0066] For the energy characteristic of the i-th subband Its relationship with fault labels Mutual information between them is defined as:

[0067] in for and The joint probability distribution, and They are respectively and The marginal probability distribution.

[0068] For each time window, calculate the energy characteristics of each sub-band separately. Corresponding fault label Mutual information values ​​between ,in The energy characteristic of the i-th sub-band within the k-th time window. This is the fault label corresponding to the k-th time window.

[0069] Weighted mutual information Calculate using the following formula:

[0070] Where K is the total number of time windows.

[0071] The weighted mutual information calculation assigns a mutual information value for each time window to represent the correlation between sub-band energy and fault type within that time window. This value is then multiplied by the rotational speed weight for that time window and summed to obtain the overall correlation between that sub-band and fault type across the entire operating range. Time windows with lower weights contribute less to the summing and averaging, thus reducing the impact of rotational speed on the mutual information calculation.

[0072] S3.4, Obtain the average speed under the current operating conditions. and rated speed Calculate dynamic threshold In the formula This is the preset baseline threshold.

[0073] When the average rotational speed deviates from the rated speed, the amplitude range of the sub-band energy changes. A fixed threshold may result in screening too few features at low speeds and too many features at high speeds. A dynamic threshold adjusts in the same direction as the rotational speed to keep the number of features screened balanced across all speeds.

[0074] S3.5 Select the subband energy with weighted mutual information greater than the dynamic threshold as the filtered feature subset.

[0075] This embodiment eliminates frequency band components that are weakly correlated with the fault type through this screening step, while retaining sub-band features that are sensitive to faults across the entire operating range. This reduces the feature dimension and alleviates the computational burden on the subsequent liquid neural network.

[0076] S3.6, combine the filtered feature subset, time-domain statistical features, frequency-domain statistical features, and rotational speed normalization factor. The vectors are concatenated to form an adaptive feature vector for the working conditions.

[0077] Time-domain statistical features are characteristic parameters obtained by statistically analyzing vibration signals in the time domain, reflecting the temporal distribution characteristics of the signal. The time-domain statistical features used in this embodiment include the following eight: peak value (reflecting the maximum amplitude of the signal), mean value (reflecting the central trend of the signal), root mean square value (reflecting the effective energy of the signal), kurtosis (reflecting the intensity of the impact component), kurtosis factor, impulse factor, waveform factor, and skewness (reflecting the asymmetry of the signal distribution).

[0078] Frequency domain statistical features are characteristic parameters obtained by statistical analysis of vibration signals in the frequency domain, reflecting the frequency distribution characteristics of the signal. The frequency domain statistical features used in this embodiment include the following eight: frequency domain peak value, frequency domain mean value, frequency domain root mean square value, frequency domain kurtosis, frequency domain skewness, centroid frequency (reflecting the center position of spectral energy), root mean square frequency (reflecting the dispersion of the spectrum), and fourth root frequency.

[0079] Speed ​​normalization factor As an auxiliary feature, it is added to the feature vector to eliminate the influence of speed variation on the feature amplitude, so that the feature vector is comparable under different operating conditions.

[0080] The filtered subband energy feature subset Time-domain statistical eigenvectors Frequency domain statistical eigenvectors Concatenating the vector with the speed normalization factor yields the adaptive feature vector for the operating condition. :

[0081] In this embodiment, through the splicing operation, the feature vector contains both time-frequency domain fault information extracted from the vibration signal and operating condition information, providing a complete input for the subsequent operating condition-linked LNN.

[0082] S4, input the working condition adaptive feature vector, the instantaneous speed value of the current time window and the speed change rate into the working condition linkage liquid neural network, and output the fault type identification result; the liquid time constant of each neuron in the liquid neural network is dynamically adjusted in real time according to the working condition parameters, and the network state is controlled by variable coefficient ordinary differential equations for iterative updates.

[0083] Liquid Neural Networks (LNNs) are a novel type of neural network that uses ordinary differential equations (ODEs) to describe the continuous-time dynamic characteristics of neuron states. LNNs leverage the liquid time constant to regulate the flow rate of information within the network and the degree of historical information forgetting, giving them an advantage in processing temporally dynamic signals. This embodiment further designs the liquid time constant to be a linkage mechanism that changes with operating conditions, enabling the network's dynamic characteristics to be adjusted in real time according to the rate of change of rotational speed.

[0084] S4.1 Calculate the average rotational speed of the current time window based on the instantaneous rotational speed signal within the current time window. The rate of change of rotational speed is calculated based on the difference between the average rotational speed of the current time window and the average rotational speed of the previous time window. .

[0085] Rate of change of rotational speed It is used to characterize the degree of drastic change in the current operating condition and reflects the acceleration / deceleration state. Its calculation method is as follows:

[0086] In the formula, The average rotational speed of the time window in the previous moment. This represents the time interval between adjacent time windows. Time indicates steady-state operating conditions. Indicates acceleration mode. Indicates deceleration condition. The larger the value, the more drastic the changes in operating conditions.

[0087] S4.2 sets a basic time constant for each neuron h in the liquid neural network. Then, the liquid time constant of each neuron at the current moment is calculated. The linkage function In the formula, β is the preset minimum time constant scaling factor, and γ is the preset speed change rate scaling parameter.

[0088] In this embodiment, the basic time constant This represents the baseline response speed of a neuron under steady-state conditions.

[0089] This linkage function indicates that: when Time (steady-state operating condition). time constant By maximizing the value, neurons retain more historical information to capture long-term trend characteristics of failures. When the exponent increases (acceleration / deceleration transient conditions), the exponent term Decrease Approaching β, time constant By taking the minimum value, neurons pay more attention to the current input and respond quickly to changes in the signal.

[0090] The speed change rate scale parameter γ is used to control the sensitivity of the time constant to the speed change rate: the smaller γ is, the more sensitive the time constant is to the speed change rate, and a small speed change rate can trigger a decrease in the time constant; the larger γ is, the less sensitive the time constant is to the speed change rate, and a decrease in the time constant is only triggered when the speed change rate is large.

[0091] S4.3, the adaptive feature vector of the working condition is used as the input I(t) at the current time, and the liquid time constant of each neuron constitutes the time-varying time constant vector. The iterative update of the neuron state x(t) is controlled by a variable coefficient ordinary differential equation.

[0092] The variable coefficient ordinary differential equation is a first-order variable coefficient ordinary differential equation, expressed as:

[0093] In the formula, x(t) is the neuron state vector, and its elements are... This represents the membrane potential state of the h-th neuron. The time constant vector of the time-varying liquid is determined in real time according to the working conditions as described in step S4.2, and N is the total number of neurons in the liquid layer. I(t) is the input feature vector at the current time, i.e., the adaptive feature vector of the working condition constructed in step S3.6. ; This is a neural network mapping function with parameter θ, used to describe the nonlinear interactions between neurons.

[0094] In this equation, The leakage term represents the natural decay of the neuron's state over time, with the decay rate determined by the time constant. control; The driving term represents the stimulating effect of the current input and other neuron states on the current neuron.

[0095] In this equation, the time constant Instead of being a fixed value, it is a time-varying vector that changes with operating conditions, allowing the network's dynamic characteristics to adjust in real time according to the rate of change of rotational speed. When the time constant is large, leakage terms... The decay rate is relatively slow, and historical information is retained in the network for a longer time; when the time constant is small, the decay rate is faster, historical information is quickly forgotten, and the network state is mainly determined by the current input.

[0096] It should be noted that the liquid neural network described in step S4.3 is a pre-trained model. During the training phase, this model optimizes the network parameters θ and the basic time constant using the backpropagation algorithm. This enables the network to accurately identify fault types under various operating conditions. Specific parameter settings during the training process will be detailed in subsequent embodiments.

[0097] S4.4 uses numerical integration to discretize and iteratively solve the ordinary differential equations, updating the state of each neuron.

[0098] As a preferred method, the fourth-order Runge-Kutta method (RK4) is used for numerical solution.

[0099] S4.5, the iteratively updated neuron state is input into the fully connected layer, mapped to a probability distribution by the Softmax activation function, and the category corresponding to the highest probability is taken as the fault type identification result.

[0100] After the liquid neural network completes its iterative update, the neuron state x(t) at the final time step is... k +1) The input fully connected layer is linearly transformed to obtain the original output logits values ​​for the four categories.

[0101] The output of the fully connected layer is mapped to a probability distribution by the Softmax activation function:

[0102] in, Let be the output logit value of the c-th neuron in the fully connected layer. The denominator is the exponential sum of the logit values ​​of all four classes, ensuring that the sum of the probabilities of the four classes is 1.

[0103] The output layer contains four neurons, corresponding to four states: normal state (label 1), inner ring fault (label 2), outer ring fault (label 3), and rolling element fault (label 4). The Softmax function maps the output to the probability distribution of these four states, and takes the category with the highest probability as the fault type identification result.

[0104] In this embodiment, the training process of the working condition linkage type liquid neural network is as follows.

[0105] The training of the liquid neural network adopts a supervised learning method. The training dataset consists of the condition-adaptive feature vectors obtained from steps S2 and S3 and their corresponding fault labels, containing samples of four states: normal state, inner race fault, outer race fault, and rolling element fault.

[0106] During training, the output layer of the liquid neural network uses the Softmax activation function to map the network output to a probability distribution of four states. The loss function used is cross-entropy loss, which measures the difference between the predicted probability distribution and the true label.

[0107] in One-hot encoding of the real label. This represents the c-th class probability value output by Softmax.

[0108] The optimizer chosen is Adam (Adaptive Moment Estimation), which adaptively adjusts the learning rate and calculates the first-order moment estimate and the second-order moment estimate on different parameter dimensions, which can effectively accelerate convergence and improve training stability.

[0109] Learnable parameters during training include: (1) Liquid Neural Network Mapping Function The parameter θ in the equation includes the connection weights and biases between neurons; (2) The base time constant of each neuron This parameter is optimized synchronously with the weights during training, so that the response characteristics of each neuron under steady-state conditions match the fault identification task.

[0110] During training, the gradient is propagated to the target region via backpropagation of the variable coefficient ordinary differential equation. This ensures that the network can obtain the optimal time-series information retention length at each steady-state operating point covered by the training data.

[0111] During parameter update, the time constant It is constrained within a preset range to avoid numerical instability.

[0112] As a preferred approach, the training parameters are set as follows: batch size of 32, maximum number of epochs of 200, and initial learning rate of 0.001. The learning rate adopts an exponential decay strategy, decreasing to 0.9 times the current value every 50 epochs to fine-tune the parameters in the later stages of training.

[0113] To prevent overfitting, an early stopping strategy is adopted during training: 20% of the training set is randomly selected as the validation set, and training is terminated when the validation set loss does not decrease for 15 consecutive rounds, and the model parameters with the lowest validation set loss are saved.

[0114] Through the above training process, the working condition-linked liquid neural network learns the mapping relationship from the working condition adaptive feature vector to the four fault states, and can accurately identify the fault type on new samples.

[0115] The foregoing has described in detail an embodiment of a rolling bearing fault diagnosis method based on WPD-LNN. Based on the rolling bearing fault diagnosis method based on WPD-LNN described in the above embodiment, the present invention also provides a rolling bearing fault diagnosis system based on WPD-LNN corresponding to the method.

[0116] Figure 2 This is a schematic block diagram of a rolling bearing fault diagnosis system based on WPD-LNN provided in an embodiment of the present invention. In this embodiment, the rolling bearing fault diagnosis system based on WPD-LNN can be divided into multiple functional modules according to the functions it performs. A module, as referred to in this invention, is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory.

[0117] The signal acquisition module is used to synchronously acquire the original vibration signal and instantaneous speed signal of the rolling bearing in operation, and perform noise reduction and normalization preprocessing respectively.

[0118] The adaptive decomposition module is used to divide the preprocessed vibration signal into several analysis segments using a sliding time window. Within each time window, based on the frequency band energy entropy minimization criterion, the optimal wavelet basis and optimal decomposition level are dynamically selected from the candidate wavelet basis set and candidate decomposition level. Adaptive wavelet packet decomposition is then performed on the vibration signal segment of the current time window to extract the energy features of each sub-band.

[0119] The feature construction module is used to calculate the average speed based on the instantaneous speed information within the current time window. Based on the average speed, speed weights are introduced to calculate the weighted mutual information between the energy of each sub-band and the fault type. Sub-band energies with weighted mutual information greater than the dynamic threshold are selected as the filtered feature subset. The feature subset, time-domain statistical features, frequency-domain statistical features, and speed normalization factor are concatenated into an adaptive feature vector for the operating condition.

[0120] The fault diagnosis module is used to input the adaptive feature vector of the working condition, the instantaneous speed value of the current time window and the speed change rate into the working condition linkage liquid neural network, and output the fault type identification result; the liquid time constant of each neuron in the liquid neural network is dynamically adjusted in real time according to the working condition parameters, and the network state is controlled by variable coefficient ordinary differential equations for iterative updates.

[0121] The rolling bearing fault diagnosis system based on WPD-LNN in this embodiment is used to implement the aforementioned rolling bearing fault diagnosis method based on WPD-LNN. Therefore, the specific implementation of this system can be found in the embodiment section of the rolling bearing fault diagnosis method based on WPD-LNN mentioned above. Thus, the specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.

[0122] Furthermore, since the WPD-LNN-based rolling bearing fault diagnosis system in this embodiment is used to implement the aforementioned WPD-LNN-based rolling bearing fault diagnosis method, its function corresponds to the function of the above method, and will not be repeated here.

[0123] Figure 3 This is a schematic diagram of a terminal 300 provided in an embodiment of the present invention, including: a processor 310, a memory 320, and a communication unit 330. The processor 310 is used to implement the flow steps of the above-described embodiment of the rolling bearing fault diagnosis method based on WPD-LNN when implementing the rolling bearing fault diagnosis program based on WPD-LNN stored in the memory 320.

[0124] This invention also provides a computer storage medium, which may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. The computer storage medium stores a rolling bearing fault diagnosis program based on WPD-LNN. When the WPD-LNN-based rolling bearing fault diagnosis program is executed by a processor, it implements the process steps of the above-described embodiment of the rolling bearing fault diagnosis method based on WPD-LNN.

[0125] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A rolling bearing fault diagnosis method based on WPD-LNN, characterized in that, Includes the following steps: S1, synchronously acquire the original vibration signal and instantaneous speed signal of the rolling bearing in operation, and perform noise reduction and normalization preprocessing respectively; S2, the preprocessed vibration signal is divided into several analysis segments by using a sliding time window. Within each time window, based on the frequency band energy entropy minimization criterion, the optimal wavelet basis and optimal decomposition level are dynamically selected from the candidate wavelet basis set and candidate decomposition level. Adaptive wavelet packet decomposition is performed on the vibration signal segment of the current time window to extract the energy features of each sub-band. S3, calculate the average speed based on the instantaneous speed information within the current time window, and calculate the weighted mutual information between the energy of each sub-band and the fault type based on the average speed and the speed weight. Select the sub-band energy with the weighted mutual information greater than the dynamic threshold as the filtered feature subset, and concatenate the feature subset, time domain statistical features, frequency domain statistical features and speed normalization factor into the working condition adaptive feature vector. S4, input the working condition adaptive feature vector, the instantaneous speed value of the current time window and the speed change rate into the working condition linkage liquid neural network, and output the fault type identification result; the liquid time constant of each neuron in the liquid neural network is dynamically adjusted in real time according to the working condition parameters, and the network state is controlled by variable coefficient ordinary differential equations for iterative updates.

2. The rolling bearing fault diagnosis method based on WPD-LNN according to claim 1, characterized in that, The candidate wavelet basis set is pre-set to include the db family wavelet, sym family wavelet and coif family wavelet, and the candidate decomposition level is pre-set to 3, 4, 5 and 6 levels.

3. The rolling bearing fault diagnosis method based on WPD-LNN according to claim 2, characterized in that, Step S2 specifically includes: S2.1, Set the window length and overlap of the sliding time window to divide the preprocessed vibration signal into several analysis segments; S2.2 Select a wavelet basis from the candidate wavelet basis set and select a number of layers from the candidate decomposition layer range. Use the wavelet basis and the number of layers to perform wavelet packet decomposition on the vibration signal segment of the current time window to obtain each sub-band under the decomposition layer, calculate the energy of each sub-band, and calculate the frequency band energy entropy based on the energy of each sub-band. S2.3, traverse all wavelet bases in the candidate wavelet base set and all layers in the candidate decomposition layers, repeat step S2.2, and obtain the frequency band energy entropy corresponding to each candidate combination; S2.4 Compare the frequency band energy entropy of all candidate combinations, and select the candidate combination with the smallest frequency band energy entropy as the optimal wavelet basis and the optimal number of decomposition layers for the current time window; S2.5, the energy of each sub-band corresponding to the optimal wavelet basis and the optimal decomposition level is used as the extracted sub-band energy features to form the sub-band energy feature vector of the current time window.

4. The rolling bearing fault diagnosis method based on WPD-LNN according to claim 3, characterized in that, Using the wavelet basis and the number of layers, wavelet packet decomposition is performed on the vibration signal segment of the current time window to obtain each sub-band under the decomposition layer. The energy of each sub-band is calculated, specifically including: Based on the wavelet basis and the number of layers, wavelet packet decomposition is performed on the vibration signal segment of the current time window to obtain the wavelet packet coefficients of each sub-band; The wavelet packet coefficients of each sub-band are reconstructed to obtain the time-domain reconstructed signal of each sub-band; Calculate the integral of the square of the time-domain reconstructed signal amplitude for each sub-band, and use it as the energy of that sub-band; By pooling the energy of each sub-band, we obtain the sub-band energy feature vector for the current time window.

5. The rolling bearing fault diagnosis method based on WPD-LNN according to claim 3, characterized in that, The frequency band energy entropy is calculated based on the energy of each sub-band, specifically including: Let the number of decomposition layers be J, then we get For each sub-band, calculate the proportion of energy in each sub-band i to the total energy of all sub-bands as the energy distribution probability. Then the frequency band energy entropy .

6. The rolling bearing fault diagnosis method based on WPD-LNN according to claim 1, characterized in that, Step S3 specifically includes: S3.1 Calculate the average rotational speed of the current time window based on the instantaneous rotational speed signal within the current time window. And obtain the minimum rotational speed during the entire sampling process. and maximum speed ; S3.2, normalize the average speed, and calculate the speed weight of the current time window based on the normalized average speed and the preset speed influence factor. ; S3.3, calculate the mutual information value between the energy of each sub-band and the corresponding fault tag within each time window. Then, the weighted mutual information of the energy of each subband is calculated. In the formula, K represents the total number of time windows. The energy characteristic of the i-th sub-band within the k-th time window. For the corresponding fault label; S3.4, Obtain the average speed under the current operating conditions. and rated speed Calculate dynamic threshold In the formula This is a preset baseline threshold. S3.5, select the subband energy with weighted mutual information greater than the dynamic threshold as the filtered feature subset; S3.6, combine the filtered feature subset, time-domain statistical features, frequency-domain statistical features, and rotational speed normalization factor. The vectors are concatenated to form an adaptive feature vector for the working conditions.

7. The rolling bearing fault diagnosis method based on WPD-LNN according to claim 6, characterized in that, The rotational speed weight for the current time window is calculated using the following formula. : In the formula, α is the preset rotational speed influence factor.

8. The rolling bearing fault diagnosis method based on WPD-LNN according to claim 1, characterized in that, Step S4 specifically includes: S4.1 Calculate the average rotational speed of the current time window based on the instantaneous rotational speed signal within the current time window. The rate of change of rotational speed is calculated based on the difference between the average rotational speed of the current time window and the average rotational speed of the previous time window. ; S4.2 sets a basic time constant for each neuron h in the liquid neural network. Then, the liquid time constant of each neuron at the current moment is calculated. The linkage function In the formula, β is the preset minimum time constant scaling factor, and γ is the preset speed change rate scaling parameter. S4.3, the adaptive feature vector of the working condition is used as the input I(t) at the current time, and the liquid time constant of each neuron constitutes the time-varying time constant vector. The iterative update of the neuron state x(t) is controlled by a variable coefficient ordinary differential equation; S4.4, the numerical integration method is used to discretize and iteratively solve the ordinary differential equations and update the state of each neuron; S4.5, the iteratively updated neuron state is input into the fully connected layer, mapped to a probability distribution by the Softmax activation function, and the category corresponding to the highest probability is taken as the fault type identification result.

9. The rolling bearing fault diagnosis method based on WPD-LNN according to claim 8, characterized in that, The variable coefficient ordinary differential equation is a first-order variable coefficient ordinary differential equation, expressed as: In the formula, Let be the neural network mapping function with parameter θ.

10. A rolling bearing fault diagnosis system based on WPD-LNN, characterized in that, include: The signal acquisition module is used to synchronously acquire the original vibration signal and instantaneous speed signal of the rolling bearing in operation, and perform noise reduction and normalization preprocessing respectively. The adaptive decomposition module is used to divide the preprocessed vibration signal into several analysis segments using a sliding time window. Within each time window, based on the frequency band energy entropy minimization criterion, the optimal wavelet basis and optimal decomposition level are dynamically selected from the candidate wavelet basis set and candidate decomposition level. Adaptive wavelet packet decomposition is performed on the vibration signal segment of the current time window to extract the energy features of each sub-band. The feature construction module is used to calculate the average speed based on the instantaneous speed information within the current time window. Based on the average speed, speed weights are introduced to calculate the weighted mutual information between the energy of each sub-band and the fault type. Sub-band energy with weighted mutual information greater than the dynamic threshold is selected as the filtered feature subset. The feature subset, time-domain statistical features, frequency-domain statistical features and speed normalization factor are concatenated into an operating condition adaptive feature vector. The fault diagnosis module is used to input the adaptive feature vector of the working condition, the instantaneous speed value of the current time window and the speed change rate into the working condition linkage liquid neural network, and output the fault type identification result; the liquid time constant of each neuron in the liquid neural network is dynamically adjusted in real time according to the working condition parameters, and the network state is controlled by variable coefficient ordinary differential equations for iterative updates.

Images