Establishment method, application method and system of envelope signal extraction model
Through the B-LSTM deep learning model and bidirectional LSTM structure, the overshoot phenomenon and noise interference problems of traditional methods in complex signal processing are solved, and high-precision fault feature extraction and adaptive envelope spectrum analysis are realized.
Patent Information
- Application Number
- CN202510672805.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-23
- Publication Date
- 2025-07-01
AI Technical Summary
When the prior art deals with complex and variable mechanical vibration signals, traditional methods are difficult to accurately capture subtle fault characteristics, and signal processing is difficult under strong noise interference, resulting in poor fault diagnosis accuracy and frequency intensity.
The B-LSTM deep learning model is used to replace the traditional Hilbert transformation and the Teager energy operator. The fault signal data set is generated in stages, combined with the multi-parameter coupled simulation formula, and the correlation relationship between the signal and the envelope spectrum line is established. The bidirectional LSTM structure is used to capture timing features and filter.
It significantly improves the analysis accuracy of strong noise signals and early weak fault signals, reduces the misjudgment rate, meets the monitoring requirements of high-reliability equipment, and realizes the extraction of adaptive envelope spectrum lines.
Smart Images
Figure CN120234597A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deep learning, and particularly relates to a method for establishing an envelope signal extraction model, an application method, and a system. Background Art
[0002] In actual signal analysis scenarios, signals often exhibit a high degree of complexity, containing multiple signal components with different frequency components and modulation characteristics. These components are intertwined and superimposed, making it extremely difficult to extract valuable information from the signals. Signal envelope spectra often carry a large amount of useful information related to fault characteristics and play an important role in mechanical fault diagnosis and signal analysis. The signal envelope spectra obtained through Hilbert transform, Teager energy operator, etc. can convert such non-linear signals into a form that is easier to understand and analyze, deeply revealing the internal structure and variation law of the signals. With its unique analysis mechanism, it shows significant advantages in dealing with non-linear and non-stationary complex signals. In a complex industrial environment, the collected signals are often accompanied by high-intensity noise, whose intensity even far exceeds that of the fault signal itself. At this time, this characteristic of the signal envelope spectrum is particularly crucial, as it can significantly improve the signal-to-noise ratio of the signal and make the analysis results more reliable.
[0003] Many signals in actual engineering, such as mechanical vibration signals, often have non-linear characteristics. However, current envelope spectrum line extraction methods, such as traditional Hilbert transform and energy operator calculation, have certain limitations when facing complex and variable actual signals. When dealing with signals with strong noise interference or signals with complex modulation characteristics, the extracted envelope spectrum lines are prone to distortion, making it difficult to accurately capture the subtle fault characteristics in the signals, resulting in poor accuracy and fault frequency intensity of subsequent fault diagnosis based on the envelope spectrum. Moreover, when the signal contains multiple signal components with different frequency components and modulation characteristics that are superimposed on each other, these methods show problems of low efficiency and poor effect in separating these complex signal components to obtain accurate envelope spectrum lines. In order to solve the above problems, scholars often pay more attention to the superimposed use of various complex signal filtering and demodulation methods to extract and separate the effective frequency characteristics from the interference sources. However, too many signal processing methods not only increase the analysis cost but sometimes also introduce artificial interference components, distorting the originally weak signal characteristics. Summary of the Invention
[0004] The object of the present invention is: aiming at the above existing problems, the present invention provides a method for establishing an envelope signal extraction model, an application method, and a system. By replacing the traditional Hilbert transform, Teager energy operator, and cubic spline interpolation method with a B-LSTM deep learning model, the core defects such as overshoot phenomenon, non-differentiability, difficult signal processing, and poor generalization ability existing in the prior art are solved, and the analysis accuracy of strong noise signals and early weak fault signals is significantly improved.
[0005] A method for establishing an envelope signal extraction model includes the following steps: Obtain a local fault signal dataset: Select a detection target, where the detection target includes multiple components. Generate a simulated signal with a single resonance frequency for the specified component with damage as the local fault signal corresponding to this component; Replace another component with damage to be measured, and repeat this step until the local fault signals of all components of the detection target are obtained, and establish a local fault signal dataset of the detection target; Obtain a dataset of fault signals with different intensities: Establish a mathematical analysis model based on the local fault signal dataset of the detection target. This mathematical analysis model includes multiple adjustable parameters, denoted as P1, P2,..., P n ; Perform feature intensity training on all adjustable parameters respectively to establish a dataset of fault signals with different intensities for all adjustable parameters; Obtain a composite fault signal dataset: First, perform a full permutation and combination on the dataset of fault signals with different intensities for each adjustable parameter to obtain all possible sets of adjustable parameters P = {P1, P2,..., P n}, and use the signal simulation formula to calculate the fault signals of all sets of adjustable parameters P to obtain the corresponding composite fault signal dataset; Establish an envelope signal extraction model: Normalize the composite fault signal dataset to obtain a standardized fault signal dataset, and use the B-LSTM model to train the standardized fault signal dataset to establish the correlation relationship between the input signal and the envelope spectrum line, thereby generating an envelope signal extraction model for the detection target.
[0006] Due to the adoption of the above technical solution, by replacing the traditional Hilbert transform, Teager energy operator, or cubic spline interpolation method with a data-driven deep learning framework, the defects such as overshoot phenomenon, non-differentiability, and difficult signal processing existing in the prior art are solved; The local fault, variable-intensity fault, and composite fault datasets are generated in stages, which can comprehensively cover the fault types and parameter combinations that may occur in the actual operation of the detection target, and improve the diversity of model training data; Through the bidirectional time-series feature capture ability of the B-LSTM model, the complex mapping relationship between the signal and the envelope spectrum line can be effectively established, providing a high-precision analysis basis for subsequent fault diagnosis.
[0007] Due to the above technical solution, the quantitative simulation of the fault intensity is realized through multi-dimensional parameter regulation. In particular, the introduction of the signal-to-noise ratio parameter can simulate the noise interference in the actual working condition and enhance the robustness of the model in the complex environment.
[0008] Furthermore, in the step of obtaining the fault signal datasets with different intensities, the specific process of feature intensity training includes: selecting an adjustable parameter, setting its value range as [X1, X2], selecting the data when the value is X1 as the original adjustment data N0, obtaining the new adjustment data N1 = N0 + Y according to the set step size Y, so that the fault feature intensity presented by the corresponding simulated signal changes, obtaining the simulated signal corresponding to the adjustment data N1 as the variable-intensity fault signal, obtaining the new adjustment data N2 = N1 + Y again according to the set step size Y, obtaining the simulated signal corresponding to the adjustment data N2 as the variable-intensity fault signal, repeating the above steps until all the variable-intensity fault signals within the value range [X1, X2] of the adjustable parameter are obtained, and establishing the basic dataset of the fault signal with different intensities for this adjustable parameter; then adding random perturbations to each adjustment data within the range of the step size Y to obtain the fault signal datasets with different intensities.
[0009] Due to the above technical solution, by combining the method of parameter step-by-step adjustment and random perturbation, it not only ensures the systematicness of parameter changes but also increases the randomness of data samples, so as to more comprehensively obtain the comprehensive influence of different characteristic frequencies and their random fluctuations on the overall characteristics of the signal, and improve the generalization recognition ability of fault characteristics with different intensities.
[0010] Furthermore, in the step of establishing the envelope signal extraction model, first divide the composite fault signal dataset into three datasets: training set, validation set and test set according to the set ratio, then perform normalization processing on the training set and record the normalization parameters, and then synchronously transform the validation set and the test set using the normalization parameters; train the constructed training set using the B-LSTM model, use the cross-entropy loss function as the loss function, select the Adam function as the optimization function, and iteratively train until the R² fitting rate of the test set reaches the target fitting value, and save the model with the best result in the validation set as the envelope signal extraction model for the detection target.
[0011] Due to the above technical solution, adopting a scientific data division and parameter synchronous conversion strategy ensures the data consistency in the model training, validation and testing links; through the combined use of the cross-entropy loss function and the Adam optimizer, it significantly improves the model convergence speed and parameter optimization efficiency.
[0012] Further, in the step of establishing the envelope signal extraction model, when using the B-LSTM model to train the standardized fault signal dataset, the B-LSTM model performs forward LSTM sequential processing and backward LSTM reverse processing on the standardized fault signal simultaneously until the forward LSTM sequential processing and the backward LSTM reverse processing are merged, capturing the bidirectional dependency relationship of the time series corresponding to the envelope signal, mapping the LSTM output at each time step to the envelope signal dimension, establishing the correlation relationship between the input signal and the envelope signal, and finally generating the envelope signal extraction model for the detection target.
[0013] Due to the above technical solution, the bidirectional LSTM structure can capture the front and rear timing characteristics of the fault signal simultaneously, effectively solve the problem that the traditional unidirectional model does not make full use of historical information, and significantly improve the ability to analyze the timing characteristics of the fault signal.
[0014] Further, the specific process of the B-LSTM model working includes: Based on the first bidirectional LSTM layer, feature extraction is performed on the input standardized fault signal from both the forward and backward directions simultaneously, and the bidirectional outputs are concatenated to obtain the first processed data; Based on the dropout layer, some neurons of the first processed data are randomly discarded to obtain the second processed data; Based on the second bidirectional LSTM layer, high-dimensional feature extraction is performed on the second processed data from both the forward and backward directions simultaneously, and the bidirectional outputs are concatenated to obtain the third processed data; Based on the fully connected layer, the high-dimensional features of the third processed data are mapped to the envelope signal dimension.
[0015] Due to the above technical solution, multi-level abstraction of features is realized through the double-layer bidirectional LSTM structure, and with the regularization effect of the dropout layer, it can not only extract deep timing features but also prevent the network from overfitting, ensuring the generalization performance of the model.
[0016] An application method of an envelope signal extraction model, using the envelope signal extraction model established by the above envelope signal extraction model establishment method, includes the following steps: collecting the detection signal data during the operation of the detection target through a sensor, inputting the detection signal data into the envelope signal extraction model, obtaining the corresponding envelope signal according to the output of the envelope signal extraction model, performing filtering processing on the envelope signal to obtain the processed envelope signal, performing Fourier transform on the processed envelope signal to obtain the envelope spectrum line, judging whether the characteristic frequency of the envelope spectrum line meets the requirements, if so, intercepting the characteristic frequency that meets the requirements in the envelope spectrum line, retaining the information data corresponding to the analysis requirements, and discarding redundant data; if not, returning to perform filtering processing on the envelope signal and readjusting the filtering parameters of the envelope signal.
[0017] Due to the adoption of the above technical solution, the true detection signal is processed by the envelope signal extraction model to achieve efficient extraction of the characteristic frequency from the original vibration signal. Combining with the dynamic filtering parameter adjustment mechanism, it solves the problem that the traditional envelope spectrum method relies on manual experience when intercepting the characteristic frequency. Combining with the feedback optimization after Fourier transform, it realizes the adaptive extraction of the envelope spectrum line, significantly reducing the misjudgment rate in scenarios such as communication system signal recognition and structural health monitoring, and meeting the monitoring requirements of high-reliability equipment.
[0018] Furthermore, after performing singular value rejection processing and filtering and noise reduction processing on the detection signal data, it is then input into the envelope signal extraction model.
[0019] Due to the adoption of the above technical solution, the abnormal interference and background noise in the signal acquisition process are effectively eliminated through the preprocessing link, significantly improving the quality of the input signal and providing clean data input for subsequent model processing.
[0020] An envelope signal extraction system includes a signal acquisition module, a data processing module, and a human-computer interaction module. The signal acquisition module is used to obtain the detection signal during the operation of the detection target, and the detection signal is the original data that can reflect the real-time operation state of the detection target. The data processing module stores the envelope signal extraction model established by the above method for establishing the envelope signal extraction model, and the data processing module is used to call the envelope signal extraction model to process the detection signal input by the signal acquisition module to obtain the envelope signal or the envelope spectrum line. The human-computer interaction module is used to receive and display the envelope signal or the envelope spectrum line output by the data processing module.
[0021] An electronic device includes a processor, and the processor is coupled with a memory. The memory is used to store a program, and when the program is executed by the processor, the system is enabled to implement the steps in the above method for establishing the envelope signal extraction model.
[0022] A computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, the steps in the above method for establishing the envelope signal extraction model are implemented.
[0023] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are as follows: 1. The present invention replaces the traditional Hilbert transform and cubic spline interpolation methods with the B-LSTM deep learning model, solves the core defects such as overshoot phenomenon, difficult processing of non-differentiable signals, and poor generalization ability existing in the prior art, and significantly improves the analysis accuracy of strong noise signals and early weak fault signals.
[0024] 2. The present invention generates a full-dimensional dataset covering partial faults, variable-intensity faults, and compound faults in stages, combines multi-parameter coupling simulation formulas, and simulates non-linear modulation, noise interference, and multi-fault concurrent scenarios in the actual operation of the detection target, enabling the model to learn the signal decoupling ability under real working conditions and showing excellent robustness in non-stationary signal analysis.
[0025] 3. The present invention utilizes the timing feature capture ability of the bidirectional LSTM structure to synchronously analyze the forward and backward dependence relationships of the signal, establishes an envelope signal extraction model, and combines dropout layer regularization and the Adam optimizer to achieve efficient extraction of low-frequency fault features and suppression of high-frequency noise.
[0026] 4. The present invention breaks through the dependence on manual parameter adjustment of traditional methods through a dynamic filtering adjustment mechanism and an adaptive feature truncation strategy, and combines feedback optimization after Fourier transform to achieve adaptive extraction of envelope spectral lines, significantly reducing the misjudgment rate in scenarios such as signal recognition in communication systems and structural health monitoring, and meeting the monitoring requirements of high-reliability equipment. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 is a flowchart of the method for establishing the envelope signal extraction model of the present invention; Figure 2 is the characteristic frequency spectrum of the detection signal directly subjected to Hilbert envelope transform; Figure 3 is the characteristic frequency spectrum of the detection signal processed by the envelope signal extraction model; Figure 4 is the characteristic frequency spectrum of the noise composite signal directly subjected to Hilbert envelope transform; Figure 5 is the characteristic frequency spectrum of the noise composite signal processed by the envelope signal extraction model; Figure 6 is a schematic structural diagram of the envelope signal extraction system. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0028] The following will describe the present invention in detail with reference to the accompanying drawings.
[0029] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0030] Embodiment 1 A method for establishing an envelope signal extraction model, which can be applied to fault detection and diagnosis in fields such as bearings, gear mechanisms, pumps, compressors, motors, automotive manufacturing and maintenance, and power systems, such as Figure 1As shown, it includes the following steps: Obtain a local fault signal dataset: Select a detection target, where the detection target includes multiple components. Generate a simulated signal with a single resonance frequency for a specified component with damage as the local fault signal corresponding to this component; Replace another component to be tested with damage and repeat this step until the local fault signals of all components of the detection target are obtained, and establish a local fault signal dataset for the detection target; Obtain different-intensity fault signal datasets: Establish a mathematical analysis model based on the local fault signal dataset of the detection target. This mathematical analysis model includes multiple adjustable parameters, denoted as P1, P2, ……, P n ; Conduct feature intensity training on all adjustable parameters respectively to establish different-intensity fault signal datasets for all adjustable parameters; Obtain a composite fault signal dataset: First, perform a full permutation and combination on the different-intensity fault signal datasets of each adjustable parameter to obtain all possible sets of adjustable parameters P = {P1, P2, ……, P n}, and use the signal simulation formula to calculate the fault signals of all sets of adjustable parameters P to obtain the corresponding composite fault signal dataset; Establish an envelope signal extraction model: Normalize the composite fault signal dataset to obtain a standardized fault signal dataset, and use the B-LSTM model to train the standardized fault signal dataset to establish the correlation between the input signal and the envelope spectrum line, thereby generating an envelope signal extraction model for the detection target.
[0031] Specifically, by replacing the traditional Hilbert transform or cubic spline interpolation method with a data-driven deep learning framework, it solves the defects such as overshoot phenomenon and difficult processing of non-differentiable signals existing in the prior art; Generating local faults, variable-intensity faults, and composite fault datasets in stages can comprehensively cover the fault types and parameter combinations that may occur during the actual operation of the detection target, improving the diversity of model training data; Through the bidirectional temporal feature capture ability of the B-LSTM model, the complex mapping relationship between the signal and the envelope spectrum line can be effectively established, providing a high-precision analysis basis for subsequent fault diagnosis.
[0032] In the step of establishing the envelope signal extraction model, first divide the composite fault signal dataset into three datasets: a training set, a validation set, and a test set according to a set ratio. Then, perform normalization processing on the training set and record the normalization parameters. Next, synchronously transform the validation set and the test set using the normalization parameters. Train the constructed training set using the B-LSTM model, use the cross-entropy loss function as the loss function, select the Adam function as the optimization function, and iteratively train until the R² fitting rate of the test set reaches the target fitting value. Save the model with the best result in the validation set as the envelope signal extraction model for the detection target. Preferably, the set ratio for dividing the training set, the validation set, and the test set is 7:2:1. It can be understood that other set ratios can also be set according to requirements.
[0033] Specifically, adopt a scientific data division and parameter synchronous transformation strategy to ensure data consistency in the model training, validation, and testing links. By combining the use of the cross-entropy loss function and the Adam optimizer, significantly improve the model convergence speed and parameter optimization efficiency.
[0034] In the step of establishing the envelope signal extraction model, when using the B-LSTM model to train the standardized fault signal dataset, the B-LSTM model performs forward LSTM sequential processing and backward LSTM reverse processing on the standardized fault signal simultaneously until the forward LSTM sequential processing and the backward LSTM reverse processing are combined to capture the bidirectional dependence relationship of the time series corresponding to the envelope signal. Map the LSTM output at each time step to the envelope signal dimension to establish the correlation relationship between the input signal and the envelope signal, and finally generate the envelope signal extraction model for the detection target.
[0035] Specifically, the bidirectional LSTM structure can simultaneously capture the front and back time series features of the fault signal, effectively solve the problem that the traditional unidirectional model does not make full use of historical information, and significantly improve the ability to analyze the time series features of the fault signal.
[0036] The specific working process of the B-LSTM model includes: Based on the first bidirectional LSTM layer, perform feature extraction on the input standardized fault signal from both the forward and backward directions simultaneously, splice the bidirectional outputs, and obtain the first processed data; Based on the dropout layer, randomly discard some neurons from the first processed data to obtain the second processed data. Preferably, randomly discard 20% of the neurons. It can be understood that the number of discarded neurons can be set arbitrarily according to requirements; Based on the second bidirectional LSTM layer, perform high-dimensional feature extraction on the second processed data from both the forward and backward directions simultaneously, splice the bidirectional outputs, and obtain the third processed data; Based on the fully connected layer, map the high-dimensional features of the third processed data to the envelope signal dimension.
[0037] Specifically, a multi-level abstraction of features is achieved through a double-layer bidirectional LSTM structure. Combined with the regularization effect of the dropout layer, it can not only extract deep temporal features but also prevent the network from overfitting, ensuring the generalization performance of the model.
[0038] Example 2 An application method of an envelope signal extraction model, using the envelope signal extraction model established by the method for establishing an envelope signal extraction model provided in Example 1, includes the following steps: Collect detection signals during the operation of the detection target through a sensor. The detection signals can be original data such as vibration signals and acceleration signals during the operation of the target device, which can reflect the real-time operation state of the detection target. Perform singular value rejection processing and filter noise reduction processing on the detection signal data, input the detection signal data into the envelope signal extraction model, and according to the envelope signal extraction model, output the corresponding envelope signal. Perform filtering processing on the envelope signal, and use a Butterworth filter for filtering processing. The advantage of the Butterworth filter is that while suppressing high-frequency noise, it can well maintain the amplitude of high-frequency effective signals. Here, a low-pass filter mode is selected and set as a high-order filter. The high-order filter has excellent characteristics. Its passband flatness can be maintained in a wider frequency range, and its attenuation speed in the stopband is faster, which enables it to more effectively suppress high-frequency components and thus obtain a purer signal.
[0039] Obtain the processed envelope signal, perform Fourier transform on the processed envelope signal to obtain an envelope spectrum line, and determine whether the characteristic frequency of the envelope spectrum line meets the requirements. If so, intercept the characteristic frequencies that meet the requirements in the envelope spectrum line, retain the information data corresponding to the analysis requirements, and discard redundant data; if not, return to perform filtering processing on the envelope signal and re-adjust the filtering parameters of the envelope signal.
[0040] Due to the adoption of the above technical solution, the real detection signals are processed by the envelope signal extraction model to achieve efficient extraction from the original vibration signal to the characteristic frequency. Combined with the dynamic filter parameter adjustment mechanism, it solves the problem that the traditional envelope spectrum method relies on manual experience in intercepting characteristic frequencies. Combined with the feedback optimization after Fourier transform, it realizes the adaptive extraction of the envelope spectrum line, significantly reduces the misjudgment rate in scenarios such as communication system signal recognition and structural health monitoring, and meets the monitoring requirements of high-reliability equipment; through the preprocessing link, abnormal interference and background noise in the signal acquisition process are effectively eliminated, significantly improving the input signal quality and providing clean data input for subsequent model processing.
[0041] Example 3 Example 3 is an application of Example 1 in the bearing field. The detection target is a bearing, and the components of the bearing include an inner ring, an outer ring, rolling elements, and a cage.
[0042] Specifically, clarify the division criteria for the key components of the bearing to ensure the pertinence of the generation of local fault signals, enable the dataset to accurately reflect the characteristic differences of faults in different components, and provide data support for subsequent model identification of specific component faults. The envelope spectrum generated by this method can clearly distinguish the modulation frequency components of different components, improving the accuracy of structural health monitoring.
[0043] The adjustable parameters include amplitude, resonance frequency, characteristic frequency, rotational frequency, random fluctuation, attenuation coefficient, and signal-to-noise ratio.
[0044] Specifically, through multi-dimensional parameter regulation, the quantification simulation of fault intensity is realized. In particular, the introduction of the signal-to-noise ratio parameter can simulate the noise interference in actual working conditions and enhance the robustness of the model in complex environments.
[0045] Select the simulation model of the rolling element bearing SKF6205. The pitch diameter of this type of bearing is 39.04 mm, the rolling element diameter is 7.94 mm, the number of rolling elements reaches 9, and the contact angle is set to 0°. When the variable motor speed is set at 1440 r / min, according to the conversion relationship between speed and frequency, the rotational frequency can be calculated as 24 Hz. On this basis, based on the envelope signal extraction model for bearings, the remaining adjustable parameters are reasonably set.
[0046] Calculate the outer race fault characteristic frequency according to the following formula (1) : (1) In formula (1), represents the number of rolling elements; represents the rotational frequency; represents the rolling element diameter; represents the pitch diameter of the bearing; represents the contact angle.
[0047] Calculate the inner race fault characteristic frequency according to the following formula (2) : (2) In formula (2), represents the number of rolling elements; represents the rotational frequency; represents the rolling element diameter; represents the pitch diameter of the bearing; represents the contact angle.
[0048] In the step of obtaining the dataset of fault signals with different intensities, the specific process of feature intensity training includes: Select an adjustable parameter, set its value range as [X1, X2], select the data when the value is X1 as the original adjustment data N0, and obtain the new adjustment data N1 = N0 + Y according to the set step size Y, so that the fault feature intensity presented by the corresponding analog signal changes. Obtain the analog signal corresponding to the adjustment data N1 as the variable-intensity fault signal. Again, obtain the new adjustment data N2 = N1 + Y according to the set step size Y, and obtain the analog signal corresponding to the adjustment data N2 as the variable-intensity fault signal. Repeat the above steps until all variable-intensity fault signals within the value range [X1, X2] of the adjustable parameter are obtained, and establish the basic dataset of fault signals with different intensities for this adjustable parameter; then add random perturbations to each adjustment data within the range of the step size Y , to obtain the dataset of fault signals with different intensities.
[0049] Specifically, by combining the parameter step-by-step adjustment method with random perturbation, it not only ensures the systematicness of parameter changes but also increases the randomness of data samples, so as to more comprehensively obtain the comprehensive influence of different characteristic frequencies and their random fluctuations on the overall characteristics of the signal, and improve the generalization recognition ability of fault characteristics with different intensities.
[0050] For example, taking the inner ring fault characteristic frequency as the adjustable parameter, in the step of obtaining the dataset of fault signals with different intensities, set the value range of the inner ring fault characteristic frequency as [100, 150], and set the step size as 1. In this way, variable-intensity fault signals with frequencies increasing in fixed values can be generated. Add random perturbations according to the following formula (3) . For each adjustment data, there is an increase amplitude of (-1, +1), which can make the adjustment data no longer follow the increasing rule or only have the integer part, such as (100.68, 101.23, 102.77, 102.97, 104.52...).
[0051] (3) In formula (3), represents generating a random number uniformly distributed in the interval [0, 1).
[0052] In the process of obtaining the composite fault signal dataset, the signal simulation formula is: (4) In formula (4), represents an amplitude modulator with a period of ; represents the amplitude of the modulation signal; represents the resonance damping coefficient; represents exciting the natural frequency of a certain component; represents the periodic decaying oscillation caused by a fault; represents random fluctuations; represents an arbitrary constant; represents the reciprocal of the characteristic frequency of the adjustable parameter of the input; represents the base of the natural logarithm; represents noise; represents time; represents the rotational frequency.
[0053] Specifically, through the signal simulation formula, the signal characteristics under the coupling action of multiple parameters in the bearing compound fault can be accurately simulated, providing a mathematical basis for the model to learn complex fault patterns.
[0054] For example, when occurs, the signal simulated by the current signal simulation formula is the inner race fault signal; occurs, the signal simulated by the current signal simulation formula is the outer race fault signal.
[0055] Taking the real rolling element bearing SKF6205 as the detection target, the vibration signal during the operation of the rolling element bearing SKF6205 is obtained through a sensor, and this vibration signal is used as the detection signal. As Figure 2 shown, it is the characteristic frequency spectrum of the direct Hilbert envelope transform of the detection signal, where the horizontal axis represents frequency and the vertical axis represents amplitude.
[0056] Then, the detection signal is input into the envelope signal extraction model, and the corresponding envelope signal is output. The envelope signal is filtered to obtain the processed envelope signal, and the Fourier transform of the processed envelope signal is performed to obtain Figure 3 , that is, the characteristic frequency spectrum of the detection signal after being processed by the envelope signal extraction model. By comparing Figure 2 and Figure 3 It can be clearly found that the trained envelope signal extraction model has successfully and accurately analyzed the frequency components of the detection signal, demonstrating the effectiveness of the model in processing signal frequency characteristics.
[0057] To further verify the performance of the envelope signal extraction model, a comparative experiment is carried out on the signal with added Gaussian white noise.
[0058] Gaussian white noise is added to the detection signal to obtain a noise composite signal. As Figure 4 shown, it is the characteristic frequency spectrum of the direct Hilbert envelope transform of the noise composite signal. It can be clearly seen from Figure 4 that no effective characteristic frequency can be extracted from it, and the signal characteristics are seriously interfered and masked by the noise.
[0059] Then, input the noise composite signal into the envelope signal extraction model, output the corresponding envelope signal, perform filtering processing on the envelope signal to obtain the processed envelope signal, and perform Fourier transform on the processed envelope signal to obtain Figure 5 , that is, the characteristic frequency spectrum of the noise composite signal after being processed by the envelope signal extraction model. It can be intuitively seen that the envelope signal extraction model successfully extracts the characteristic frequency effectively. Even under noise interference, it can still accurately identify and separate the key frequency characteristics of the signal, providing a solid and reliable data basis for subsequent signal analysis and fault diagnosis.
[0060] Example 4 Example 4 is an application of Example 1 in the field of automobile manufacturing and maintenance. The detection target is an automobile engine, and the engine components include a crankcase, an oil pump, a cylinder head, etc. The adjustable parameters include amplitude, resonance frequency, characteristic frequency, rotational frequency, random fluctuation, attenuation coefficient, and signal-to-noise ratio.
[0061] Install sensors on different detection targets. After preprocessing the collected signals, according to the envelope signal extraction model established in Example 1, input the preprocessed engine signals into the trained envelope signal extraction model. The envelope signal extraction model outputs the envelope signal, and further analyzes the characteristics of the envelope signal to achieve fast and accurate detection of engine faults.
[0062] Example 5 Example 5 is an application of Example 1 in the field of power systems. The detection target is the power system, and the detection target is a generator. The generator components include key components such as a rotor, a stator, and a bearing. The adjustable parameters include vibration signals, current signals, and voltage signals.
[0063] Install sensors on different detection targets, preprocess the collected signals, and according to the method for establishing the envelope signal extraction model provided in Example 1, use a power system signal dataset containing different fault types for training to obtain the corresponding envelope signal extraction model. Input the preprocessed power system signals into the trained envelope signal extraction model. The envelope signal extraction model outputs the envelope signal, and further analyzes the characteristics of the envelope signal to achieve fast and accurate detection of power equipment faults.
[0064] Example 6 An envelope signal extraction system includes a signal acquisition module, a data processing module, and a human-computer interaction module. The signal acquisition module is used to obtain detection signals during the operation of a detection target, and the detection signals are raw data that can reflect the real-time operation state of the detection target. The data processing module stores an envelope signal extraction model established by the method for establishing the envelope signal extraction model provided in Embodiment 1. The data processing module is used to call the envelope signal extraction model to process the detection signals input by the signal acquisition module to obtain envelope signals or envelope spectra. The human-computer interaction module is used to receive and display the envelope signals or envelope spectra output by the data processing module. The human-computer interaction module can be an intelligent terminal with a display screen, and users can send control instructions to the signal acquisition module and the data processing module through the intelligent terminal.
[0065] As Figure 6 shown, taking the application example in the bearing field provided in Embodiment 3 as an example, sensors matching the bearing are arranged on the bearing.
[0066] When the bearing is running, the control terminal controls the signal acquisition module to start, and the signal acquisition module obtains the detection signals during the operation of the detection target through the sensors.
[0067] The detection signal data can be processed offline. First, deploy the envelope signal extraction model provided in Embodiment 1 to an embedded device. The embedded device can select hardware platforms such as single-chip microcontrollers and FPGAs as the device infrastructure according to specific requirements. The detection signal data is input into the embedded device for data processing. After the feature processing and analysis of the envelope signal extraction model, the final processed result is output to the display for visual presentation.
[0068] The detection signal data can also be processed online. First, deploy the envelope signal extraction model provided in Embodiment 1 to an online analysis module. The detection signal data is input into the embedded device through WIFI networking. The embedded device is connected to the Web side through WIFI, and the detection signal data is stored in the server database through the Web side via the TCP / IP network transmission protocol. Users can, through the control terminal, retrieve the detection signal data to be processed from the server database into the online analysis module. After the feature processing and analysis of the envelope signal extraction model, the final processed result is output to the display for visual presentation.
[0069] Embodiment 7 An electronic device includes a processor, and the processor is coupled to a memory. The memory is used to store a program. When the program is executed by the processor, the system is enabled to implement the steps in the method for establishing the envelope signal extraction model provided in Embodiment 1.
[0070] Embodiment 8 A computer-readable storage medium has a computer program stored thereon, and when the computer program is executed by a processor, it implements the steps in the method for establishing the envelope signal extraction model provided in Embodiment 1.
[0071] Specific embodiments are applied in this article to elaborate on the principles and implementation manners of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be noted that for those of ordinary skill in the art, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.
[0072] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the present invention is usually placed during use. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention.
[0073] In the description of the present invention, it should also be noted that unless otherwise clearly specified and limited, the terms "set", "installed", "connected", "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.
Claims
1. A method for establishing an envelope signal extraction model, characterized in that It includes the following steps: Obtain a local fault signal dataset: Select a detection target, where the detection target includes multiple components. Generate a simulated signal with a single resonance frequency for a specified component with damage as the local fault signal corresponding to this component; Replace another component with damage to be measured and repeat this step until the local fault signals of all components of the detection target are obtained, and establish a local fault signal dataset for the detection target; Obtain fault signal datasets of different intensities: Establish a mathematical analysis model based on the local fault signal dataset of the detection target. This mathematical analysis model includes multiple adjustable parameters, denoted as P1, P2, ……, P n ; Conduct feature intensity training on all adjustable parameters respectively to establish fault signal datasets of different intensities for all adjustable parameters; Obtain the compound fault signal dataset: First, perform a full permutation and combination on the fault signal datasets with different intensities for each adjustable parameter to obtain all possible sets of adjustable parameters \(P = \{P_1, P_2,\cdots, P\}\), n and use the signal simulation formula to calculate the fault signals for all sets of adjustable parameters \(P\) , and obtain the corresponding compound fault signal dataset; Establish an envelope signal extraction model: Normalize the composite fault signal dataset to obtain a standardized fault signal dataset. Use the B-LSTM model to train the standardized fault signal dataset to establish the correlation between the input signal and the envelope spectrum line, thereby generating an envelope signal extraction model for the detection target.
2. The method for establishing an envelope signal extraction model according to claim 1, wherein, In the step of obtaining fault signal datasets with different intensities, the specific process of feature intensity training includes: Select an adjustable parameter, set its value range as [X1, X2], select the data when the value is X1 as the original adjustment data N0, and obtain a new adjustment data N1 = N0 + Y according to the set step size Y, so that the fault feature intensity presented by the corresponding simulated signal changes. Obtain the simulated signal corresponding to the adjustment data N1 as a variable-intensity fault signal, and again obtain a new adjustment data N2 = N1 + Y according to the set step size Y, and obtain the simulated signal corresponding to the adjustment data N2 as a variable-intensity fault signal. Repeat the above steps until all variable-intensity fault signals within the value range [X1, X2] of this adjustable parameter are obtained, and establish a basic dataset of fault signals with different intensities for this adjustable parameter; Then add random perturbations to each adjustment data within the range of the step size Y to obtain a fault signal dataset with different intensities.
3. The method for establishing an envelope signal extraction model according to claim 1, characterized in that, In the step of establishing the envelope signal extraction model, first divide the composite fault signal dataset into three datasets: a training set, a validation set, and a test set according to a set ratio, then normalize the training set and record the normalization parameters, and then synchronously transform the validation set and the test set using the normalization parameters; Train using the B-LSTM model with the constructed training set, use the cross-entropy loss function as the loss function, select the Adam function as the optimization function, and iteratively train until the R² fitting rate of the test set reaches the target fitting value. Save the model with the best result in the validation set as the envelope signal extraction model for the detection target.
4. The method for establishing an envelope signal extraction model according to claim 1, characterized in that In the step of establishing the envelope signal extraction model, when using the B-LSTM model to train the standardized fault signal dataset, perform forward LSTM sequential processing and backward LSTM reverse processing on the standardized fault signal simultaneously through the B-LSTM model until the forward LSTM sequential processing and the backward LSTM reverse processing are merged to capture the bidirectional dependence relationship of the time series corresponding to this envelope signal. Map the output of the LSTM at each time step to the envelope signal dimension to establish the correlation between the input signal and the envelope signal, and finally generate an envelope signal extraction model for the detection target.
5. The method for establishing an envelope signal extraction model according to claim 4, characterized in that The specific process of the B-LSTM model working includes: Based on the first bidirectional LSTM layer, perform feature extraction on the input standardized fault signal from both the forward and backward directions, splice the bidirectional outputs, and obtain the first processed data; Based on the dropout layer, randomly discard some neurons from the first processed data to obtain the second processed data; Based on the second bidirectional LSTM layer, perform high-dimensional feature extraction on the second processed data from both the forward and reverse directions simultaneously, concatenate the bidirectional outputs to obtain the third processed data; Based on the fully connected layer, map the high-dimensional features of the third processed data to the envelope signal dimension.
6. A method for applying an envelope signal extraction model, using the envelope signal extraction model established by the method for establishing an envelope signal extraction model according to any one of claims 1-5, characterized in that, It includes the following steps: collect the detection signal during the operation of the detection target through a sensor to obtain detection signal data, input the detection signal data into the envelope signal extraction model, obtain the corresponding envelope signal according to the output of the envelope signal extraction model, perform filtering processing on the envelope signal to obtain the processed envelope signal, perform Fourier transform on the processed envelope signal to obtain the envelope spectrum line, determine whether the characteristic frequency of the envelope spectrum line meets the requirements. If so, intercept the characteristic frequency that meets the requirements in the envelope spectrum line, retain the information data corresponding to the analysis requirements, and discard the redundant data; if not, return to perform filtering processing on the envelope signal and re-adjust the filtering parameters of the envelope signal.
7. The application method of the envelope signal extraction model according to claim 6, characterized in that, After performing singular value rejection processing and filtering and noise reduction processing on the detection signal data, input it into the envelope signal extraction model.
8. An envelope signal extraction system, characterized in that It includes a signal acquisition module, a data processing module, and a human-computer interaction module. The signal acquisition module is used to obtain the detection signal during the operation of the detection target, and the detection signal is the original data that can reflect the real-time operation state of the detection target; the data processing module stores the envelope signal extraction model established by the method for establishing the envelope signal extraction model according to any one of claims 1-5, and the data processing module is used to call the envelope signal extraction model to process the detection signal input by the signal acquisition module to obtain the envelope signal or the envelope spectrum line; the human-computer interaction module is used to receive and display the envelope signal or the envelope spectrum line output by the data processing module.
9. An electronic device, comprising a processor, the processor being coupled to a memory for storing a program, characterized in that When the program is executed by the processor, the system is enabled to implement the steps in the method for establishing the envelope signal extraction model according to any one of claims 1 to 5.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps in the method for establishing the envelope signal extraction model according to any one of claims 1-5.