Flywheel fault detection method, device and equipment and medium

Abstract

The invention discloses a flywheel fault detection method and device, equipment and a medium, and relates to the field of flywheels, a convolutional neural network model capable of performing fault risk assessment according to state parameters is trained and generated in advance, and during application, the obtained flywheel state parameters are directly input into the convolutional neural network model, so that the fault risk assessment efficiency is improved. And the convolutional neural network model outputs an evaluation result of the fault risk according to the state parameters, and finally executes a corresponding protection strategy according to the evaluation result to prevent expansion of the fault. An original signal input into the convolutional neural network model comprises state parameters of multiple aspects of the flywheel, fault detection of the flywheel can be realized from multiple aspects such as dynamic characteristics, and the accuracy and reliability of a fault diagnosis result are improved. The fault state of the flywheel energy storage system is effectively and accurately detected by using the convolutional neural network model, the diagnosis result is fed back to the control system and timely adjusted, and the reliability and safety of the operation of the whole system are improved.

Description

Technical Field

[0001] The present invention relates to the field of flywheels, and in particular to a method, device, equipment and medium for detecting flywheel faults. Background Art

[0002] Flywheel energy storage has the advantages of high energy efficiency conversion, high power density, long life, fast response, low maintenance cost and environmental friendliness, and has been widely used in the fields of aerospace, power systems, rail transit and UPS (Uninterruptible Power Supply). However, flywheel energy storage depends on the rotation of the flywheel. The normal operating speed of the flywheel is between 10,000 revolutions and tens of thousands of revolutions. As a high-speed rotating body, the flywheel will cause great damage when it disintegrates or becomes unstable. Therefore, there are unsafe risks when the flywheel is applied. How to effectively realize the fault diagnosis of the flywheel to timely avoid safety risks is a technical problem that needs to be solved urgently at present. Summary of the Invention

[0003] The purpose of the present invention is to provide a method, device, equipment and medium for detecting flywheel faults, which can effectively realize the detection of flywheel faults and improve the safety and reliability of the flywheel.

[0004] To solve the above technical problems, the present invention provides a method for detecting flywheel faults, including:

[0005] Obtaining the state parameters of the flywheel; the state parameters include at least two of the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters and dynamic performance parameters;

[0006] Inputting the state parameters into a pre-constructed convolutional neural network model to evaluate the fault risk of the flywheel by using the pre-constructed convolutional neural network model; the fault risk of the flywheel includes the probability of the flywheel having an over-temperature fault in the shafting, the probability of the flywheel having a static imbalance fault, the probability of the flywheel having a dynamic imbalance fault, the probability of the flywheel having a physical scratch fault, the probability of the flywheel having a shafting immersion fault, the probability of the flywheel having a vacuum degree fault, the probability of the flywheel having a first-order vibration fault, the probability of the flywheel having a second-order vibration fault, and the probability of the flywheel having a third-order vibration fault, at least two of them;

[0007] Executing corresponding protection strategies based on the evaluation results of the convolutional neural network model.

[0008] Optionally, before inputting the state parameters into the pre-constructed convolutional neural network model, it further includes:

[0009] Preprocessing the state parameters;

[0010] Inputting the state parameters into a pre-constructed convolutional neural network model includes:

[0011] Taking the preprocessed state parameters as the original signal and inputting them into a pre-constructed convolutional neural network model.

[0012] Optionally, when the state parameters include vibration parameters, the original signal corresponding to the vibration parameters in the state parameters includes a first original sub-signal and a second original sub-signal. Preprocessing the state parameters includes:

[0013] Performing an s-transform on the vibration parameters in the state parameters and determining the s-transform spectrogram obtained by the s-transform as the first original sub-signal of the vibration parameters;

[0014] Performing an fft-transform on the vibration parameters in the state parameters and determining the two-dimensional grayscale image obtained by the fft-transform as the second original sub-signal of the vibration parameters.

[0015] Optionally, obtaining the state parameters of the flywheel includes:

[0016] Performing sliding window sampling on the state parameters of the flywheel based on a preset sliding window length and a preset sampling channel;

[0017] Inputting the preprocessed state parameters into a pre-constructed convolutional neural network model includes:

[0018] Dividing the preprocessed state parameters into each preset batch; where each of the preset batches has a corresponding batch number;

[0019] Constructing a three-dimensional list based on the sampling channel corresponding to the state parameters, the preset batch corresponding to the state parameters, and the original signal corresponding to the state parameters;

[0020] Inputting the three-dimensional list into the pre-constructed convolutional neural network model batch by batch according to the batch numbers of each preset batch.

[0021] Optionally, performing corresponding protection strategies based on the evaluation results of the convolutional neural network model includes:

[0022] If the occurrence probability of a fault in the fault risk is greater than a first preset threshold, then adjust the crossing speed of the flywheel across the resonance region according to the probability of the flywheel having a first-order vibration fault, the probability of the flywheel having a second-order vibration fault, and the probability of the flywheel having a third-order vibration fault;

[0023] If the occurrence probability of a fault in the fault risk is greater than a second preset threshold, then adjust the suspension gap of the flywheel and / or reduce the rotational speed of the flywheel to a preset value;

[0024] If the occurrence probability of a fault in the fault risk is greater than a third preset threshold, then control the flywheel to stop.

[0025] Optionally, the pre-constructed convolutional neural network model includes an input layer, a convolutional layer, an activation layer, a pooling layer, a fully connected layer, and an output layer; the process of the convolutional neural network model evaluating the fault risk of the flywheel specifically includes:

[0026] Receive the state parameters of the flywheel through the input layer;

[0027] Perform a convolution operation on the state parameters based on the convolution kernels in the convolutional layer to extract the response features in the state parameters;

[0028] Output the response features to the activation layer to perform non-linear processing on the response features by using the activation function in the activation layer;

[0029] Perform downsampling on the response characteristics after non-linear processing based on the pooling layer to extract the local features in the response features;

[0030] Integrate all the local features output by the pooling layer into a global feature through the fully connected layer, and map the global feature to each fault category to output the occurrence probability of the fault corresponding to each fault category.

[0031] Optionally, performing a convolution operation on the state parameters based on the convolution kernels in the convolutional layer to extract the response features of the state parameters includes:

[0032] Adopt a one-dimensional convolution method to perform single-channel feature extraction on the time series data corresponding to a single sampling channel of the state parameters based on the sampling time sequence;

[0033] Adopt a spatial-temporal separable convolution method to obtain the fusion features of several sampling channels of the state parameters based on the sampling channels and the sampling time sequence;

[0034] Combine all the single-channel features and the fusion features to obtain the response features of the state parameters.

[0035] To solve the above technical problems, the present invention also provides a fault detection device for a flywheel, including:

[0036] A parameter acquisition unit for acquiring the state parameters of the flywheel; the state parameters include at least two of the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters, and dynamic performance parameters;

[0037] A fault diagnosis unit, configured to input the state parameters into a pre-constructed convolutional neural network model, so as to evaluate the fault risk of the flywheel by using the pre-constructed convolutional neural network model; the fault risk of the flywheel includes at least two of the probability of the flywheel having an over-temperature fault in the shafting, the probability of the flywheel having a static unbalance fault, the probability of the flywheel having a dynamic unbalance fault, the probability of the flywheel having a physical scratch fault, the probability of the flywheel having a shafting immersion fault, the probability of the flywheel having a vacuum degree fault, the probability of the flywheel having a first-order vibration fault, the probability of the flywheel having a second-order vibration fault, and the probability of the flywheel having a third-order vibration fault;

[0038] An execution protection unit, configured to execute corresponding protection strategies based on the evaluation result of the convolutional neural network model.

[0039] To solve the above technical problems, the present invention also provides an electronic device, including:

[0040] A memory, configured to store a computer program;

[0041] A processor, configured to implement the steps of the flywheel fault detection method as described above.

[0042] To solve the above technical problems, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the flywheel fault detection method as described above are implemented.

[0043] The present invention provides a flywheel fault detection method, which pre-trains and generates a convolutional neural network model capable of evaluating fault risks according to state parameters. During application, the obtained state parameters of the flywheel are directly input into the convolutional neural network model, and the convolutional neural network model outputs an evaluation result of the fault risk according to the state parameters. Finally, corresponding protection strategies are executed according to this evaluation result to prevent the expansion of faults. The original signals input into the convolutional neural network model include state parameters in multiple aspects of the flywheel, and can realize the fault detection of the flywheel from multiple aspects such as dynamic characteristics, improve the accuracy and reliability of the fault diagnosis result, effectively realize the accurate detection of the fault state of the flywheel energy storage system by using the convolutional neural network model, feedback the diagnosis result to the control system and make timely adjustments, and improve the reliability and safety of the operation of the entire system.

[0044] The present invention also provides a flywheel fault detection device, an electronic device, and a computer-readable storage medium, which have the same beneficial effects as the above flywheel fault detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the attached drawings required in the prior art and the embodiments. Obviously, the attached drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other attached drawings can be obtained based on these attached drawings.

[0046] Figure 1 Schematic flowchart of a flywheel fault detection method provided by the present invention;

[0047] Figure 2 Schematic flowchart of another flywheel fault detection method provided by the present invention;

[0048] Figure 3 Schematic logic diagram of a flywheel fault detection method provided by the present invention;

[0049] Figure 4 Schematic diagram of a vibration parameter provided by the present invention;

[0050] Figure 5 Schematic diagram of an s-transform spectrum obtained after performing an s-transform on a vibration parameter provided by the present invention;

[0051] Figure 6 Schematic diagram of a frequency-domain signal obtained after performing an fft transform on a vibration parameter provided by the present invention;

[0052] Figure 7 Schematic diagram of a two-dimensional grayscale image obtained after performing an fft transform on a vibration parameter provided by the present invention;

[0053] Figure 8 Schematic diagram of the structure of an electronic device provided by the present invention. Detailed implementation manners

[0054] The core of the present invention is to provide a flywheel fault detection method, device, equipment and medium, which can prevent the expansion of faults, improve the accuracy and reliability of fault diagnosis results, effectively realize the precise detection of the fault state of the flywheel energy storage system by using a convolutional neural network model, feed the diagnosis results back to the control system and make timely adjustments, and improve the reliability and safety of the operation of the entire system.

[0055] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the attached drawings in the embodiments of the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0056] Please refer to Figure 1 , Figure 1 which is a schematic flow chart of a method for detecting faults of a flywheel provided by the present invention; Please refer to Figure 2 , Figure 2 which is a schematic flow chart of another method for detecting faults of a flywheel provided by the present invention; To solve the above technical problems, the present invention provides a method for detecting faults of a flywheel, including:

[0057] S11: Obtain the state parameters of the flywheel; The state parameters include at least two of the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters, and dynamic performance parameters.

[0058] It is not difficult to understand that considering the lack of flywheel fault diagnosis from the perspective of dynamic characteristics at present, when the present application conducts flywheel fault diagnosis, at least two or more state parameters including the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters, and dynamic performance parameters of the flywheel will be used as input features. A more preferred embodiment is to use the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters, and dynamic performance parameters as the state parameters of the flywheel to be obtained. Correspondingly, when pre-training a CNN (Convolutional Neural Networks) model, at least two, multiple, or all of the state parameters of the flywheel including the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters, and dynamic performance parameters will be used as the input features of the training set for training, so that the CNN can automatically learn the feature associations between the input features, thereby improving the accuracy and reliability of the final fault diagnosis. The specific type and implementation manner of the flywheel are not particularly limited in the present application. The specific acquisition methods of various state parameters are not particularly limited in the present application, and can be implemented by setting sensors, etc. For example, a current sensor can be used to detect the control current of the flywheel, and a displacement sensor can be used to detect the displacement parameters. Specifically, temperature sensors, pressure sensors, vibration displacement sensors, vibration velocity sensors, vibration acceleration sensors, Hall sensors, eddy current sensors, etc. can be used to collect some state parameters.

[0059] It should be noted that the state parameters of the flywheel include some structural characteristic parameters, dynamic characteristic parameters, and some related control parameters, not limited to these state parameters provided in this application. The inherent parameters of the flywheel refer to the basic parameters of the flywheel itself, specifically including the mass of the flywheel, the length of the shafting, the radius of the flywheel, the axial thickness of the flywheel, the moment of inertia, the theoretical natural frequency, the designed suspension gap, etc. These data are inherent parameters and serve as globally shared information in the entire fault diagnosis process. The control current of the flywheel includes the control current of the magnetic suspension bearing, etc. There are 12 control currents of the magnetic suspension bearing in the flywheel. The displacement parameters include the axial end displacement and the radial displacement of the flywheel. The vibration parameters include the axial vibration velocity and / or acceleration and the radial vibration velocity and / or acceleration of the flywheel. The speed parameters include the rotational speed and / or acceleration when the flywheel rotates. The environmental parameters include parameters such as temperature, vacuum degree, and pressure. The environmental parameters can specifically include the cavity temperature, vacuum degree, cavity pressure, and environmental temperature. To better reflect the local state of the flywheel, variables of the same type need to be collected at different positions of the flywheel. For example, for the designed suspension gap of the flywheel, two different positions can be set to detect 2 designed suspension gaps. For the axial end displacement, 2 different acquisition positions are set. For the radial displacement, 3 different acquisition positions are set. For the cavity temperature, 3 different acquisition positions are set. For the axial vibration velocity / acceleration, 2 different acquisition positions are set. For the radial vibration velocity / acceleration, 3 different acquisition positions are set. Further, the state parameters not only include some parameters that can be directly collected, but also include parameters such as the dynamic performance parameters of the flywheel obtained after calculation based on the collected parameters. The dynamic performance parameters include the kinetic energy of the flywheel obtained by calculation and the torque of the flywheel obtained by calculation, etc.

[0060] S12: Input the state parameters into a pre-constructed convolutional neural network model to evaluate the fault risk of the flywheel using the pre-constructed convolutional neural network model; the fault risk of the flywheel includes at least two of the probability of the flywheel having an over-temperature fault of the shafting, the probability of the flywheel having a static imbalance fault, the probability of the flywheel having a dynamic imbalance fault, the probability of the flywheel having a physical scraping fault, the probability of the flywheel having a shafting immersion fault, the probability of the flywheel having a vacuum degree fault, the probability of the flywheel having a first-order vibration fault, the probability of the flywheel having a second-order vibration fault, and the probability of the flywheel having a third-order vibration fault.

[0061] It can be understood that considering that there are very few current diagnostic methods for flywheel failures and most of the diagnostic results are inaccurate, the present application pre-trains and generates a CNN model capable of evaluating the flywheel failure risk based on the input state parameters. The CNN model can output the types and probabilities of different failures according to the input state parameters. Flywheel system failures include mechanical failures, electrical failures, and communication failures. Among them, the body failure of the energy storage flywheel is the most important part of mechanical failures. In the present application, the CNN model is mainly used to predict the occurrence probability of flywheel body failures. One embodiment is that the CNN model outputs at least two probability results among the probabilities of the flywheel having an over-temperature failure of the shafting, the flywheel having a static imbalance failure, the flywheel having a dynamic imbalance failure, the flywheel having a physical scratch failure, the flywheel having an immersion failure of the shafting, the flywheel having a vacuum degree failure, the flywheel having a first-order vibration failure, the flywheel having a second-order vibration failure, and the flywheel having a third-order vibration failure; a preferred embodiment is that the CNN model outputs all nine probability results among the probabilities of the flywheel having an over-temperature failure of the shafting, the flywheel having a static imbalance failure, the flywheel having a dynamic imbalance failure, the flywheel having a physical scratch failure, the flywheel having an immersion failure of the shafting, the flywheel having a vacuum degree failure, the flywheel having a first-order vibration failure, the flywheel having a second-order vibration failure, and the flywheel having a third-order vibration failure.

[0062] It should be noted that the basic principle of flywheel energy storage is to convert electrical energy into the kinetic energy of the flywheel for storage and then convert the kinetic energy back into electrical energy for release when needed. There are three working modes: charging mode, discharging mode, and standby mode. In related technologies, there has been relatively much research on the bearing fault diagnosis of high-speed rotating machinery, but less research on the faults of the bearings of energy storage flywheels. Through the analysis of the dynamic characteristics of the flywheel, the energy storage flywheel belongs to high-speed rotating machinery. In addition to having some commonalities of general rotating machinery, it also has its own characteristics. The similarities are that the energy storage flywheel has fault phenomena such as misalignment and imbalance that general bearings have. The difference is that the bearings of the suspended energy storage flywheel are mostly non-contact structures with relatively high rotation speeds, and their precession effect, gyroscopic effect, vibration characteristics, stress characteristics, etc. are more significantly affected by factors such as flywheel material, shape, temperature, mass, fluid field, and external forces. The most direct external manifestation caused by these comprehensive factors is the vibration and instability of the flywheel. Based on this, the present application classifies the faults of the flywheel body into 9 types, namely over-temperature fault of the shafting, static unbalance fault, dynamic unbalance fault, physical scratching fault, shafting immersion fault, vacuum degree fault, first-order vibration fault, second-order vibration fault, and third-order vibration fault. Therefore, the present application defines the output structure of the CNN model as the probabilities of different types of faults occurring in the flywheel, enabling the CNN model to perform fault classification based on the input state parameters and at the same time determine the occurrence probabilities of different fault types of the flywheel. The CNN outputs the types and probabilities of different faults. The fault diagnosis method based on CNN can predict the faults in the early, middle, and late stages of flywheel operation and effectively prevent the expansion of faults according to the protection grading measures. CNN uses feature extraction. On the premise that the pre-training is relatively accurate, sensitivity analysis is carried out using on-site data, and the fault types can be diagnosed relatively accurately.

[0063] S13: Execute corresponding protection strategies based on the evaluation results of the convolutional neural network model.

[0064] It is not difficult to understand that in order to ensure the safety and reliability of flywheel operation, after diagnosing the flywheel fault state using CNN, active defense operations will be carried out according to the fault state when necessary to avoid catastrophic consequences. The corresponding protection strategies can be set and adjusted according to the occurrence probabilities of various fault types in the evaluation results and the actual application requirements of the flywheel. The present application does not make special limitations here.

[0065] Furthermore, the CNN model is mainly used for fault diagnosis of the flywheel body. There may also be other types of mechanical faults, electrical faults, and communication faults in the flywheel. Mechanical faults include: bearing faults (wear / cracking), rotor imbalance, shaft bending or breakage, and mechanical looseness; electrical faults include: motor faults (winding short circuit, insulation damage, rotor bar breakage), sensor faults, power electronics faults, control algorithm failures, and communication faults; environmental faults include: overheating, external shock and vibration. Some of these faults are diagnosed by the controller software itself or other modular fault diagnosis units. The diagnostic systems based on the CNN model and other types of mechanical faults, electrical faults, and communication faults constitute the fault diagnosis network of the flywheel control system. When implementing the protection strategy, combined with the probabilities of the diagnostic results of other fault diagnosis systems, further hierarchical protection measures are taken to ensure the stability of the magnetic levitation flywheel system to the greatest extent. The specific implementation methods of other types of fault diagnosis systems are not particularly limited in this application.

[0066] It should be noted that the typical faults of the energy storage flywheel include faults in aspects such as the temperature of the shafting, physical scratching of the shafting, immersion of the shafting in liquid, rotor vibration mode and amplitude, etc. For the temperature of the shafting, the change in the rotor temperature can be monitored by the amplitude of the frequency points at specific frequencies on the flywheel and the telescopic amount of the rotating shaft, so as to realize the monitoring of the shafting temperature. For the physical scratching of the shafting, after the shafting is scratched, it will cause a change in the damping of the shafting, which will excite the amplitude of the anti-whirling frequency points, while the amplitude of the direct-whirling frequency points will weaken or even disappear; at the same time, it will excite the amplitudes of 0.5 times the same frequency and 2 times the frequency. Therefore, the judgment of this fault can be carried out by analyzing the frequency data of the radial vibration displacement sensor. The direct-whirling frequency point refers to the frequency point where the vibration trajectory (whirling) of the rotor is the same as the rotation direction when the rotor rotates. That is, the precession direction of the vibration vector is the same as the rotation direction of the rotor. The anti-whirling frequency point refers to the frequency point where the precession direction of the rotor vibration trajectory is opposite to the rotation direction. That is, the rotation direction of the vibration vector is opposite to the rotation direction of the rotor. For the immersion of the shafting in liquid, the through-flow cooling of the flywheel may cause the liquefied accumulation of the cooling medium, resulting in partial immersion of the flywheel shafting in the liquid. When the flywheel runs, it will cause a significant increase in the amplitude in the frequency band of 0.3 to 0.7 times the natural frequency. For the rotor vibration mode and amplitude, because the rotor has first-order, second-order, third-order and even higher vibration frequencies, in addition to measuring the telescopic amount of the rotating shaft, vibration displacement sensors and acceleration sensors need to be installed at other different positions of the rotating shaft to detect the bending degree of the shafting, so as to obtain the high-order modal frequencies. Before the rotor speed reaches the first-order bending mode frequency, the rotor vibration mode already has a bending mode (at this time, the vibration intensity is not enough to cause the bearing control to become unstable). Due to the clearance of the protective bearing, it causes a large range of revolution of the shafting, which may cause irreversible bending damage to the shafting in the bending mode (operation instability drop occurs under the operating conditions near the bending mode frequency). Therefore, it becomes very important to detect the first-order bending mode frequency to determine whether the flywheel is faulty. Based on this, in the training and actual fault diagnosis of this application, state parameters including the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters and dynamic performance parameters are used for fault detection. At the same time, the first-order bending mode frequency, second-order bending mode frequency and third-order bending mode frequency of the current flywheel are obtained by pre-training through the setting of the training set for subsequent fault diagnosis.

[0067] As a specific embodiment, the fault diagnosis system of the flywheel is as Figure 2As shown in the figure, it includes a flywheel state monitoring and fault diagnosis system, a CNN model, and a flywheel control and protection system. The flywheel state monitoring and fault diagnosis system senses the state of the flywheel, combines the real-time working conditions of the flywheel control system, and the CNN outputs the fault type and the probability of occurrence. The flywheel control and protection system makes necessary protection instructions according to the diagnosis results of the CNN and the flywheel feedback to avoid the further expansion of the fault. The process of inputting state parameters into the CNN model specifically includes processes such as sampling, data preprocessing, and data calculation and reshaping, and completes fault diagnosis through processes such as data preprocessing in cooperation with the CNN model.

[0068] Starting from being close to engineering practice, the present invention proposes a fault diagnosis method for an active magnetic suspension flywheel, realizes the accurate detection of the fault state of the flywheel energy storage system, the diagnosis result can be fed back to the control system of the flywheel and makes timely adjustments, and improves the reliability of the flywheel system operation. It is applicable to the fault diagnosis of high-speed magnetic suspension flywheels and also applicable to the fault detection of other rotating machinery, with a wide range of applications. When diagnosing, collect as much information such as state parameters as possible, select different types and different numbers of sensors according to the shape of the flywheel body, collect different variables and sample state parameters at different sampling rates, and design a unique sampling hardware configuration. In actual application, appropriate variable collection and CNN layer design can be selected according to different application scenarios, and appropriate fault diagnosis types can be adaptively selected. For different diagnosis results of the CNN, hierarchical protection measures are proposed to prevent the expansion of the fault. For different faults, different protection strategies can be designed, the protection measures are more clear, the overall reliability and safety of the flywheel system are improved, and the maintenance cost is reduced at the same time.

[0069] Based on the above embodiments: Please refer to Figure 3 , Figure 3 which is the flow logic diagram of a flywheel fault detection method provided by the present invention.

[0070] As an optional embodiment, before inputting the state parameters into the pre-constructed convolutional neural network model, it further includes:

[0071] Preprocess the state parameters;

[0072] Inputting the state parameters into the pre-constructed convolutional neural network model includes:

[0073] Taking the preprocessed state parameters as the original signal and inputting them into the pre-constructed convolutional neural network model.

[0074] It can be understood that, for the convenience of processing by the CNN model and considering the design of the input layer of the CNN model, the collected state parameters need to be preprocessed. Specifically, data that needs to be sampled and calculated, except for the basic parameters such as the inherent parameters of the flywheel, is classified and reshaped. The preprocessing mainly includes outlier removal, filtering, debiasing, normalization, etc. For the vibration parameters, that is, vibration signals, the preprocessing includes fft transformation (Fast Fourier Transform) and s transformation (Stockwell transformation). The preprocessed data above is finally used as the available original signal and input into the CNN model. The specific implementation method of the preprocessing is not particularly limited in this application.

[0075] Specifically, by further preprocessing the data corresponding to the sampled state parameters, the accuracy and reliability of the original signal finally input into the CNN model are improved. At the same time, through preprocessing, the state parameters are all converted into a format recognizable by the CNN model, ensuring the effective extraction of the input state parameters by the CNN model and ensuring the accuracy and reliability of the final fault diagnosis result.

[0076] Please refer to Figure 4 , Figure 4 for a schematic diagram of a vibration parameter provided by the present invention; please refer to Figure 5 , Figure 5 for a schematic diagram of an s-transform spectrum obtained after performing an s-transform on a vibration parameter provided by the present invention; please refer to Figure 6 , Figure 6 for a schematic diagram of a frequency-domain signal obtained after performing an fft transform on a vibration parameter provided by the present invention; please refer to Figure 7 , Figure 7 for a schematic diagram of a two-dimensional grayscale image obtained after performing an fft transform on a vibration parameter provided by the present invention; As an optional embodiment, when the state parameters include vibration parameters, the original signal corresponding to the vibration parameters in the state parameters includes a first original sub-signal and a second original sub-signal. Preprocessing the state parameters includes:

[0077] Performing an s-transform on the vibration parameters in the state parameters, and determining the s-transform spectrum obtained by the s-transform as the first original sub-signal of the vibration parameters;

[0078] Performing an fft transform on the vibration parameters in the state parameters, and determining the two-dimensional grayscale image obtained by the fft transform as the second original sub-signal of the vibration parameters.

[0079] It is not difficult to understand that when using sensors to sample state parameters, the sampled data are all time-domain signals. Since the vibration signal itself has multiple frequency components superimposed, in order to improve the accuracy and reliability of the fault diagnosis results, for the vibration signal in the state parameters, the s-transform and fft-transform are used to convert the vibration signal from the time domain to the frequency domain for the CNN model to analyze. By obtaining the s-transform spectrogram and two-dimensional grayscale image corresponding to the vibration signal, the energy distribution of each frequency component is directly displayed, realizing the accurate positioning of the fault characteristic frequency, distinguishing the normal vibration signal from the abnormal vibration signal, and converting the physical essence of the vibration signal into an intuitive frequency spectrum feature through the frequency-energy mapping relationship to achieve more accurate fault diagnosis. For the same vibration signal, the s-transform and fft-transform are simultaneously used to generate two corresponding original sub-signals, further improving the accuracy and reliability of the fault diagnosis.

[0080] As a specific embodiment, when obtaining state parameters, the sliding window sampling method is used to obtain the sampling data of state parameters. The sampling rate and sliding window length of the sliding window sampling are determined separately according to the variable type. The specific configuration method of the data acquisition system is shown in Table 1, where samp and nbu are both configurable specific data volumes, and torque, kinetic energy, and rotational acceleration are calculated based on other sampling data. The update speed of the calculated values is consistent with the data update speed of the highest sampling rate. After the calculation is completed, it is input into the CNN model through a corresponding channel. For other sampling data, the number of channels is the number of sampling channels.

[0081] Table 1 Data Acquisition Table

[0082] It should be noted that for the convenience of the CNN model to recognize, after preprocessing, the data needs to be further classified and reshaped. Specifically, after classifying the collected data (including sampling data and calculated values), each type of data is combined into a multi-channel time array, that is, all current data are integrated into a multi-channel time array. The multi-channel time data includes time series corresponding to each channel one by one. For example, the data sampled from each channel corresponding to the current are all integrated into the form of a time series sequence, and then the time series of 12 channels are combined together to generate a multi-channel time array and converted into a format recognizable by the CNN.

[0083] Furthermore, for vibration parameters, classification and reshaping need to be carried out after preprocessing. Vibration parameters such as axial vibration velocity, axial vibration acceleration, radial vibration velocity, and radial vibration acceleration are respectively subjected to s-transform and fft-transform. After the s-transform, the s-transform spectrogram, which is an image, is directly passed to the input layer of the CNN. To reduce the number of neurons, the s-transform result can also be further converted into a grayscale image; after the fft-transform, it is converted into a 2D grayscale image and passed to the input layer of the CNN. As Figures 4 to 7 shown, after performing the s-transform on the vibration signal shown in Figure 4 , the s-transform spectrogram shown in Figure 5 is obtained. The s-transform spectrogram is a two-dimensional image. Its horizontal axis represents time, the vertical axis represents frequency, and the color (or grayscale) represents the energy intensity (the modulus value of the s-transform result) of the signal at the corresponding time-frequency point. The enlarged part is the result of magnifying some pixel points in the image. After performing the fft-transform on the vibration signal shown in Figure 4 , the frequency-domain signal shown in Figure 6 is obtained, and then it is converted into the 2D grayscale image shown in Figure 7 . The rows of the 2D grayscale image represent frequency, the columns represent time, and the grayscale of each pixel represents the amplitude intensity. The enlarged part is the result of magnifying some pixel points in the image. The converted grayscale image has a depth of 8 bits or 16 bits according to the vibration frequency range. The deeper the image depth, the higher the resolution and the richer the information represented. For the image information obtained after these s-transforms and fft-transforms, it is necessary to combine it into a three-dimensional array in the format of a multi-channel image and input it into the CNN model.

[0084] It should be further noted that all the above-mentioned acquired data need to be converted into a three-dimensional array in the format of multi-channel time series or multi-channel image to facilitate better feature extraction by the CNN. However, it should be noted that the results obtained by different transformation methods have different numerical ranges and physical meanings, and appropriate data normalization processing is required to ensure that the CNN can treat each input feature fairly. Otherwise, some features may have too much or too little impact on model training due to too large or too small numerical ranges (eliminating dimensional differences). To maintain consistency in the time dimension, the data length of the variable with the highest sampling rate is used as the standard, and the data with a lower sampling rate remains unchanged after sampling and fills in the missing data. In addition, the flywheel mass m, the shafting length l, the radius R, the flywheel thickness , the moment of inertia J, and the theoretical natural frequency F req are used as global shared information.

[0085] It is not difficult to understand that when constructing a CNN model, the number of neurons in its input layer is determined by the "area" of the two-dimensional grayscale image (the amount of data determined by the length and width), the "area" of the s-transform image, and the total amount of data in all channels corresponding to the sliding window length of the time series variable. The number of neurons in the input layer is determined according to the situation of the input data. The input neurons of this diagnosis are within 10 4 levels.

[0086] Specifically, by performing different processing on different signals (such as vibration and current, etc.), the original information is retained to the greatest extent, and the vibration signal is reshaped into a "pseudo-image" format and input to the CNN for operations such as feature extraction of the CNN model. The signal processing method is refined to ensure a high accuracy rate for magnetic levitation fault diagnosis through signal processing.

[0087] As an optional embodiment, obtain the state parameters of the flywheel, including:

[0088] Perform sliding window sampling on the state parameters of the flywheel based on a preset sliding window length and a preset sampling channel;

[0089] Use the preprocessed state parameters as the original signal and input them into a pre-constructed convolutional neural network model, including:

[0090] Divide the preprocessed state parameters into each preset batch; among them, there is a batch number corresponding to each preset batch;

[0091] Construct a three-dimensional list based on the sampling channel corresponding to the state parameters, the preset batch corresponding to the state parameters, and the original signal corresponding to the state parameters;

[0092] Input the three-dimensional list into the pre-constructed convolutional neural network model in batches according to the batch numbers of each preset batch.

[0093] It is understandable that due to the large amount of data required to be collected in the entire diagnostic process, the number of neurons in the input layer is excessive, leading to an increased risk of overfitting. To avoid this situation, the rolling window data can be further reshaped into a three-dimensional matrix: [channels, batches, [length, width]]. The data is input in batches by adding elements of batches in the three-dimensional matrix. Taking the vibration data as an example of being input into the CNN in 3 batches, for the grayscale image corresponding to the fft of the vibration signal collected by channel 1, it is reshaped into [1, 1, [255, 255]], representing the data of the first batch of channel 1 within the sliding window, and the data volume is 255×255 (pixel points). Suppose the fft grayscale images of 10 channels of the vibration signal are input into the CNN in 3 batches in sequence, and the data volume of each batch is evenly divided. Then the number of neurons corresponding to the input layer of the fft grayscale image of the vibration signal at this time should be 3×10×255×255. In addition, the s-transform of the vibration signal is a color image, and each pixel point of the color image requires 3 colors (RGB) to express. At this time, the s-transform can be represented by [type A channels, batches, [length, width, type B channels]], where type A channels are sampling channels, and the number of channels depends on the sampling design. Type B channels are color channels, and the total number of channels is 3 (corresponding to R, G, and B respectively). For example, [1, 1, [255, 255, 3]] represents the vibration data of the first batch of channel 1 within the sliding window, and the data volume is 255×255 (pixel points), corresponding to color channel 3 (the channel corresponding to color B). The complete information description of these 255×255 pixel points should be [1, 1, [255, 255, 1]], [1, 1, [255, 255, 2]], [1, 1, [255, 255, 3]]. At this time, the total data volume of the s-transform of the first batch of channel 1 is 1×255×255×3. Suppose the s-transforms of 10 channels of the vibration signal are input into the CNN in 3 batches in sequence, and the data volume of each batch is evenly divided. Then the number of neurons corresponding to the input layer of the s-transform of the vibration signal at this time should be 3×10×255×255×3. However, in practice, the data volume of a certain batch of a certain channel is determined by the total length and total width of the picture, rather than 255×255 (the actual data volume of 255×255 may only be a local area of the entire picture). Calculate the number of neurons corresponding to the vibration signal according to the above method, and then add the number of neurons corresponding to the 27-channel time series (converting multi-channel time series data into "pseudo-images" for processing, and the number of channels is the "height of the pseudo-image") variable, which is the total number of neurons in the input layer. Such processing can reduce the risk of overfitting and reduce the occupancy of hardware memory (the specific number of neurons depends on the actual signal sampling rate and the sliding window length).

[0094] Specifically, by inputting data in batches into the CNN model, the risk of overfitting is avoided, the hardware memory occupancy is reduced, the signal processing method is further refined, the accuracy and reliability of the entire fault diagnosis process are improved, and a design basis is provided for the design of the input layer in the CNN model.

[0095] As an alternative embodiment, corresponding protection strategies are executed based on the evaluation results of the convolutional neural network model, including:

[0096] If the occurrence probability of a fault in the fault risk is greater than the first preset threshold, the crossing speed of the flywheel across the resonance region is adjusted according to the probability of the flywheel having a first-order vibration fault, the probability of the flywheel having a second-order vibration fault, and the probability of the flywheel having a third-order vibration fault;

[0097] If the occurrence probability of a fault in the fault risk is greater than the second preset threshold, the suspension gap of the flywheel is adjusted and / or the rotational speed of the flywheel is reduced to a preset value;

[0098] If the occurrence probability of a fault in the fault risk is greater than the third preset threshold, the flywheel is controlled to stop.

[0099] It is not difficult to understand that for different fault situations, kinetic energy active defense operations are also set in this application, and according to the severity of the fault, the protection measures for the flywheel system are divided into three levels: primary preventive protection, secondary warning protection, and tertiary emergency protection, realizing hierarchical protection measures. The specific types and implementation methods of the first preset threshold, the second preset threshold, and the third preset threshold are not particularly limited in this application, and the triggering conditions for the protection measures at each level are not limited to the solutions described in this embodiment, and can be adjusted and set according to the actual application of the flywheel.

[0100] It should be noted that when the first-level preventive protection is triggered, the control system of the flywheel needs to ensure that the flywheel operates in the non-resonant region and quickly crosses the resonant region during the acceleration and deceleration processes. Appropriate crossing speeds are set in the resonant region mainly based on the probabilities of the first-order vibration fault, second-order vibration fault, and third-order vibration fault (the user can customize the probability threshold). In addition, key components are regularly inspected and replaced, and the functions of various sensors are detected to be in good condition. When the second-level early warning protection is triggered, according to the severity (probability value) of the early / mid-stage faults, operations such as speed limitation of the flywheel, adjustment of the magnetic levitation force of the flywheel, cooling, and limitation of charge and discharge power are performed to keep the rotor in a working state as much as possible. For example, for the degree of scratching and imbalance faults, that is, the probability of physical scratching faults of the flywheel, the probability of static imbalance faults of the flywheel, and the probability of dynamic imbalance faults of the flywheel, the magnetic levitation controller is adjusted to finely tune its suspension gap, and the speed is appropriately reduced according to the working conditions (avoiding the resonance point). For the vacuum degree reduction fault, that is, the probability of vacuum degree faults of the flywheel, it is controlled whether to connect the vacuum pump; for the undervoltage fault, a supercapacitor buffer is connected; for multi-modal vibrations, that is, the probability of the first-order vibration fault of the flywheel, the probability of the second-order vibration fault of the flywheel, and the probability of the third-order vibration fault of the flywheel, high-order harmonics are injected to suppress the vibration mode; for the over-temperature fault, that is, the probability of the shaft system over-temperature fault of the flywheel, the liquid nitrogen micro-spray cooling system is enabled. Second-level faults are generally recoverable faults, and suspension can be maintained during the fault. When the third-level emergency protection is triggered, operations such as deceleration braking (supplemented by hydraulic mechanical braking if necessary) and shutdown need to be performed, or energy discharge measures are executed. For example, when the grid-side converter of the flywheel system is in a fault state and cannot operate, the kinetic energy of the flywheel is consumed on the DC-side energy consumption braking resistor; when the machine-side converter of the flywheel system is in a fault state and cannot operate, the kinetic energy of the flywheel is consumed on the AC-side energy consumption braking resistor. At the same time, the backup power supply or energy storage system is started to ensure the power supply of key equipment; according to the probability of the fault, when it is clear that one or more of the nine faults occur; or the mechanical faults and electrical faults described above can all enter the energy discharge or braking state and the shutdown state.

[0101] Specifically, by setting hierarchical protection measures to execute different fault strategies according to different fault states, the protection of the flywheel is achieved, and through differentiated protection strategies, the safety efficiency is maximized.

[0102] As an optional embodiment, the pre-constructed convolutional neural network model includes an input layer, a convolutional layer, an activation layer, a pooling layer, a fully connected layer, and an output layer; the process of the convolutional neural network model evaluating the fault risk of the flywheel specifically includes:

[0103] Receiving the state parameters of the flywheel through the input layer;

[0104] Perform a convolution operation on the state parameters using the convolution kernels in the convolutional layer to extract the response features in the state parameters;

[0105] Output the response features to the activation layer to perform non-linear processing on the response features using the activation function in the activation layer;

[0106] Perform downsampling on the response characteristics after non-linear processing based on the pooling layer to extract the local features in the response features;

[0107] Integrate all the local features output by the pooling layer into a global feature through the fully connected layer, and map the global feature to each fault category to output the fault occurrence probability corresponding to each fault category.

[0108] It can be understood that CNN mainly includes five parts: the input layer, the convolutional layer, the pooling layer, the fully connected layer, and the output layer. An activation layer can also be added after the convolutional layer. Its execution is mainly divided into two stages: convolution and classification. For the input layer, the number of neurons in the CNN input layer depends on the total number of elements in the matrix and multi-dimensional array after data processing. To enhance data visualization and better convolution effect, data classification and reshaping are required, such as reshaping a one-dimensional array or a two-dimensional array into the form of a multi-dimensional tensor. The specific data classification and the processing process during data input are as described above. The number of neurons in the output layer is consistent with the number of fault types, and the probability corresponding to each fault type is output.

[0109] Among them, the convolutional layer is a key component of the CNN model. The convolutional layer uses a series of filters and local regions of the input, and uses a convolutional kernel of a certain size to obtain response features through convolution operation with the input features of the previous layer. Convolution operation is performed to extract the features of the original signal. Assume that the previous convolution is the l-th layer and the current convolution is the l - 1 layer. If it contains M feature maps and N filters, then the mapped feature value of the j-th output of the upper layer to the current layer is , and its value is:

[0110] ;

[0111] Among them, is the mapping of the i-th output of the upper layer, is the convolutional kernel of the j-th filter connecting the i-th output mapping, is the bias term of the j-th filter, is the convolution operation.

[0112] Furthermore, to enhance the model's expressive power, an activation layer is introduced after the convolutional layer. The application of the activation function enables the network to learn more complex patterns and features. The rectified linear unit (ReLU) function can be chosen as the activation function, as it has simple calculations and can effectively alleviate the vanishing gradient problem. After being processed by the activation function, the becomes , which is expressed as:

[0113] ;

[0114] That is . Among them, f is the ReLU activation function, which can not only make the network converge faster but also has a very small computational amount. Its mathematical calculation formula is:

[0115] .

[0116] It should be noted that to reduce the computational amount and prevent overfitting, the CNN model generally contains a pooling layer. The pooling layer is also known as the downsampling layer and is generally located after the convolutional layer. The pooling function can adopt average pooling and max pooling. Among them, average pooling divides the input features into several regions and then calculates the average value within the divided regions. Max pooling outputs the maximum value within the divided regions. After the pooling operation, the feature map output by the l-th pooling layer (pooling occurs after the previous layer of convolution) is calculated as:

[0117] ;

[0118] In the formula, is the j-th output map, is the bias of the j-th filter, and down is the downsampling function. The pooling layer reduces the dimension of the features extracted by the convolutional layer for subsequent recognition operations. The pooling layer reduces the amount of data and retains the internal local features of the input channel data variables and their mutual relationships. After several convolutions and poolings, the local features of the flywheel fault are continuously strengthened.

[0119] The convolutional layer and the pooling layer can learn the features of the samples and then expand the multi-dimensional features into a one-dimensional vector to prepare for sample classification. The one-dimensional feature vector is then connected to the fully connected layer, where all the features between the next layer and the previous layer are interconnected. Finally, the features of the fully connected layer are recognized through the Softmax regression model, and its expression is:

[0120] ;

[0121] In the formula, represents the Softmax regression model parameters, Represents the input features of the i-th sample, and K is the total number of recognized sample types.

[0122] Finally, it is the design of the output layer. The number of neurons in the output layer is the same as the number of fault types of the flywheel, that is, 9. Each neuron outputs the probability of the corresponding fault, and each neuron and the fault type should be specified one by one in advance. However, it should be noted that the accuracy of the CNN output result mainly depends on the data level (data quality) and the model level (the design of the convolutional, pooling, and fully connected layers of the CNN). In this embodiment, the number of convolutional layers is designed to be 3 (one activation function layer is added after each convolutional layer), the number of pooling layers is 2 (the third convolution does not need to go through pooling and is directly passed to the fully connected layer), the number of fully connected layers is 2, and in addition, there is 1 layer for both the input layer and the output layer, with a total of 9 layers of structure.

[0123] Specifically, by refining the structural design method of the CNN, a design method for each layer in the CNN model is provided, providing a reliable design basis for the design of the CNN model.

[0124] As an alternative embodiment, based on the convolutional kernel in the convolutional layer, a convolutional operation is performed on the state parameters to extract the response features of the state parameters, including:

[0125] Adopt the method of one-dimensional convolution to perform single-channel feature extraction on the time series data corresponding to a single sampling channel of the state parameters based on the sampling time sequence;

[0126] Adopt the method of spatio-temporal separable convolution to obtain the fusion features of several sampling channels of the state parameters based on the sampling channel and the sampling time sequence;

[0127] Combine all single-channel features and fusion features to obtain the response features of the state parameters.

[0128] It is not difficult to understand that the design of the convolution kernel in the convolutional layer is very important. The convolution kernel in the convolutional layer measures the "contribution degree", that is, the weight, of each variable in the input layer to the fault probability in the fault diagnosis of flywheel energy storage. The CNN convolution kernel design method provided in this embodiment includes three parts. The first part is independent first and then fusion, the second part is multi-channel coverage, and the third part is multi-scale feature fusion. The first part is mainly to extract single-channel time features, aiming to extract the time-series channel features of each sensor itself (such as current, temperature, displacement, etc.), using 1D convolution and sliding along the time axis. The extraction of single-channel time features refers to the corresponding change trends (local and long-term) of the variables in this channel, such as sudden changes in current, inflection points of temperature, and speeds of displacement (i.e., accelerations of displacement), etc. The second part is mainly to use spatio-temporal separable convolution, first convolving along the channel axis (each acquisition channel), and then convolving along the time axis (acquisition time series), aiming to automatically learn the weight combination between multi-channel signals, realize cross-channel feature fusion, and explore the potential correlations between signals among multiple sensors, with few parameters and high efficiency. Taking displacement, temperature, and acceleration as examples (the depth of the correlation between variables depends on the receptive field of the CNN), spatial convolution will find the relationships among the three variables of displacement, temperature, and acceleration. For example, when the temperature is high, the end will elongate, and when the acceleration is high, the shaft will bend and the end will contract inward, and the displacement is affected by temperature and acceleration, etc. Multi-channel coverage solves the internal relationships of multiple variables. Among them, the contribution degree of each variable to the fault is different, and this is the weight. The entire convolutional layer adopts a parallel branch design. After obtaining the single-channel and multi-channel features through the first part and the second part respectively, multi-scale feature fusion can be carried out. Multi-scale fusion means merging the features output by each channel (such as splicing, weighted summation, etc.). Specifically, the features after convolution can be fused by means of channel splicing or addition to enhance the expression ability of the features, and then downsampling or upsampling is used to align the feature dimensions of different branches to further highlight the multi-variable features and the correlations between features. The three parts of the entire convolutional layer can be implemented through dilated convolution (atrous convolution) to strengthen the extraction of multi-scale time-series pattern features.

[0129] To further explore the internal relationships between multiple channels, an SE module (Squeeze-and-Excitation) can be added after the activation function to achieve non-linear decoupling with the ReLU activation function and improve the quality of fault features through channel attention. Introducing the SE module after the activation function enhances the expression ability of the features through the dynamic channel attention mechanism.

[0130] As a specific embodiment, such as Figure 3As shown in the figure, the CNN module convolves the input signal, and through the receptive field function and convolution kernel, it extracts the preliminary fault features of the upper-layer data; then through the pooling layer, it downsamples the data and selectively retains the local features; the fully connected layer combines each local feature into an overall feature and classifies it, maps the integrated feature to different categories, and finally the connection layer outputs the probability of the flywheel fault.

[0131] Specifically, this embodiment provides a design method for a multi-channel sequence convolution kernel. Without reducing the resolution, it uses dilated convolution to implement the convolution kernel, expands the receptive field, and fully captures the mutual relationship between multi-channel data. It refines the design process of the CNN, and through reasonable design, the entire magnetic levitation fault diagnosis has a high accuracy rate.

[0132] To solve the above technical problems, the present invention also provides a flywheel fault detection device, including:

[0133] A parameter acquisition unit, configured to acquire the state parameters of the flywheel; the state parameters include at least two of the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters, and dynamic performance parameters;

[0134] A fault diagnosis unit, configured to input the state parameters into a pre-constructed convolutional neural network model to evaluate the fault risk of the flywheel by using the pre-constructed convolutional neural network model; the fault risk of the flywheel includes at least two of the probability of the flywheel having an over-temperature fault in the shafting, the probability of the flywheel having a static imbalance fault, the probability of the flywheel having a dynamic imbalance fault, the probability of the flywheel having a physical scratch fault, the probability of the flywheel having a shafting immersion fault, the probability of the flywheel having a vacuum degree fault, the probability of the flywheel having a first-order vibration fault, the probability of the flywheel having a second-order vibration fault, and the probability of the flywheel having a third-order vibration fault;

[0135] An execution protection unit, configured to execute corresponding protection strategies based on the evaluation result of the convolutional neural network model.

[0136] As an optional embodiment, it further includes:

[0137] A preprocessing unit, configured to preprocess the state parameters before inputting the state parameters into a pre-constructed convolutional neural network model;

[0138] The fault diagnosis unit includes:

[0139] An input unit, configured to input the state parameters after preprocessing as the original signal into a pre-constructed convolutional neural network model.

[0140] As an alternative embodiment, when the state parameter includes a vibration parameter, the original signal corresponding to the vibration parameter in the state parameter includes a first original sub-signal and a second original sub-signal, and the preprocessing unit includes:

[0141] A first transformation unit, configured to perform an s-transform on the vibration parameter in the state parameter, and determine the s-transform spectrogram obtained by the s-transform as the first original sub-signal of the vibration parameter;

[0142] A second transformation unit, configured to perform an fft transform on the vibration parameter in the state parameter, and determine the two-dimensional grayscale image obtained by the fft transform as the second original sub-signal of the vibration parameter.

[0143] As an alternative embodiment, the parameter acquisition unit includes:

[0144] A sliding window sampling unit, configured to perform sliding window sampling on the state parameters of the flywheel based on a preset sliding window length and a preset sampling channel;

[0145] The input unit includes:

[0146] A batching unit, configured to divide the state parameters after preprocessing into respective preset batches; wherein, there is a batch number corresponding to each of the preset batches;

[0147] A three-dimensional list construction unit, configured to construct a three-dimensional list based on the sampling channel corresponding to the state parameter, the preset batch corresponding to the state parameter, and the original signal corresponding to the state parameter;

[0148] An input subunit, configured to input the three-dimensional list in batches into a pre-constructed convolutional neural network model according to the batch numbers of the respective preset batches.

[0149] As an alternative embodiment, the execution protection unit includes:

[0150] A first-level protection subunit, configured to, if the occurrence probability of a fault in the fault risk is greater than a first preset threshold, adjust the crossing speed of the flywheel across the resonance region according to the probability of the flywheel having a first-order vibration fault, the probability of the flywheel having a second-order vibration fault, and the probability of the flywheel having a third-order vibration fault;

[0151] A second-level protection subunit, configured to, if the occurrence probability of a fault in the fault risk is greater than a second preset threshold, adjust the suspension gap of the flywheel and / or reduce the rotational speed of the flywheel to a preset value;

[0152] A third-level protection subunit, configured to, if the occurrence probability of a fault in the fault risk is greater than a third preset threshold, control the flywheel to stop.

[0153] As an optional embodiment, the pre-built convolutional neural network model includes an input layer, a convolution layer, an activation layer, a pooling layer, a fully connected layer, and an output layer; the flywheel fault detection device also includes a model execution unit, which includes:

[0154] An input layer unit, configured to receive state parameters of the flywheel through an input layer;

[0155] A convolution layer unit, configured to perform a convolution operation on the state parameter based on a convolution kernel in the convolution layer to extract a response feature from the state parameter;

[0156] an activation layer unit, configured to output the response feature to an activation layer, so as to perform nonlinear processing on the response feature using an activation function in the activation layer;

[0157] A pooling layer unit, configured to downsample the response characteristics after nonlinear processing based on the pooling layer to extract local features from the response characteristics;

[0158] The output layer unit is used to integrate all local features output by the pooling layer into a global feature through a fully connected layer, and map the global feature to each fault category to output the fault occurrence probability corresponding to each fault category.

[0159] As an optional embodiment, the convolutional layer unit includes:

[0160] A single-channel extraction unit is used to extract single-channel features from the time series data corresponding to a single sampling channel of the state parameter based on the sampling timing by using a one-dimensional convolution method;

[0161] A multi-channel extraction unit, configured to obtain fusion features of the plurality of sampling channels of the state parameter based on the sampling channels and sampling timing by adopting a space-time separation convolution method;

[0162] A fusion unit is used to combine all the single-channel features and the fusion features to obtain the response features of the state parameters.

[0163] For an introduction to a flywheel fault detection device provided by the present invention, please refer to the embodiment of the flywheel fault detection method described above, and the present invention will not be described in detail here.

[0164] Please refer to Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device provided by the present invention. To solve the above technical problems, the present invention also provides an electronic device, including:

[0165] Memory 21, for storing computer programs;

[0166] The processor 22 is configured to implement the steps of the flywheel fault detection method as described above.

[0167] Among them, the processor 22 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 22 may be implemented in at least one of the following hardware forms: DSP (Digital Signal Processor), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 22 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the wake state, also known as the central processor; the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 22 may integrate a GPU (graphics processing unit), and the GPU is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 22 may further include an AI (Artificial Intelligence) processor, and the AI processor is used to process computational operations related to machine learning.

[0168] The memory 21 may include one or more computer-readable storage media, and the computer-readable storage media may be non-transitory. The memory 21 may further include high-speed random access memory and non-volatile memory, such as one or more disk storage devices and flash storage devices. In this embodiment, the memory 21 is at least used to store the following computer programs. After the computer programs are loaded and executed by the processor 22, the relevant steps of the flywheel fault detection method disclosed in any of the foregoing embodiments can be implemented. In addition, the resources stored in the memory 21 may also include an operating system and data, etc., and the storage method may be temporary storage or permanent storage. Among them, the operating system may include Windows, Unix, Linux, etc. The data may include, but is not limited to, data of the flywheel fault detection method, etc.

[0169] In some embodiments, the electronic device may further include a display screen, an input / output interface 25, a communication interface 24, a power supply 23, and a communication bus 26.

[0170] Those skilled in the art can understand that Figure 8 the structure shown in

[0171] For the introduction of an electronic device provided by the present invention, please refer to the embodiments of the above-mentioned flywheel fault detection method, and the present invention will not be elaborated herein.

[0172] To solve the above technical problems, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned flywheel fault detection method are implemented.

[0173] It can be understood that if the method in the above embodiments is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and executes all or part of the steps of the methods described in the various embodiments of the present application. Specifically, the computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, and mobile hard disks, etc., or any type of medium or device suitable for storing instructions and data, etc. The present application does not make special limitations here.

[0174] For the introduction of a computer-readable storage medium provided by the present invention, please refer to the embodiments of the above-mentioned flywheel fault detection method, and the present invention will not be elaborated herein.

[0175] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For the relevant parts, please refer to the description of the method part. Professionals can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0176] The foregoing description of the disclosed embodiments enables those skilled in the art to practice or use the present invention. 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 the present invention. Thus, the present invention is not intended 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 method for detecting faults of a flywheel, characterized in that, Including: Obtaining the state parameters of the flywheel; the state parameters include at least two of the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters, and dynamic performance parameters; Inputting the state parameters into a pre-constructed convolutional neural network model to evaluate the fault risk of the flywheel by using the pre-constructed convolutional neural network model; the fault risk of the flywheel includes at least two of the probability of the flywheel having an over-temperature fault in the shafting, the probability of the flywheel having a static imbalance fault, the probability of the flywheel having a dynamic imbalance fault, the probability of the flywheel having a physical scratch fault, the probability of the flywheel having a shafting immersion fault, the probability of the flywheel having a vacuum degree fault, the probability of the flywheel having a first-order vibration fault, the probability of the flywheel having a second-order vibration fault, and the probability of the flywheel having a third-order vibration fault; Executing corresponding protection strategies based on the evaluation results of the convolutional neural network model.

2. The method for detecting a fault of a flywheel according to claim 1, characterized in that, Before inputting the state parameters into the pre-constructed convolutional neural network model, it further includes: Preprocessing the state parameters; Inputting the state parameters into the pre-constructed convolutional neural network model includes: Taking the preprocessed state parameters as the original signal and inputting them into the pre-constructed convolutional neural network model.

3. The method for detecting a fault of a flywheel according to claim 2, wherein, When the state parameters include vibration parameters, the original signal corresponding to the vibration parameters in the state parameters includes a first original sub-signal and a second original sub-signal. Preprocessing the state parameters includes: Performing an s-transform on the vibration parameters in the state parameters and determining the s-transform spectrogram obtained by the s-transform as the first original sub-signal of the vibration parameters; Performing an fft-transform on the vibration parameters in the state parameters and determining the two-dimensional grayscale image obtained by the fft-transform as the second original sub-signal of the vibration parameters.

4. The method for detecting a fault of a flywheel according to claim 3, wherein, Obtaining the state parameters of the flywheel includes: Performing sliding window sampling on the state parameters of the flywheel based on a preset sliding window length and a preset sampling channel; Inputting the preprocessed state parameters into the pre-constructed convolutional neural network model includes: Dividing the preprocessed state parameters into each preset batch; where each of the preset batches has a corresponding batch number; Constructing a three-dimensional list based on the sampling channel corresponding to the state parameters, the preset batch corresponding to the state parameters, and the original signal corresponding to the state parameters; Inputting the three-dimensional list into the pre-constructed convolutional neural network model in batches according to the batch numbers of each preset batch.

5. The method for detecting a fault of a flywheel according to claim 1, wherein Executing corresponding protection strategies based on the evaluation results of the convolutional neural network model includes: If the occurrence probability of a fault in the fault risk is greater than a first preset threshold, then adjusting the crossing speed of the flywheel across the resonance region according to the probability of the flywheel having a first-order vibration fault, the probability of the flywheel having a second-order vibration fault, and the probability of the flywheel having a third-order vibration fault; If the occurrence probability of a fault in the fault risk is greater than a second preset threshold, then adjusting the suspension gap of the flywheel and / or reducing the rotational speed of the flywheel to a preset value; If the occurrence probability of a fault in the fault risk is greater than a third preset threshold, then controlling the flywheel to stop.

6. The method for detecting a fault of a flywheel according to any one of claims 1 to 5, characterized in that, The pre - constructed convolutional neural network model includes an input layer, a convolutional layer, an activation layer, a pooling layer, a fully - connected layer, and an output layer; the process of the convolutional neural network model evaluating the fault risk of the flywheel specifically includes: Receiving the state parameters of the flywheel through the input layer; Performing a convolution operation on the state parameters based on the convolution kernels in the convolutional layer to extract the response features in the state parameters; Outputting the response features to the activation layer to perform non - linear processing on the response features by using the activation function in the activation layer; Based on the pooling layer, downsampling the response characteristics after non - linear processing to extract the local features in the response features; Integrating all the local features output by the pooling layer into a global feature through the fully - connected layer and mapping the global feature to each fault category to output the probability of the occurrence of the fault corresponding to each fault category.

7. The method for detecting a fault of a flywheel according to claim 6, wherein, Performing a convolution operation on the state parameters based on the convolution kernels in the convolutional layer to extract the response features of the state parameters, including: Adopting a one - dimensional convolution method to perform single - channel feature extraction on the time - series data corresponding to a single sampling channel of the state parameters based on the sampling time series; Adopting a space - time separable convolution method to obtain the fusion features of several sampling channels of the state parameters based on the sampling channels and the sampling time series; Combining all the single - channel features and the fusion features to obtain the response features of the state parameters.

8. A failure detection device for a flywheel, characterized in that, Including: A parameter acquisition unit for acquiring the state parameters of the flywheel; the state parameters include at least two of the inherent parameters of the flywheel, the control current of the flywheel, vibration parameters, displacement parameters, environmental parameters, speed parameters, and dynamic performance parameters; A fault diagnosis unit for inputting the state parameters into the pre - constructed convolutional neural network model to evaluate the fault risk of the flywheel by using the pre - constructed convolutional neural network model; the fault risk of the flywheel includes at least two of the probability of the flywheel having an over - temperature fault in the shafting, the probability of the flywheel having a static imbalance fault, the probability of the flywheel having a dynamic imbalance fault, the probability of the flywheel having a physical scratch fault, the probability of the flywheel having a shafting immersion fault, the probability of the flywheel having a vacuum degree fault, the probability of the flywheel having a first - order vibration fault, the probability of the flywheel having a second - order vibration fault, and the probability of the flywheel having a third - order vibration fault; An execution protection unit for executing corresponding protection strategies based on the evaluation results of the convolutional neural network model.

9. An electronic device, characterized in that, Including: A memory for storing computer programs; A processor for implementing the steps of the flywheel fault detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer - readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements the steps of the flywheel fault detection method according to any one of claims 1 to 7.

Images