Wind power digital twin robustness detection method based on time-frequency rotating door perception matching

Abstract

The application discloses a wind power digital twin robustness detection method based on time-frequency rotating door perception matching, which comprises obtaining a wind turbine detection signal generated by a digital twin and a wind turbine detection signal collected by a sensor, wherein the wind turbine detection signal is the temperature, radial vibration, axial vibration, current, voltage or power of the wind turbine; the rotating door algorithm SDA is used to respectively preprocess the two kinds of wind turbine detection signals, so that the feature frequency amplitude error of the preprocessed wind turbine detection signal is less than a preset threshold; the receptive field of the corresponding convolutional neural network of each wind turbine detection signal is updated based on the preprocessed wind turbine detection signal to determine the convolutional neural network convolutional layer number, and the difference between the convolutional neural network convolutional layer numbers of the two kinds of wind turbine detection signals is taken as the robustness detection result output of the digital twin. The application aims to improve the data processing capability, solve the spectrum deviation and aliasing when the signal is smoothly changed, and the misdiagnosis possibly caused by signal mutation, and improve the accuracy of robustness judgment.

Description

Technical Field

[0001] The present invention relates to the technical field of digital twin technology for the service quality of wind turbine groups, and in particular to a wind power digital twin robustness detection method based on time-frequency revolving door perception matching. Background Art

[0002] During the long-term operation of wind turbines, due to the complex and ever-changing environment they operate in, they often face a variety of failures and uncertainties. These failures not only affect electricity production efficiency but can also threaten the stability of the power grid. With the rapid development of the global wind power industry, the number and scale of wind turbines continue to expand. Effectively evaluating and improving the service quality of wind turbines to ensure the long-term stable operation of wind farms has become a critical technical challenge that urgently needs to be addressed. The core concept of digital twins is to fully leverage physical data and mechanisms to construct ultra-fidelity virtual replicas. Through interaction and iteration between the virtual and physical worlds, the physical world can be monitored, controlled, and optimized. Testing and verification technology for wind turbine fleet service quality digital twin systems is crucial for improving wind turbine fleet reliability and performance. They can predict failures, optimize operations, diagnose problems, and provide intelligent management to enhance wind turbine fleet reliability and performance, reduce downtime and maintenance costs, and ultimately achieve more efficient wind energy utilization. However, in practice, wind turbine operating environments are extremely complex. External factors such as wind speed fluctuations, humidity, and temperature can affect turbine performance. Furthermore, internal issues such as sensor accuracy degradation and signal noise can gradually reduce the system's detection and diagnostic capabilities. Therefore, robustness has become a key metric in the application of digital twin technology in wind turbines. Robustness refers to a system's ability to maintain stable performance in the face of external interference, uncertainty, or internal changes. For wind turbine digital twin systems, robustness is reflected not only in their adaptability to the external environment but also in their ability to cope with complex situations such as equipment aging and failure. A highly robust system ensures efficient and stable performance despite sensor accuracy degradation, noise interference, or changes in turbine status. Summary of the Invention

[0003] The technical problem to be solved by the present invention is as follows: In response to the above-mentioned problems of the prior art, a wind power digital twin robustness detection method based on time-frequency revolving door perception matching is provided. The present invention aims to improve data processing capabilities to cope with the problem that the wind power digital twin system usually adopts a high sampling rate when collecting signals to ensure the accuracy of the detection and verification system, resulting in a large amount of data. It solves the misdiagnosis that may be caused by spectrum deviation and aliasing when the signal changes steadily and signal mutation, and improves the accuracy of robustness judgment.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A wind turbine digital twin robustness detection method based on time-frequency rotating door perception matching, comprising:

[0006] In step S1, a wind turbine detection signal generated by a digital twin and a wind turbine detection signal collected by a sensor are obtained, and the wind turbine detection signal is a temperature, a radial vibration, an axial vibration, a current, a voltage or a power of the wind turbine;

[0007] In step S2, the rotating door algorithm SDA is used to preprocess the two wind turbine detection signals respectively to make the feature frequency amplitude error of the preprocessed wind turbine detection signal less than a preset threshold;

[0008] In step S3, the receptive field of the respective corresponding convolutional neural network is updated based on the preprocessed wind turbine detection signal to determine the convolutional neural network convolutional layer number, and the difference between the convolutional neural network convolutional layer numbers of the two wind turbine detection signals is taken as the robustness detection result of the digital twin.

[0009] Optionally, the calculation function expression of the feature frequency amplitude error in step S2 is:

[0010]

[0011] In the above formula, e max is the feature frequency amplitude error, f ε,max is the feature frequency amplitude of the preprocessed wind turbine detection signal, and f max is the feature frequency amplitude of the original wind turbine detection signal.

[0012] Optionally, step S2 includes:

[0013] In step S2.1, the initial door width of the rotating door algorithm SDA is set;

[0014] In step S2.2, the rotating door algorithm SDA is used to process the two wind turbine detection signals respectively;

[0015] In step S2.3, the two wind turbine detection signals are converted into frequency domain signals by fast Fourier transform and the feature frequency amplitude of the original wind turbine detection signal is extracted, the preprocessed wind turbine detection signal is converted into a frequency domain signal by fast Fourier transform after the processing in step S2.2 and the feature frequency amplitude of the preprocessed wind turbine detection signal is extracted, the feature frequency amplitude error of the two preprocessed wind turbine detection signals is calculated according to the feature frequency amplitude of the preprocessed wind turbine detection signal and the feature frequency amplitude of the original wind turbine detection signal respectively, if the two feature frequency amplitude errors are less than the preset threshold, then jump to step S3; otherwise, the door width of the rotating door algorithm SDA is updated, and jump to step S2.2.

[0016] Optionally, after using the revolving door algorithm SDA to process the two fan detection signals respectively in step S2.2, the two fan detection signals are further processed respectively using an abnormal algorithm, including for each current data point, based on a given abnormal parameter V E The length of the current data point is 2V above and below E The abnormal range of the next data point is used to determine whether an abnormality occurs. If the next data point exceeds the abnormal range, the next data point is deleted and the data point is reconstructed using linear interpolation based on the first two data points of the deleted data point.

[0017] Optionally, step S2.2 further includes using data segment merging to process the two wind turbine detection signals respectively, including detecting a climbing segment with an increase in data points according to a change in the data points, and performing climbing segment merging according to the following formula:

[0018]

[0019] In the above formula, l is the current time; n is the total time; x l+1 is the amplitude of the vibration signal at time l+1; x l is the amplitude of the vibration signal at time l; b(l) is the reverse trend data segment between two identical trends whose amplitudes are both less than the preset threshold.

[0020] Optionally, step S2.2 further includes using small amplitude jitter data segment detection merging to process the two fan detection signals respectively, including: finding a data segment in the fan detection signal that continuously shows an upward and downward change within a time period, and the amplitude of the upward and downward trend segments is less than a preset threshold as a small amplitude jitter data segment; Its function expression is:

[0021]

[0022] In the above formula, are respectively small jitter data segments The 1st to cth monotonic data segments, and there is any dth monotonic data segment The function expression is:

[0023]

[0024] In the above formula, and are the dth monotonic data segments respectively The starting and ending points of each small jitter data segment The two adjacent monotonic data segments are traversed in sequence, and the equivalent revolving door starting point and door width parameters of the small-amplitude jitter data segment are calculated according to the following formula:

[0025]

[0026] X c is the signal amplitude at time t c ; X is the amplitude of the specific point; g is the signal amplitude at time t g ; X is the signal amplitude of the rotating door point of the rotating door algorithm SDA at time t c ; g is the point time; X is the time of the specific point; X is the time of the rotating door point of the rotating door algorithm SDA; X is the amplitude of the lower support point of the rotating door of the rotating door algorithm SDA; X is the time of the upper support point of the rotating door of the rotating door algorithm SDA; X is the time of the lower support point of the rotating door of the rotating door algorithm SDA; X is the amplitude of the upper support point of the rotating door of the rotating door algorithm SDA; ε' is a small amplitude jitter data segment door width parameter X0 is the starting point signal amplitude; the specific point is set as the door point of the rotating door of the rotating door algorithm SDA

[0027] According to the signal amplitudes V of the current two monotonic data segments, the upper rotating door signal amplitude V h , and the lower rotating door signal amplitude V l obtained by detecting the equivalent rotating door starting point and the door width parameter based on the small amplitude jitter data segment, it is judged whether the detection condition V of the rotating door algorithm is established h <V<V l , and then the two monotonic data segments obtained by traversal are merged.

[0028] Optionally, the function expression of the convolutional neural network convolution layer number is determined by updating the receptive field of each corresponding convolutional neural network based on the preprocessed fan detection signal in step S3 as follows:

[0029]

[0030] In the above formula, n k-PRAA is the data quantity of the preprocessed fan detection signal, and τ is a related constant of different fault features; n kis the amount of original signal data under different faults; ε is the gate width of the revolving door algorithm; r l is the receptive field size of the lth convolutional layer; r l-1 is the receptive field size of the l-1th convolutional layer; f l is the convolution kernel size of the lth convolution layer; s i is the step size of the i-th convolutional layer; the convolutional neural network is composed of multiple convolutional layers and pooling layers.

[0031] In addition, the present invention also provides a wind power digital twin robustness detection system based on time-frequency revolving door perception matching, including an interconnected microprocessor and a memory, and the microprocessor is programmed or configured to execute the wind power digital twin robustness detection method based on time-frequency revolving door perception matching.

[0032] In addition, the present invention also provides a computer-readable storage medium, which stores a computer program or instruction, and the computer program or instruction is programmed or configured to execute the wind power digital twin robustness detection method based on time-frequency revolving door perception matching through a processor.

[0033] In addition, the present invention also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the wind power digital twin robustness detection method based on time-frequency revolving door perception matching through a processor.

[0034] Compared with the prior art, the present invention mainly has the following advantages:

[0035] During long-term wind turbine operation, constantly changing operating conditions lead to varying speeds, further complicating the signals used for digital twin system evaluation. These speed variations can also lead to frequency modulation and spectral blurring of vibration signals, further complicating digital twin system evaluation. To ensure the accuracy of detection and verification systems, digital twin systems typically use high sampling rates when acquiring signals, resulting in a significant amount of data processing. Due to the large amount of data, the existing robustness detection and verification method requires a large amount of storage memory, which makes the evaluation process very slow. This phenomenon brings data pressure to the entire process. The present invention includes obtaining the wind turbine detection signal generated by the digital twin and the wind turbine detection signal collected by the sensor, and the wind turbine detection signal is the temperature, radial vibration, axial vibration, current, voltage or power of the wind turbine; the two wind turbine detection signals are preprocessed by the rotating door algorithm SDA so that the characteristic frequency amplitude error of the preprocessed wind turbine detection signal is less than the preset threshold; based on the preprocessed wind turbine detection signal, the receptive field of each corresponding convolutional neural network is updated to determine the number of convolutional layers of the convolutional neural network, and the difference in the number of convolutional layers of the convolutional neural network of the two wind turbine detection signals is output as the robustness detection result of the digital twin. Through the above method, the problem of limited data processing capability of the existing digital twin system evaluation index algorithm can be solved, and the data processing capability can be improved to cope with the problem of large data volume caused by the high sampling rate usually adopted when collecting signals in order to ensure the accuracy of the detection and verification system.

[0036] 2. The amplitude and frequency of the signal change with the change of the working state of the motor, and the characteristic frequency of the signal in the steady state is very clear. When the working conditions change steadily, the spectrum shows slight deviation and aliasing compared with the steady state. When the working conditions suddenly change, the change in the characteristic frequency of the signal is more obvious. At the same time, the signal mutation may lead to misdiagnosis of normal motors. The present invention determines the number of convolutional layers of the convolutional neural network by updating the receptive fields of the corresponding convolutional neural networks based on the preprocessed fan detection signals, and outputs the difference in the number of convolutional layers of the convolutional neural networks of the two fan detection signals as the robustness detection result of the digital twin. It can solve the spectrum deviation and aliasing when the signal changes steadily and the misdiagnosis caused by the signal mutation, thereby improving the accuracy of the robustness judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of the basic process of the method of the embodiment of the present invention.

[0038] Figure 2 Schematic diagram of the basic principle of the method of the embodiment of the present invention.

[0039] Figure 3 Schematic diagram of the revolving door algorithm SDA in an embodiment of the present invention.

[0040] Figure 4 The schematic diagram of the abnormality algorithm in the embodiment of the application.

[0041] Figure 5 The schematic diagram of the data segment merging and wavelet dithering data segment processing in the embodiment of the application.

[0042] Figure 6 The time domain and frequency domain digital twin signals in the stable state and the transient state in the embodiment of the application, wherein (a) is the stable state, and (b) is the transient state.

[0043] Figure 7 The schematic diagram of the original vibration signal data collected in the embodiment of the application.

[0044] Figure 8 The schematic diagram of the data volume comparison in the embodiment of the application.

[0045] Figure 9 The time domain signal comparison of the amplitude error preprocessed in the embodiment of the application of different characteristic frequencies, wherein (a) is the time domain signal after preprocessing of the original signal, (b) is the time domain signal after preprocessing of the signal with an amplitude error of 0.3%, (c) is the time domain signal after preprocessing of the signal with an amplitude error of 2.2%, and (d) is the time domain signal after preprocessing of the signal with an amplitude error of 4.0%.

[0046] Figure 10 The diagnostic rate of the same data under different convolution layer numbers in the embodiment of the application. DETAILED DESCRIPTION

[0047] As shown in Figure 1 and Figure 2 The wind power digital twin robustness detection method based on time-frequency rotating door perception matching in the embodiment includes:

[0048] Step S1, obtaining a wind turbine detection signal generated by a digital twin and a wind turbine detection signal collected by a sensor, the wind turbine detection signal being temperature, radial vibration, axial vibration, current, voltage or power of the wind turbine;

[0049] Step S2, preprocessing the two kinds of wind turbine detection signals respectively by a rotating door algorithm SDA so that the characteristic frequency amplitude error of the preprocessed wind turbine detection signal is less than a preset threshold;

[0050] Step S3, updating the receptive field of the respective corresponding convolutional neural network based on the preprocessed wind turbine detection signal to determine the convolutional neural network convolutional layer number, and taking the difference between the convolutional neural network convolutional layer numbers of the two kinds of wind turbine detection signals as the robustness detection result of the digital twin and outputting.

[0051] The calculation function expression of the characteristic frequency amplitude error in step S2 of the embodiment is:

[0052]

[0053] In the above formula, e max is the characteristic frequency amplitude error, f ε,max is the characteristic frequency amplitude of the fan detection signal after preprocessing, f max is the characteristic frequency amplitude of the original fan detection signal.

[0054] The revolving door algorithm SDA is a well-known algorithm. The revolving door algorithm SDA was originally proposed for trend fitting of signals. The principle of the revolving door algorithm SDA is as follows: Figure 3 As shown in the figure, the revolving door algorithm (SDA) selects a gate width at the starting point of each iteration during the data point retrieval process. Subsequent data points are connected to this gate width as the upper and lower boundaries. When one boundary reaches its maximum value, it is fixed, while the other boundary continues to update as the data points change. When the two boundaries are parallel, the current measurement data point is selected as the end point of the iteration, and a new iteration begins. For the current measurement data point, the revolving door algorithm detects the following:

[0055] V h <V<V l ,

[0056] In the above formula, V is the signal amplitude of the current detection point; V h is the upper revolving door signal amplitude; V l is the amplitude of the lower revolving door signal. In this embodiment, step S2 includes:

[0057] Step S2.1, setting the initial door width of the revolving door algorithm SDA;

[0058] Step S2.2, using the revolving door algorithm SDA to process the two types of fan detection signals respectively;

[0059] Step S2.3, convert the two fan detection signals into frequency domain signals through fast Fourier transform and extract the characteristic frequency amplitude of the original fan detection signal, convert the processed signals in step S2.2 into frequency domain signals through fast Fourier transform and extract the characteristic frequency amplitude of the preprocessed fan detection signal, calculate the characteristic frequency amplitude errors of the two preprocessed fan detection signals according to the characteristic frequency amplitude of the preprocessed fan detection signal and the characteristic frequency amplitude of the original fan detection signal respectively, if both characteristic frequency amplitude errors are less than the preset threshold, jump to step S3; otherwise, update the door width of the revolving door algorithm SDA and jump to step S2.2.

[0060] In step S2.2 of this embodiment, after using the revolving door algorithm SDA to process the two fan detection signals respectively, the embodiment further includes using an abnormal algorithm to process the two fan detection signals respectively, including for each current data point, based on a given abnormal parameter V E The length of the current data point is 2V above and below E The abnormal range of the next data point is used to determine whether an abnormality occurs. If the next data point exceeds the abnormal range, the next data point is deleted and the data point is reconstructed using linear interpolation based on the two data points before the deleted data point. The abnormality algorithm can be used to identify erroneous data points after the signal is preprocessed by this type of method. Figure 4 As shown, the abnormal parameter V E Determines the tolerance of each data point for abnormal conditions, based on the upper and lower lengths of the current data point being 2V E The abnormal range is used to determine whether the next data point is abnormal. For example, at 3 seconds on the horizontal axis, if the fourth data point is outside the abnormal range of the previous point, it is determined to be a data anomaly and deleted. Then, based on the linear relationship between the two data points before it, linear interpolation is used to reconstruct the data point. The reconstructed data point is the reconstructed point.

[0061] After obtaining the pre-processed data segments using the revolving door algorithm (SDA) and the anomaly algorithm, some data segments with the same trend or continuous fluctuations still exist. For these data segments, the parameter and resolution adaptive algorithm mainly performs data segment merging and processing of slightly jittered data segments. Step S2.2 of this embodiment also includes using data segment merging to process the two types of wind turbine detection signals separately, including detecting the ramp-up segments with increased data points based on the changes in the data points, and performing ramp segment merging according to the following formula:

[0062]

[0063] In the above formula, l is the current time; n is the total time; x l+1 is the amplitude of the vibration signal at time l+1; x l is the amplitude of the vibration signal at time l; b(l) is the reverse trend data segment between two identical trends whose amplitudes are both less than the preset threshold, which is called a bump event. The data segment merging stage mainly processes the redundant points that still exist after the revolving door algorithm is processed, and merges these points. Figure 5 As shown in the figure, the data points with abscissas between 0s and 3.5s represent a continuous uphill segment. During the continuous backward detection process, a downhill segment with abscissas between 3.5s and 7s is found. The data point at 3.5s is set as the end point of the previous uphill segment and may also be the starting point of the next downhill segment. Similarly, the data point at 7s is also set as the end point of the previous segment and may be the starting point of the next segment.

[0064] In the above merging process, some unimportant slope events are still retained, that is, the rise and fall changes continue to occur within a time period, and the amplitude of the rise and fall trend segments are not large. Such data segments are defined as small-amplitude jitter data segments. Therefore, step S2.2 of this embodiment also includes using small-amplitude jitter data segment detection merging to process the two fan detection signals respectively, including: finding data segments in the fan detection signals that continue to rise and fall within a time period, and the amplitude of the rise and fall trend segments are both less than a preset threshold, as small-amplitude jitter data segments. Its function expression is:

[0065]

[0066] In the above formula, are respectively small jitter data segments The 1st to cth monotonic data segments, and there is any dth monotonic data segment The function expression is:

[0067]

[0068] In the above formula, and are the dth monotonic data segments respectively The starting and ending points of each small jitter data segment The two adjacent monotonic data segments are traversed in sequence, and the equivalent revolving door starting point and door width parameters of the small-amplitude jitter data segment are calculated according to the following formula:

[0069]

[0070]

[0071] In the above formula, X c t c The signal amplitude at time t; For a specific point The amplitude of X g t g The signal amplitude at time t; The revolving door point of the revolving door algorithm SDA The signal amplitude of t c is the current moment; t g It is the gate time; For a specific point moment; The revolving door point of the revolving door algorithm SDA moment; The lower support point of the revolving door algorithm SDA The amplitude of The upper support point of the revolving door algorithm SDA moment; The lower support point of the revolving door algorithm SDA moment; The upper support point of the revolving door algorithm SDA The amplitude of ε' is the gate width parameter of the small-amplitude jitter data segment X0 is the amplitude of the starting point signal; the specific point Set as the gate point of the revolving door algorithm SDA

[0072] The signal amplitudes V of the current two monotonic data segments and the upper revolving door signal amplitude V are obtained based on the equivalent revolving door starting point and door width parameters detected based on the small-amplitude jitter data segment. h And the lower revolving door signal amplitude V l , determine the detection condition V of the revolving door algorithm h <V<V l If it is true, the two monotonic data segments obtained by traversal will be merged.

[0073] Updating the door width of the revolving door algorithm SDA in step S2.3 refers to adding the preset door width accuracy to the door width of the revolving door algorithm SDA. For example, in this embodiment, the door width accuracy is 0.0001, then the expression for updating the door width of the revolving door algorithm SDA is: ε′=ε+0.0001, in the above formula, ε′ is the door width of the revolving door algorithm SDA after update, and ε is the door width of the revolving door algorithm SDA before update. Then, the updated door width ε′ of the revolving door algorithm SDA can be used as the new door width ε of the revolving door algorithm SDA to jump to step S2.2 to continue iteratively executing the revolving door algorithm SDA.

[0074] The amplitude and frequency of the signal change with the working state of the wind turbine. Figure 6 are the time domain and frequency domain digital twin signals in the steady state (a) and transient state (b) of this embodiment. Figure 6As shown, the signal characteristic frequency in the steady state (a) is very clear. When the working conditions change steadily, the spectrum in the transient state (b) shows slight deviation and aliasing compared with the steady state (a). When the working conditions suddenly change, the change in the signal characteristic frequency is more obvious. At the same time, the signal mutation may lead to misdiagnosis of normal wind turbines, reducing the accuracy of the judgment on the robustness of the digital twin system. In order to solve the above technical problems, the method of this embodiment includes using step S2 to preprocess the two wind turbine detection signals respectively through the revolving door algorithm SDA so that the characteristic frequency amplitude error of the preprocessed wind turbine detection signal is less than the preset threshold; and using step S3 to update the receptive field of each corresponding convolutional neural network based on the preprocessed wind turbine detection signal to determine the number of convolutional layers of the convolutional neural network, and output the difference in the number of convolutional layers of the convolutional neural network of the two wind turbine detection signals as the robustness detection result of the digital twin, so as to improve the accuracy of the judgment on the robustness of the digital twin system.

[0075] Figure 7 Schematic diagram of the original vibration signal data collected in this embodiment. Figure 7 As shown, in this embodiment, a vibration signal of 10,000 data points is selected to explore the compression of the data volume before and after processing using different signal frequency domain characteristic frequency amplitude errors. Figure 8 This is a schematic diagram of data volume comparison in this embodiment. Figure 8 As shown in the figure, when the characteristic frequency amplitude error begins to increase, the data volume begins to drop sharply, and when the error reaches 0.3%, the data volume does not change significantly. However, since the signal frequency domain is composed of multiple characteristic frequencies, and their amplitudes are different, when the gate width is greater than the corresponding value, the characteristic frequency of the corresponding amplitude will be processed. Figure 8 It can be seen that as the error increases, the data volume shows a "platform phenomenon". Figure 9 The comparison of time domain signals after amplitude error preprocessing of different characteristic frequencies in the embodiment of the present invention is shown in Figure 1, where (a) is the time domain signal after preprocessing of the original signal, (b) is the time domain signal after preprocessing of the signal with an amplitude error of 0.3%, (c) is the time domain signal after preprocessing of the signal with an amplitude error of 2.2%, and (d) is the time domain signal after preprocessing of the signal with an amplitude error of 4.0%. Figure 9 It can be seen that with the increase of the characteristic frequency amplitude error, the vibration signal gradually becomes sparse. While reducing the noise, the fault information will continue to be lost. Therefore, considering the reduction of data volume and the retention of fault characteristics, this paper selects the frequency domain characteristic frequency amplitude error of 0.3% as the threshold for adaptive selection of the gate width. Step S2 pre-processes the two fan detection signals respectively through the rotating door algorithm SDA. The goal is to make the characteristic frequency amplitude error of the pre-processed fan detection signal e max Less than 0.3%;

[0076] In step S3, the number of convolutional layers is calculated for both the digital twin signal and the original model that meet the conditions. In step S3 of this embodiment, the receptive fields of the corresponding convolutional neural networks are updated based on the preprocessed wind turbine detection signals to determine the number of convolutional layers of the convolutional neural network. The function expression is:

[0077]

[0078] In the above formula, n k-PRAA is the data volume of the pre-processed fan detection signal, τ is the correlation constant of different fault characteristics; n k is the amount of original signal data under different faults; ε is the gate width of the revolving door algorithm; r l is the receptive field size of the lth convolutional layer; r l-1 is the receptive field size of the l-1th convolutional layer; f l is the convolution kernel size of the lth convolution layer; s i is the step size of the i-th convolutional layer; the convolutional neural network (CNN) is composed of multiple convolutional layers and pooling layers.

[0079] In order to use convolutional neural network for recognition, it is necessary to convert the pre-processed fan detection signal into an image. For example, after the pre-processed fan detection signal is converted into a two-dimensional image using GASF (a type of Gram Angular Field GAF) or SDP (Symmetric Point Plot) method, the image size is n×n, where n=n k-PRAA / 2. The two-dimensional image is then input into the convolutional neural network, and the receptive fields of the corresponding convolutional neural networks are updated based on the preprocessed fan detection signals to determine the number of convolutional layers in the convolutional neural network. The matching relationship between the preprocessed fan detection signals and the diagnostic perception of the convolutional neural network is obtained by combining the above-mentioned convolution kernel receptive field calculation formula:

[0080]

[0081] In the above formula, the left side is the size of the two-dimensional image obtained based on the preprocessed fan detection signal; the right side is the number of layers and convolution kernel size corresponding to the receptive field range of the convolutional neural network.

[0082] To verify the effectiveness of the wind power digital twin robustness detection method based on time-frequency revolving door perceptual matching, this embodiment uses Case Western Reserve bearing data to verify the above method. The frequency domain data of the Case Western Reserve bearing data is preprocessed using the revolving door algorithm (SDA) and exception handling methods, and then converted into a GASF two-dimensional image to facilitate fault identification by the neural network model. The convolutional neural network in this embodiment uses only a 3×3 convolution kernel, and both the padding and stride are set to 1. This embodiment is mainly used to verify that the missing number of convolutional layers is related to recognition ability. The structure of the neural network model is shown in Table 1.

[0083] Table 1: Neural network model results

[0084] Number of layers content 1 3x3-8 convolution kernel 2 3x3-8 convolution kernel Pooling layer 2x2 stride 2 pooling layer 3 3x3-16 convolution kernel 4 3x3-16 convolution kernel Pooling layer 2x2 stride 2 pooling layer 5 3x3-32 convolution kernel 6 3x3-32 convolution kernel 7 3x3-32 convolution kernel 8 3x3-32 convolution kernel Pooling layer 2x2 stride 2 pooling layer 9 3x3-32 convolution kernel 10 3x3-32 convolution kernel 11 3x3-32 convolution kernel 12 3x3-32 convolution kernel Pooling layer 2x2 stride 2 pooling layer 13 3x3-64 convolution kernel 14 3x3-64 convolution kernel 15 3x3-64 convolution kernel 16 3x3-64 convolution kernel Pooling layer 2x2 stride 2 pooling layer 17 3x3-64 convolution kernel 18 3x3-64 convolution kernel 19 3x3-64 convolution kernel 20 3x3-64 convolution kernel Pooling layer 2x2 stride 2 pooling layer Fully connected layer 32-node fully connected layer Fully connected layer 8-node fully connected layer

[0085] After the Case Western Reserve bearing data passes through the revolving door algorithm, anomaly algorithm, data segment merging algorithm, and small jitter detection algorithm, the loop termination condition e is set. max <0.3%, the diagnostic rate can be obtained as shown in Table 2 and Figure 10 shown.

[0086] Table 2: e max Case Western Reserve bearing data diagnostic rate <0.3%

[0087] Number of convolutional layers 3 6 7 9 13 16 Diagnosis rate 92.0％ 96.7％ 97.0％ 98.5％ 98.0％ 96.7％

[0088] See Table 2 and Figure 10 It can be seen that the diagnostic rate of the same data in the wind power digital twin robustness detection method based on time-frequency revolving door perception matching in this embodiment is different under different numbers of convolutional layers, and the diagnostic rate is the highest at 9 layers.

[0089] In summary, the method of this embodiment realizes the robustness judgment of the wind turbine digital twin system through time-frequency preprocessing of the wind turbine signal, spectrum feature extraction and adaptive updating of the convolutional neural network. Specifically, it includes obtaining the wind turbine detection signal generated by the digital twin and the wind turbine detection signal collected by the sensor, and preprocessing the two wind turbine detection signals respectively through the revolving door algorithm SDA so that the characteristic frequency amplitude error of the preprocessed wind turbine detection signal is less than the preset threshold; based on the preprocessed wind turbine detection signal, the receptive field of each corresponding convolutional neural network is updated to determine the number of convolutional layers of the convolutional neural network, and the difference in the number of convolutional layers of the convolutional neural network of the two wind turbine detection signals is used as the robustness detection result of the digital twin. The above method can solve the problem of limited data processing capability of the existing digital twin system evaluation index algorithm, and improve the data processing capability to cope with the problem that the wind power digital twin system usually adopts a high sampling rate when collecting signals to ensure the accuracy of the detection and verification system, resulting in a large amount of data. The method of this embodiment updates the receptive field of each corresponding convolutional neural network based on the preprocessed wind turbine detection signal to determine the number of convolutional layers of the convolutional neural network, and outputs the difference in the number of convolutional layers of the convolutional neural network of the two wind turbine detection signals as the robustness detection result of the digital twin. It can solve the misdiagnosis caused by spectral deviation and aliasing when the signal changes steadily and signal mutation, and improve the accuracy of robustness judgment.

[0090] In addition, the embodiment further provides a wind power digital twin robustness detection system based on time-frequency rotating door perception matching, comprising a microprocessor and a memory connected with each other, the microprocessor being programmed or configured to execute the wind power digital twin robustness detection method based on time-frequency rotating door perception matching.

[0091] In addition, the embodiment further provides a computer readable storage medium, wherein a computer program or instructions are stored, the computer program or instructions being programmed or configured to execute the wind power digital twin robustness detection method based on time-frequency rotating door perception matching by a processor.

[0092] In addition, the embodiment further provides a computer program product, comprising a computer program or instructions, the computer program or instructions being programmed or configured to execute the wind power digital twin robustness detection method based on time-frequency rotating door perception matching by a processor.

[0093] Those skilled in the art should understand that the technical solutions provided by the embodiments of the present application can be in the form of a method, a system, or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer readable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer usable program code. The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The functions specified in one or more flows and / or blocks These computer program instructions can also be stored in a computer readable memory that can cause the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0094] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A wind power digital twin robustness detection method based on time-frequency revolving door perception matching, characterized by include: Step S1: Acquire a fan detection signal generated by the digital twin and a fan detection signal collected by a sensor, wherein the fan detection signal is the temperature, radial vibration, axial vibration, current, voltage, or power of the fan; Step S2, preprocessing the two fan detection signals respectively by using a revolving door algorithm SDA so that the characteristic frequency amplitude error of the preprocessed fan detection signal is less than a preset threshold; Step S3: Based on the preprocessed wind turbine detection signals, the receptive fields of the corresponding convolutional neural networks are updated to determine the number of convolutional layers of the convolutional neural network. The difference in the number of convolutional layers of the convolutional neural network between the two wind turbine detection signals is output as the robustness test result of the digital twin. Step S2 includes: Step S2.1, setting the initial door width of the revolving door algorithm SDA; Step S2.2, using the revolving door algorithm SDA to pre-process the two types of fan detection signals respectively; Step S2.3, converting the two fan detection signals into frequency domain signals through fast Fourier transform and extracting the characteristic frequency amplitude of the original fan detection signal, converting the fan detection signal preprocessed in step S2.2 into frequency domain signals through fast Fourier transform and extracting the characteristic frequency amplitude of the preprocessed fan detection signal, respectively calculating the characteristic frequency amplitude errors of the two preprocessed fan detection signals based on the characteristic frequency amplitudes of the preprocessed fan detection signal and the characteristic frequency amplitudes of the original fan detection signal, if both characteristic frequency amplitude errors are less than a preset threshold, skipping to step S3; otherwise, updating the door width of the revolving door algorithm SDA and skipping to step S2.2; In step S3, based on the preprocessed fan detection signal, the receptive fields of the corresponding convolutional neural networks are updated respectively to determine the number of convolutional layers of the convolutional neural network. The function expression is: ， ， In the above formula, is the data volume of the pre-processed fan detection signal, is the correlation constant of different fault characteristics; is the amount of original signal data under different faults; is the door width of the revolving door algorithm; For the The receptive field size of the convolutional layer; For the The receptive field size of the convolutional layer; For the The convolution kernel size of the convolution layer; For the The convolutional neural network is composed of multiple convolutional layers and pooling layers.

2. The wind power digital twin robustness detection method based on time-frequency revolving door perception matching according to claim 1 is characterized in that: The calculation function expression of the characteristic frequency amplitude error in step S2 is: ， In the above formula, is the characteristic frequency amplitude error, is the characteristic frequency amplitude of the fan detection signal after preprocessing, is the characteristic frequency amplitude of the original fan detection signal.

3. The wind power digital twin robustness detection method based on time-frequency revolving door perception matching according to claim 1 is characterized in that: After using the revolving door algorithm SDA to process the two fan detection signals respectively in step S2.2, the two fan detection signals are also processed respectively using the abnormal algorithm, including for each current data point, based on the given abnormal parameters The length of the upper and lower parts of the current data point is The abnormal range of the next data point is used to determine whether an abnormality occurs. If the next data point exceeds the abnormal range, the next data point is deleted and the data point is reconstructed using linear interpolation based on the first two data points of the deleted data point.

4. The wind power digital twin robustness detection method based on time-frequency revolving door perception matching according to claim 1 is characterized in that: Step S2.2 also includes using data segment merging to process the two wind turbine detection signals respectively, including detecting the climbing segment with increased data points according to the change of data points, and performing climbing segment merging according to the following formula: ， In the above formula, For the current moment; is the total time; for The amplitude of the vibration signal at each moment; for The amplitude of the vibration signal at each moment; It is a reverse trend data segment where the amplitude between two identical trends is less than the preset threshold.

5. The wind power digital twin robustness detection method based on time-frequency revolving door perception matching according to claim 1 is characterized in that: Step S2.2 also includes using small amplitude jitter data segment detection and merging to process the two fan detection signals respectively, including: finding the data segment in the fan detection signal that continuously rises and falls within a time period, and the amplitude of the rising and falling trend segments is less than the preset threshold as the small amplitude jitter data segment; , its function expression is: ， In the above formula, ~ are respectively small jitter data segments The 1st to cth monotonic data segments, and there is any dth monotonic data segment The function expression is: ， In the above formula, and Respectively Monotonic data segments The starting and ending points of each small jitter data segment The two adjacent monotonic data segments are traversed in sequence, and the equivalent revolving door starting point and door width parameters of the small-amplitude jitter data segment are calculated according to the following formula: ， ， ， In the above formula, for The signal amplitude at time t; For a specific point The amplitude of for The signal amplitude at time ; The revolving door point of the revolving door algorithm SDA The signal amplitude of For the current moment; It is the gate time; For a specific point moment; The revolving door point of the revolving door algorithm SDA moment; The lower support point of the revolving door algorithm SDA The amplitude of The upper support point of the revolving door algorithm SDA moment; The lower support point of the revolving door algorithm SDA moment; The upper support point of the revolving door algorithm SDA The amplitude of The gate width parameter for the small jitter data segment is the signal amplitude at the starting point; Set as the gate point of the revolving door algorithm SDA ; The signal amplitudes of the current two monotonic data segments are obtained based on the equivalent revolving door starting point and door width parameters detected based on the small-amplitude jitter data segment , upper revolving door signal amplitude and the lower revolving door signal amplitude , determine the detection conditions of the revolving door algorithm If it is true, the two monotonic data segments obtained by traversal will be merged.

6. A wind power digital twin robustness detection system based on time-frequency revolving door perception matching, comprising an interconnected microprocessor and a memory, characterized in that: The microprocessor is programmed or configured to execute the wind power digital twin robustness detection method based on time-frequency revolving door perception matching as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program or instruction stored therein, characterized in that: The computer program or instruction is programmed or configured to execute the wind power digital twin robustness detection method based on time-frequency revolving door perception matching as described in any one of claims 1 to 5 through a processor.

8. A computer program product comprising a computer program or instructions, characterized in that The computer program or instruction is programmed or configured to execute the wind power digital twin robustness detection method based on time-frequency revolving door perception matching as described in any one of claims 1 to 5 through a processor.

Images