Non-intrusive servo motor driver fault prediction method and device

Abstract

The invention discloses a non-intrusive servo motor driver fault prediction method and a non-intrusive servo motor driver fault prediction device, belongs to the field of power equipment prediction, and aims to solve the problem of fault prediction of a servo motor driver. Obtaining a two-dimensional image of the voltage signal; performing bidirectional empirical mode decomposition on the two-dimensional image according to BEMD to obtain a plurality of intrinsic mode function components and residual terms; on the basis of a pre-trained servo motor driver fault prediction model, prediction images of the modal components and the residual terms are obtained for the multiple intrinsic mode function components and the residual terms; performing reconstruction and error calculation on the predicted image to obtain an error evaluation index corresponding to the reconstructed image; and finally, judging the error evaluation index according to a preset fault threshold to determine a prediction result of the fault type of the servo motor driver. The method improves the spatial-temporal feature fusion capability, and enables the prediction result to be high in precision and to be real-time.

Description

Technical Field

[0001] The present invention relates to a non-invasive servo motor drive fault prediction method, belonging to the technical field of power equipment prediction. Background Art

[0002] With the development of power electronic devices towards high power density and high frequency, the health status monitoring of servo motor drives faces severe challenges. Traditional fault prediction methods mainly rely on the analysis of one-dimensional time series signals such as output voltage, and there are the following technical bottlenecks:

[0003] (1) Insufficient identification of one-dimensional signal features: Existing monitoring systems mostly use oscilloscope waveform display or fast Fourier transform. Although they can reflect basic parameters such as voltage amplitude and frequency, it is difficult to effectively capture multi-dimensional coupling features in the dynamic process. For example, transient anomalies such as voltage spikes and harmonic distortions are easily masked by noise in the time-domain waveform, and although spectral analysis can separate frequency-domain components, the time-domain dynamic correlation is lost. In addition, the long-term operation trend analysis requires manual comparison of historical waveforms, lacking automated feature extraction means, resulting in delayed fault warnings.

[0004] (2) Limitations of two-dimensional image processing methods: To improve the visualization effect, some studies have tried to convert time series signals into two-dimensional images. Recurrence plots and wavelet transforms are commonly used techniques, but they have significant defects: recurrence plots have insufficient feature contrast due to a single color gradient, and wavelet transforms lose edge information due to fixed basis functions. Although the Markov transfer field can retain some time series correlations, artifacts and noise are introduced during the conversion process.

[0005] (3) Existing prediction models have problems with insufficient spatio-temporal feature fusion ability: Current prediction models can be divided into two categories: one-dimensional models can handle time series dependencies but ignore spatial features; two-dimensional CNN models can extract local image features but are difficult to model long-term degradation trends.

[0006] The above defects seriously restrict the accuracy and real-time performance of servo motor drive fault prediction, and the application of servo motor drives in new energy and power systems is very critical. Their failures may lead to system instability or shutdown. Therefore, there is an urgent need for a highly reliable prediction scheme that can fuse spatio-temporal features and take into account computational efficiency to ensure meeting the requirements of practical applications. Summary of the Invention

[0007] The object of the present invention is to provide a non-invasive fault prediction method for a servo motor drive to solve the problems in the background art. A non-invasive sampling method is used to collect one-dimensional voltage signals and convert them into two-dimensional images based on the Gramian angular field. The two-dimensional empirical mode decomposition (BEMD) is combined to solve the redundant noise problem generated by the two-dimensional images. Then, the convolutional long short-term memory network (ConvLSTM) is used to predict each decomposed component separately, and the prediction results are reconstructed and the errors are calculated to further improve the spatio-temporal feature fusion ability, so that the prediction results are accurate and real-time, and the stable operation of the equipment to which the servo motor drive belongs can be better ensured.

[0008] To solve the above technical problems, the present invention is implemented by the following technical solutions.

[0009] In a first aspect, the present invention provides a non-invasive fault prediction method for a servo motor drive, including:

[0010] Obtain the voltage signal output by the servo motor drive;

[0011] Perform polar coordinate mapping on the voltage signal using the Gramian angular field function to obtain a two-dimensional image of the voltage signal;

[0012] Perform bidirectional empirical mode decomposition on the two-dimensional image of the voltage signal according to the two-dimensional empirical mode decomposition (BEMD) to obtain multiple intrinsic mode function components and a residual term of the voltage signal;

[0013] Based on a pre-trained fault prediction model of the servo motor drive, predict the multiple intrinsic mode function components and the residual term to obtain a prediction image for each intrinsic mode function component and the residual term;

[0014] Reconstruct and calculate the error of the prediction image for each intrinsic mode function component and the residual term to obtain an error evaluation index corresponding to the reconstructed image;

[0015] According to the comparison result between a preset fault threshold and the error evaluation index corresponding to the reconstructed image, obtain the prediction result of the fault type of the servo motor drive.

[0016] Optionally, obtaining the voltage signal output by the servo motor drive includes:

[0017] Sample the voltage signal of the servo motor drive in a non-invasive manner at a fixed sampling frequency and interval to obtain sample data of the voltage;

[0018] Normalize the sample data of the voltage to obtain normalized time-series data;

[0019] Among them, the sample data is one-dimensional time-series data, the sampling frequency is set to 5 MHz, the sampling interval is set to 12 hours, and the length of the sample data is 2500 points.

[0020] Normalize the sample data of the voltage. The expression for the standardized time-series data is as follows:

[0021] (1)

[0022] In the formula, is the standardized time-series data, is the maximum value in the sample data, is the minimum value in the sample data, is the sample data.

[0023] Optionally, use the Gram angular field function to perform polar coordinate mapping on the voltage signal to obtain a two-dimensional image of the voltage signal, including:

[0024] Convert the standardized time-series data from the Cartesian coordinate system to the polar coordinate system to obtain the polar coordinates of all data points in the time-series data;

[0025] Use trigonometric functions to perform triangular difference calculation on the polar coordinates of all data points to obtain the angle difference between any two data points;

[0026] Construct a Gram angular field angle matrix based on the angle difference between any two data points;

[0027] Perform pixel gray-scale conversion on the elements of the Gram angular field angle matrix to obtain a two-dimensional image of the voltage signal;

[0028] Among them, the expression for obtaining the polar coordinates of all data points in the time-series data is:

[0029] (2)

[0030] In the formula, is the polar angle corresponding to the i-th time point, is the radius in the polar coordinates, is the timestamp, represents the number of time points in the time-series data, is the arccosine result of the standardized time-series data;

[0031] The expression for constructing a Gram angular field angle matrix based on the angle difference between any two data points is:

[0032] G = [ s i n ( φ 1 − φ 1 ) s i n ( φ 1 − φ 2 ) . . . s i n ( φ 1 − φ n ) s i n ( φ 2 − φ 1 ) s i n ( φ 2 − φ 2 ) . . . s i n ( φ 2 − φ n ) . . . . . . . . . . . . s i n ( φ n − φ 1 ) s i n ( φ n − φ 2 ) . . . s i n ( φ n − φ n ) ] (3)

[0033] In the formula, is the included angle value from the 1st to the n-th time point.

[0034] Optionally, perform bidirectional empirical mode decomposition on the two-dimensional image of the voltage signal according to bidirectional empirical mode decomposition (BEMD) to obtain multiple intrinsic mode function components and a residual term of the voltage signal, including:

[0035] Use the sliding window technique to extract extreme values from the two-dimensional image. After the sliding window traverses all regions of the two-dimensional image, all local maximum points and minimum points extracted by the window are obtained;

[0036] Perform surface fitting on all local maximum points and minimum points respectively to obtain the initial estimates of the upper and lower envelope surfaces;

[0037] Perform interpolation calculation on the initial estimates of the upper and lower envelope surfaces to obtain the upper envelope surface and the lower envelope surface of the two-dimensional image;

[0038] Calculate the average value of the upper envelope surface and the lower envelope surface of the two-dimensional image to obtain the average envelope surface;

[0039] Separate the two-dimensional image according to the average envelope surface to obtain the result image;

[0040] Use the local mean detrending method to perform iterative screening on the result image to obtain the updated multiple intrinsic mode function components and the residual term of the voltage signal;

[0041] Among them, the expression for using the sliding window technique to extract extreme values from the two-dimensional image. After the sliding window traverses all regions of the two-dimensional image, all local maximum points and minimum points extracted by the window are:

[0042] (4)

[0043] In the formula, N(x,y) represents the set of numbers of 4*4 rectangular windows centered on (x,y), I max (x,y) and I min (x,y) represent the local maximum point and the local minimum point;

[0044] The expression for calculating the average value of the upper envelope surface and the lower envelope surface of the two-dimensional image to obtain the average envelope surface is:

[0045] M 1 ( x , y ) = [ U ( x , y ) + L ( x , y ) ] 2 (5)

[0046] In the formula, is the upper envelope surface of the image, is the lower envelope surface of the upper envelope surface of the image; is the average envelope surface;

[0047] The expression for separating the two-dimensional image according to the average envelope surface to obtain the result image is:

[0048] (6)

[0049] Wherein, is the result image;

[0050] The expression for iterative screening of the result image is:

[0051] (7)

[0052] Wherein, SD is the iteration termination condition, is the temporary result image obtained in the k-th iteration.

[0053] Optionally, the servo motor drive fault prediction model is constructed by a CNN feature extraction module and a ConvLSTM spatio-temporal fusion module;

[0054] Among them, the gating mechanism and state update of the ConvLSTM unit are both through convolution operations. The gating mechanism includes an input gate, a forget gate, and an output gate; the state update includes a cell state and a hidden state;

[0055] The input gate is used to control the amount of voltage signal sample data entering the cell state;

[0056] The forget gate is used to control the discard ratio of historical degradation features in the cell state;

[0057] The output gate is used to control the output intensity of transient harmonic components in the cell state;

[0058] The cell state is used to control the forgetting of historical degradation features through the forget gate and the input update of the amount of voltage signal sample data through the input gate;

[0059] The hidden state is used to process the next data in the spatio-temporal sequence data;

[0060] Optionally, the expression of the input gate is:

[0061] (8)

[0062] Wherein, is the input gate, is the sigmoid activation function, is the convolutional kernel weight matrix of the input data of the input gate, is the convolutional kernel weight matrix of the hidden state data of the input gate, is the input at the current moment, is the hidden state at the previous moment, is the bias term of the input gate, * represents the convolution operation, is the Hadamard product operation;

[0063] The expression of the forget gate is:

[0064] (9)

[0065] In the formula, is the forget gate, is the convolution kernel weight matrix of the input data of the forget gate, is the convolution kernel weight matrix of the hidden state data of the forget gate, is the bias term of the forget gate;

[0066] The expression of the output gate is:

[0067] (10)

[0068] In the formula, is the output gate, is the convolution kernel weight matrix of the input data of the output gate, is the convolution kernel weight matrix of the hidden state data of the output gate, is the bias term of the output gate;

[0069] The expression of the cell state is:

[0070] (11)

[0071] In the formula, is the cell state, is the convolution kernel weight matrix of the input data of the cell state, is the convolution kernel weight matrix of the hidden state of the cell state, is the bias term of the output gate;

[0072] The expression of the hidden state is:

[0073] (12)

[0074] In the formula, is the hidden state at the current moment.

[0075] Optionally, the training method of the servo motor drive fault prediction model includes:

[0076] Obtain the historical data set of the voltage signal;

[0077] Perform data partitioning on the historical data set of the voltage signal to obtain a training set and a test set;

[0078] Use the Gramian angular field method to perform image conversion on the one-dimensional voltage time series signal in the training set to obtain a two-dimensional image of the historical voltage signal;

[0079] Perform bidirectional empirical mode decomposition on the two-dimensional image of the historical voltage signal according to the BEMD decomposition method to obtain multiple intrinsic mode function components and a residual term of the historical voltage signal;

[0080] Input the multiple intrinsic mode function components into the prediction model based on the convolutional long short-term memory network for feature extraction and training to obtain the model parameters required for fault prediction;

[0081] Update the prediction model based on the convolutional long short-term memory network according to the model parameters required for fault prediction to obtain an updated fault prediction model for the servo motor drive; among them, each intrinsic mode function component and residual term obtained by BEMD decomposition are used as training samples and input into the basic prediction model constructed based on the convolutional long short-term memory network (ConvLSTM) for spatial feature extraction and temporal feature modeling processing; the Adam gradient descent algorithm is used to optimize the parameters of the training model, and iterative updates are performed according to the set number of training times and update strategy during the training process until the stop condition is met, then stop training, save all the current training parameters, and obtain a relatively stable updated fault prediction model for the servo motor drive;

[0082] Input the test set into the updated fault prediction model of the servo motor drive for prediction to obtain the prediction results;

[0083] Substitute the prediction results and the true results of the test set into the calculation of the root mean square error (RMSE) and the mean absolute percentage error (MAPE) respectively to obtain the root mean square error value and the mean absolute percentage error value;

[0084] Adopt an evaluation strategy to quantitatively evaluate the root mean square error value and the mean absolute percentage error value to determine the servo motor drive fault prediction model with the optimal performance.

[0085] Optionally, perform reconstruction and error calculation on the predicted values of each intrinsic mode function component and residual term to obtain the error evaluation index corresponding to the reconstructed image, including:

[0086] Perform weighted superposition on all the predicted mode function components and residual terms according to the composition rule of the original image to obtain a reconstructed fault prediction image;

[0087] Perform error evaluation on the reconstructed fault prediction image and the actual observed image to obtain the error evaluation index corresponding to the reconstructed image;

[0088] The error metrics include: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Pearson Correlation Coefficient (PCC).

[0089] Optionally, based on the comparison result between the preset fault threshold and the error evaluation metric corresponding to the reconstructed image, a prediction result of the servo motor drive fault type is obtained, including:

[0090] By comparing the error evaluation metric corresponding to the reconstructed image with a preset fault identification threshold or performing trend analysis on consecutive multiple all modal function components according to a set high-error trend, a prediction result of the servo motor drive fault type is obtained;

[0091] When the error metric exceeds the upper threshold under the corresponding working condition or consecutive multiple modal function components show a high-error trend, the servo motor drive fault is identified and relevant error metrics and fault characteristics corresponding to the relevant modal function components are extracted;

[0092] Based on the matching result between the predefined fault label database and the fault characteristics, a prediction result of the servo motor drive fault type is obtained.

[0093] In a second aspect, the present invention provides a non-invasive servo motor drive fault identification device, including: a data acquisition and conversion module, a two-dimensional modal decomposition module, a prediction module, a reconstruction and error calculation module, and an identification module;

[0094] The data acquisition module is used to acquire the voltage signal output by the servo motor drive and perform polar coordinate mapping on the voltage signal using the Gramian angular field function to obtain a two-dimensional image of the voltage signal;

[0095] The two-dimensional modal decomposition module is used to perform bidirectional empirical mode decomposition on the two-dimensional image of the voltage signal according to the two-dimensional empirical mode decomposition BEMD to obtain multiple intrinsic mode function components and a residual term of the voltage signal;

[0096] The prediction module is used to predict multiple intrinsic mode function components and the residual term based on a pre-trained servo motor drive fault prediction model to obtain a prediction image of each intrinsic mode function component and the residual term;

[0097] The reconstruction and error calculation module is used to reconstruct and calculate the error of the prediction image of each intrinsic mode function component and the residual term to obtain an error evaluation metric corresponding to the reconstructed image;

[0098] The recognition module is used to obtain the prediction result of the servo motor driver fault type according to the comparison result between the preset fault threshold and the error evaluation index corresponding to the reconstructed image.

[0099] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0100] 1. A non-invasive sampling method is adopted to collect one-dimensional voltage signals and convert them into two-dimensional images based on the Gram angle field. The two-dimensional empirical mode decomposition (BEMD) is combined to solve the redundant noise problem generated by the two-dimensional images. Then, the convolutional long short-term memory network (ConvLSTM) is used to predict each decomposed component separately, and the prediction results are reconstructed and the errors are calculated, further improving the spatio-temporal feature fusion ability, making the prediction results accurate and real-time, so as to better ensure the stable operation of the equipment to which the servo motor driver belongs.

[0101] 2. In the servo motor driver fault prediction, by combining the CNN feature extraction module and the ConvLSTM spatio-temporal fusion module, the one-dimensional voltage signal is first converted, and then the ConvLSTM is used to directly process the time-series data with spatial structure. At the same time, the gating mechanism of the ConvLSTM can selectively retain the key information in the CNN features, and has a flexible dynamic feature construction ability, so as to realize the dynamic adjustment of the weights of the CNN and the ConvLSTM under various working conditions and have strong robustness.

[0102] 3. A cross-modal fusion diagnosis system is developed, which combines multiple error indexes, high-error mode trend analysis, and multi-level and multi-index judgments on the servo motor driver fault based on the existing fault label database, realizing early detection and accurate classification of the servo motor driver fault, and systematically improving the intelligent level of the servo motor driver fault diagnosis, providing core guarantee for high-reliability power electronic systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 The figure shows the flow chart of the non-invasive servo motor driver fault prediction method of the present invention;

[0104] Figure 2 The figure shows an embodiment diagram of the polar coordinate mapping of the voltage signal of the present invention;

[0105] Figure 3 The figure shows the flow chart of the bidirectional empirical mode decomposition of the two-dimensional image of the voltage signal of the present invention;

[0106] Figure 4 The figure shows a prediction result diagram of the servo motor driver fault prediction model of the present invention;

[0107] Figure 5The following figure shows the second prediction result diagram of the servo motor driver fault prediction model of the present invention;

[0108] Figure 6 The following figure shows the third prediction result diagram of the servo motor driver fault prediction model of the present invention;

[0109] Figure 7 The following figure shows the fourth prediction result diagram of the servo motor driver fault prediction model of the present invention;

[0110] Figure 8 It is a prediction evaluation index table based on prediction algorithms of different dimensions;

[0111] Figure 9 It is an intuitive diagram of prediction evaluation indexes based on prediction algorithms of different dimensions. Specific implementation manners

[0112] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0113] Embodiment 1

[0114] This embodiment provides a non-invasive servo motor driver fault prediction method, as Figure 1 shown including:

[0115] Step 1: Obtain the voltage signal output by the servo motor driver;

[0116] Step 2: Use the Gram angular field function to perform polar coordinate mapping on the voltage signal to obtain a two-dimensional image of the voltage signal;

[0117] Step 3: Perform bidirectional empirical mode decomposition on the two-dimensional image of the voltage signal according to the two-dimensional empirical mode decomposition BEMD to obtain multiple intrinsic mode function components and a residual term of the voltage signal;

[0118] Step 4: Based on the pre-trained servo motor driver fault prediction model, predict the multiple intrinsic mode function components and the residual term to obtain a prediction image of each intrinsic mode function component and the residual term;

[0119] Step 5: Reconstruct and calculate the error of the prediction image of each intrinsic mode function component and the residual term to obtain an error evaluation index corresponding to the reconstructed image;

[0120] Step 6: According to the comparison result between the preset fault threshold and the error evaluation index corresponding to the reconstructed image, obtain the prediction result of the servo motor driver fault type.

[0121] Optionally, obtaining the voltage signal output by the servo motor driver in step 1 includes:

[0122] Step 1.1: Sample the voltage signal of the servo motor driver in a non-invasive manner at a fixed sampling frequency and interval to obtain the sample data of the voltage.

[0123] Step 1.2: Normalize the sample data of the voltage to obtain the standardized time-series data.

[0124] In this embodiment, the sample data is one-dimensional time-series data, the sampling frequency is set to 5 MHz, the sampling interval is set to 12 hours, and the length of the sample data is 2,500 points.

[0125] The expression for normalizing the sample data of the voltage to obtain the standardized time-series data is:

[0126] (1)

[0127] In the formula, is the standardized time-series data, is the maximum value in the sample data, is the minimum value in the sample data, is the sample data.

[0128] Optionally, in step 2, the Gram angle field function is used to perform polar coordinate mapping on the voltage signal to obtain a two-dimensional image of the voltage signal, as Figure 2 shown including:

[0129] Step 2.1: Convert the standardized time-series data from the Cartesian coordinate system to the polar coordinate system to obtain the polar coordinates of all data points in the time-series data.

[0130] Step 2.2: Use trigonometric functions to perform triangular difference calculation on the polar coordinates of all data points to obtain the angle difference between any two data points.

[0131] Step 2.3: Construct a Gram angle field angle matrix according to the angle difference between any two data points.

[0132] Step 2.4: Perform pixel gray-scale conversion on the elements of the Gram angle field angle matrix to obtain a two-dimensional image of the voltage signal.

[0133] Among them, the expression for obtaining the polar coordinates of all data points in the time-series data is:

[0134] (2)

[0135] In the formula, is the polar angle corresponding to the i-th time point, is the radius in the polar coordinates, is the timestamp, represents the number of time points in the time-series data, The arccosine result of standardized time series data;

[0136] According to the angular difference between any two data points, the expression for constructing the Gram angle field angular matrix is:

[0137] G = [ s i n ( φ 1 − φ 1 ) s i n ( φ 1 − φ 2 ) . . . s i n ( φ 1 − φ n ) s i n ( φ 2 − φ 1 ) s i n ( φ 2 − φ 2 ) . . . s i n ( φ 2 − φ n ) . . . . . . . . . . . . s i n ( φ n − φ 1 ) s i n ( φ n − φ 2 ) . . . s i n ( φ n − φ n ) ] (3)

[0138] In the formula, is the included angle value at the 1st to the nth time points.

[0139] Optionally, in step 3, the two-dimensional empirical mode decomposition BEMD is used to perform two-way empirical mode decomposition on the two-dimensional image of the voltage signal, and multiple intrinsic mode function components and residual terms of the voltage signal are obtained. As Figure 3 shown, it includes:

[0140] Step 3.1: Use the sliding window technique to extract the extreme values of the two-dimensional image. After the sliding window traverses all regions of the two-dimensional image, all local maximum points and minimum points extracted by the window are obtained;

[0141] Step 3.2: Respectively perform surface fitting on all local maximum points and minimum points to obtain the initial estimates of the upper and lower envelope surfaces;

[0142] Step 3.3: Perform interpolation calculation on the initial estimates of the upper and lower envelope surfaces to obtain the upper envelope surface and the lower envelope surface of the two-dimensional image; local maximum points and minimum points are selected through a 4*4 rectangular sliding window

[0143] Step 3.4: Calculate the average value of the upper envelope surface and the lower envelope surface of the two-dimensional image to obtain the average envelope surface;

[0144] Step 3.5: Separate the two-dimensional image according to the average envelope surface to obtain the result image;

[0145] Step 3.6: Use local mean detrending to perform iterative screening on the result image to obtain the updated multiple intrinsic mode function components and residual terms of the voltage signal.

[0146] In this embodiment, it is defined that (1) I(x, y) is a two-dimensional image on the rectangular region {(x, y)|x = 1, 2, …, W; y = 1, 2, …, H}. All local maximum points and minimum points of I(x, y) are selected by the neighborhood comparison method, and local maximum points and minimum points are selected through a 4*4 rectangular sliding window; the specific calculation formula is as follows:

[0147] (4)

[0148] Wherein, N(x,y) represents the set of numbers of 4*4 rectangular windows centered on (x,y), and I max (x,y) and I min (x,y) represent local maximum points and local minimum points.

[0149] (2) Perform surface fitting on all local maximum points and minimum points respectively, and obtain the upper envelope surface of the image through interpolation calculation and the lower envelope surface , calculate the average value according to the upper and lower envelope surfaces of the image to obtain the average envelope surface , and its calculation formula is as follows:

[0150] M 1 ( x , y ) = [ U ( x , y ) + L ( x , y ) ] 2 (5)

[0151] Wherein, is the upper envelope surface of the image, is the lower envelope surface of the upper envelope surface of the image; is the average envelope surface.

[0152] (3) Separate the average envelope surface M1(x,y) from the original two-dimensional image to obtain the result image , and its calculation formula is as follows:

[0153] (6)

[0154] Wherein, is the result image;

[0155] (4) For Repeat the above steps k times to obtain H 1k (x,y), calculate the screening termination condition SD. If H 1k (x,y) satisfies the screening termination condition, the calculation formula of the screening termination condition SD is as follows:

[0156] (7)

[0157] Wherein, SD is the screening termination condition, is The temporary result image obtained in the kth iteration. The value range of SD is the threshold of the preset standard deviation change rate of the residual signal, generally set between 0.2 and 0.3;

[0158] (5) is the first-order intrinsic mode function BIMF component , subtract from the original image to obtain the first-order difference image . The calculation formula of is as follows:

[0159] (9)

[0160] wherein is obtained by subtracting from the original image to obtain the first-order difference image.

[0161] (6) Taking the first-order difference image as the new original signal and repeating steps (1) to (5) can obtain the second-order BIMF component . Repeating the above steps n times until the nth-order BIMF component and its residual are obtained. and are calculated as follows:

[0162] (10)

[0163] wherein is the second-order BIMF component, and are the nth-order BIMF component and its residual term obtained by iterative screening n times; among them, when is a monotonic function, the BEMD decomposition process terminates. The entire original image can be expressed as:

[0164] (11)

[0165] wherein is the ith-order BIMF component, is the upper limit of the number of iterations when the final iteration stops.

[0166] Optionally, the servo motor drive fault prediction model is constructed by a CNN feature extraction module and a ConvLSTM spatio-temporal fusion module;

[0167] wherein, the gating mechanism and state update of the ConvLSTM unit are both through convolution operations. The gating mechanism includes an input gate, a forget gate, and an output gate; the state update includes a cell state and a hidden state;

[0168] The input gate is used to control the amount of voltage signal sample data entering the cell state;

[0169] The forget gate is used to control the discard ratio of historical degraded features in the cell state;

[0170] The output gate is used to control the output intensity of transient harmonic components in the cell state;

[0171] The cell state is used to control the forgetting of historical degraded features through the forget gate and the input update of the voltage signal sample data volume through the input gate;

[0172] The hidden state is used to process the next data in the spatio-temporal sequence data;

[0173] Optionally, the expression of the input gate is:

[0174] (12)

[0175] In the formula, is the input gate, is the sigmoid activation function, is the convolutional kernel weight matrix of the input data of the input gate, is the convolutional kernel weight matrix of the hidden state data of the input gate, is the input at the current moment, is the hidden state at the previous moment, is the bias term of the input gate, * represents the convolution operation, is the Hadamard product operation;

[0176] The expression of the forget gate is:

[0177] (13)

[0178] In the formula, is the forget gate, is the convolutional kernel weight matrix of the input data of the forget gate, is the convolutional kernel weight matrix of the hidden state data of the forget gate, is the bias term of the forget gate;

[0179] The expression of the output gate is:

[0180] (14)

[0181] In the formula, is the output gate, is the convolutional kernel weight matrix of the input data of the output gate, is the convolutional kernel weight matrix of the hidden state data of the output gate, is the bias term of the output gate;

[0182] The expression of the cell state is:

[0183] (15)

[0184] In the formula, is the cell state, is the convolutional kernel weight matrix of the input data of the cell state, is the convolutional kernel weight matrix for the cell state hidden state, is the bias term of the output gate;

[0185] The expression for the hidden state is:

[0186] (16)

[0187] In the formula, is the hidden state at the current moment.

[0188] Optionally, the training method of the servo motor driver fault prediction model includes:

[0189] Obtain the historical data set of voltage signals;

[0190] Perform data partitioning on the historical data set of voltage signals to obtain a training set and a test set;

[0191] Adopt the Gramian Angular Field method to perform image conversion on the one-dimensional voltage time series signal in the training set to obtain the two-dimensional image of the historical voltage signal;

[0192] Perform bidirectional empirical mode decomposition on the two-dimensional image of the historical voltage signal according to the BEMD decomposition method to obtain multiple intrinsic mode function components and a residual term of the historical voltage signal;

[0193] Input multiple intrinsic mode function components into the prediction model based on the convolutional long short-term memory network for feature extraction and training to obtain the model parameters required for fault prediction;

[0194] Update the prediction model based on the convolutional long short-term memory network according to the model parameters required for fault prediction to obtain the updated servo motor driver fault prediction model; among them, each intrinsic mode function component and residual term obtained by BEMD decomposition are used as training samples and input into the basic prediction model constructed based on the convolutional long short-term memory network (ConvLSTM) for spatial feature extraction and temporal feature modeling processing; the Adam gradient descent algorithm is used to optimize the parameters of the training model, and iterative updates are performed according to the set number of training times and update strategy during the training process until the stop condition is met, then stop training, save all the current training parameters, and obtain the relatively stable updated servo motor driver fault prediction model

[0195] Input the test set into the updated servo motor driver fault prediction model for prediction to obtain the prediction result;

[0196] Substitute the prediction result and the real result of the test set into the calculation of the root mean square error (RMSE) and the mean absolute percentage error (MAPE) respectively to obtain the root mean square error value and the mean absolute percentage error value;

[0197] Quantitatively evaluate the evaluation strategy for the root mean square error value and the mean absolute percentage error value to determine the servo motor drive fault prediction model with the optimal performance.

[0198] Step 5.1 of this embodiment: Set the number of iterations of the ConcLSTM model to 200, the learning rate to 0.001, the number of hidden neurons to 64, and divide the data set into a training set and a test set according to a ratio of 8:2.

[0199] Step 5.2: To ensure that the lost information is interference noise rather than key information in the original image, substitute the original image into the ConvLSTM model for prediction and substitute the BIMF component and the residual image R into the ConvLSTM model for prediction respectively, and compare the evaluation indexes after prediction. The changes of the loss function during the training of the BIMF component and the residual image R are as Figures 4 to 7 shown, indicating that the model converges rapidly;

[0200] As shown in the appendix Figure 4 shown, the abscissa (X-axis) represents the time step, that is, the serial number of the time sequence sampling point, and the ordinate (Y-axis) represents the feature intensity value extracted after the Gram angle field (GAF) transformation. Figure 4 The overall curve in the figure shows a smooth and slowly bending change trend, indicating that the output characteristics of the servo motor drive are relatively stable in this state, the feature fluctuation amplitude is small, and it belongs to the normal operation condition.

[0201] As shown in the appendix Figure 5 shown, the definitions of the abscissa and the ordinate are the same as those in the appendix Figure 4 consistent. Compared with the appendix Figure 4 , in the appendix Figure 5 there are obvious local fluctuations and bending phenomena in the curve, indicating that there are slight abnormalities in the output process of the servo motor drive in the corresponding situation, resulting in an increase in the dynamic response disturbance of the feature signal, which belongs to the initial abnormal condition.

[0202] As shown in the appendix Figure 6 shown, the abscissa is the time step and the ordinate is the GAF feature intensity value. In the appendix Figure 6 the curve shows the characteristics of sharp bending and short-time high-frequency oscillation, indicating that the internal fault of the servo motor drive intensifies at this stage, resulting in frequent and large-amplitude instantaneous mutations in the output signal characteristics, which belongs to the moderate fault condition.

[0203] As shown in the appendix Figure 7 shown, the abscissa is the time step and the ordinate is the GAF feature intensity value. In the appendix Figure 7 the overall curve of the figure shows an irregular, high-frequency and violent oscillation trend. Compared with the previous figures, the feature changes are more disordered and violent, indicating that the servo motor drive has entered a serious deterioration state, the system output signal is unstable, and it belongs to the critical fault or failure stage.

[0204] By comparing the attached Figure 4 to the attached Figure 7 It can be seen that as the operating state of the servo motor drive gradually degrades, the GAF feature curve changes from the initial smooth bend (attached Figure 4 ) to having local perturbations (attached Figure 5 ), short-term strong oscillations (attached Figure 6 ), and finally develops into irregular and large-amplitude fluctuations (attached Figure 7 ).

[0205] In addition, this embodiment also conducts performance verification and comparative experiments in combination with actual applications. The specific methods include:

[0206] Designed prediction and comparison algorithm models in different dimensions to verify the effectiveness of the GADF-BEMD-ConvLSTM prediction model. The comparative algorithms adopted in the present invention are RELM, BiLSTM, and AM-BiGRU. In all comparative models, the data preprocessing work of the data set is completed by permutation entropy and ICEEMDAN. Among them, the parameter settings of the comparative prediction model RELM are: the number of iterations is 1000, the number of hidden layer nodes is 50, the learning rate is 0.01, and the regularization coefficient λ is 0.01. The number of iterations of the comparative prediction models BiLSTM and AM-BiGRU is 1000, the number of hidden neurons is 200, the learning rate is 0.005, and the Adam gradient descent algorithm is selected as the optimizer. The comparative prediction results are as Figure 8 shown. The prediction results of each group of prediction models are presented in a more intuitive way, as Figure 9 shown.

[0207] From Figure 8 and Figure 9 it can be seen that under six working conditions, the average values of the prediction evaluation indexes RMSE and MAPE based on the RELM model are 4.0652 and 4.161 respectively; the average values of the prediction evaluation indexes RMSE and MAPE based on the BiLSTM model are 2.9301 and 2.965 respectively; the average values of the prediction evaluation indexes RMSE and MAPE based on the AM-BiGRU model are 2.7196 and 2.698 respectively; the average values of the prediction evaluation indexes RMSE and MAPE based on the GADF-BEMD-ConvLSTM model are 1.6455 and 1.503 respectively. Compared with the RELM, BILSTM, and AM-BiGRU models, the average RMSE of the GADF-BEMD-ConvLSTM model is reduced by 59.52%, 43.84%, and 39.49% respectively, and the average MAPE is reduced by 63.88%, 49.31%, and 44.3% respectively. Refer to Figure 9From Figures a and b, the differences in the prediction evaluation metrics can be visually observed. It can also be seen that the GADF-BEMD-ConvLSTM model has a higher prediction accuracy compared to other one-dimensional prediction models, and can establish a prediction model with smaller prediction errors.

[0208] Optionally, in step 5, the predicted values of each intrinsic mode function component and the residual term are reconstructed and the error is calculated to obtain the error evaluation metrics corresponding to the reconstructed image, including:

[0209] Step 5.1: All the predicted mode function components and the residual term are weighted and superimposed according to the composition rule of the original image to obtain the reconstructed fault prediction image;

[0210] Step 5.2: The reconstructed fault prediction image and the actual observed image are evaluated for error to obtain the error evaluation metrics corresponding to the reconstructed image;

[0211] The error metrics include but are not limited to: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Pearson Correlation Coefficient (PCC).

[0212] Optionally, in step 6, based on the comparison result between the preset fault threshold and the error evaluation metrics corresponding to the reconstructed image, the prediction result of the servo motor drive fault type is obtained, including:

[0213] Step 6.1: Compare the error evaluation metrics corresponding to the reconstructed image with the preset fault identification threshold or perform trend analysis on consecutive multiple all mode function components according to the set high error trend to obtain the prediction result of the servo motor drive fault type;

[0214] Step 6.2: When the error metric exceeds the upper threshold under the corresponding working condition or consecutive multiple mode function components show a high error trend, identify the servo motor drive fault and extract the relevant error metrics and the fault characteristics corresponding to the relevant mode function components;

[0215] Step 6.3: According to the matching result between the predefined fault label database and the fault characteristics, obtain the prediction result of the servo motor drive fault type.

[0216] Example 2

[0217] This embodiment provides a non-invasive servo motor driver fault identification device, including: a data acquisition and conversion module, a two-dimensional modal decomposition module, a prediction module, a reconstruction and error calculation module, and an identification module;

[0218] The data acquisition module is used to acquire the voltage signal output by the servo motor driver, and perform polar coordinate mapping on the voltage signal using the Gram angular field function to obtain a two-dimensional image of the voltage signal;

[0219] The two-dimensional modal decomposition module is used to perform bidirectional empirical mode decomposition on the two-dimensional image of the voltage signal according to the two-dimensional empirical mode decomposition (BEMD) to obtain multiple intrinsic mode function components and a residual term of the voltage signal;

[0220] The prediction module is used to predict the multiple intrinsic mode function components and the residual term based on a pre-trained servo motor driver fault prediction model to obtain a prediction image of each intrinsic mode function component and the residual term;

[0221] The reconstruction and error calculation module is used to reconstruct and calculate the error of the prediction image of each intrinsic mode function component and the residual term to obtain an error evaluation index corresponding to the reconstructed image;

[0222] The identification module is used to obtain a prediction result of the servo motor driver fault type according to the comparison result between a preset fault threshold and the error evaluation index corresponding to the reconstructed image.

[0223] In summary, the present invention uses a non-invasive sampling method to collect one-dimensional voltage signals and converts them into two-dimensional images based on the Gram angular field. It combines the two-dimensional empirical mode decomposition (BEMD) to solve the redundant noise problem generated by the two-dimensional images. Then, it uses the convolutional long short-term memory network (ConvLSTM) to predict each decomposed component separately, and reconstructs and calculates the error of the prediction results, further improving the spatio-temporal feature fusion ability, making the prediction results accurate and real-time, so as to better ensure the stable operation of the equipment to which the servo motor driver belongs.

[0224] In the servo motor driver fault prediction, by combining the CNN feature extraction module and the ConvLSTM spatio-temporal fusion module, the one-dimensional voltage signal is first converted, and then the ConvLSTM is used to directly process the time-series data with a spatial structure. At the same time, the gating mechanism of the ConvLSTM can selectively retain the key information in the CNN features, and has a flexible dynamic feature construction ability, so as to realize the dynamic adjustment of the weights of the CNN and the ConvLSTM under various working conditions and have strong robustness.

[0225] Develop a cross-modal fusion diagnosis system, combine multiple error metrics, high-error mode trend analysis, and perform multi-level and multi-index judgments on servo motor drive faults based on the existing fault label database, so as to achieve early detection and accurate classification of servo motor drive faults, systematically improve the intelligent level of servo motor drive fault diagnosis, and provide core guarantee for high-reliability power electronic systems.

[0226] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0227] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the processes and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for realizing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0228] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device realizes the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0229] These 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 generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for realizing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0230] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative rather than restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit of the present invention and the scope protected by the claims. All of these fall within the protection scope of the present invention.

Claims

1. A non-invasive fault prediction method for servo motor drivers, characterized in that, Including: Obtain the voltage signal output by the servo motor driver; Perform polar coordinate mapping on the voltage signal using the Gramian angular field function to obtain a two-dimensional image of the voltage signal; Perform bidirectional empirical mode decomposition on the two-dimensional image of the voltage signal according to the two-dimensional empirical mode decomposition (BEMD) to obtain multiple intrinsic mode function components and a residual term of the voltage signal; Based on a pre-trained servo motor driver fault prediction model, predict multiple intrinsic mode function components and the residual term to obtain a prediction image for each intrinsic mode function component and the residual term; Reconstruct and calculate the error of the prediction image for each intrinsic mode function component and the residual term to obtain an error evaluation index corresponding to the reconstructed image; According to the comparison result between a preset fault threshold and the error evaluation index corresponding to the reconstructed image, obtain the prediction result of the servo motor driver fault type.

2. The non-intrusive servo motor drive fault prediction method according to claim 1, characterized in that, Obtain the voltage signal output by the servo motor driver, including: Sample the voltage signal of the servo motor driver in a non-invasive manner at a fixed sampling frequency and interval to obtain sample data of the voltage; Normalize the sample data of the voltage to obtain standardized time series data, where the sample data is one-dimensional time series data; The expression for normalizing the sample data of the voltage to obtain standardized time series data is: （1） Wherein, is the standardized time series data, is the maximum value in the sample data, is the minimum value in the sample data, is the sample data.

3. The non-invasive servo motor drive fault prediction method according to claim 2, characterized in that, Perform polar coordinate mapping on the voltage signal using the Gramian angular field function to obtain a two-dimensional image of the voltage signal, including: Convert the standardized time series data from the Cartesian coordinate system to the polar coordinate system to obtain the polar coordinates of all data points in the time series data; Perform trigonometric difference calculation on the polar coordinates of all data points using trigonometric functions to obtain the angle difference between any two data points; Construct a Gramian angular field angle matrix according to the angle difference between any two data points; Perform pixel gray level conversion on the elements of the Gramian angular field angle matrix to obtain a two-dimensional image of the voltage signal; Among them, the expression for obtaining the polar coordinates of all data points in the time series data is: （2） Wherein, is the polar angle corresponding to the i-th time point, is the radius in polar coordinates, is the timestamp, represents the number of time points in the time series data, is the arccosine result of the standardized time series data; The expression for constructing a Gramian angular field angle matrix according to the angle difference between any two data points is: （3） In the formula, is the included angle value at the 1st to the nth time points.

4. The non-intrusive servo motor driver fault prediction method according to claim 1, characterized in that, Perform bidirectional empirical mode decomposition on the two-dimensional image of the voltage signal according to the two-dimensional empirical mode decomposition (BEMD) to obtain multiple intrinsic mode function components and a residual term of the voltage signal, including: Use the sliding window technique to extract the extreme values of the two-dimensional image until the sliding window traverses all regions of the two-dimensional image, and then obtain all local maximum points and minimum points extracted by the window; Perform surface fitting on all local maximum points and minimum points respectively to obtain the initial estimates of the upper and lower envelope surfaces; Perform interpolation calculation on the initial estimates of the upper and lower envelope surfaces to obtain the upper envelope surface and the lower envelope surface of the two-dimensional image; Calculate the average value of the upper envelope surface and the lower envelope surface of the two-dimensional image to obtain the average envelope surface; Separate the two-dimensional image according to the average envelope surface to obtain the result image; Perform iterative screening on the result image to obtain multiple updated intrinsic mode function components and a residual term of the voltage signal; Among them, the expression for using the sliding window technique to extract the extreme values of the two-dimensional image until the sliding window traverses all regions of the two-dimensional image and then obtaining all local maximum points and minimum points extracted by the window is: （4） where N(x, y) represents the set of numbers of 4×4 rectangular windows centered at (x, y), I max (x, y) and I min (x, y) represent local maximum points and local minimum points; Calculate the average value of the upper envelope surface and the lower envelope surface of the two-dimensional image, and the expression of the average envelope surface is obtained as follows: （5） In the formula, is the upper envelope surface of the image, is the lower envelope surface of the upper envelope surface of the image; is the average envelope surface; Separate the two-dimensional image according to the average envelope surface, and the expression of the resulting image is obtained as follows: （6） In the formula, is the result image; The expression for iterative screening of the resulting image is as follows: （7） where SD is the iteration termination condition, is the temporary result image obtained in the k-th iteration.

5. The non-invasive servo motor drive fault prediction method according to claim 1, characterized in that, The servo motor drive fault prediction model is constructed by a CNN feature extraction module and a ConvLSTM spatio-temporal fusion module; Among them, the gating mechanism and state update of the ConvLSTM unit are both through convolution operations. The gating mechanism includes an input gate, a forget gate, and an output gate; the state update includes a cell state and a hidden state; The input gate is used to control the amount of voltage signal sample data entering the cell state; The forget gate is used to control the discard ratio of historical degradation features in the cell state; The output gate is used to control the output intensity of transient harmonic components in the cell state; The cell state is used to control the forgetting of historical degradation features through the forget gate and the input update of the amount of voltage signal sample data through the input gate; The hidden state is used to process the next data in the spatio-temporal sequence data.

6. The non-invasive servo motor drive fault prediction method according to claim 5, characterized in that the input The expression of the gate is as follows: （8） Wherein, is the input gate, is the sigmoid activation function, is the convolutional kernel weight matrix of the input data of the input gate, is the convolutional kernel weight matrix of the hidden state data of the input gate, is the input at the current moment, is the hidden state at the previous moment, is the bias term of the input gate, * represents the convolution operation, is the Hadamard product operation; The expression of the forget gate is as follows: （9） In the formula, is the forgetting gate, is the convolutional kernel weight matrix of the input data of the forgetting gate, is the convolutional kernel weight matrix of the hidden state data of the forgetting gate, is the bias term of the forgetting gate; The expression of the output gate is as follows: （10） wherein, is the output gate, is the convolutional kernel weight matrix of the input data of the output gate, is the convolutional kernel weight matrix of the hidden state data of the output gate, is the bias term of the output gate; The expression of the cell state is as follows: （11） In the formula, is the cell state, is the convolution kernel weight matrix of the cell state input data, is the convolution kernel weight matrix of the cell state hidden state, is the bias term of the output gate; The expression of the hidden state is as follows: （12） In the formula, is the hidden state at the current moment.

7. The non-invasive servo motor drive fault prediction method according to claim 1, characterized in that The training method of the servo motor drive fault prediction model includes: Obtain the historical data set of voltage signals; Divide the historical data set of voltage signals to obtain a training set and a test set; Use the Gramian angular field method to perform image conversion on the one-dimensional voltage time series signals in the training set to obtain the two-dimensional image of the historical voltage signals; Perform bidirectional empirical mode decomposition on the two-dimensional image of the historical voltage signals according to the BEMD decomposition method to obtain multiple intrinsic mode function components and a residual term of the historical voltage signals; Input multiple intrinsic mode function components into the prediction model based on the convolutional long short-term memory network for feature extraction and training to obtain the model parameters required for fault prediction; Update the prediction model based on the convolutional long short-term memory network according to the model parameters required for fault prediction to obtain an updated servo motor drive fault prediction model; Input the test set into the updated servo motor drive fault prediction model for prediction to obtain a prediction result; Substitute the prediction results and the true results of the test set into the calculation of the root mean square error and the mean absolute percentage error respectively to obtain the root mean square error value and the mean absolute percentage error value; Adopt an evaluation strategy to quantitatively evaluate the root mean square error value and the mean absolute percentage error value to determine the servo motor drive fault prediction model with the optimal performance.

8. The non-intrusive servo motor drive fault prediction method according to claim 1, characterized in that, Reconstruct and calculate the error of the predicted value of each intrinsic mode function component and the residual term to obtain the error evaluation index corresponding to the reconstructed image, including: Perform weighted superposition on all the modal function components and the residual term obtained by prediction according to the composition rule of the original image to obtain the reconstructed fault prediction image; Perform error evaluation on the reconstructed fault prediction image and the actual observed image to obtain the error evaluation index corresponding to the reconstructed image; The error indexes include: root mean square error, mean absolute error, mean absolute percentage error, and correlation coefficient.

9. The non-invasive servo motor drive fault prediction method according to claim 8, characterized in that, Based on the comparison result between the preset fault threshold and the error evaluation index corresponding to the reconstructed image, the prediction result of the servo motor drive fault type is obtained, including: By comparing the error evaluation index corresponding to the reconstructed image with the preset fault identification threshold or performing trend analysis on consecutive multiple all modal function components according to the set high error trend, the prediction result of the servo motor drive fault type is obtained; When the error index exceeds the upper threshold of the corresponding working condition or consecutive multiple modal function components show a high error trend, the servo motor drive fault is identified and the relevant error index and the fault characteristics corresponding to the relevant modal function components are extracted; Based on the matching result between the predefined fault label database and the fault characteristics, the prediction result of the servo motor drive fault type is obtained.

10. A non-intrusive servo motor drive fault identification device, characterized in that, Including: Data acquisition and conversion module, two-dimensional modal decomposition module, prediction module, reconstruction and error calculation module, and identification module; The data acquisition module is used to acquire the voltage signal output by the servo motor drive and perform polar coordinate mapping on the voltage signal by using the Gram angular field function to obtain the two-dimensional image of the voltage signal; The two-dimensional modal decomposition module is used to perform bidirectional empirical mode decomposition on the two-dimensional image of the voltage signal according to the two-dimensional empirical mode decomposition BEMD to obtain multiple intrinsic mode function components and a residual term of the voltage signal; The prediction module is used to predict multiple intrinsic mode function components and a residual term based on the pre-trained servo motor drive fault prediction model to obtain the prediction images of each intrinsic mode function component and the residual term; The reconstruction and error calculation module is used to reconstruct and calculate the error of the prediction images of each intrinsic mode function component and the residual term to obtain the error evaluation index corresponding to the reconstructed image; The identification module is used to obtain the prediction result of the servo motor drive fault type based on the comparison result between the preset fault threshold and the error evaluation index corresponding to the reconstructed image.

Images