Phase modifier carbon brush overcurrent identification method and system based on GoogleNet network, and electronic equipment

Abstract

The invention discloses a GoogleNet network-based phase modifier carbon brush overcurrent identification method and system and electronic equipment, and the method comprises the steps: processing current data in a source domain data set and a target domain data set, and obtaining a source domain time-frequency image set and a target domain time-frequency image set; the method comprises the following steps: deploying a first GoogleNet network and a second GoogleNet network; training the first GoogleNet network by using the source domain time-frequency image set; transferring parameters of the trained first GoogleNet network to a second GoogleNet network, and training the second GoogleNet network by using the target domain time-frequency image set to obtain a carbon brush overcurrent identification model; and performing over-current identification on the current current data of the carbon brush by using the carbon brush over-current identification model to obtain an over-current identification result of the carbon brush. According to the invention, the overcurrent identification precision and efficiency are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of phase regulator fault diagnosis, and in particular relates to a phase regulator carbon brush overcurrent identification method, system and electronic equipment based on a GoogleNet network. Background Art

[0002] With the continuous development and upgrading of the power system, the phase regulator is an important reactive power compensation device in the power grid, and its operating status has a vital impact on the stability and reliability of the power grid. Carbon brushes are key current transmission components in the phase regulator, and their working status is directly related to the performance and life of the phase regulator.

[0003] Most existing carbon brush monitoring methods rely on manual periodic inspections or single sensor monitoring, which is not only inefficient, time-consuming and labor-intensive, but also difficult to respond to abnormal conditions that may occur in the carbon brushes in a timely manner. Manual inspections are easily affected by human factors, making it difficult to ensure the accuracy of monitoring results. In addition, due to the lack of real-time monitoring methods, existing carbon brush monitoring methods are weak in fault warning capabilities, increasing the risk of power grid operation.

[0004] With the development of artificial intelligence technology, diagnostic methods based on machine learning have begun to be studied and applied to carbon brush anomaly detection. Existing diagnostic methods based on machine learning usually require a large amount of labeled sample data to train the model. However, the carbon brushes of a phase regulator work in a healthy state most of the time, and it is difficult to obtain sufficient data to support the establishment of a machine learning model, resulting in serious performance degradation. At the same time, the constructed general model is difficult to accurately diagnose different carbon brushes in the phase regulator. Summary of the invention

[0005] The purpose of the present invention is to provide a phase regulator carbon brush overcurrent identification method, system and electronic equipment based on the GoogleNet network to solve the problems of low efficiency and difficulty in ensuring accuracy of existing carbon brush monitoring methods, as well as the lack of sample data in artificial intelligence-based monitoring methods, which leads to reduced diagnostic performance.

[0006] The present invention solves the above technical problems through the following technical solutions: A method for identifying overcurrent of carbon brushes of a phase regulator based on GoogleNet network, comprising:

[0007] Acquire current data of different carbon brushes in the phase condenser, and generate a source domain data set and a target domain data set; wherein the current data in the source domain data set is the current experimental data of a part of the carbon brushes in the phase condenser, and the current data in the target domain data set is the current measured data of another part of the carbon brushes in the phase condenser;

[0008] Processing the current data in the source domain data set and the target domain data set to obtain a time-frequency graph, and then obtaining a source domain time-frequency image set and a target domain time-frequency image set;

[0009] Deploy the first GoogleNet network and the second GoogleNet network with the same architecture;

[0010] Using the source domain time-frequency image set to train the first GoogleNet network;

[0011] Migrating the parameters of the trained first GoogleNet network to the second GoogleNet network, and then training the second GoogleNet network using the target domain time-frequency image set to obtain a carbon brush overcurrent recognition model;

[0012] The carbon brush overcurrent identification model is used to perform overcurrent identification on the current current data of the carbon brush to obtain an overcurrent identification result of the carbon brush.

[0013] Further, the current data in the source domain data set and the target domain data set are processed using synchronous compression wavelet transform, specifically including:

[0014] The wavelet coefficients of the current data in the source domain data set and the target domain data set are calculated. The specific formula is:

[0015] ;

[0016] in, is the wavelet coefficient of the current data; is the scale parameter; is the translation parameter; is the current data; is the mother wavelet function; Indicates time;

[0017] The wavelet coefficients of the current data are converted from the time domain to the frequency domain by Fourier transform to obtain the time-frequency representation of the current data. The specific formula is:

[0018] ;

[0019] in, is the time-frequency representation of the current data; is the angular frequency; i is the imaginary number sign; is the Fourier transform of the current data; is the Fourier transform of the mother wavelet function;

[0020] The instantaneous frequency of the current data is calculated according to the time-frequency representation of the current data. The specific formula is:

[0021] ;

[0022] in, is the instantaneous frequency of the current data; is the symbol of partial derivative;

[0023] According to the instantaneous frequency of the current data, the wavelet coefficients of the current data are converted from the time scale plane to the time frequency plane, and the time frequency plane is divided into different frequency intervals;

[0024] After selecting a center frequency in each frequency interval, the wavelet coefficients in each frequency interval are compressed;

[0025] According to the center frequency, the compressed wavelet coefficients are reconstructed using inverse wavelet transform to obtain a time-frequency diagram corresponding to the current data; wherein the discrete state expression of the time-frequency diagram is:

[0026] ;

[0027] in, The frequency interval The center frequency of is the center frequency The time-frequency image value of ; is the frequency interval; are discrete points of different scales in the time-frequency plane; For The scale interval at .

[0028] Furthermore, the first GoogleNet network and the second GoogleNet network each include an initial module, an Inception module and an output module;

[0029] The initial module includes an input layer, a first convolutional layer, a first maximum pooling layer, a first normalization layer, a second convolutional layer, a third convolutional layer, a second normalization layer and a second maximum pooling layer connected in sequence;

[0030] The Inception module includes multiple Inception layers; each Inception layer includes a first branch, a second branch, a third branch and a fourth branch; the first branch is a convolutional layer; the second branch is two convolutional layers connected in sequence; the third branch is two convolutional layers connected in sequence; the fourth branch is a maximum pooling layer and a convolutional layer connected in sequence;

[0031] The output module includes a global average pooling layer, a Dropout layer, a fully connected layer, a Softmax layer and a classification layer which are connected in sequence.

[0032] Further, the first GoogleNet network is trained using the source domain time-frequency image set, specifically including:

[0033] Dividing the source domain time-frequency image set into a first training set and a first test set according to a preset ratio;

[0034] Input the time-frequency graph in the first training set into the first GoogleNet network to obtain the first predicted label;

[0035] A loss function is used to calculate the loss error between the first predicted label and the true label of the corresponding time-frequency graph, and the parameters of the first GoogleNet network are updated by back propagation, and iterative training is continuously performed; wherein the true label is 0 or 1, 0 indicates a normal state, and 1 indicates an overcurrent state;

[0036] When the number of iterations reaches a first preset number, the training is stopped, and the performance of the trained first GoogleNet network is tested using the first test set;

[0037] Determine whether the recognition accuracy of the trained first GoogleNet network reaches a first accuracy threshold; if so, output the trained first GoogleNet network; if not, repeat the step of training the first GoogleNet network using the first training set until the recognition accuracy of the trained first GoogleNet network reaches the first accuracy threshold.

[0038] Further, the target domain time-frequency image set is used to train the second GoogleNet network, specifically including:

[0039] Dividing the target domain time-frequency image set into a second training set and a second test set according to a preset ratio;

[0040] Freeze the first few Inception layers of the Inception module in the second GoogleNet network, and initialize the parameters of the last few Inception layers of the Inception module;

[0041] The second training set is used to train the last few Inception layers of the Inception module to obtain the second predicted label;

[0042] The loss function is used to calculate the loss error between the second predicted label and the true label of the corresponding time-frequency graph, and the parameters of the last few Inception layers of the Inception module are updated by back propagation, and iterative training is continued; wherein the true label is 0 or 1, 0 indicates a normal state, and 1 indicates an overcurrent state;

[0043] When the number of iterations reaches a second preset number, the training is stopped, and the performance of the trained second GoogleNet network is tested using the second test set;

[0044] Determine whether the recognition accuracy of the trained second GoogleNet network reaches the second accuracy threshold; if so, output the carbon brush overcurrent recognition model; if not, repeat the step of training the last few Inception layers of the Inception module using the second training set until the recognition accuracy of the trained second GoogleNet network reaches the second accuracy threshold.

[0045] Furthermore, the loss function is expressed as:

[0046] ;

[0047] in, is the loss error, m is the number of input time-frequency graphs; is the true label of the i-th input time-frequency graph; is the predicted label of the i-th input frequency spectrum.

[0048] Furthermore, the identification method also includes carbon brush abnormality diagnosis, which specifically includes:

[0049] While collecting the current current data of the carbon brush, collect the current temperature data of the carbon brush;

[0050] When the carbon brush is in an overcurrent state, the temperature rise rate of the carbon brush is calculated based on the current temperature data of the carbon brush;

[0051] Determine whether the temperature rise rate of the carbon brush exceeds the temperature rise rate threshold; if so, the carbon brush is in an abnormal state; otherwise, the carbon brush may have a hidden danger of an abnormal state.

[0052] Based on the same concept, the present invention also provides a phase regulator carbon brush overcurrent identification system based on GoogleNet network, comprising:

[0053] A data acquisition module is used to acquire current data of different carbon brushes in the phase condenser and generate a source domain data set and a target domain data set; wherein the current data in the source domain data set is the current experimental data of a part of the carbon brushes in the phase condenser, and the current data in the target domain data set is the current measured data of another part of the carbon brushes in the phase condenser;

[0054] A data processing module, used to process the current data in the source domain data set and the target domain data set to obtain a time-frequency diagram, and then obtain a source domain time-frequency image set and a target domain time-frequency image set;

[0055] A model training module is used to deploy a first GoogleNet network and a second GoogleNet network of the same architecture; train the first GoogleNet network using the source domain time-frequency image set; migrate the parameters of the trained first GoogleNet network to the second GoogleNet network, and then train the second GoogleNet network using the target domain time-frequency image set to obtain a carbon brush overcurrent recognition model;

[0056] The abnormality diagnosis module is used to perform overcurrent identification on the current current data of the carbon brush using the carbon brush overcurrent identification model to obtain an overcurrent identification result of the carbon brush.

[0057] Based on the same concept, the present invention also provides an electronic device, including a memory, a processor and a computer program / instruction stored in the memory, wherein the processor executes the computer program / instruction to implement the phase-shifting carbon brush overcurrent identification method based on the GoogleNet network as described above.

[0058] Based on the same concept, the present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the phase-modulator carbon brush overcurrent identification method based on the GoogleNet network as described above.

[0059] Compared with the prior art, the advantages of the present invention are:

[0060] The present invention utilizes a carbon brush overcurrent recognition model to realize real-time recognition of overcurrent, which not only improves the recognition efficiency but also improves the accuracy of carbon brush overcurrent recognition compared with manual periodic inspection or single sensor monitoring; at the same time, the carbon brush overcurrent recognition model is constructed based on the GoogleNet network, which not only reduces the number of network parameters but also increases the depth and width of the network, so that the carbon brush overcurrent recognition model can extract more current fault features under the same amount of calculation, further improving the recognition accuracy and efficiency.

[0061] Based on the similarity of the structures between different carbon brushes in a phase regulator, the present invention uses the current experimental data of a part of the carbon brushes to train the first GoogleNet network, and then transfers the parameters of the trained first GoogleNet network to the second GoogleNet network through a transfer learning algorithm. Then, the current measured data of another part of the carbon brushes is used to train the second GoogleNet network to obtain a carbon brush overcurrent recognition model, so that the carbon brush overcurrent recognition model can complete the model training through a small amount of current measured data, effectively reducing the training data demand of the GoogleNet network, ensuring the model performance with less current measured data, and being able to achieve accurate diagnosis of different carbon brushes in the phase regulator. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only one embodiment of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0063] Figure 1 is a flow chart of a method for identifying overcurrent of a phase regulator carbon brush based on a GoogleNet network in an embodiment of the present invention;

[0064] Figure 2 is an architecture diagram of a first GoogleNet network and a second GoogleNet network in an embodiment of the present invention;

[0065] Figure 3 is a training flow chart of the first GoogleNet network and the second GoogleNet network in an embodiment of the present invention;

[0066] Figure 4 is a flow chart of carbon brush abnormality diagnosis in an embodiment of the present invention;

[0067] Figure 5 is a training accuracy curve diagram in an embodiment of the present invention;

[0068] Figure 6 is a test accuracy curve diagram in an embodiment of the present invention;

[0069] Figure 7 It is a structural block diagram of a phase regulator carbon brush overcurrent identification system based on the GoogleNet network in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The following is a clear and complete description of the technical solutions in the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] The technical solution of the present application is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0072] Embodiment 1

[0073] Figure 1 The present invention shows a method for identifying overcurrent of a phase regulator carbon brush based on a GoogleNet network. Figure 1 As shown, the phase regulator carbon brush overcurrent identification method of the present invention comprises the following steps:

[0074] Step S1: obtaining current data of different carbon brushes in a phase regulator, and generating a source domain data set and a target domain data set.

[0075] In order to improve the performance of the model, an overcurrent experiment is performed on a part of the carbon brushes in the condenser (i.e., the selected carbon brushes), and the current data of different carbon brushes in normal state and overcurrent state are obtained, that is, the current experimental data of different carbon brushes are obtained, and the source domain data set is generated according to the current experimental data of different carbon brushes. Each sample in the source domain data set includes the current experimental data and its true label. The true label is 0 or 1, 0 represents the normal state, and 1 represents the overcurrent state.

[0076] In order to make the model better applied to the diagnosis of carbon brush overcurrent in engineering practice, the current data of another part of the carbon brushes in the phase regulator (i.e., the carbon brushes not selected in the overcurrent experiment) in engineering practice are collected by sensors, and the current data of different carbon brushes in normal state and overcurrent state are obtained, that is, the current measured data of different carbon brushes are obtained, and the target domain data set is generated according to the current measured data of different carbon brushes. Each sample in the target domain data set includes the current measured data and its true label, and the true label is 0 or 1, 0 represents the normal state, and 1 represents the overcurrent state.

[0077] The current data of the source domain dataset and the target domain dataset include the current data of all carbon brushes in the phase regulator, and the number of carbon brushes in the source domain dataset is greater than that in the target domain dataset. In this embodiment, the number of samples in the source domain dataset is 1000, and the number of samples in the target domain dataset is 300.

[0078] Step S2: Process the current data (i.e., current experimental data and current measured data) in the source domain data set and the target domain data set to obtain a time-frequency diagram, and then obtain a source domain time-frequency image set and a target domain time-frequency image set.

[0079] In a specific embodiment of the present invention, the current data in the source domain data set and the target domain data set are processed using Synchrosqueezed Wavelet Transform (SWT). SWT is a method that combines wavelet transform and rearrangement algorithm to compress and rearrange time-frequency coefficients. The energy of the time-frequency spectrum is concentrated by the synchronous compression operator, and the concentration of the instantaneous frequency is enhanced, thereby achieving high-resolution time-frequency expression of the original signal, which better solves the fuzzy phenomenon that occurs in traditional time-frequency analysis.

[0080] The current data in the source domain data set and the target domain data set are processed by synchronous compression wavelet transform, specifically including:

[0081] Step S2.1: Calculate the wavelet coefficients of the current data in the source domain data set and the target domain data set. The specific formula is:

[0082] (1)

[0083] in, is the wavelet coefficient of the current data; is the scale parameter; is the translation parameter; is the current data; is the mother wavelet function; Indicates time.

[0084] Step S2.2: Convert the wavelet coefficients of the current data from the time domain to the frequency domain through Fourier transform to obtain the time-frequency representation of the current data. The specific formula is:

[0085] (2)

[0086] in, is the time-frequency representation of the current data; is the angular frequency; i is the imaginary number sign; is the Fourier transform of the current data; is the Fourier transform of the mother wavelet function.

[0087] Step S2.3: Calculate the instantaneous frequency of the current data according to the time-frequency representation of the current data. The specific formula is:

[0088] (3)

[0089] in, is the instantaneous frequency of the current data; is the symbol of partial derivative.

[0090] Step S2.4: According to the instantaneous frequency of the current data, the wavelet coefficients of the current data are converted from the time scale plane to the time frequency plane, and the time frequency plane is divided into different frequency intervals.

[0091] Step S2.5: After selecting a center frequency in each frequency interval, compress the wavelet coefficients in each frequency interval.

[0092] Step S2.6: According to the center frequency, the compressed wavelet coefficients are reconstructed using inverse wavelet transform to obtain a time-frequency diagram corresponding to the current data.

[0093] In some embodiments, the function expression of the time-frequency diagram includes a discrete state expression and a continuous state expression. The discrete state expression of the time-frequency diagram is:

[0094] (4)

[0095] in, The frequency interval The center frequency of is the center frequency The time-frequency image value of ; is the frequency interval; are discrete points of different scales in the time-frequency plane, and the essence of the discrete points is the scale parameter; For The scale interval at .

[0096] The continuous state expression of the time-frequency diagram is:

[0097] (5)

[0098] in, Continuous frequency The time-frequency image value of ; is the set of non-zero wavelet coefficients; is the Dirac function.

[0099] Therefore, the samples in the source domain time-frequency image set and the target domain time-frequency image set are both time-frequency images and their true labels. In order to eliminate the scale difference of current data, accelerate model convergence, and improve model performance, before calculating the wavelet coefficients of the current data in the source domain data set and the target domain data set, each current data in the source domain data set and the target domain data set can also be normalized.

[0100] Step S3: deploy a first GoogleNet network and a second GoogleNet network with the same architecture.

[0101] The GoogleNet network is improved on the basis of the classic convolutional neural network (CNN). The total number of layers of the GoogleNet network is 100 and the network depth is 22, which effectively increases the depth and width of the network so that the GoogleNet network can extract richer feature information.

[0102] like Figure 2 As shown, the first GoogleNet network and the second GoogleNet network of the present invention both include an initial module, an Inception module and an output module.

[0103] The initial module includes an input layer (Image Input Layer), a 7×7 convolution layer (Convolution-7×7), the first maximum pooling layer (Max Pooling Layer), the first normalization layer (Normalization Layer), a 3×3 convolution layer (Convolution-3×3), a 3×3 convolution layer (Convolution-3×3), the second normalization layer (Normalization Layer) and the second maximum pooling layer (Max Pooling Layer) connected in sequence.

[0104] The Inception module includes 9 Inception layers; each Inception layer includes the first branch, the second branch, the third branch and the fourth branch; the first branch is a 1×1 convolution layer (Convolution-1×1); the second branch is two 3×3 convolution layers (Convolution-3×3) connected in sequence; the third branch is two 5×5 convolution layers (Convolution-5×5) connected in sequence; the fourth branch is a maximum pooling layer (Max Pooling Layer) and a 1×1 convolution layer (Convolution-1×1) connected in sequence.

[0105] In this embodiment, all convolutional layers in the initial module and the Inception module are connected with a ReLU layer as an activation function.

[0106] In some embodiments, the Inception module processes input data in different ways through four branches. The outputs of these branches are concatenated and fused in the feature dimension and then transmitted to the next layer of Inception module. The Inception module can reduce the dimension of the input features by introducing a 1×1 convolution layer before the 3×3 and 5×5 convolution layers, thereby significantly reducing the computational complexity of the convolution operation when the number of input fault features is large.

[0107] Specifically, the output module includes a global average pooling layer, a dropout layer, a fully connected layer, a softmax layer, and a classification layer, which are connected in sequence.

[0108] Step S4: Train the first GoogleNet network using the source domain time-frequency image set.

[0109] In a specific embodiment of the present invention, Figure 3As shown, the first GoogleNet network is trained using the source domain time-frequency image set, specifically including:

[0110] Step S4.1: Divide the source domain time-frequency image set into a first training set and a first test set according to a preset ratio;

[0111] Step S4.2: Input the time-frequency graph in the first training set into the first GoogleNet network to obtain a first predicted label;

[0112] Step S4.3: Use the loss function to calculate the loss error between the first predicted label and the true label of the corresponding time-frequency graph, and update the parameters of the first GoogleNet network through back propagation, and continue iterative training.

[0113] Among them, the expression of the loss function is:

[0114] (6)

[0115] in, is the loss error, m is the number of input time-frequency graphs; is the true label of the i-th input time-frequency graph; is the predicted label of the i-th input frequency spectrum.

[0116] Step S4.4: When the number of iterations reaches a first preset number, stop training, and use the first test set to perform a performance test on the trained first GoogleNet network;

[0117] Step S4.5: According to the performance test results, determine whether the recognition accuracy of the trained first GoogleNet network reaches the first accuracy threshold; if so, it indicates that the model performance is qualified, and the trained first GoogleNet network (i.e., the source domain recognition model) is output; if not, it indicates that the model performance is unqualified, and go to step S4.2.

[0118] In this embodiment, the iterative training of the first GoogleNet network adopts the Adam optimizer, and the initial learning rate is 0.001; the batch size is 128, and the first GoogleNet network processes 128 samples in each training; the first preset number of times is set to 400, among which the learning rate parameter is set to 0.001 for the first 200 training times, and the parameter is set to 0.0001 for the next 200 training times.

[0119] Step S5: Migrate the parameters of the trained first GoogleNet network to the second GoogleNet network, and then use the target domain time-frequency image set to train the second GoogleNet network to obtain a carbon brush overcurrent recognition model.

[0120] In this embodiment, the parameters of the Inception module in the first GoogleNet network after training are transferred to the second GoogleNet network. Figure 3 As shown, the second GoogleNet network is trained using the target domain time-frequency image set, specifically including:

[0121] Step S5.1: Divide the target domain time-frequency image set into a second training set and a second test set according to a preset ratio;

[0122] Step S5.2: Freeze the first three Inception layers of the Inception module in the second GoogleNet network, and initialize the parameters of the last six Inception layers of the Inception module;

[0123] Step S5.3: Use the second training set to train the last 6 Inception layers of the Inception module to obtain a second predicted label;

[0124] Step S5.4: Use the loss function to calculate the loss error between the second predicted label and the true label of the corresponding time-frequency graph, and update the parameters of the last 6 Inception layers of the Inception module through back propagation, and continue iterative training;

[0125] Step S5.5: When the number of iterations reaches a second preset number, stop training, and use the second test set to perform a performance test on the trained second GoogleNet network;

[0126] Step S5.6: According to the performance test results, determine whether the recognition accuracy of the trained second GoogleNet network reaches the second accuracy threshold; if so, it indicates that the model performance is qualified, and the carbon brush overcurrent recognition model is output; if not, it indicates that the model performance is unqualified, and go to step S5.3.

[0127] When training the second GoogleNet network, only the parameters of the last few Inception layers of the Inception module are trained, achieving better performance with less training data. The second GoogleNet network is trained using the target domain time-frequency image set to learn the current characteristics of the carbon brush under specific working conditions, thereby improving the accuracy and pertinence of overcurrent diagnosis, providing customized overcurrent diagnosis for each unknown carbon brush, and enhancing the model's adaptability to different carbon brush characteristics and the accuracy of diagnosis. This personalized training process helps to improve the reliability and maintenance efficiency of the entire system.

[0128] Step S6: using the carbon brush overcurrent identification model to perform overcurrent identification on the current current data of the carbon brush to obtain an overcurrent identification result of the carbon brush.

[0129] The current current data of the carbon brush in the condenser is collected, and the current current data is input into the carbon brush overcurrent identification model after SWT transformation to obtain the overcurrent identification result of the carbon brush. In order to further diagnose the abnormality of the carbon brush, the current temperature data of the carbon brush is collected at the same time as the current current data of the carbon brush.

[0130] Step S7: Carbon brush abnormality diagnosis.

[0131] After the overcurrent of the carbon brush is identified, abnormality diagnosis is also performed based on the overcurrent identification result of the carbon brush. Figure 4 As shown in the figure, carbon brush abnormality diagnosis specifically includes:

[0132] Step S7.1: judging whether the carbon brush is in an overcurrent state according to the overcurrent identification result of the carbon brush;

[0133] Step S7.2: When the carbon brush is in an overcurrent state, the temperature rise rate of the carbon brush is calculated according to the current temperature data of the carbon brush;

[0134] Step S7.3: Determine whether the temperature rise rate of the carbon brush exceeds the temperature rise rate threshold; if so, the carbon brush is in an abnormal state; otherwise, the carbon brush may have a hidden danger of abnormal state, and continue to monitor the current state of the carbon brush.

[0135] It should be noted that in the training of the GoogleNet network, the high accuracy of the GoogleNet network is supported by a large amount of data, which often requires a large amount of labeled data as a training set for training. There are multiple carbon brushes in the phase regulator. If a large amount of current data is collected for training for each carbon brush, the cost is too high, the amount of data is large, and the calculation takes a long time. In order to solve this technical problem, the present invention first conducts a large-scale overcurrent experiment on different selected carbon brushes, obtains a large amount of current experimental data of different selected carbon brushes, and constructs and trains the first GoogleNet network with this current experimental data; according to the symmetry of different carbon brushes in the phase regulator and the same working principle, the parameters of the trained first GoogleNet network are migrated to the second GoogleNet network, and a small amount of current measured data of unselected carbon brushes is obtained, and the second GoogleNet network after parameter migration is trained using a small amount of current measured data to achieve fine-tuning of the second GoogleNet network, so as to obtain an overcurrent recognition model for all carbon brushes in the phase regulator under specific working conditions. The present invention can effectively reduce the amount of training data required for overcurrent identification models of all carbon brushes in a phase regulator, thereby reducing the model training cost; at the same time, through the overcurrent identification models of different carbon brushes, carbon brushes with abnormal conditions can be discovered in a timely manner, thereby improving the accuracy of the carbon brush abnormality diagnosis results.

[0136] The TL-GoogleNe model established by the present invention through the GoogleNet network, transfer learning principle and SWT has a significant improvement in the accuracy of overcurrent identification compared with the GoogleNet model based on CWT (Continuous Wavelet Transform) and the GoogleNet model based on SWT. Figure 5 and Figure 6 shown.

[0137] Embodiment 2

[0138] like Figure 7 As shown, the phase regulator carbon brush overcurrent identification system 100 based on the GoogleNet network provided in the embodiment of the present invention includes a data acquisition module 101, a data processing module 102, a model training module 103 and an abnormality diagnosis module 104.

[0139] The data acquisition module 101 is used to acquire the current data of different carbon brushes in the phase condenser and generate a source domain data set and a target domain data set; wherein the current data in the source domain data set is the current experimental data of a part of the carbon brushes in the phase condenser, and the current data in the target domain data set is the current measured data of another part of the carbon brushes in the phase condenser.

[0140] The data processing module 102 is used to process the current data in the source domain data set and the target domain data set to obtain a time-frequency diagram, and then obtain a source domain time-frequency image set and a target domain time-frequency image set.

[0141] The model training module 103 is used to deploy a first GoogleNet network and a second GoogleNet network of the same architecture; train the first GoogleNet network using the source domain time-frequency image set; migrate the parameters of the trained first GoogleNet network to the second GoogleNet network, and then train the second GoogleNet network using the target domain time-frequency image set to obtain a carbon brush overcurrent recognition model.

[0142] The abnormality diagnosis module 104 is used to perform overcurrent identification on the current current data of the carbon brush using the carbon brush overcurrent identification model to obtain an overcurrent identification result of the carbon brush.

[0143] In a specific embodiment of the present invention, the abnormal diagnosis module 104 is also used for carbon brush abnormality diagnosis, specifically for determining whether the carbon brush is in an overcurrent state based on the overcurrent identification result of the carbon brush; when the carbon brush is in an overcurrent state, calculating the temperature rise rate of the carbon brush based on the current temperature data of the carbon brush; determining whether the temperature rise rate of the carbon brush exceeds the temperature rise rate threshold; if so, the carbon brush is in an abnormal state; otherwise, the carbon brush may have a hidden danger of an abnormal state, and continue to monitor the current state of the carbon brush.

[0144] Embodiment 3

[0145] An embodiment of the present invention also provides an electronic device, which includes: a memory, a processor and a computer program / instructions stored in the memory, and the processor executes the computer program / instructions to implement the phase-shifting carbon brush overcurrent identification method based on the GoogleNet network in the embodiment of the present application.

[0146] Although not shown, the electronic device includes a processor that can perform various appropriate operations and processes according to the programs and / or data stored in the read-only memory (ROM) or the programs and / or data loaded from the storage part into the random access memory (RAM). The processor can be a multi-core processor or can include multiple processors. In some embodiments, the processor can include a general-purpose main processor and one or more special coprocessors, such as a central processing unit, a graphics processing unit (GPU), a neural network processor (NPU), a digital signal processor (DSP), etc. In RAM, various programs and data required for device operation are also stored. The processor, ROM, and RAM are connected to each other via a bus. The input / output (I / O) interface is also connected to the bus.

[0147] The processor and memory are used together to execute the program / instructions stored in the memory. When the program / instructions are executed by the computer, the methods, steps or functions described in the above embodiments can be implemented.

[0148] Although not shown, an embodiment of the present invention further provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the method for identifying overcurrent of a phase regulator carbon brush based on a GoogleNet network in an embodiment of the present application.

[0149] Readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0150] What is disclosed above is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or modifications within the technical scope disclosed in the present invention, which should be covered within the protection scope of the present invention.

Claims

1. A method for identifying overcurrent of carbon brush of phase regulator based on GoogleNet network, characterized in that: The identification method comprises: Acquire current data of different carbon brushes in the phase condenser, and generate a source domain data set and a target domain data set; wherein the current data in the source domain data set is the current experimental data of a part of the carbon brushes in the phase condenser, and the current data in the target domain data set is the current measured data of another part of the carbon brushes in the phase condenser; Processing the current data in the source domain data set and the target domain data set to obtain a time-frequency graph, and then obtaining a source domain time-frequency image set and a target domain time-frequency image set; Deploy the first GoogleNet network and the second GoogleNet network with the same architecture; Using the source domain time-frequency image set to train the first GoogleNet network; Migrating the parameters of the trained first GoogleNet network to the second GoogleNet network, and then training the second GoogleNet network using the target domain time-frequency image set to obtain a carbon brush overcurrent recognition model; The carbon brush overcurrent identification model is used to perform overcurrent identification on the current current data of the carbon brush to obtain an overcurrent identification result of the carbon brush.

2. The method for identifying overcurrent of carbon brushes of a phase regulator based on GoogleNet network according to claim 1 is characterized in that: The current data in the source domain data set and the target domain data set are processed by synchronous compression wavelet transform, specifically including: The wavelet coefficients of the current data in the source domain data set and the target domain data set are calculated. The specific formula is: ； in, is the wavelet coefficient of the current data; is the scale parameter; is the translation parameter; is the current data; is the mother wavelet function; Indicates time; The wavelet coefficients of the current data are converted from the time domain to the frequency domain by Fourier transform to obtain the time-frequency representation of the current data. The specific formula is: ； in, is the time-frequency representation of the current data; is the angular frequency; i is the imaginary number sign; is the Fourier transform of the current data; is the Fourier transform of the mother wavelet function; The instantaneous frequency of the current data is calculated according to the time-frequency representation of the current data. The specific formula is: ； in, is the instantaneous frequency of the current data; is the symbol of partial derivative; According to the instantaneous frequency of the current data, the wavelet coefficients of the current data are converted from the time scale plane to the time frequency plane, and the time frequency plane is divided into different frequency intervals; After selecting a center frequency in each frequency interval, the wavelet coefficients in each frequency interval are compressed; According to the center frequency, the compressed wavelet coefficients are reconstructed using inverse wavelet transform to obtain a time-frequency diagram corresponding to the current data; wherein the discrete state expression of the time-frequency diagram is: ； in, The frequency interval The center frequency of is the center frequency The time-frequency image value of is the frequency interval; are discrete points of different scales in the time-frequency plane; For The scale interval at .

3. The method for identifying overcurrent of carbon brushes of a phase regulator based on GoogleNet network according to claim 1 is characterized in that: The first GoogleNet network and the second GoogleNet network both include an initial module, an Inception module and an output module; The initial module includes an input layer, a first convolutional layer, a first maximum pooling layer, a first normalization layer, a second convolutional layer, a third convolutional layer, a second normalization layer and a second maximum pooling layer connected in sequence; The Inception module includes multiple Inception layers; each Inception layer includes a first branch, a second branch, a third branch and a fourth branch; the first branch is a convolutional layer; the second branch is two convolutional layers connected in sequence; the third branch is two convolutional layers connected in sequence; the fourth branch is a maximum pooling layer and a convolutional layer connected in sequence; The output module includes a global average pooling layer, a Dropout layer, a fully connected layer, a Softmax layer and a classification layer which are connected in sequence.

4. The method for identifying overcurrent of carbon brushes of a phase regulator based on GoogleNet network according to claim 1 is characterized in that: Training the first GoogleNet network using the source domain time-frequency image set specifically includes: Dividing the source domain time-frequency image set into a first training set and a first test set according to a preset ratio; Input the time-frequency graph in the first training set into the first GoogleNet network to obtain the first predicted label; A loss function is used to calculate the loss error between the first predicted label and the true label of the corresponding time-frequency graph, and the parameters of the first GoogleNet network are updated by back propagation, and iterative training is continuously performed; wherein the true label is 0 or 1, 0 indicates a normal state, and 1 indicates an overcurrent state; When the number of iterations reaches a first preset number, the training is stopped, and the performance of the trained first GoogleNet network is tested using the first test set; Determine whether the recognition accuracy of the trained first GoogleNet network reaches a first accuracy threshold; if so, output the trained first GoogleNet network; if not, repeat the step of training the first GoogleNet network using the first training set until the recognition accuracy of the trained first GoogleNet network reaches the first accuracy threshold.

5. The method for identifying overcurrent of carbon brushes of a phase regulator based on GoogleNet network according to claim 1 is characterized in that: The second GoogleNet network is trained using the target domain time-frequency image set, specifically including: Dividing the target domain time-frequency image set into a second training set and a second test set according to a preset ratio; Freeze the first few Inception layers of the Inception module in the second GoogleNet network, and initialize the parameters of the last few Inception layers of the Inception module; The second training set is used to train the last few Inception layers of the Inception module to obtain the second predicted label; The loss function is used to calculate the loss error between the second predicted label and the true label of the corresponding time-frequency graph, and the parameters of the last few Inception layers of the Inception module are updated by back propagation, and iterative training is continued; wherein the true label is 0 or 1, 0 indicates a normal state, and 1 indicates an overcurrent state; When the number of iterations reaches a second preset number, the training is stopped, and the performance of the trained second GoogleNet network is tested using the second test set; Determine whether the recognition accuracy of the trained second GoogleNet network reaches the second accuracy threshold; if so, output the carbon brush overcurrent recognition model; if not, repeat the step of training the last few Inception layers of the Inception module using the second training set until the recognition accuracy of the trained second GoogleNet network reaches the second accuracy threshold.

6. The method for identifying overcurrent of carbon brushes of a phase regulator based on GoogleNet network according to claim 4 or 5, characterized in that: The expression of the loss function is: ； in, is the loss error, m is the number of input time-frequency graphs; is the true label of the i-th input time-frequency graph; is the predicted label of the i-th input frequency spectrum.

7. The method for identifying overcurrent of carbon brushes of a phase regulator based on GoogleNet network according to claim 1 is characterized in that: The identification method also includes carbon brush abnormality diagnosis, which specifically includes: While collecting the current current data of the carbon brush, collect the current temperature data of the carbon brush; When the carbon brush is in an overcurrent state, the temperature rise rate of the carbon brush is calculated based on the current temperature data of the carbon brush; Determine whether the temperature rise rate of the carbon brush exceeds the temperature rise rate threshold; if so, the carbon brush is in an abnormal state; otherwise, the carbon brush may have a hidden danger of an abnormal state.

8. A phase regulator carbon brush overcurrent identification system based on GoogleNet network, characterized in that: The identification system comprises: A data acquisition module is used to acquire current data of different carbon brushes in the phase condenser and generate a source domain data set and a target domain data set; wherein the current data in the source domain data set is the current experimental data of a part of the carbon brushes in the phase condenser, and the current data in the target domain data set is the current measured data of another part of the carbon brushes in the phase condenser; A data processing module, used to process the current data in the source domain data set and the target domain data set to obtain a time-frequency diagram, and then obtain a source domain time-frequency image set and a target domain time-frequency image set; A model training module is used to deploy a first GoogleNet network and a second GoogleNet network of the same architecture; train the first GoogleNet network using the source domain time-frequency image set; migrate the parameters of the trained first GoogleNet network to the second GoogleNet network, and then train the second GoogleNet network using the target domain time-frequency image set to obtain a carbon brush overcurrent recognition model; The abnormality diagnosis module is used to perform overcurrent identification on the current current data of the carbon brush using the carbon brush overcurrent identification model to obtain an overcurrent identification result of the carbon brush.

9. An electronic device comprising a memory, a processor, and a computer program / instruction stored in the memory, characterized in that: The processor executes the computer program / instruction to implement the phase regulator carbon brush overcurrent identification method based on the GoogleNet network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the method for identifying overcurrent of carbon brushes of a phase regulator based on a GoogleNet network as described in any one of claims 1 to 7 is implemented.

Images