Far and near field multi-axis sensing array coupled converter transformer monitoring method and system

Abstract

The invention provides a far and near field multi-axis sensing array coupled converter transformer monitoring method and system. The system comprises a near field four-axis sensing array, a far field four-axis sensing array, a signal-to-noise improving unit, an original point processing unit, an AD conversion unit and a background display control seat unit. Acquiring near-field and far-field noise signals without excitation, and coupling the near-field noise signals; taking the far-field noise signal as input, taking the coupled near-field noise signal as output, and training a deep network model; acquiring a to-be-detected signal and a field noise signal under excitation, and coupling the to-be-detected signal; taking the field noise signal as the input of the trained deep network model to obtain a reconstructed coupling noise signal; the method comprises the following steps: obtaining a denoised signal to be detected and a denoised coupling signal by using a coupling noise signal, carrying out signal processing to obtain a corresponding analog signal, converting the analog signal into a corresponding digital signal, and carrying out visual monitoring. The fault recognition rate and the operation and maintenance efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system fault detection, and in particular to a converter transformer monitoring method and system using a far-field and near-field multi-axis sensing array coupling under low signal-to-noise ratio for converter transformer multi-parameter fault monitoring. Background Art

[0002] As the core hub equipment for voltage conversion and power distribution in the power system, power converters are crucial for the stable operation of the power grid. A single unit is often worth millions to tens of millions of yuan, and a failure can lead to the following consequences:

[0003] (1) Economic losses of electricity: A single unplanned outage can cause economic losses of tens to hundreds of thousands of yuan per hour, and the fault repair cycle can take weeks to months.

[0004] (2) System stability risk: Converter transformer failure may trigger a chain reaction, leading to regional power grid disconnection or voltage collapse (for example, in the 2021 Texas blackout in the United States, converter transformer failure exacerbated the power grid paralysis).

[0005] (3) Safety hazards: Short-circuit faults can cause explosions due to a sudden rise in oil temperature (such as the 2019 Sao Paulo substation explosion in Brazil), endangering personnel safety.

[0006] In the existing technology, although a variety of diagnostic systems have been developed, the following technical bottlenecks still exist:

[0007] (1) Oil chromatography three-ratio method: The misjudgment rate of compound faults (such as overheating + discharge) is as high as 30%, and the fault location cannot be located.

[0008] (2) Vibration threshold method: The existing standard (such as 100Hz energy proportion > 80%) is difficult to adapt to different types of converter transformers. During normal operation of a certain 330kV converter transformer, the 200Hz component can account for up to 40%.

[0009] (3) Dependence on manual experience: 90% of grassroots operation and maintenance units still rely on expert experience and judgment, resulting in poor diagnostic timeliness (average time > 48 hours).

[0010] In addition, current research mainly focuses on relatively ideal and stable magnetic interference environments. Eliminating strong interference and high transient noise in actual complex working conditions remains a difficult and challenging problem. Summary of the Invention

[0011] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method and system for monitoring a commutation transformer with a far-field and near-field multi-axis sensing array coupled under low signal-to-noise ratio.

[0012] According to one aspect of the present invention, a method for monitoring a commutation transformer coupled with a near- and far-field multi-axis sensing array is provided, comprising:

[0013] In a non-excitation scenario, respectively acquiring a near-field noise signal and a far-field noise signal, coupling the near-field noise signal to obtain a coupled near-field noise signal to form training data;

[0014] Providing a deep network model combining a noise-guided network and a temporal convolutional network, taking the far-field noise signal as input and the coupled near-field noise signal as output, and training the deep network model;

[0015] In an excitation scenario, respectively obtaining a multi-source detection signal of a commutation transformer and a field noise signal in a strong noise environment, and coupling the multi-source detection signal of the commutation transformer to obtain a coupled signal;

[0016] Using the on-site noise signal as input to the trained deep network model, and outputting a reconstructed coupled noise signal;

[0017] Subtracting the coupled noise signal from the commutation transformer multi-source detection signal and the coupled signal respectively to obtain a denoised detection signal and a coupled signal;

[0018] Performing signal processing on the de-noised signal to be detected and the coupled signal to obtain a corresponding analog signal;

[0019] The analog signal is converted into a corresponding digital signal and visually monitored.

[0020] According to another aspect of the present invention, a near-field and far-field multi-axis sensor array coupled commutation transformer monitoring system is provided, comprising: a near-field four-axis sensor array unit, a far-field four-axis sensor array unit, a signal-to-noise enhancement unit, an origin processing unit, an AD conversion unit, and a background display and control seat unit; wherein:

[0021] The near-field four-axis sensor array unit collects noise signals for constructing a training data set as model output data without excitation; and collects multi-source signals to be detected from the commutation transformer in a strong noise environment under excitation and transmits them to the origin processing unit;

[0022] The far-field four-axis sensing array unit collects noise signals for constructing a training data set without excitation, and the noise signals serve as model input data in the training data; and captures field noise signals under excitation and transmits them to the signal-to-noise enhancement unit;

[0023] The signal-to-noise enhancement unit uses the training data set to train a deep network model that combines a noise guidance network and a temporal convolutional network, and uses the on-site noise signal as the input of the trained deep network model to reconstruct the coupled noise signal of the origin processing unit and transmit it to the origin processing unit;

[0024] The origin processing unit couples the noise signal collected by the near-field four-axis sensor array unit to construct a training data set, wherein the coupled noise signal serves as the model output data in the training data; couples the received multi-source signal to be detected from the converter transformer, and uses the coupled noise signal reconstructed by the signal-to-noise enhancement unit to obtain a denoised signal to be detected and a coupled signal; performs signal processing on the denoised signal to be detected and the coupled signal to obtain a corresponding analog signal and output it to the AD conversion unit;

[0025] The AD conversion unit converts the analog signal into a corresponding digital signal and outputs it to the background display and control seat unit;

[0026] The background display and control seat unit issues monitoring instructions and / or visualizes the digital signal.

[0027] Due to the adoption of the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art:

[0028] The commutation transformer monitoring method and system coupled with near- and far-field multi-axis sensor arrays provided by the present invention deploys near-field and far-field four-axis sensor arrays, uses the noise signals captured by the far-field four-axis sensor array to train a deep learning model, dynamically reconstructs and eliminates the noise components of the origin processing unit, and significantly improves the signal-to-noise ratio under low signal-to-noise ratio conditions.

[0029] The commutation transformer monitoring method and system provided by the present invention, which couples far-field and near-field multi-axis sensor arrays, combines a noise guidance network (CNN) with a temporal convolutional network (TCN), introduces generalized Gaussian noise to enhance the generalization capability of the model, achieves accurate modeling and separation of complex noise, and avoids the misjudgment problem of traditional threshold methods.

[0030] The present invention provides a commutation transformer monitoring method and system coupled with a far-field and near-field multi-axis sensing array. The system collects four-channel signals through a near-field four-axis sensor (partial discharge, vibration, and dual magnetic field) and couples them into an equally weighted integrated signal (coupling weight 0.25). Combined with the far-field four-axis noise data, the system realizes the synchronous analysis of multi-source parameters of magnetic field, sound wave, and vibration.

[0031] The commutator transformer monitoring method and system coupled with far- and near-field multi-axis sensor arrays provided by the present invention solve the accuracy and reliability problems of commutator transformer monitoring under low signal-to-noise ratio conditions through deep learning-driven noise suppression, multi-source signal fusion, robust hardware design, and dynamic diagnosis mechanisms. This significantly improves the fault recognition rate and operation and maintenance efficiency, and is suitable for the health management of key equipment in complex power grid environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0033] Figure 1 This is an organizational diagram of a commutation transformer monitoring system coupled with a far-field and near-field multi-axis sensing array in a preferred embodiment of the present invention.

[0034] Figure 2 The figure is a workflow diagram of a commutation transformer monitoring method coupled with a far-field and near-field multi-axis sensing array in a preferred embodiment of the present invention.

[0035] Figure 3 Schematic diagram of an 8*8 planar array structure in a preferred embodiment of the present invention.

[0036] Figure 4 This is a diagram of the deep network model architecture that combines a noise-guided network and a temporal convolutional network in a preferred embodiment of the present invention.

[0037] Figure 5 Schematic diagram of the deep network model training process in a preferred embodiment of the present invention.

[0038] Figure 6 Schematic diagram of a monitoring instruction set in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. It should be noted that those skilled in the art may make various modifications and improvements without departing from the scope of the present invention, and these modifications and improvements fall within the scope of protection of the present invention.

[0040] Existing power converter fault diagnosis systems often suffer from the following technical issues: a high misjudgment rate for complex faults and an inability to locate the fault location; difficulty adapting to different types of converter transformers; and reliance on expert judgment, resulting in poor diagnostic timeliness. Furthermore, existing technologies primarily focus on relatively ideal and stable magnetic interference environments, and eliminating strong interference and high transient noise in complex actual operating conditions remains a difficult and challenging task.

[0041] In response to the above problems, an embodiment of the present invention provides a commutation transformer monitoring system coupled with a far-field and near-field multi-axis sensor array. The system addresses the problem of accurate denoising of magnetic field and acoustic wave signals under strong noise interference. First, an origin processing unit and a near-field four-axis sensor array unit are constructed to detect the multi-source parameters of the commutation transformer in a strong noise environment. The near-field four-axis sensor array unit is distributed in a ring configuration in the near-field sensing area of ​​the origin unit. Under the hardware synchronization trigger mechanism, the four-channel commutation transformer parameters are collected and transmitted back to the origin processing unit. The origin processing unit is responsible for coupling multi-source signals; then a deep neural network model combining a noise guidance network and a bidirectional long short-term memory network is constructed. , the constructed deep neural network model is trained using the noise signal captured by the far-field four-axis sensor array unit deployed at a long distance; in the multi-source parameter detection process of the converter transformer, the noise obtained by the far-field four-axis sensor array unit is used to reconstruct the noise component of the origin processing unit, and then the noisy signal received by the origin processing unit is subtracted from the reconstructed noise to obtain the denoised signal to be detected; then the coupled signal and the four-channel signal to be detected are clipped, filtered, and amplified, and then the overvoltage and overcurrent protection are performed to obtain the analog signal output result; based on the four-channel signal to be detected and the coupled signal, the partial discharge fault of the converter transformer, the internal vibration fault of the winding, etc. are simultaneously judged.

[0042] Specifically, if Figure 1 As shown, the commutation transformer monitoring system coupled with near- and far-field multi-axis sensor arrays provided in this embodiment may include: a near-field four-axis sensor array unit, a far-field four-axis sensor array unit, a signal-to-noise enhancement unit, an origin processing unit, an AD conversion unit, and a background display and control seat unit; wherein:

[0043] The near-field four-axis sensor array unit collects noise signals for constructing a training data set as model output data without excitation; under excitation, it collects multiple-source signals to be detected from the commutation transformer in a strong noise environment and transmits them to the origin processing unit;

[0044] The far-field four-axis sensor array unit collects noise signals for constructing a training data set without excitation. This signal serves as the model input data in the training data. Under excitation, it captures the on-site noise signal and transmits it to the signal-to-noise enhancement unit.

[0045] The signal-to-noise enhancement unit uses the training data set to train a deep network model that combines a noise guidance network and a temporal convolutional network. The on-site noise signal is used as the input of the trained deep network model to reconstruct the coupling noise signal of the origin processing unit and transmit it to the origin processing unit.

[0046] The origin processing unit couples the noise signal collected by the near-field four-axis sensor array unit to construct a training data set, wherein the coupled noise signal serves as the model output data in the training data; couples the received multi-source signal to be detected from the commutation transformer, and uses the coupled noise signal reconstructed by the signal-to-noise enhancement unit to obtain a denoised signal to be detected and a coupled signal; performs signal processing on the denoised signal to be detected and the coupled signal to obtain a corresponding analog signal and output it to the AD conversion unit;

[0047] AD conversion unit converts analog signals into corresponding digital signals and outputs them to the backstage display and control seat unit;

[0048] The background display and control seat unit issues monitoring instructions and / or visual digital signals.

[0049] In some preferred embodiments, the above system may further include:

[0050] The origin processing unit couples the noise signal collected by the near-field four-axis sensor array unit, and the coupled noise signal is used as the model output data in the training data.

[0051] In some preferred embodiments, the above system may further include:

[0052] A planar array structure is defined, where the origin processing unit is deployed at the geometric center of the planar array structure, the near-field four-axis sensing array units are distributed in a ring configuration in the near-field sensing area of ​​the origin processing unit, and the far-field four-axis sensing array units are deployed at the four corner points of the planar array structure.

[0053] In some preferred embodiments, the near-field four-axis sensing array unit and the far-field four-axis sensing array unit both include: a partial discharge MEMS ultrasonic probe, a vibration detection probe, and two tunnel magnetoresistive sensors; wherein:

[0054] Partial discharge MEMS ultrasonic probe, used to collect partial discharge ultrasonic electrical signals or noise signals;

[0055] Vibration detection probe, used to collect vibration signals or noise signals;

[0056] Two tunnel magnetoresistive sensors are used to collect magnetic field electrical signals or noise signals respectively.

[0057] In some preferred embodiments, the signal-to-noise enhancement unit may further include a deep network model combining a noise guidance network and a temporal convolutional network, and may further include: an input layer, a convolutional layer, a pooling layer, a fully connected layer, a noise guidance layer, a TCN layer, and an output layer; wherein:

[0058] An input layer for receiving a time series noise signal collected by a far-field four-axis sensor array; further preferably, the time series noise signal is 200×7×4, including a time step of 200, 7 feature dimensions, and 4 sensor channels;

[0059] A convolution layer is used to extract spatial features through a convolution kernel, perform a two-dimensional convolution operation on the input signal, and output a feature map; further preferably, the convolution kernel is 3×3×32, and the output feature map is a 198×5×32 feature map with 32 channels;

[0060] The pooling layer uses a maximum pooling operation to downsample the feature map. Further preferably, the feature map is downsampled by 2×2 to compress the feature map dimension to 99×2×32, retaining significant features and reducing computational complexity.

[0061] The fully connected layer flattens the pooled features, performs nonlinear transformation through the weight matrix, and outputs high-level semantic features. Preferably, the flattening is performed into a 63552-dimensional vector, the weight matrix is ​​32×32×1, and the output semantic features are 32-dimensional.

[0062] The noise guidance layer introduces generalized Gaussian noise into the feature space to perturb high-level semantic features and enhance the robustness of the model to noise distribution. Further preferably, the 32-dimensional high-level semantic features are perturbed with a probability of 0.25.

[0063] The TCN layer uses a temporal convolution kernel to model temporal features through causal convolution and output temporal correlation features. Preferably, a 32-channel temporal convolution kernel with a dilation factor of 2 is used, and the output features are 64×1 dimensions.

[0064] The output layer maps the TCN layer output to a coupled noise signal through a linear transformation to reconstruct the noise component of the origin processing unit; further preferably, the coupled noise signal is 64×1 dimensional.

[0065] In some preferred embodiments, the noise guidance layer introduces generalized Gaussian noise to improve the model generalization ability and reconstruction accuracy; wherein the generalized Gaussian noise can be further expressed as:

[0066]

[0067] Where G a,β (x) is the noise value of the x-th neuron; Γ() is the Gamma function; α is the scale parameter, which ranges from [1,∞]; β is the edge parameter, which ranges from [0,1].

[0068] Furthermore, in the above preferred embodiment, the parameters of the input layer can be 200*7*4, the parameters of the convolutional layer can be 3*3*32, the parameters of the fully connected layer can be 32*32*1, the parameters of the TCN layer can be 32*32*2, and the parameters of the output layer can be 64*1.

[0069] In some preferred embodiments, the origin processing unit may further include: a coupling unit, a filtering unit, a clipping unit, a gain unit, a protection unit, and a matching unit connected in sequence; wherein:

[0070] A coupling unit is used to couple the commutation variable multi-source detection signals collected by the near-field four-axis sensor array unit to obtain a coupled signal; further, in a preferred embodiment, the coupling weights used by the coupling unit can be 0.25 respectively;

[0071] A filtering unit is used to subtract the coupled noise signal reconstructed by the signal-to-noise enhancement unit from the commutation transformer multi-source to-be-detected signal and the coupled signal, respectively, to obtain the denoised to-be-detected signal and the coupled signal;

[0072] A clipping unit is used to reduce the amplitude of the denoised signal to be detected and the coupled signal to achieve a controllable range;

[0073] A gain unit, used for amplifying the clipped signal to be detected and the coupled signal;

[0074] A voltage stabilizing unit, used for stabilizing the input voltage of the input power;

[0075] Protection unit, used to protect the amplified signal to be detected and the coupled signal from overvoltage (threshold set to 12V) and overcurrent (threshold set to 20mA)

[0076] The matching unit is used to perform signal impedance matching on the over-protected signal and output an analog signal.

[0077] In a specific application example, the coupling unit model can be zynq-7z100; the gain unit can use the TIOPA2188 amplifier, and the offset voltage threshold can be set to 0.5μV; the voltage regulator unit can use Shengbang Micro SGM2036, and the voltage after regulation can be 3.3V; the protection unit can use the Chint NZ8 series, and the safety output contact capacity can be set to 6A.

[0078] In the case of the above constituent units, further, in a specific application example, the planar array structure can be defined as an 8*8 planar array structure, the origin processing unit is deployed at the geometric center (coordinates are [4,4]), the deployment coordinates of the partial discharge MEMS ultrasonic probe, vibration detection probe and two tunnel magnetoresistive sensors of the near-field four-axis sensing unit are [3,3], [4,5], [4,3], [5,5] (coordinate system grid units), and the partial discharge MEMS ultrasonic probe, vibration detection probe and two tunnel magnetoresistive sensors of the far-field four-axis sensing array are deployed at four corner points respectively.

[0079] Based on the same inventive concept, an embodiment of the present invention further provides a commutation transformer monitoring method coupled with far-field and near-field multi-axis sensing arrays.

[0080] Specifically, if Figure 2 As shown, the commutation transformer monitoring method provided by this embodiment using a far-field and near-field multi-axis sensing array coupling may include:

[0081] S1, in a non-excitation scenario, obtain near-field noise signals and far-field noise signals respectively, couple the near-field noise signals, and obtain coupled near-field noise signals to form training data;

[0082] S2, providing a deep network model that combines a noise-guided network and a temporal convolutional network, taking the far-field noise signal as input and the coupled near-field noise signal as output, and training the deep network model;

[0083] S3, in an excitation scenario, respectively obtaining a multi-source detection signal of the commutation transformer in a strong noise environment and a field noise signal, and coupling the multi-source detection signal of the commutation transformer to obtain a coupled signal;

[0084] S4, takes the on-site noise signal as the input of the trained deep network model and outputs the reconstructed coupled noise signal;

[0085] S5, subtracting the coupled noise signal from the multi-source to-be-detected signal and the coupled signal of the converter transformer to obtain the denoised to-be-detected signal and the coupled signal;

[0086] S6, performing signal processing on the denoised signal to be detected and the coupled signal to obtain a corresponding analog signal;

[0087] S7 converts the analog signal into the corresponding digital signal and performs visual monitoring.

[0088] In some preferred embodiments, the above may further include:

[0089] Couple the near-field noise signal obtained in S1 to obtain a coupled near-field noise signal;

[0090] In S2, the coupled near-field noise signal is used as output data to train the deep network model.

[0091] In some preferred embodiments, the above S2, combining the deep network model of the noise-guided network and the temporal convolutional network, may further include: an input layer, a convolutional layer, a pooling layer, a fully connected layer, a noise-guided layer, a TCN layer, and an output layer; wherein:

[0092] Input layer, used to receive far-field noise signals;

[0093] Convolution layer, used to extract spatial features of far-field noise signals through convolution kernels and output feature maps;

[0094] Pooling layer, which uses the maximum pooling operation to downsample the feature map;

[0095] The fully connected layer is used to flatten the pooled feature map and perform nonlinear transformation through the weight matrix to output high-level semantic features;

[0096] The noise guidance layer is used to introduce generalized Gaussian noise into the feature space to perturb high-level semantic features;

[0097] The TCN layer uses a temporal convolution kernel to model the temporal features of the perturbed high-level semantic features through causal convolution and outputs temporal correlation features;

[0098] The output layer is used to map the temporal correlation feature output into a coupled noise signal through linear transformation and reconstruct the coupled noise signal.

[0099] In some preferred embodiments, the noise guidance layer introduces generalized Gaussian noise to improve the generalization ability of the model and the reconstruction accuracy; wherein the generalized Gaussian noise is expressed as:

[0100]

[0101] Where G a,β (x) is the noise value of the x-th neuron; Γ() is the Gamma function; α is the scale parameter, which ranges from [1,∞]; β is the edge parameter, which ranges from [0,1].

[0102] In some preferred embodiments, the above S6, performing signal processing on the denoised signal to be detected and the coupled signal, may further include:

[0103] S61, reducing the amplitude of the denoised signal to be detected and the coupled signal to achieve a controllable range;

[0104] S62, amplifying the clipped signal to be detected and the coupled signal;

[0105] S63 , performing overvoltage and overcurrent protection on the amplified signal to be detected and the coupled signal, completing signal processing, and obtaining a corresponding analog signal.

[0106] In some preferred embodiments, the above method may further include:

[0107] S8, based on the signal to be detected and the coupled signal in the digital signal, the partial discharge fault of the commutation transformer and the internal vibration fault of the winding are judged at the same time, and a monitoring instruction is issued.

[0108] In some preferred embodiments, the above S8, determining the commutation transformer partial discharge fault and the winding internal vibration fault, may further include:

[0109] S81: When the peak value of the monitored coupling signal exceeds the range of -5V to +5V, it indicates a fault and requires maintenance.

[0110] S82, when the peak value of the monitored coupling signal does not exceed the range of -5v to +5v, the peak value of the signal to be detected is judged; if the peak value of the ultrasonic partial discharge signal in the signal to be detected is in the range of -3.3v to +3.3v, it is considered that the solid insulation degradation fault is caused; if the peak value of the vibration signal in the signal to be detected is in the range of -2.8v to +2.8v, it is considered that the iron core is loose or the winding is deformed; if the average peak value of the magnetic field signal in the signal to be detected is in the range of -3.1v to +3.1v, it is considered that the contact is poor or the winding is broken.

[0111] In some preferred embodiments, the above S8, monitoring instructions, may further include: starting noise baseline acquisition, starting signal-to-noise enhancement model training, starting four-channel detection of the converter transformer, starting real-time noise reconstruction, performing coupled signal denoising, enabling signal conditioning and protection, and / or performing comprehensive fault diagnosis.

[0112] In some preferred embodiments, the above-mentioned coupled signal denoising may further include: a coupling instruction, a clipping instruction, and a filtering instruction;

[0113] The aforementioned enabling signal conditioning and protection may further include: a gain instruction and an overvoltage protection instruction;

[0114] The above-mentioned comprehensive fault diagnosis may further include coupling comprehensive analysis, partial discharge threshold verification, vibration spectrum analysis, magnetic field imbalance detection, coupling signal weight calibration and sending operation and maintenance warnings.

[0115] It should be noted that the steps in the method provided by the present invention can be implemented by using the corresponding components in the system. Those skilled in the art can refer to the technical solution of the system to implement the step flow of the method, and can also refer to the technical solution of the method to implement the composition of the system. That is, the embodiments in the system and the embodiments in the method can be understood as preferred examples of each other, and will not be elaborated here.

[0116] The technical solution provided by the above embodiment of the present invention is further described in detail below with reference to a specific application example and the accompanying drawings.

[0117] like Figure 1 As shown in the figure, the commutation transformer monitoring system with near-field and far-field multi-axis sensor array coupling under low signal-to-noise ratio involved in this specific application example includes a near-field four-axis sensor array unit, a far-field four-axis sensor array, a signal-to-noise enhancement unit, an origin processing unit, an AD conversion unit, and a background display and control unit. Among them:

[0118] Both the near-field and far-field four-axis sensing arrays include a partial discharge MEMS ultrasonic probe, a vibration detection probe, and two tunnel magnetoresistive sensors.

[0119] The signal-to-noise enhancement unit includes a deep network model that combines a noise-guided network and a temporal convolutional network (TCN). The constructed network model is trained using four-channel noise signals captured by a far-field four-axis sensor array. During the multi-source parameter detection of the converter transformer, the noise component of the origin processing unit is reconstructed using the noise of the far-field four-axis sensor array. The reconstructed noise is then subtracted from the noisy signal received by the origin processing unit to obtain the denoised signal to be detected.

[0120] The origin processing unit includes a coupling unit, a filtering unit, a clipping unit, a gain unit, a protection unit and a matching unit connected in sequence. Among them:

[0121] The coupling unit is used to couple the four-channel detection signals consisting of the partial discharge ultrasonic electrical signal, vibration signal and two groups of magnetic field electrical signals collected by the near-field four-axis sensor array unit, with the coupling weights of 0.25 respectively.

[0122] The filtering unit performs a process of subtracting the coupling noise reconstructed by the signal-to-noise enhancement unit from the four-channel detection signals and the coupling signal.

[0123] The clipping unit is used to reduce the amplitude of the filtered four-channel detection signal and the coupled signal to achieve a controllable range.

[0124] The gain unit is used to amplify the filtered four-channel detection signals and the coupled signal.

[0125] The voltage stabilizing unit is used to stabilize the input power supply.

[0126] The protection unit and the matching unit are used to perform overvoltage and overcurrent protection on the amplified four-channel detection signals and the coupled signals and then output analog signals.

[0127] The AD conversion unit is used to convert analog signals into digital signals and connect to the background display and control seat unit.

[0128] The background display and control seat unit is used to issue instructions such as data collection, early warning analysis, and visualize the four-channel test signals and coupled signals after AD conversion.

[0129] In order to determine the optimal measurement position of each sensor under noise interference, considering the size of the actual sensor acquisition module, the non-magnetic base used to arrange the auxiliary sensor array is evenly divided into grids and the following grids are established: Figure 3 The coordinate system shown defines an 8x8 planar array structure, with the origin processing unit deployed at the geometric center (coordinates [4,4]). Utilizing the advantages of high computational accuracy and rapid modeling of finite element simulation, simulations were performed using the computer software ANSYS-Maxwell. The near-field sensor array was designed using the detection point most similar to the spatial signal at the origin processing unit. The simulated signals from these 6x6 measurement locations were used as the near-field four-axis sensor array signals and data screening was performed. The Flamenche distance was introduced to perform one-dimensional curve similarity matching between the origin processor signal and the near-field four-axis array signal. The coordinates for the partial discharge MEMS ultrasonic probe, vibration detection probe, and two tunnel magnetoresistive sensors of the near-field four-axis sensor array unit were determined to be [3,3], [4,5], [4,3], and [5,5] (coordinate system grid units). The MEMS ultrasonic probe, vibration detection probe, and two tunnel magnetoresistive sensors of the far-field four-axis sensor array were deployed at the four corner points.

[0130] like Figure 4 As shown in the figure, the network model in the signal-noise enhancement unit includes an input layer (200*7*4), a convolutional layer (3*3*32), a pooling layer, a fully connected layer (32*32*1), a noise guidance layer, a TCN layer (32*32*2), and an output layer (64*1).

[0131] Convolutional neural networks use convolution operations to extract local features, while reducing the number of parameters and computational complexity through operations such as weight sharing and pooling. This can better extract the internal features of the data, reduce the risk of inter-layer connections and overfitting in the network, and improve classification accuracy. Noise guidance can be applied to CNNs to increase their classification accuracy. The noise-enhanced CNN module we use improves the generalization ability and classification accuracy of CNNs by injecting generalized Gaussian noise into them.

[0132] The injected generalized Gaussian noise formula is:

[0133]

[0134] Among them, Γ() is the Gamma function, α and β ranges are [1,∞] and [0,1], which are the scale parameter and edge parameter respectively, G a,β (x) is the noise value of the x-th neuron.

[0135] TCN uses dilated convolution to capture long-term dependencies, avoid the gradient vanishing problem, and meet lightweight requirements. It is conducive to real-time monitoring, its structure supports parallelization, and its inference speed is better than that of recurrent networks.

[0136] The network model training data set in the signal-to-noise enhancement unit is the four-channel noise signal captured by the far-field four-axis sensor array without excitation, and the output is the coupled noise signal of the four-channel noise signal captured by the near-field four-axis sensor array without excitation; in the commutation transformer parameter detection, the input of the network model is the four-channel noise signal captured by the far-field four-axis sensor array under excitation, and the output is the coupled noise signal of the reconstructed origin processing unit.

[0137] Under no excitation, the near-field and far-field four-axis sensor arrays collect noise signals and use the coupled noise signal of the origin processing unit coupled to the near-field four-axis sensor array unit as output and the four-channel noise signal collected by the far-field four-axis sensor array as input to construct a training set. When obtaining the data set, the programmable power supply is turned off, the operation state of the converter is stopped, and the far-field four-axis sensor array noise signal and the coupled noise signal of the origin processing unit are obtained. The obtained unexcited and interfered training set is input into the proposed noise guidance network and TCN network to learn the mapping relationship between the origin coupling noise signal and the far-field four-axis sensor array noise signal. The training platform configuration of the model is as follows: CPUi9-10900F, GPURTX3060, operating system is WIN10-64 bit, model framework is Tensorflow, and the training process (training curve) of the model is as follows Figure 5 shown.

[0138] The model's loss function and root-mean-square error (RMSE) decrease with increasing iterations. These two metrics stabilize after approximately 30 iterations, indicating rapid model convergence. After multiple iterations, a trained network model is obtained that clearly reflects the mapping relationship between coupled noise and the spatial noise collected by the far-field four-axis sensor array.

[0139] Under excitation, the programmable power supply is turned on, the commutator enters the operating state, and the signals collected by the near-field four-axis sensor array are transmitted back to the origin processing unit. Due to the presence of noise, the coupled signals and four-channel detection signals detected are combination signals of pure signals and interference noise.

[0140] After the denoising results are subjected to signal clipping and signal amplification in turn, the amplified four-channel test signals and coupled signals are subjected to overvoltage and overcurrent protection to obtain the analog signal output result. The analog signal output signal is transmitted to the background display and control seat unit after AD conversion. The display and control seat unit simultaneously judges the partial discharge fault of the converter transformer and the internal vibration fault of the winding based on the four-channel test signals and coupled signals, and issues a series of monitoring instructions.

[0141] If the display and control seat unit detects that the coupling signal peak exceeds the range of -5v to +5v, it indicates that there is a fault and needs to be repaired; if the coupling signal peak does not exceed the range of -5v to +5v, the peak value of the four-channel test signal is judged. If the ultrasonic partial discharge signal peak is in the range of -3.3v to +3.3v, it is considered that the solid insulation degradation fault is caused; if the vibration signal peak is in the range of -2.8v to +2.8v, it is considered that the iron core is loose or the winding is deformed; if the average peak value of the two magnetic field signals is in the range of -3.1v to +3.1v, it is considered that the contact is poor or the winding is broken.

[0142] like Figure 6 As shown, the instructions issued by the display and control unit in sequence are: start noise baseline acquisition, start signal-to-noise improvement model training, start commutation transformer four-channel detection, start real-time noise reconstruction, perform coupled signal denoising and include sub-instructions such as coupling instruction, clipping instruction, and filtering instruction; enable signal conditioning and protection and include sub-instructions such as gain instruction and overvoltage protection instruction; perform comprehensive fault diagnosis and the following sub-instructions:

[0143] Coupling comprehensive analysis: judge the fault condition based on the coupling signal amplitude.

[0144] Partial discharge threshold verification: Determines the risk of solid insulation degradation based on the ultrasonic signal amplitude (-3.3V to +3.3V).

[0145] Vibration spectrum analysis: Combines the frequency domain characteristics of vibration signals (such as the characteristic frequency of loose core) to diagnose mechanical faults.

[0146] Magnetic field imbalance detection: Identify poor contact or winding deformation through the difference in dual magnetic field signals (X / Y axis).

[0147] Coupling signal weight calibration: Dynamically adjust the four-channel coupling weight (0.25 equal weight or adaptive weight) to optimize the overall signal reliability.

[0148] Issue operation and maintenance warnings: trigger warning instructions (such as shutdown for maintenance, gain adjustment, model retraining) based on the diagnosis results and push them to the operation and maintenance interface.

[0149] Based on the technical solution of the above-mentioned system, the steps of the commutation transformer monitoring method using the system to couple the corresponding far-field and near-field multi-axis sensor arrays include:

[0150] Step 1: Under no excitation (i.e., in the non-working state), the near-field and far-field four-axis sensor arrays collect noise signals and use the coupled noise signal of the origin processing unit coupled to the near-field four-axis sensor array unit as output and the four-channel noise signal collected by the far-field four-axis sensor array as input to construct a training set.

[0151] Step 2: Train the network model of the signal-to-noise boosting unit using the dataset constructed in step 1 and save it.

[0152] Step 3: Under stimulation (i.e., the compiler is in operation), the near-field four-axis sensor array unit collects the four-channel detection signals of the commutation parameters and transmits them back to the origin processing unit.

[0153] Step 4: Under excitation, the far-field four-axis sensor array collects the on-site noise signal and inputs it into the signal-to-noise enhancement unit to reconstruct the coupled noise signal of the origin processing unit.

[0154] Step 5: The four-channel No. 2 coupled signal of the origin processing unit is respectively subtracted from the coupled noise signal reconstructed in step 4 to obtain a denoising result.

[0155] Step 6: After performing signal clipping and signal amplification on the denoising results, the amplified four-channel test signal and coupled signal are protected against overvoltage and overcurrent to obtain the analog signal output result. The analog signal output signal is transmitted to the background display and control seat unit after AD conversion.

[0156] Step 7: The display and control seat unit determines the partial discharge fault of the commutation transformer and the internal vibration fault of the winding based on the four-channel detection signals and the coupling signal, and issues a series of monitoring instructions.

[0157] Furthermore, step 7 further includes:

[0158] If the display and control seat unit detects that the coupling signal peak exceeds the range of -5v to +5v, it indicates that there is a fault and needs to be repaired; if the coupling signal peak does not exceed the range of -5v to +5v, the peak value of the four-channel test signal is judged. If the ultrasonic partial discharge signal peak is in the range of -3.3v to +3.3v, it is considered that the solid insulation degradation fault is caused; if the vibration signal peak is in the range of -2.8v to +2.8v, it is considered that the iron core is loose or the winding is deformed; if the average peak value of the two magnetic field signals is in the range of -3.1v to +3.1v, it is considered that the contact is poor or the winding is broken.

[0159] Furthermore, the monitoring instructions issued by the background display and control seat unit include: starting noise baseline collection, starting signal-to-noise improvement model training, starting four-channel detection of the converter transformer, starting real-time noise reconstruction, performing coupled signal denoising, enabling signal conditioning and protection, and performing comprehensive fault diagnosis.

[0160] Furthermore, the coupled signal denoising includes sub-instructions including a coupling instruction, a clipping instruction, and a filtering instruction;

[0161] Furthermore, enabling signal conditioning and protection includes sub-instructions of a gain instruction and an overvoltage protection instruction;

[0162] Furthermore, the execution of comprehensive fault diagnosis includes sub-instructions such as coupling comprehensive analysis, partial discharge threshold verification, vibration spectrum analysis, magnetic field imbalance detection, coupling signal weight calibration, and issuing operation and maintenance warnings.

[0163] The above-mentioned embodiment of the present invention provides a converter transformer monitoring system and method coupled with a far-field and far-field multi-axis sensor array. By deploying a near-field and far-field four-axis sensor array, the noise signal captured by the far-field four-axis sensor array is used to train a deep learning model, dynamically reconstruct and eliminate the noise component of the origin processing unit, and significantly improve the signal-to-noise ratio under low signal-to-noise ratio. Combining the noise guidance network (CNN) with the temporal convolutional network (TCN), the generalized Gaussian noise enhancement model generalization capability is introduced to achieve accurate modeling and separation of complex noise, avoiding the misjudgment problem of the traditional threshold method. Four-channel signals are collected by near-field four-axis sensors (partial discharge, vibration, dual magnetic field), and coupled into a comprehensive signal with equal weight (coupling weight 0.25). Combined with the far-field four-axis noise data, the synchronous analysis of multi-source parameters of magnetic field, sound wave, and vibration is achieved. In summary, this system and method solves the accuracy and reliability problems of converter transformer monitoring under low signal-to-noise ratio through deep learning-driven noise suppression, multi-source signal fusion, robust hardware design and dynamic diagnosis mechanism, significantly improves fault recognition rate and operation and maintenance efficiency, and is suitable for key equipment health management in complex power grid environments.

[0164] Matters not mentioned in the above embodiments of the present invention are well known in the art.

[0165] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various variations or modifications within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A method for monitoring commutation transformers by coupling near- and far-field multi-axis sensing arrays, characterized in that: include: In a non-excitation scenario, respectively acquiring a near-field noise signal and a far-field noise signal, coupling the near-field noise signal to obtain a coupled near-field noise signal to form training data; Providing a deep network model combining a noise-guided network and a temporal convolutional network, taking the far-field noise signal as input and the coupled near-field noise signal as output, and training the deep network model; In an excitation scenario, respectively obtaining a multi-source detection signal of a commutation transformer and a field noise signal in a strong noise environment, and coupling the multi-source detection signal of the commutation transformer to obtain a coupled signal; Using the on-site noise signal as input to the trained deep network model, and outputting a reconstructed coupled noise signal; Subtracting the coupled noise signal from the commutation transformer multi-source detection signal and the coupled signal respectively to obtain a denoised detection signal and a coupled signal; Performing signal processing on the de-noised signal to be detected and the coupled signal to obtain a corresponding analog signal; The analog signal is converted into a corresponding digital signal and visually monitored.

2. The commutation transformer monitoring method of the far-field and near-field multi-axis sensing array coupling according to claim 1 is characterized in that: The deep network model combining the noise-guided network and the temporal convolutional network includes: an input layer, a convolutional layer, a pooling layer, a fully connected layer, a noise-guided layer, a TCN layer, and an output layer; wherein: The input layer is used to receive far-field noise signals; The convolution layer is used to extract spatial features of the far-field noise signal through a convolution kernel and output a feature map; The pooling layer downsamples the feature map using a maximum pooling operation; The fully connected layer is used to flatten the pooled feature map and perform nonlinear transformation through the weight matrix to output high-level semantic features; The noise guidance layer is used to introduce generalized Gaussian noise into the feature space to perturb the high-level semantic features; The TCN layer uses a temporal convolution kernel to perform temporal feature modeling on the perturbed high-level semantic features through causal convolution and outputs temporal correlation features; The output layer is used to map the time series correlation feature output into a coupled noise signal through linear transformation, and reconstruct the coupled noise signal.

3. The commutation transformer monitoring method of the near-field and far-field multi-axis sensing array coupling according to claim 2 is characterized in that: The noise guidance layer introduces generalized Gaussian noise to improve the generalization ability and reconstruction accuracy of the model; wherein the generalized Gaussian noise is expressed as: Where G a,β (x) is the noise value of the x-th neuron; Γ() is the Gamma function; α is the scale parameter, which ranges from [1,∞]; β is the edge parameter, which ranges from [0,1].

4. The commutation transformer monitoring method of the far-field and near-field multi-axis sensing array coupling according to claim 1 is characterized in that: The signal processing of the denoised signal to be detected and the coupled signal includes: reducing the amplitude of the denoised signal to be detected and the coupled signal to achieve a controllable range; Amplifying the clipped signal to be detected and the coupled signal; The amplified signal to be detected and the coupled signal are protected against overvoltage and overcurrent, and signal processing is completed to obtain a corresponding analog signal.

5. The commutation transformer monitoring method using a far-field and near-field multi-axis sensing array coupling according to any one of claims 1 to 4, characterized in that: Also includes: According to the signal to be detected and the coupled signal in the digital signal, the partial discharge fault of the commutation transformer and the internal vibration fault of the winding are judged at the same time, and a monitoring instruction is issued.

6. The commutation transformer monitoring method of the near-field and far-field multi-axis sensing array coupling according to claim 5 is characterized in that: The method of determining a commutation transformer partial discharge fault and a winding internal vibration fault includes: When the peak value of the monitored coupling signal exceeds the range of -5v to +5v, it indicates that there is a fault and maintenance is required; When the peak value of the monitored coupling signal does not exceed the range of -5v to +5v, the peak value of the signal to be detected is judged; if the peak value of the ultrasonic partial discharge signal in the signal to be detected is in the range of -3.3v to +3.3v, it is considered that the solid insulation degradation fault is caused; if the peak value of the vibration signal in the signal to be detected is in the range of -2.8v to +2.8v, it is considered that the iron core is loose or the winding is deformed; if the average peak value of the magnetic field signal in the signal to be detected is in the range of -3.1v to +3.1v, it is considered that the contact is poor or the winding is broken.

7. The commutation transformer monitoring method of the near-field and far-field multi-axis sensing array coupling according to claim 5 is characterized in that: The monitoring instructions include: starting noise baseline acquisition, starting signal-to-noise enhancement model training, starting four-channel detection of the converter transformer, starting real-time noise reconstruction, performing coupled signal denoising, enabling signal conditioning and protection, and / or performing comprehensive fault diagnosis.

8. The commutation transformer monitoring method of the near-field and far-field multi-axis sensing array coupling according to claim 7 is characterized in that: The execution of coupled signal denoising includes: coupling instructions, clipping instructions and filtering instructions; The enabling signal conditioning and protection includes: a gain instruction and an overvoltage protection instruction; The comprehensive fault diagnosis includes coupling comprehensive analysis, partial discharge threshold verification, vibration spectrum analysis, magnetic field imbalance detection, coupling signal weight calibration and issuing operation and maintenance warnings.

9. A commutation transformer monitoring system coupled with a near-field and far-field multi-axis sensor array, characterized in that: include: Near-field four-axis sensor array unit, far-field four-axis sensor array unit, signal-to-noise enhancement unit, origin processing unit, AD conversion unit and background display and control seat unit; Among them: The near-field four-axis sensor array unit collects noise signals for constructing a training data set as model output data without excitation; and collects multi-source signals to be detected from the commutation transformer in a strong noise environment under excitation and transmits them to the origin processing unit; The far-field four-axis sensing array unit collects noise signals for constructing a training data set without excitation, and the noise signals serve as model input data in the training data; and captures field noise signals under excitation and transmits them to the signal-to-noise enhancement unit; The signal-to-noise enhancement unit uses the training data set to train a deep network model that combines a noise guidance network and a temporal convolutional network, and uses the on-site noise signal as the input of the trained deep network model to reconstruct the coupled noise signal of the origin processing unit and transmit it to the origin processing unit; The origin processing unit couples the noise signal collected by the near-field four-axis sensor array unit to construct a training data set, wherein the coupled noise signal serves as the model output data in the training data; couples the received multi-source signal to be detected from the converter transformer, and uses the coupled noise signal reconstructed by the signal-to-noise enhancement unit to obtain a denoised signal to be detected and a coupled signal; performs signal processing on the denoised signal to be detected and the coupled signal to obtain a corresponding analog signal and output it to the AD conversion unit; The AD conversion unit converts the analog signal into a corresponding digital signal and outputs it to the background display and control seat unit; The background display and control seat unit issues monitoring instructions and / or visualizes the digital signal.

10. The commutation transformer monitoring system coupled with near- and far-field multi-axis sensing arrays according to claim 9, characterized in that: A planar array structure is defined, wherein the origin processing unit is deployed at the geometric center of the planar array structure, the near-field four-axis sensing array units are distributed in a ring configuration in the near-field sensing area of ​​the origin processing unit, and the far-field four-axis sensing array units are respectively deployed at the four corner points of the planar array structure.

11. The commutation transformer monitoring system coupled with a near- and far-field multi-axis sensing array according to claim 9, characterized in that: The near-field four-axis sensing array unit and the far-field four-axis sensing array unit both include: a partial discharge MEMS ultrasonic probe, a vibration detection probe, and two tunnel magnetoresistive sensors; wherein: The partial discharge MEMS ultrasonic probe is used to collect partial discharge ultrasonic electrical signals or noise signals; The vibration detection probe is used to collect vibration signals or noise signals; The two tunnel magnetoresistive sensors are used to collect magnetic field electrical signals or noise signals respectively.

12. The commutation transformer monitoring system coupled with near- and far-field multi-axis sensing arrays according to claim 9, characterized in that: The origin processing unit includes: a coupling unit, a filtering unit, a clipping unit, a gain unit, a protection unit and a matching unit connected in sequence; wherein: The coupling unit is used to couple the commutation transformer multi-source detection signals collected by the near-field four-axis sensing array unit to obtain a coupled signal; The filtering unit is configured to subtract the coupled noise signal reconstructed by the signal-to-noise enhancement unit from the commutation transformer multi-source to-be-detected signal and the coupled signal, respectively, to obtain the denoised to-be-detected signal and the coupled signal; The clipping unit is used to reduce the amplitude of the denoised signal to be detected and the coupled signal to achieve a controllable range; The gain unit is used to amplify the clipped signal to be detected and the coupled signal; The voltage stabilizing unit is used to stabilize the input power supply; The protection unit is used to protect the amplified signal to be detected and the coupled signal from overvoltage and overcurrent; The matching unit is used to perform signal impedance matching on the over-protected signal and output an analog signal.

Images