Cable joint partial discharge magnetic signal denoising method and system based on asymmetric magnetic sensing

Abstract

The application discloses a cable joint partial discharge magnetic signal denoising method and system based on asymmetric magnetic sensing, and the method comprises the following steps: deploying an asymmetric magnetic sensing module array at a cable joint, including a hub magnetic sensing module and a plurality of branch magnetic sensing modules, wherein the hub magnetic sensing module is closest to the cable joint; when the cable is in a non-working state, training a neural network by using magnetic field noise signals captured by the branch magnetic sensing modules and the hub magnetic sensing module to obtain a noise guide network model; when the cable is in a working state, inputting magnetic field noise signals captured by the branch magnetic sensing modules into the noise guide network model to predict the magnetic field noise captured by the hub magnetic sensing module, subtracting the predicted noise from the noise-containing magnetic signals captured by the hub magnetic sensing module to obtain denoised cable joint partial discharge magnetic signals. The application can realize dynamic and accurate denoising of partial discharge signals under strong magnetic field interference of the cable joint.

Description

Technical Field

[0001] This invention belongs to the field of cable joint discharge fault detection technology, specifically relating to a method and system for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing. Background Technology

[0002] Magnetic field signal detection plays a crucial role in the quality inspection, diagnosis, and prediction of key components in machines and equipment. The most significant factor affecting the accuracy of magnetic field measurements is background noise interference; for example, the interference intensity of the Earth's magnetic field is very high and cannot be ignored in practical magnetic field measurements.

[0003] For magnetic field detection of target equipment on cable joints, magnetic field interference from electrical and mechanical equipment is more significant. Discharge signal detection under strong magnetic field interference is easily affected and unstable, making accurate judgment difficult. Therefore, achieving rapid and accurate signal denoising under strong interference is crucial for magnetic field measurement. Physical shielding tools such as shielding barrels can achieve good denoising results, but they are complex to operate and expensive. Background noise cancellation is easier to implement, but it only ensures accuracy to a certain extent. Furthermore, current research mainly focuses on relatively ideal and stable magnetic interference environments; eliminating strong interference and high transient noise in real-world complex operating conditions remains a difficult and challenging problem. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing, achieving dynamic and accurate denoising of partial discharge signals under conditions of strong magnetic field interference at cable joints.

[0005] The present invention adopts the following technical solution.

[0006] The first aspect of this invention proposes a method for denoising partial discharge magnetic signals at cable joints based on asymmetric magnetic sensing, comprising:

[0007] An array of asymmetric magnetic sensing modules is deployed at the cable joint, including a hub magnetic sensing module and several branch magnetic sensing modules arranged asymmetrically relative to the hub magnetic sensing module, wherein the hub magnetic sensing module is closest to the cable joint.

[0008] When the cable is not in operation, the magnetic field noise signal captured by the branch magnetic sensing module and the hub magnetic sensing module is used to train the neural network to obtain the noise-guided network model.

[0009] When the cable is in operation, the magnetic field noise signal captured by the branch magnetic sensing module is input into the noise guidance network model to predict the magnetic field noise captured by the hub magnetic sensing module. The noisy magnetic signal captured by the hub magnetic sensing module is subtracted from the predicted noise to obtain the denoised partial discharge magnetic signal of the cable joint.

[0010] Preferably, the hub magnetic sensing module is deployed near the cable joint, and the straight line between the two is the center line of the asymmetric magnetic sensing module array, with the location of the hub magnetic sensing module serving as the center position of the asymmetric magnetic sensing module array.

[0011] On the center line, the interval between the branch magnetic sensing module and the hub magnetic sensing module on the side closer to the cable joint is d1, and the interval between the branch magnetic sensing module and the hub magnetic sensing module on the side farther from the cable joint is d2, and the ratio of d1 to d2 is dynamically adjusted according to the real-time current of the cable.

[0012] N branch magnetic sensing modules are arranged at intervals on both sides of the central position, and the angles between the branch magnetic sensing modules on both sides and the central line are α and β, respectively, satisfying α≠β. The values ​​of α and β are set based on the gradient characteristics of the magnetic field distribution of the corresponding side cable, forming an asymmetric offset layout, and N is greater than or equal to 2.

[0013] Preferably, the ratio of d1 to d2 is dynamically adjusted according to the real-time current of the cable, as follows:

[0014]

[0015] in: This is the normalized value of the real-time cable current. Magnetic field gradient coefficient adapted to cable connector type; The optimal operating reference temperature for the TMR sensor module; This refers to the real-time ambient temperature at the cable joint.

[0016] Preferably, the included angle Set as:

[0017]

[0018] in, This represents the rate of change of the magnetic field gradient in the region corresponding to the cable joint. This serves as a reference value for the magnetic field gradient; This is a correction factor for the corresponding side connector type; The optimal operating reference temperature for the TMR sensor module; This refers to the real-time ambient temperature at the cable joint.

[0019] Preferably, both the branch magnetic sensing module and the hub magnetic sensing module are TMR sensing modules.

[0020] Preferably, the step of training a neural network using the magnetic field noise signals captured by the branch magnetic sensing module and the hub magnetic sensing module to obtain a noise-guided network model includes:

[0021] Using the magnetic field noise signal captured by the branch magnetic sensing module as input and the magnetic field noise signal captured by the hub magnetic sensing module as output, a neural network is trained to obtain a guided network model. The guided network model includes an input layer, two hidden layers, and an output layer, and its prediction process is as follows:

[0022]

[0023] in, This is the weight matrix; For bias terms; The input vector; For predicting output; This is the original output; This is the output of the first hidden layer; This is the output of the second hidden layer; and These are the activation values ​​for the first and second hidden layers, respectively. and Here are the learnable parameters; Sigmoid is the Sigmoid function; U is the upper limit value. This is a parameterized ReLU activation function.

[0024] Preferably, the neural network is trained using the following gradient-aware dynamically weighted Huber loss function:

[0025]

[0026] in, Dynamic weights; The main party responsible for Huber's losses; This is a gradient consistency penalty term; The actual magnetic field noise captured by the hub magnetic sensing module corresponding to the i-th sample; The noise of the hub magnetic sensing module predicted by the network for the i-th sample; To dynamically adjust parameters; This is the penalty coefficient; This represents the number of samples.

[0027] Preferably, the Huber loss subject is:

[0028]

[0029] in, For the first Dynamically adjust parameters during round training.

[0030] Preferably, the gradient consistency penalty term is:

[0031]

[0032] in, for noise input vector of the branch magnetic sensing module The gradient; for right The gradient; It is a norm.

[0033] Preferably, the dynamic adjustment parameter is:

[0034]

[0035] in, For the first All samples in the training batch standard deviation This is the adaptation coefficient.

[0036] Preferably, the gradient-aware dynamically weighted Huber loss function and the gradient sign-aware AdamW optimizer are combined with an adaptive early stopping mechanism based on the second derivative of the validation set loss to dynamically adjust the number of training rounds.

[0037] A second aspect of this invention provides a cable joint partial discharge magnetic signal denoising system based on asymmetric magnetic sensing, comprising:

[0038] An asymmetric magnetic sensing module array is deployed at the cable joint;

[0039] The data acquisition module is used to acquire magnetic field noise signals captured by the branch magnetic sensing module and the hub magnetic sensing module when the cable is in a non-working state; and to acquire noise signals captured by the branch magnetic sensing module and noisy magnetic signals captured by the hub magnetic sensing module when the cable is in a working state.

[0040] The network training module is used to train the neural network using the magnetic field noise signals captured by the branch magnetic sensing module and the hub magnetic sensing module when the cable is in a non-working state, so as to obtain a noise-guided network model.

[0041] The real-time noise reconstruction module is used to input the magnetic field noise signal captured by the branch magnetic sensing module into the noise steering network model when the cable is in operation, and to predict the magnetic field noise captured by the hub magnetic sensing module.

[0042] The noise reduction module is used to subtract the predicted noise from the noisy magnetic signal captured by the hub magnetic sensing module to obtain the denoised partial discharge magnetic signal of the cable joint.

[0043] The control and coordination module is used to control and coordinate the runtime sequence of the asymmetric magnetic sensing module array, data acquisition module, and noise-guided network model, and to monitor the system's operating status in real time.

[0044] A third aspect of the present invention provides a terminal comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the method.

[0045] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0046] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0047] This invention separates noise through unexcited training and excited prediction reconstruction, adapting to complex electromagnetic environments;

[0048] This invention employs an asymmetric topology array of magnetic sensing modules, matching the distribution of the branched magnetic sensing modules to the actual magnetic field gradient distribution characteristics around the cable joint. This enables more accurate and comprehensive capture of background noise signals from different directions. The asymmetric array layout of the magnetic sensing modules, combined with a neural network model, enhances the magnetic field signal resolution capability. The separation of hardware and software modules facilitates system expansion and maintenance. Therefore, this invention has practical value for denoising partial discharge signals from precise cable joint magnetic fields under strong noise interference, providing a new solution to practical application problems such as the susceptibility to interference, unstable quality, and difficulty in accurate judgment of partial discharge signals under strong magnetic field interference.

[0049] This invention features a refined and adaptable layout design for the asymmetric magnetic sensing module array. It clarifies the centerline positioning relationship between the hub magnetic sensing module and the cable connector, including dynamically adjusting the intervals d1 and d2 between the branch magnetic sensing modules and the hub magnetic sensing module on the centerline based on the real-time cable current. An asymmetric offset layout is achieved by setting N branch magnetic sensing modules on each side of the center position with an included angle α≠β, the angle value being preset based on the cable's magnetic field distribution gradient characteristics. The dynamically adjusted intervals adapt to magnetic field changes under different currents, the asymmetric layout conforms to the actual cable magnetic field gradient distribution, and the centerline positioning ensures the hub module accurately captures the target signal. This improves the correlation and comprehensiveness of noise signal capture between the branch modules and the hub module, providing more accurate and comprehensive data source support for subsequent noise-guided network model training and noise prediction.

[0050] The network structure and prediction process of the guided network model of this invention can alleviate the gradient vanishing / exploding problem in deep network training and enhance the model's ability to express negative features.

[0051] This invention employs a gradient-aware, dynamically weighted Huber loss function to train the network, dynamically... The combination with gradient penalty enables the network to not only accurately predict noise values ​​but also match noise trends. Compared to traditional Huber loss, the subtraction operation between the noisy signal and predicted noise in the subsequent hub module is more thorough, improving the signal-to-noise ratio of the denoised partial discharge signal, thus enhancing noise prediction accuracy and directly optimizing the denoising effect; dynamic It avoids the sudden increase / decrease of gradients caused by fixed thresholds. Dynamic weights enable the network to dynamically pay attention to samples with different error levels during training. Combined with the early stopping mechanism of the second derivative of the validation set loss, it can effectively prevent the network from overfitting or gradient explosion, enhance training stability, and adapt to complex electromagnetic working conditions such as multiple interference sources and time-varying noise in cable joints. Attached Figure Description

[0052] Figure 1 The specific steps of the method of the present invention are as follows;

[0053] Figure 2 This describes the internal structure of the TMR sensing module of the present invention.

[0054] Figure 3 This is a flowchart of the network training process of the present invention;

[0055] Figure 4 This is a flowchart of the signal reconstruction process of the present invention;

[0056] Figure 5 This is a schematic diagram illustrating the specific components of the system of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of this invention.

[0058] Embodiment 1 of this invention provides a method for denoising partial discharge magnetic signals from cable joints based on asymmetric magnetic sensing. It utilizes an asymmetric hub-and-branch architecture TMR array to optimize the denoising of partial discharge magnetic signals from cable joints. Figure 1 As shown, the method includes:

[0059] Step 1: Deploy an asymmetric magnetic sensing module array at the cable joint, including a hub magnetic sensing module and several branch magnetic sensing modules arranged asymmetrically relative to the hub magnetic sensing module, wherein the hub magnetic sensing module is closest to the cable joint.

[0060] More preferably, the hub magnetic sensing module is deployed near the cable joint, and the straight line between the two is the center line of the asymmetric magnetic sensing module array, with the location of the hub magnetic sensing module serving as the center position of the asymmetric magnetic sensing module array.

[0061] On the centerline, the interval between the branch magnetic sensing module and the hub magnetic sensing module on the side closer to the cable joint is d1, and the interval between the branch magnetic sensing module and the hub magnetic sensing module on the side farther from the cable joint is d2. The ratio of d1 to d2 is dynamically adjusted according to the real-time current of the cable, and the adjustment range is 0.8:1 to 1.2:1.

[0062]

[0063] in: The spacing between the branch module and the hub module on the center line near the cable joint side; The spacing between the branch module and the hub module on the center line away from the cable joint side;

[0064] This is the normalized value of the real-time cable current. , The real-time operating current of the cable is collected by a current transformer; This is the rated current of the cable. ;

[0065] The magnetic field gradient coefficient adapted to the cable joint type can be obtained through pre-calibration or by looking up a table based on the joint type, and its value is generally in the range of 0.9. 1.1;

[0066] The optimal operating reference temperature for the TMR sensor module can be determined based on the sensor calibration environment, and is usually 20℃ or 25℃.

[0067] This refers to the real-time ambient temperature at the cable joint.

[0068] Figure 5The topology diagram of the asymmetric hub-and-branch architecture TMR sensor module array is shown. The spacing between the near-end and far-end branch modules on the center line is designed according to the above logic: B5 is the branch module on the center line (connecting the hub magnetic sensor module and the cable joint) closer to the cable joint, and its spacing from the hub magnetic sensor module is d1; B6 is the branch module on the center line farther from the cable joint, and its spacing from the hub magnetic sensor module is d2. In actual engineering applications, the target adjustment range of the ratio of d1 to d2 is dynamically adjusted within the range of 0.8:1 to 1.2:1 according to the real-time current of the cable to adapt to the magnetic field changes under different currents. At the same time, B5 and B6 can accurately capture magnetic field noise signals at different positions in the direction of the center line, providing a more adaptable data source for the subsequent noise-guided network model.

[0069] In practical implementation, an asymmetric hub-and-branch architecture TMR sensor module array is deployed at the cable joint. The specific design of this invention is as follows: Figure 5 The diagram illustrates an asymmetric hub-and-branch tunnel magnetoresistive (TMR) sensor array topology. A two-dimensional Cartesian coordinate system (x,y) is established, with the array's geometric center as the origin O(0,0), and the coordinate unit being the sensor unit spacing. The hub TMR sensor module is located at the closest point aligned with the origin O(0,0) and the cable connector along a straight line, serving as the array's signal processing center node. The branch sensors are distributed, with six branch TMR sensor modules asymmetrically distributed within the coordinate system. Specifically:

[0070] The hub TMR sensing module is located at the center of the asymmetric array, i.e., at the closest point on the same straight line as the cable connector, and its coordinates are defined as (0,0). N branch magnetic sensing modules are spaced apart on each side of the center, with the angles between each branch magnetic sensing module and the center line being α and β, respectively, satisfying α≠β. The values ​​of α and β are pre-set based on the gradient characteristics of the cable's magnetic field distribution, forming an asymmetric offset layout. N is greater than or equal to 2.

[0071] One side angle The calculation formula is:

[0072]

[0073] The other side angle The calculation formula is:

[0074]

[0075] in, The rate of change of the magnetic field gradient in the region on one side of the cable joint; The rate of change of the magnetic field gradient in the region on the other side of the cable joint; This serves as a reference value for the magnetic field gradient; For the correction factor of the connector type on one side; Correction factor for the connector type on the other side; The reference temperature for magnetic field gradient detection; This represents the real-time ambient temperature at the cable joint. The reference value for the magnetic field gradient can be determined based on the geomagnetic field background gradient or through on-site calibration; the correction factor is used to correct for local structural asymmetry at the joint, and its value is typically in the range of 0.85. 1.15.

[0076] The aforementioned asymmetric offset layout strategy can optimize the accuracy of noise signal acquisition by precisely matching the angle value with the magnetic field gradient. A steep gradient results in a larger angle, which can cover a wider area; a gentle gradient results in a smaller angle, which can focus on key areas. This ensures that both left and right branch modules can efficiently capture the magnetic field noise signal of the corresponding area, providing highly correlated and complete input data for the subsequent noise-guided network model. The noise prediction error can be reduced by 15%-20%, improving the denoising effect of the partial discharge magnetic signal.

[0077] Both the branch magnetic sensing module and the hub magnetic sensing module are TMR sensing modules; the TMR module includes:

[0078] TMR sensors are used to detect magnetic field signals and output differential voltage;

[0079] Instrumentation amplifiers are used to amplify differential voltage signals with high precision.

[0080] The operational amplifier and the dynamic compensation circuit together form an adjustable-gain low-pass filter circuit to filter out high-frequency noise and achieve zero-point drift calibration of the signal. The dynamic compensation circuit includes a resistor network (containing one full-scale adjustable resistor R and three auxiliary fixed resistors R1, R2, R3). A relay controlled by the motor drive chip switches different combinations of resistors (R, R1, R2, R3) to change the resistance value of the operational amplifier's feedback network.

[0081] like Figure 2 As shown, the TMR sensing module includes:

[0082] TMR sensors are used to detect magnetic field signals and output differential voltage;

[0083] The AD620 instrumentation amplifier with integrated temperature compensation incorporates a miniature temperature sensor at its input to monitor ambient temperature in real time. Its input also receives the differential voltage, which is amplified with high precision and used as the output signal of the TMR sensor module. The AD620 uses this temperature data for two main purposes: first, it corrects the amplifier's gain and zero-point drift caused by temperature changes by fine-tuning the internal bias voltage or feedback parameters; second, it compensates for the temperature-induced deviation of the TMR sensor's output differential voltage. Working in conjunction with a dynamic compensation circuit and operational amplifier, it adjusts the circuit resistance and operational amplifier feedback coefficient to ultimately counteract the dual effects of temperature fluctuations on both the sensor and amplifier, ensuring the accuracy of the amplified signal and providing accurate data for subsequent processing.

[0084] The dynamic compensation circuit (4 resistors) using an 8421 code resistor network consists of an adjustable resistor group (including one full-scale adjustable resistor R=10kΩ and auxiliary fixed resistors R1, R2, and R3) controlled by a double-pole double-throw relay. The dynamic adjustment of the resistance value is achieved through a motor drive chip.

[0085] The differential voltage output by the TMR sensor is connected to the input terminal of the instrumentation amplifier, and the output terminal of the instrumentation amplifier is connected to the input terminal of the operational amplifier, so that the amplified differential voltage signal is transmitted to the operational amplifier.

[0086] The resistor network of the dynamic compensation circuit serves as the feedback network of the operational amplifier, connected between the output and inverting input of the operational amplifier. By changing the equivalent resistance value of the resistor network using a motor driver chip, the feedback coefficient of the operational amplifier can be adjusted, thereby performing gain calibration and zero-point drift compensation on the signal from the instrumentation amplifier.

[0087] The operational amplifier can adjust the resistance of the resistor network (by changing the resistance of R through the motor driver chip, in combination with R1, R2, and R3) to perform zero-point drift correction and temperature compensation-related signal calibration on the signal from the instrumentation amplifier, and finally output a stable signal after interference elimination through the operational amplifier.

[0088] The dynamic adjustment of the resistance value aims to eliminate signal deviations caused by temperature drift and geomagnetic field interference. The core strategy is to adjust the combined resistance value of each resistor in the resistor network through the motor driver chip based on the real-time monitored deviation signal (such as the zero-point offset of the operational amplifier output and the signal error related to ambient temperature).

[0089] First, determine the range of total resistance values ​​of the resistor network that need to be adjusted based on the type of deviation (such as zero-point offset caused by temperature) and the magnitude of the deviation.

[0090] Then, by switching the double-pole double-throw relay, the full-scale adjustable resistor R is combined with auxiliary fixed resistors R1, R2, and R3 in different ways (for example, when the deviation is small, R is used in combination with R1 and R2 for fine adjustment; when the deviation is large, R is adjusted alone and combined with R3 for compensation), thereby changing the overall feedback characteristics of the resistor network.

[0091] Finally, based on the signal accuracy of the operational amplifier output, the resistance value of R is repeatedly fine-tuned until the deviation is eliminated.

[0092] For example, when changes in ambient temperature cause a zero-point shift in the signal, the system will trigger an adjustment by increasing or decreasing the resistance value of R through the motor driver chip, while optimizing the total resistance value of the resistor network using R1 and R2, thereby changing the feedback coefficient of the operational amplifier, correcting the zero point of the amplified signal, and ensuring a stable output signal.

[0093] The operational amplifier (LM741), working in conjunction with the dynamic compensation circuit, provides dual intelligent compensation for temperature drift and geomagnetic field interference, eliminating zero-point drift in the circuit.

[0094] A microcontroller dynamically adjusts the resistance value of the compensation circuit based on the temperature sensor readings integrated in the module and the DC offset of the signal output, using a lookup table method or control algorithm. This corrects the signal drift caused by temperature changes and geomagnetic field interference in real time, ensuring the stability of the output signal.

[0095] Step 2: When the cable is in a non-working state, use the magnetic field noise signals captured by the branch magnetic sensing module and the hub magnetic sensing module to train the neural network and obtain the noise-guided network model.

[0096] More preferably, in the absence of excitation, the noise-guided network model is trained using noise signals captured by the asymmetric sensing module array. Specifically, the magnetic field noise signal captured by the branch magnetic sensing module is used as input, and the magnetic field noise signal captured by the hub magnetic sensing module is used as output to train the neural network, thereby obtaining the guided network model, such as... Figure 3 As shown, it includes:

[0097] (1) Data acquisition stage:

[0098] When the cable is in a non-working state (no excitation condition), the ambient magnetic field noise signal is synchronously collected through six branch TMR sensing modules (B1-B6);

[0099] Record the corresponding magnetic field noise response of the hub TMR sensing module (H0) (sampling frequency not less than 10kHz, duration ≥60s).

[0100] (2) Training set construction:

[0101] Construct a supervised learning training set, i.e., input feature matrix It consists of environmental noise signals collected by the TMR sensor modules of each branch. ;

[0102] in This represents the time-series data of the environmental magnetic field noise signal collected by the i-th branch sensing module, where n represents the number of data points in a single sampling window.

[0103] Target output vector It consists of the environmental noise response collected by the hub's TMR sensing module;

[0104] (3) Guiding network training:

[0105] We employ a gradient-aware dynamically weighted Huber loss function (which dynamically adjusts the δ parameter to match the standard deviation of the gradient distribution in the current batch) and a gradient sign-aware AdamW optimizer (which triggers a sign update mechanism when the absolute value of the gradient exceeds a threshold), combined with an adaptive early stopping mechanism based on the second derivative of the validation set loss, to dynamically adjust the number of training rounds to 100-800 rounds.

[0106] The overall structure of the guided network model includes: an input layer, two hidden layers, and an output layer;

[0107] The input layer has 6 input channels, corresponding to the magnetic field noise signals of the six TMR sensing modules (B1-B6).

[0108] There are 2 hidden layers, and each layer has 128 neurons;

[0109] The output dimension of the output layer is 1, corresponding to the magnetic field noise signal (H0) of the hub TMR sensing module.

[0110] The core layer in the network that has the function of weight calculation includes a first hidden layer (fully connected), a second hidden layer (fully connected, the current layer), and an output layer (fully connected), forming a three-layer fully connected structure;

[0111] The input layer is only responsible for receiving the magnetic field noise signals from the six branch TMR sensing modules and does not participate in the fully connected weighting calculation.

[0112] The first hidden layer receives the input layer signal and completes the first fully connected operation, outputting the activation value a1;

[0113] The second hidden layer receives the output of the first hidden layer and completes the second fully connected operation through the residual structure (adding a1 to its own linear output);

[0114] The output layer receives the output of the second hidden layer and outputs the noise prediction signal of the hub module after completing the third fully connected operation.

[0115]

[0116] in, This is the weight matrix; For bias terms; The input vector represents the noise signal captured by the six-branch TMR sensing modules. For the predicted output (predicted noise signal of the hub TMR sensing module); This is the output of the first hidden layer; This is the output of the second hidden layer; and These are the activation values ​​for the first and second hidden layers, respectively. and Here are the learnable parameters; Sigmoid is the Sigmoid function; U is the upper limit value. This is a parameterized ReLU activation function.

[0117]

[0118] in α ( and ) are learnable parameters.

[0119] In the above prediction process, the activation value a1 of the first hidden layer is directly skipped from the weight calculation of the current layer (the second hidden layer) and added to the layer output z2, alleviating the gradient vanishing / exploding problem in deep network training. In the activation value calculation of the second hidden layer, the activation value a1 of the first hidden layer is directly added to the linear output z2 of the second layer (i.e., z2+a1), forming a residual structure; PReLU learns the negative half-axis slope α (α>0) to solve the neuron death problem of ReLU where the output of the negative half-axis is always 0, enhancing the model's ability to express negative features; a Sigmoid function and an upper limit U are added to the prediction output layer to convert the original output y raw Map to the interval [0, U].

[0120] Let the noise input of the branch module of the i-th sample in the network training be... The actual noise of the hub module is The network prediction noise is The loss function formula is as follows:

[0121]

[0122] The definitions of each core item are as follows:

[0123] Dynamically adjust parameters ( (For the current training round)

[0124]

[0125] in, For the first All samples in the training batch standard deviation The fit coefficient is used for balancing. The dynamic range is typically between 1.0 and 2.0;

[0126] Huber's main losses :

[0127]

[0128] Gradient consistency penalty :

[0129]

[0130] in, To predict the effect of noise on the input gradient, For the true noise to the input gradient, This is the penalty coefficient, typically ranging from 0.1 to 1.0, and can be determined through validation set performance tuning. It is an L1 or L2 norm.

[0131] Dynamic weights for:

[0132]

[0133] in, The actual noise of the hub magnetic sensing module corresponding to the j-th sample; Let be the network prediction noise corresponding to the j-th sample.

[0134] The gradient-aware, dynamically weighted Huber loss function described above will dynamically... Combined with gradient penalty, the network can not only accurately predict noise values ​​but also match noise trends. Compared to traditional Huber loss, the subtraction operation between the noisy signal and predicted noise in the subsequent hub module is more thorough, improving the signal-to-noise ratio of the denoised partial discharge signal, thus improving noise prediction accuracy and directly optimizing the denoising effect; dynamic It avoids the sudden increase / decrease of gradients caused by fixed thresholds. Dynamic weights enable the network to dynamically pay attention to samples with different error levels during training. Combined with the early stopping mechanism of the second derivative of the validation set loss, it can effectively prevent the network from overfitting or gradient explosion, enhance training stability, and adapt to complex electromagnetic working conditions such as multiple interference sources and time-varying noise in cable joints.

[0135] Step 3: When the cable is in operation, input the magnetic field noise signal captured by the branch magnetic sensing module into the noise steering network model to predict the magnetic field noise captured by the hub magnetic sensing module. Subtract the predicted noise from the noisy magnetic signal captured by the hub magnetic sensing module to obtain the denoised partial discharge magnetic signal of the cable joint.

[0136] More preferably, under excitation, the noise component of the hub TMR sensing module is reconstructed using the noise of the branch TMR sensing module, and the reconstructed noise is subtracted from the noisy magnetic signal captured by the hub TMR sensing module to obtain the denoised magnetic field signal to be detected, such as... Figure 4 As shown, it specifically includes:

[0137] (1) Data acquisition stage:

[0138] When the cable is in operation (excitation condition), the field magnetic signals are synchronously acquired through six branch TMR sensing modules (B1-B6);

[0139] Record the corresponding magnetic field signal of the hub TMR sensing module (H0) (sampling frequency not less than 10kHz, duration ≥60s).

[0140] (2) Noise reconstruction stage:

[0141] The field magnetic signals synchronously acquired by the six branch TMR sensing modules (B1-B6) are input into the trained guidance network model and the field predicted noise signal of the hub TMR sensing module is output.

[0142] (3) Magnetic signal reconstruction stage:

[0143] The noisy magnetic signal captured by the hub TMR sensing module is subtracted from the reconstruction noise to obtain the denoised magnetic field signal to be detected.

[0144] It is understandable that the branch module only captures magnetic field noise signals. When the cable joint is far away from the layout and not in operation, only magnetic field noise signals are captured. When in operation, the layout that is far away can also be considered to capture magnetic field noise signals.

[0145] Embodiment 2 of the present invention provides a cable joint partial discharge magnetic signal denoising system based on asymmetric magnetic sensing, the system comprising:

[0146] An asymmetric magnetic sensing module array is deployed at the cable connector; specifically, it employs... Figure 5 The asymmetric hub-and-branch architecture TMR sensor module array shown:

[0147] Hub TMR Sensing Module (H0):

[0148] Located at the center of the asymmetric array (coordinates (0,0)), it is the closest point on the same straight line as the cable connector. It includes a TMR magnetic sensor (such as TMR2003), an instrumentation amplifier (AD620), an operational amplifier (LM741), and a resistor network (including one 10kΩ adjustable resistor and an auxiliary fixed resistor). It detects the magnetic field signal and outputs a differential voltage, which is then transmitted to the data acquisition module after zero-drift elimination and amplification.

[0149] Branch TMR sensor modules (B1-B6):

[0150] There are 6 modules, each containing the same hardware configuration as the hub module (TMR magnetic sensor, AD620, LM741 and resistor network).

[0151] The data acquisition module is used to acquire magnetic field noise signals captured by the branch magnetic sensing module and the hub magnetic sensing module when the cable is in a non-operating state; and to acquire noise signals captured by the branch magnetic sensing module and noisy magnetic signals captured by the hub magnetic sensing module when the cable is in an operating state; specifically including:

[0152] (1) Data acquisition under no-excitation operating conditions:

[0153] When the cable is not in operation, the ambient magnetic field noise signal is synchronously acquired through 6 branch TMR sensor modules (sampling frequency ≥10kHz, duration ≥60s). The noise response signal of the hub TMR sensor module is also recorded synchronously.

[0154] (2) Data collection under excitation conditions:

[0155] While the cable is in operation, the field magnetic signals are synchronously acquired through 6 branch TMR sensor modules (sampling frequency ≥ 10kHz, duration ≥ 60s). The noisy magnetic signal from the hub TMR sensor module is also recorded synchronously.

[0156] The network training module is used to train a neural network model by using magnetic field noise signals captured by the branch magnetic sensing module and the hub magnetic sensing module when the cable is not in operation, thus obtaining a noise-guided network model. Specifically, it receives noise signals (B1-B6) from six branch TMR sensing modules. It contains a two-layer fully connected neural network with 128 neurons per layer and ReLU activation function. It outputs the noise prediction value from the hub TMR sensing module (H0).

[0157] The real-time noise reconstruction module is used to input the magnetic field noise signal captured by the branch magnetic sensing module into the noise steering network model when the cable is in operation, and predict the magnetic field noise captured by the hub magnetic sensing module. Specifically, under the excitation condition, the field magnetic signals collected by the 6 branch TMR sensing modules are input into the trained noise steering network model, and the noise prediction value of the hub TMR sensing module is output.

[0158] The noise reduction module is used to subtract the predicted noise from the noisy magnetic signal captured by the hub magnetic sensing module to obtain the denoised partial discharge magnetic signal of the cable joint; specifically, it subtracts the reconstructed noise component from the noisy magnetic signal captured by the hub TMR sensing module to obtain the denoised magnetic field signal to be detected.

[0159] The control and coordination module controls and coordinates the runtime sequence of the asymmetric magnetic sensing module array, data acquisition module, and noise-guided network model, and monitors the system's operating status in real time. Specifically, it coordinates the runtime sequence of the TMR sensing module array, data acquisition system, and noise-guided network. It also monitors the system's operating status in real time, including sensor fault detection and network model performance evaluation.

[0160] Embodiment 3 of the present invention provides a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the method.

[0161] Embodiment 4 of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0162] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0163] This invention separates noise through unexcited training and excited prediction reconstruction, adapting to complex electromagnetic environments;

[0164] This invention employs an asymmetric topology array of magnetic sensing modules, matching the distribution of the branched magnetic sensing modules to the actual magnetic field gradient distribution characteristics around the cable joint. This enables more accurate and comprehensive capture of background noise signals from different directions. The asymmetric array layout of the magnetic sensing modules, combined with a neural network model, enhances the magnetic field signal resolution capability. The separation of hardware and software modules facilitates system expansion and maintenance. Therefore, this invention has practical value for denoising partial discharge signals from precise cable joint magnetic fields under strong noise interference, providing a new solution to practical application problems such as the susceptibility to interference, unstable quality, and difficulty in accurate judgment of partial discharge signals under strong magnetic field interference.

[0165] This invention features a refined and adaptable layout design for the asymmetric magnetic sensing module array. It clarifies the centerline positioning relationship between the hub magnetic sensing module and the cable connector, including dynamically adjusting the intervals d1 and d2 between the branch magnetic sensing modules and the hub magnetic sensing module on the centerline based on the real-time cable current. An asymmetric offset layout is achieved by setting N branch magnetic sensing modules on each side of the center position with an included angle α≠β, the angle value being preset based on the cable's magnetic field distribution gradient characteristics. The dynamically adjusted intervals adapt to magnetic field changes under different currents, the asymmetric layout conforms to the actual cable magnetic field gradient distribution, and the centerline positioning ensures the hub module accurately captures the target signal. This improves the correlation and comprehensiveness of noise signal capture between the branch modules and the hub module, providing more accurate and comprehensive data source support for subsequent noise-guided network model training and noise prediction.

[0166] The guided network model of this invention can alleviate the gradient vanishing / exploding problem in deep network training and enhance the model's ability to express negative features.

[0167] This invention employs a gradient-aware, dynamically weighted Huber loss function to train the network, dynamically... The combination with gradient penalty enables the network to not only accurately predict noise values ​​but also match noise trends. Compared to traditional Huber loss, the subtraction operation between the noisy signal and predicted noise in the subsequent hub module is more thorough, improving the signal-to-noise ratio of the denoised partial discharge signal, thus enhancing noise prediction accuracy and directly optimizing the denoising effect; dynamic It avoids the sudden increase / decrease of gradients caused by fixed thresholds. Dynamic weights enable the network to dynamically pay attention to samples with different error levels during training. Combined with the early stopping mechanism of the second derivative of the validation set loss, it can effectively prevent the network from overfitting or gradient explosion, enhance training stability, and adapt to complex electromagnetic working conditions such as multiple interference sources and time-varying noise in cable joints.

[0168] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0169] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0170] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0171] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for denoising partial discharge magnetic signals in cable joints based on asymmetric magnetic sensing, characterized in that, include: An array of asymmetric magnetic sensing modules is deployed at the cable joint, including a hub magnetic sensing module and several branch magnetic sensing modules arranged asymmetrically relative to the hub magnetic sensing module, wherein the hub magnetic sensing module is closest to the cable joint. The hub magnetic sensing module is deployed near the cable joint, and the straight line between the two is the center line of the asymmetric magnetic sensing module array. The location of the hub magnetic sensing module is the center position of the asymmetric magnetic sensing module array. On the center line, the distance between the branch magnetic sensing module and the hub magnetic sensing module on the side closer to the cable joint is d1, and the distance between the branch magnetic sensing module and the hub magnetic sensing module on the side farther from the cable joint is d2. The ratio of d1 to d2 is dynamically adjusted according to the real-time current of the cable. N branch magnetic sensing modules are arranged at intervals on both sides of the center position, and the angles between the branch magnetic sensing modules on both sides and the center line are α and β, respectively, satisfying α≠β. The values ​​of α and β are set based on the gradient characteristics of the magnetic field distribution of the cable on the corresponding side, forming an asymmetric offset layout, where N is greater than or equal to 2. When the cable is not in operation, the magnetic field noise signal captured by the branch magnetic sensing module and the hub magnetic sensing module is used to train the neural network to obtain the noise-guided network model. When the cable is in operation, the magnetic field noise signal captured by the branch magnetic sensing module is input into the noise guidance network model to predict the magnetic field noise captured by the hub magnetic sensing module. The noisy magnetic signal captured by the hub magnetic sensing module is subtracted from the predicted noise to obtain the denoised partial discharge magnetic signal of the cable joint.

2. The method for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing according to claim 1, characterized in that: The ratio of d1 to d2 is dynamically adjusted according to the real-time current of the cable, as follows: in: This is the normalized value of the real-time cable current. Magnetic field gradient coefficient adapted to cable connector type; The optimal operating reference temperature for the TMR sensor module; This refers to the real-time ambient temperature at the cable joint.

3. The method for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing according to claim 1, characterized in that: included angle Set as: in, This represents the rate of change of the magnetic field gradient in the region corresponding to the cable joint. This serves as a reference value for the magnetic field gradient; This is a correction factor for the corresponding side connector type; The optimal operating reference temperature for the TMR sensor module; This refers to the real-time ambient temperature at the cable joint.

4. The method for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing according to claim 1, characterized in that: Both the branch magnetic sensing module and the hub magnetic sensing module are TMR sensing modules.

5. The method for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing according to claim 1, characterized in that: The method of training a neural network using magnetic field noise signals captured by the branch magnetic sensing module and the hub magnetic sensing module to obtain a noise-guided network model includes: Using the magnetic field noise signal captured by the branch magnetic sensing module as input and the magnetic field noise signal captured by the hub magnetic sensing module as output, a neural network is trained to obtain a guided network model. The guided network model includes an input layer, two hidden layers, and an output layer, and its prediction process is as follows: in, This is the weight matrix; For bias terms; The input vector; For predicting output; This is the original output; This is the output of the first hidden layer; This is the output of the second hidden layer; and These are the activation values ​​for the first and second hidden layers, respectively. and Here are the learnable parameters; Sigmoid is the Sigmoid function; U is the upper limit value. This is a parameterized ReLU activation function.

6. The method for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing according to claim 1, characterized in that: The neural network is trained using the following gradient-aware dynamically weighted Huber loss function: in, Dynamic weights; The main party responsible for Huber's losses; This is a gradient consistency penalty term; The actual magnetic field noise captured by the hub magnetic sensing module corresponding to the i-th sample; The noise of the hub magnetic sensing module predicted by the network for the i-th sample; To dynamically adjust parameters; This is the penalty coefficient; This represents the number of samples.

7. The method for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing according to claim 6, characterized in that: The subject of the Huber loss is: in, For the first Dynamically adjust parameters during round training.

8. The method for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing according to claim 6, characterized in that: The gradient consistency penalty term is: in, for noise input vector of the branch magnetic sensing module The gradient; for right The gradient; It is a norm.

9. The method for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing according to claim 6, characterized in that: The dynamically adjusted parameters are: in, For the first All samples in the training batch standard deviation This is the adaptation coefficient.

10. The method for denoising partial discharge magnetic signals of cable joints based on asymmetric magnetic sensing according to claim 6, characterized in that: The gradient-aware dynamically weighted Huber loss function and gradient sign-aware AdamW optimizer, combined with the validation set loss second derivative adaptive early stopping mechanism, dynamically adjust the number of training rounds.

11. A cable joint partial discharge magnetic signal denoising system based on asymmetric magnetic sensing, used to implement the method described in any one of claims 1-10, characterized in that, The system includes: An asymmetric magnetic sensing module array is deployed at the cable joint; The data acquisition module is used to acquire magnetic field noise signals captured by the branch magnetic sensing module and the hub magnetic sensing module when the cable is in a non-working state; and to acquire noise signals captured by the branch magnetic sensing module and noisy magnetic signals captured by the hub magnetic sensing module when the cable is in a working state. The network training module is used to train the neural network using the magnetic field noise signals captured by the branch magnetic sensing module and the hub magnetic sensing module when the cable is in a non-working state, so as to obtain a noise-guided network model. The real-time noise reconstruction module is used to input the magnetic field noise signal captured by the branch magnetic sensing module into the noise steering network model when the cable is in operation, and to predict the magnetic field noise captured by the hub magnetic sensing module. The noise reduction module is used to subtract the predicted noise from the noisy magnetic signal captured by the hub magnetic sensing module to obtain the denoised partial discharge magnetic signal of the cable joint. The control and coordination module is used to control and coordinate the runtime sequence of the asymmetric magnetic sensing module array, data acquisition module, and noise-guided network model, and to monitor the system's operating status in real time.

12. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-10.

Images