CNN-PINN fusion network-based ADSS optical cable potential measurement method, device and medium

Abstract

The invention discloses an ADSS optical cable potential measurement method and device based on a CNN-PINN fusion network, and a medium, and the method comprises the steps: S1, collecting a space electric field vector around an optical cable and environment temperature and humidity parameters, and constructing a three-channel tensor; s2, extracting a high-dimensional feature vector; s3, a PINN physical constraint module is embedded into a Poisson equation, field intensity gradient constraint and dielectric compensation; s4, outputting a three-dimensional space potential distribution field and an electric corrosion risk level; according to the method, CNN and PINN are fused and improved, and the bottleneck of the traditional technology is thoroughly broken through: high-gradient distortion is captured through UNet, key area perception is enhanced through CBAM, and the measurement error is reduced compared with that of a traditional contact sensor through PINN physical constraint; environmental interference is decoupled through three channels, humidity drift is suppressed through dielectric compensation, and reconstruction precision stability in a strong interference environment is guaranteed through frequency domain noise reduction; and the output layer generates three-dimensional potential field distribution, calculates an electric corrosion risk level integrating a space gradient and a time-varying characteristic in real time, and supports a corrosion protection decision.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power systems, specifically to an ADSS optical cable potential measurement method, equipment, and medium based on a CNN-PINN fusion network. Background Technology

[0002] Existing ADSS optical cable space potential measurement technology has fundamental limitations: contact-based single-point measurements, such as piezoelectric sensors, require direct contact with the optical cable surface, which distorts the original electric field distribution and leads to large errors; analytical methods such as finite element simulation can construct physical models, but their computational complexity is extremely high and cannot meet the needs of dynamic monitoring; emerging pure data-driven models such as traditional CNNs improve efficiency, but ignore the constraints of Maxwell's equations, are prone to non-physical interpretations in complex environments such as rain and fog, and are difficult to capture high-gradient distortion features around insulators.

[0003] High-voltage transmission corridors are often subjected to harsh conditions such as high humidity and strong electromagnetic interference, and existing technologies lack effective countermeasures: there is a strong coupling effect between environmental parameters and electric field distribution, but traditional methods have not established a quantitative compensation model, which leads to the drift of dielectric constant in rainy and foggy weather, causing measurement deviations; environmental electromagnetic noise, especially broadband interference from corona discharge, can overwhelm the effective signal, and conventional filtering methods cannot specifically focus on characteristic frequency bands; simplified empirical models, such as the assumption of a fixed dielectric constant, ignore the humidity diffusion hysteresis effect, resulting in inaccurate dynamic response.

[0004] Current methods are insufficient to support real-time risk prevention and control: the deployment of contact sensors is difficult, it is impossible to achieve dynamic reconstruction of the three-dimensional potential field with high-frequency updates, and there is a lack of an indicator system for quantifying the risk of electrical corrosion; offline simulation and online monitoring are disconnected, and early warning decisions lag behind the occurrence of faults, which cannot meet the response requirements of ADSS optical cable electrical corrosion protection.

[0005] Therefore, it is essential to design an ADSS optical cable potential measurement method, equipment, and medium based on a CNN-PINN fusion network. Summary of the Invention

[0006] The purpose of this invention is to provide an ADSS optical cable potential measurement method, device, and medium based on a CNN-PINN fusion network, to address the fundamental limitations of existing ADSS optical cable spatial potential measurement technologies mentioned in the background: contact-based single-point measurements, such as piezoelectric sensors, require direct contact with the optical cable surface, distorting the original electric field distribution and leading to large errors; analytical methods such as finite element simulation can construct physical models, but their computational complexity is extremely high, failing to meet the needs of dynamic monitoring; emerging pure data-driven models, such as traditional CNNs, while improving efficiency, ignore Maxwell's equation constraints, are prone to non-physical interpretations in complex environments such as rain and fog, and struggle to capture high-gradient distortion characteristics around insulators, especially considering the harsh conditions of high humidity and strong electromagnetic interference often encountered in high-voltage transmission corridors. Existing technologies lack effective countermeasures: there is a strong coupling effect between environmental parameters and electric field distribution, but traditional methods have not established quantitative compensation models, leading to measurement deviations caused by dielectric constant drift in rainy and foggy weather; environmental electromagnetic noise, especially broadband interference from corona discharge, can overwhelm effective signals, and conventional filtering methods cannot specifically focus on characteristic frequency bands; simplified empirical models, such as the assumption of a fixed dielectric constant ignoring the humidity diffusion hysteresis effect, result in inaccurate dynamic responses, and current methods are unable to support real-time risk prevention and control: deployment of contact sensors is difficult, making it impossible to achieve high-frequency updated dynamic reconstruction of the three-dimensional potential field, and there is a lack of an index system for quantifying electrolytic corrosion risk; offline simulation and online monitoring are disconnected, and early warning decisions lag behind the occurrence of faults, failing to meet the response requirements of ADSS optical cable electrolytic corrosion protection.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] Firstly, a method for measuring the potential of ADSS optical cables based on a CNN-PINN fusion network is provided, including the following steps:

[0009] S1: Acquire the spatial electric field vector (Ex, Ey, Ez) around the optical cable through a non-contact MEMS sensor array, simultaneously obtain the ambient temperature and humidity parameters through temperature and humidity sensors, construct a three-channel tensor, and collect measured field strength data through a beam splitter and optical power meter;

[0010] S2: Input the three-channel tensor into the CNN feature extraction module to extract high-dimensional feature vectors. The CNN feature extraction module adopts the UNet structure for multi-scale feature fusion, which includes a 4-level downsampling encoding layer and a 4-level upsampling decoding layer. The 4-level downsampling is responsible for extracting multi-scale global features, and the 4-level upsampling is responsible for restoring spatial details. A CBAM attention mechanism module is inserted after each convolutional layer to realize dynamic feature weighting of channels and spatial dimensions.

[0011] S3: Input the high-dimensional feature vector extracted by the CNN feature extraction module and the measured field strength data into the PINN physical constraint module. Introduce the following into the loss function: Poisson equation residual term, field strength gradient constraint term and environmental parameter dielectric compensation term. The PINN physical constraint module adopts an adaptive residual weight allocation strategy.

[0012] S4: Through end-to-end joint training, output the three-dimensional spatial potential distribution field Φ(x,y,z,t) to realize dynamic potential measurement and calculate and predict the risk level of electrolytic erosion in real time.

[0013] As a further technical solution of the present invention, the step of constructing the three-channel tensor in S1 is as follows:

[0014] The electric field in space is decomposed into three physically defined components, and a vectorized model of the electric field is performed:

[0015]

[0016] Normalization process:

[0017] θ′=π / θ+0.5

[0018]

[0019] Where |E| is the original electric field strength; θ is the electric field direction angle; To increase the vertical tilt angle; Significantly improves the modeling accuracy of the potential field beneath ultra-high voltage conductors; normalization eliminates directional singularities.

[0020] As a further technical solution of the present invention, the three-channel tensor is:

[0021] Channel 1: Original electric field strength:

[0022] |E|

[0023] Channel 2: Electric field direction angle:

[0024] θ = arctan(Ey / Ex)

[0025] Channel 3: Combined Influencing Factors of Ambient Humidity and Temperature

[0026] K = (RH / 25℃) * T -0.3

[0027] Where RH is the ambient relative humidity; T is the ambient temperature.

[0028] As a further technical solution of the present invention, the network structure of the UNet CNN feature extraction module is added in S2:

[0029]

[0030] Where t is the subpixel convolution operator; C 1×1 It is a 1×1 depthwise separable convolution; This represents the output feature of the k-th level encoder. The output features of the k-th level decoder are: depthwise separable convolution reduces parameters, differential fusion enhances edge abrupt features, and subpixel convolution eliminates ringing effects.

[0031] As a further technical solution of the present invention, the processing flow of the CBAM attention mechanism module in S2 includes:

[0032] Channel attention: Mc(F) = σ(MLP(AvgPoo1(F)) + MLP(MaxPoo1(F)))

[0033] Spatial attention: Ms(F)=σ(f7×7([AvgPoo1(F);MaxPoo1(F)]))

[0034] Where F is the input feature map; σ is the Sigmoid activation function;

[0035] The attention mechanism employs a dual-domain (spectral-spatial) attention approach.

[0036] M f (F)=F -1 [σ(F(F)⊙W f )]

[0037]

[0038] Among them, M f (F) represents the frequency domain attention output, F(·) represents the Fourier transform operator, and F -1 (·) represents the inverse Fourier transform operator, ⊙ represents element-wise multiplication, and W f For a frequency domain bandpass filter, f is the frequency variable, f c Center frequency; frequency band of frequency domain attention-focused corona discharge characteristics; risk zone where the Laplacian operator enhances the potential gradient exceeding 25 kV / cm;

[0039] Introducing potential gradient stabilization techniques to solve gradient explosion in potential field reconstruction:

[0040]

[0041] Where is the gradient of the loss function with respect to the parameters, ‖·‖2 is the L2 norm operator, γ is the gradient clipping threshold, ε is the minimum positive number, and min(1,·) is the threshold cutoff function; which improves the training convergence speed.

[0042] During the ADSS optical cable potential field reconstruction process, the gradient explosion problem is mainly caused by the high-order derivative oscillations triggered by the rapid potential jump around the insulator. This numerical instability will seriously disrupt the model training process. The gradient stabilization technology innovatively achieves dynamic control through a three-pronged mechanism: first, it monitors the gradient amplitude changes in backpropagation in real time, and immediately triggers the intervention mechanism when an abnormal growth trend is detected; second, it uses adaptive scaling technology to precisely limit the gradient change range, and constrains the update of runaway parameters within a physically reasonable threshold; and finally, it retains the effective learning signal through an intelligent truncation strategy.

[0043] As a further technical solution of the present invention, the Poisson equation residual term introduced by the loss function of the PINN physical constraint module in S3 is as follows:

[0044]

[0045] Among them, R phys For the residual term of the Poisson equation, ∈ corr The dielectric constant is dynamically corrected. Let p be the potential gradient and p be the space charge density;

[0046] Field strength gradient constraint term:

[0047]

[0048] Among them, R curl For the field strength irrotational constraint term, Here, E is the curl operator, and E is the electric field intensity vector.

[0049] Environmental parameter dielectric compensation item:

[0050]

[0051] Where, ∈ corr denoted as the dynamically corrected dielectric constant; ∈0 is the vacuum dielectric constant; χ0 is the intrinsic polarizability of the material; RH is the ambient relative humidity; τ is the humidity diffusion time constant; and γ is the hysteresis factor.

[0052] As a further technical solution of the present invention, the adaptive residual weight allocation strategy of the PINN physical constraint module in S3 is as follows:

[0053] Residual weight adaptive algorithm:

[0054]

[0055] Where, λ k (t) represents the residual weight of the k-th term, λ k(0) represents the initial weight value, a is the time decay coefficient, t is the training time step, b is the transition gain, erf(·) is the Gaussian error function, and λ k (∞) represents the asymptotically stable weight.

[0056] As a further technical solution of the present invention, the environmental parameter dielectric compensation term is dynamically corrected by temperature and humidity:

[0057] ∈ corr = 1 + 0.02(RH - 50%)

[0058] Where, ∈ corr The dielectric constant is dynamically corrected, and RH is the ambient relative humidity.

[0059] As a further technical solution of the present invention, the potential field φ(x,y,z,t) output in S4 is used for:

[0060] Locate the surface potential distortion area of ​​the ADSS optical cable;

[0061] Predicted risk level of electrical erosion:

[0062]

[0063] Among them, Risk represents the risk level of electrical erosion, ∫ s |·| represents the surface integral operator. Let ds be the potential gradient vector and ds be the area element.

[0064] As a further technical solution of the present invention, the environmental parameter quantification and fusion includes establishing a composite environmental correction factor for humidity, temperature, and air pressure:

[0065]

[0066] Where k is the composite environmental correction factor, T is the ambient temperature, P is the atmospheric pressure, 0.7 is the humidity attenuation base, 65% is the humidity reference base, 1013 is the standard atmospheric pressure, and 0.01 is the air pressure compensation coefficient.

[0067] Dynamic synergy between environmental perception and data augmentation: The composite environmental correction factor generates a quantitative environmental interference intensity index by fusing humidity, temperature and air pressure parameters in real time. This index directly affects the core noise generation module of the data augmentation strategy, dynamically adjusting the noise injection intensity of the training samples: enhancing the noise amplitude in high humidity and low pressure environments to simulate the real interference spectrum, and reducing the noise intensity in dry and stable conditions.

[0068] Data augmentation strategies:

[0069]

[0070] τ = 100 μs

[0071] Where σ is the noise amplitude, E peak t represents the current peak electric field, t represents the current timestamp, t0 represents the noise center time, τ represents the pulse width constant, and 0.1 represents the amplitude adjustment coefficient. To meet the specific needs of power scenarios, the robustness of the model under overvoltage conditions is improved.

[0072] A three-stage training strategy is adopted:

[0073] Phase 1: Train only the CNN part, freeze PINN, learning rate 10. -3 ;

[0074] Phase 2: Jointly train CNN+PINN with a learning rate of 10. -4 ;

[0075] Phase 3: Fine-tuning the physical residuals with a learning rate of 10%. -5 .

[0076] In a second aspect, an electronic device is provided, including a processor, a memory, and a computer program stored in the memory, wherein when the computer program is executed by the processor, it implements the ADSS optical cable potential measurement method based on a CNN-PINN fusion network as described in any one of the first aspects.

[0077] Thirdly, a computer-readable medium is provided having a computer program stored thereon, which, when executed by a processor, implements the ADSS optical cable potential measurement method based on a CNN-PINN fusion network as described in any one of the first aspects.

[0078] Compared with existing technologies, the advantages of this ADSS optical cable potential measurement method, equipment, and medium based on CNN-PINN fusion network are:

[0079] By integrating and improving CNN and PINN, the traditional technical bottlenecks are completely broken: the non-contact MEMS sensor array combined with UNet's multi-scale feature fusion capability accurately captures high-gradient distortion features around the insulator; the CBAM attention mechanism dynamically weights key areas in the channel and spatial dimensions, significantly improving the ability to perceive local electric field distortion; PINN, by embedding Poisson equation residuals and field strength gradient constraints, forces the output to conform to electrostatic laws, eliminates the non-physical understanding generated by pure data-driven models in rain and fog environments, and reduces measurement errors compared to traditional contact sensors;

[0080] To address the harsh environment of power transmission corridors, including high humidity and strong electromagnetic interference, an environment-electric field coupling modeling system is implemented: a three-channel input tensor is constructed to decouple multi-source information; a dynamic dielectric constant compensation model is used to correct material properties in real time to solve the humidity diffusion hysteresis effect; and a frequency domain attention mechanism is used to focus on the corona discharge characteristic frequency band, effectively suppressing broadband environmental noise and ensuring the accuracy and stability of reconstruction under strong interference conditions.

[0081] The output layer generates a three-dimensional potential field distribution and calculates the electro-erosion risk level in real time based on the comprehensive spatial gradient and time-varying characteristics, providing decision support for ADSS optical cable electro-erosion protection. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0083] Figure 2 A schematic diagram of the CNN feature extraction module with UNet added;

[0084] Figure 3 This is a schematic diagram of the improved CNN feature extraction module network structure;

[0085] Figure 4 A schematic diagram of the improved CNN-PINN fusion network structure;

[0086] Figure 5 A schematic diagram illustrating the accuracy of the improved CNN-PINN fusion network;

[0087] Figure 6 A diagram illustrating the accuracy comparison of the improved CNN-PINN fusion network under different environments;

[0088] Figure 7 A schematic diagram illustrating the measurement error of the improved CNN-PINN fusion network. Detailed Implementation

[0089] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] Please see the appendix Figure 1 - Appendix Figure 7 The present invention provides an embodiment of an ADSS optical cable potential measurement method based on a CNN-PINN fusion network, comprising the following steps:

[0091] S1: Acquire the spatial electric field vector (Ex, Ey, Ez) around the optical cable through a non-contact MEMS sensor array, simultaneously obtain the ambient temperature and humidity parameters through temperature and humidity sensors, construct a three-channel tensor, and collect measured field strength data through a beam splitter and optical power meter;

[0092] The steps to construct a three-channel tensor are as follows:

[0093] The electric field in space is decomposed into three physically defined components, and a vectorized model of the electric field is performed:

[0094]

[0095] Normalization process:

[0096] θ′=π / θ+0.5

[0097]

[0098] Where |E| is the original electric field strength; θ is the electric field direction angle; To increase the vertical tilt angle; Significantly improves the modeling accuracy of the potential field beneath ultra-high voltage conductors; normalization process eliminates directional singularities;

[0099] The three-channel tensor is:

[0100] Channel 1: Original electric field strength:

[0101] |E|

[0102] Channel 2: Electric field direction angle:

[0103] θ = arctan(Ey / Ex)

[0104] Channel 3: Combined Influencing Factors of Ambient Humidity and Temperature

[0105] K = (RH / 25℃) * T -0.3

[0106] Where RH represents the ambient relative humidity; T represents the ambient temperature;

[0107] S2: Input the three-channel tensor into the CNN feature extraction module to extract high-dimensional feature vectors. The CNN feature extraction module adopts the UNet structure for multi-scale feature fusion, which includes a 4-level downsampling encoding layer and a 4-level upsampling decoding layer. The 4-level downsampling is responsible for extracting multi-scale global features, and the 4-level upsampling is responsible for restoring spatial details. A CBAM attention mechanism module is inserted after each convolutional layer to realize dynamic feature weighting of channels and spatial dimensions.

[0108] The network structure of the CNN feature extraction module with UNet is as follows: Figure 2 As shown:

[0109]

[0110] Where t is the subpixel convolution operator; C 1×1 It is a 1×1 depthwise separable convolution; This represents the output feature of the k-th level encoder. C represents the output feature of the k-th level decoder; 1×1 for; for; For example, depthwise separable convolution reduces parameters, differential fusion enhances edge abrupt features, and subpixel convolution eliminates ringing effects.

[0111] The processing flow of the CBAM attention mechanism module includes:

[0112] Channel attention: Mc(F) = σ(MLP(AvgPoo1(F)) + MLP(MaxPoo1(F)))

[0113] Spatial attention: Ms(F)=σ(f7×7([AvgPoo1(F);MaxPoo1(F)]))

[0114] Where F is the input feature map; σ is the Sigmoid activation function;

[0115] The attention mechanism employs a dual-domain (spectral-spatial) attention approach.

[0116] M f (F)=F -1 [σ(F(F)⊙W f )]

[0117]

[0118] Among them, M f (F) represents the frequency domain attention output, F(·) represents the Fourier transform operator, and F -1 (·) represents the inverse Fourier transform operator, ⊙ represents element-wise multiplication, and W f For a frequency domain bandpass filter, f is the frequency variable, f c Center frequency; frequency domain attention-focused corona discharge characteristic frequency band; Laplacian operator enhances the risk region with potential gradient exceeding 25kV / cm; improved CNN feature extraction module network structure as follows: Figure 3 As shown;

[0119] Introducing potential gradient stabilization techniques to solve gradient explosion in potential field reconstruction:

[0120]

[0121] Where is the gradient of the loss function with respect to the parameters, ‖·‖2 is the L2 norm operator, γ is the gradient clipping threshold, ε is the minimum positive number, and min(1,·) is the threshold cutoff function; which improves the training convergence speed.

[0122] In the ADSS optical cable potential field reconstruction process, the gradient explosion problem is mainly caused by the high-order derivative oscillations triggered by the rapid potential jump around the insulator. This numerical instability will seriously disrupt the model training process. The gradient stabilization technology innovatively achieves dynamic control through a three-pronged mechanism: first, it monitors the gradient amplitude changes in backpropagation in real time, and immediately triggers the intervention mechanism when an abnormal growth trend is detected; second, it uses adaptive scaling technology to precisely limit the gradient change range and constrain the update of runaway parameters within a physically reasonable threshold; and finally, it retains the effective learning signal through an intelligent truncation strategy.

[0123] S3: Input the high-dimensional feature vector extracted by the CNN feature extraction module and the measured field strength data into the PINN physical constraint module. Introduce the following into the loss function: Poisson equation residual term, field strength gradient constraint term and environmental parameter dielectric compensation term. The PINN physical constraint module adopts an adaptive residual weight allocation strategy.

[0124] The Poisson equation residual term introduced by the loss function of the PINN physical constraint module:

[0125]

[0126] Among them, R phys For the residual term of the Poisson equation, ∈ corr The dielectric constant is dynamically corrected. Let p be the potential gradient and p be the space charge density;

[0127] Field strength gradient constraint term:

[0128]

[0129] Among them, R curl For the field strength irrotational constraint term, Here, E is the curl operator, and E is the electric field intensity vector.

[0130] Environmental parameter dielectric compensation item:

[0131]

[0132] Where, ∈ corr denoted as the dynamically corrected dielectric constant; ∈0 is the vacuum dielectric constant; χ0 is the intrinsic polarizability of the material; RH is the ambient relative humidity; τ is the humidity diffusion time constant; γ is the hysteresis factor.

[0133] The adaptive residual weight allocation strategy of the PINN physical constraint module is as follows:

[0134] Residual weight adaptive algorithm:

[0135]

[0136] Where, λ k (t) represents the residual weight of the k-th term, λ k (0) represents the initial weight value, a is the time decay coefficient, t is the training time step, b is the transition gain, erf(·) is the Gaussian error function, and λ k (∞) represents the asymptotically stable weight;

[0137] The environmental parameter dielectric compensation term is dynamically corrected by temperature and humidity:

[0138] ∈ corr = 1 + 0.02(RH - 50%)

[0139] Where, ∈ corr The dielectric constant is dynamically adjusted, and RH represents the ambient relative humidity; the improved CNN-PINN fusion network structure is as follows: Figure 4 As shown;

[0140] S4: Through end-to-end joint training, output the three-dimensional spatial potential distribution field Φ(x,y,z,t) to realize dynamic potential measurement and calculate and predict the risk level of electrolytic erosion in real time;

[0141] The output potential field Φ(x,y,z,t) is used for:

[0142] Locate the surface potential distortion area of ​​the ADSS optical cable;

[0143] Predicted risk level of electrical erosion:

[0144]

[0145] Among them, Risk represents the risk level of electrical erosion, ∫ s |·| represents the surface integral operator. Let ds be the potential gradient vector, and ds be the area element.

[0146] The quantitative integration of environmental parameters includes establishing a composite environmental correction factor for humidity, temperature, and air pressure.

[0147]

[0148] Where k is the composite environmental correction factor, T is the ambient temperature, P is the atmospheric pressure, 0.7 is the humidity attenuation base, 65% is the humidity reference base, 1013 is the standard atmospheric pressure, and 0.01 is the air pressure compensation coefficient.

[0149] Dynamic synergy between environmental perception and data augmentation: The composite environmental correction factor generates a quantitative environmental interference intensity index by fusing humidity, temperature and air pressure parameters in real time. This index directly affects the core noise generation module of the data augmentation strategy, dynamically adjusting the noise injection intensity of the training samples: enhancing the noise amplitude in high humidity and low pressure environments to simulate the real interference spectrum, and reducing the noise intensity in dry and stable conditions.

[0150] Data augmentation strategies:

[0151]

[0152] τ = 100 μs

[0153] Where σ is the noise amplitude, E peak t represents the current peak electric field, t represents the current timestamp, t0 represents the noise center time, τ represents the pulse width constant, and 0.1 represents the amplitude adjustment coefficient. To meet the specific needs of power scenarios, the robustness of the model under overvoltage conditions is improved.

[0154] A three-stage training strategy is adopted:

[0155] Phase 1: Train only the CNN part, freeze PINN, learning rate 10. -3 ;

[0156] Phase 2: Jointly train CNN+PINN with a learning rate of 10. -4 ;

[0157] Phase 3: Fine-tuning the physical residuals with a learning rate of 10%. -5 .

[0158] One embodiment of the present invention is an electronic device comprising a processor, a memory, and a computer program stored in the memory. When the computer program is executed by the processor, it implements an ADSS optical cable potential measurement method based on a CNN-PINN fusion network.

[0159] One embodiment of the present invention provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements an ADSS optical cable potential measurement method based on a CNN-PINN fusion network.

[0160] In summary, this invention, by integrating and improving CNN and PINN, completely breaks through the bottlenecks of traditional technology: the non-contact MEMS sensor array combined with the multi-scale feature fusion capability of UNet accurately captures high-gradient distortion features around the insulator; the CBAM attention mechanism dynamically weights key areas in the channel and spatial dimensions, significantly improving the ability to perceive local electric field distortion; PINN, by embedding Poisson equation residuals and field strength gradient constraints, forces the output to conform to electrostatic laws, eliminating the non-physical aspects generated by pure data-driven models in rain and fog environments, thus reducing measurement errors compared to traditional contact sensors.

[0161] To address the harsh environment of power transmission corridors, including high humidity and strong electromagnetic interference, an environment-electric field coupling modeling system is implemented: a three-channel input tensor is constructed to decouple multi-source information; a dynamic dielectric constant compensation model is used to correct material properties in real time to solve the humidity diffusion hysteresis effect; and a frequency domain attention mechanism is used to focus on the corona discharge characteristic frequency band, effectively suppressing broadband environmental noise and ensuring the accuracy and stability of reconstruction under strong interference conditions.

[0162] The output layer generates a three-dimensional potential field distribution and calculates the electro-erosion risk level in real time based on the comprehensive spatial gradient and time-varying characteristics, providing decision support for ADSS optical cable electro-erosion protection.

[0163] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An ADSS optical cable potential measurement method based on CNN-PINN fusion network, characterized in that: Includes the following steps: S1: Acquire the spatial electric field vector (Ex, Ey, Ez) around the optical cable through a non-contact MEMS sensor array, simultaneously obtain the ambient temperature and humidity parameters through temperature and humidity sensors, construct a three-channel tensor, and collect measured field strength data through a beam splitter and optical power meter; S2: Input the three-channel tensor into the CNN feature extraction module to extract high-dimensional feature vectors. The CNN feature extraction module adopts the UNet structure for multi-scale feature fusion, which includes 4 levels of downsampling encoding layers and 4 levels of upsampling decoding layers. After each convolutional layer, the CBAM attention mechanism module is inserted to realize dynamic feature weighting of channel and spatial dimensions. S3: Input the high-dimensional feature vector extracted by the CNN feature extraction module and the measured field strength data into the PINN physical constraint module. Introduce the following into the loss function: Poisson equation residual term, field strength gradient constraint term and environmental parameter dielectric compensation term. The PINN physical constraint module adopts an adaptive residual weight allocation strategy. S4: Through end-to-end joint training, output the three-dimensional spatial potential distribution field Φ(x,y,z,t) to realize dynamic potential measurement and calculate and predict the risk level of electrolytic erosion in real time.

2. The ADSS optical cable potential measurement method based on CNN-PINN fusion network according to claim 1, characterized in that: The steps for constructing the three-channel tensor in S1 are as follows: The electric field in space is decomposed into three physically distinct components: Normalization process: i ′ =π / θ+0.5 Where |E| is the original electric field strength; θ is the electric field direction angle; It is the vertical tilt angle.

3. The ADSS optical cable potential measurement method based on CNN-PINN fusion network according to claim 1, characterized in that: The three-channel tensor is: Channel 1: Original electric field strength: |E| Channel 2: Electric field direction angle: θ = arctan(Ey / Ex) Channel 3: Combined Influencing Factors of Ambient Humidity and Temperature K=(RH / 25℃)*T -0.3 Where RH is the ambient relative humidity; T is the ambient temperature.

4. The ADSS optical cable potential measurement method based on CNN-PINN fusion network according to claim 1, characterized in that: The network structure of the CNN feature extraction module with UNet added in S2 is as follows: Where t is the subpixel convolution operator; C 1×1 It is a 1×1 depthwise separable convolution; This represents the output feature of the k-th level encoder. This represents the output feature of the k-th level decoder.

5. The ADSS optical cable potential measurement method based on CNN-PINN fusion network according to claim 1, characterized in that: The processing flow of the CBAM attention mechanism module in S2 includes: Channel attention: Mc(F) = σ(MLP(AvgPoo1(F)) + MLP(MaxPoo1(F)) Spatial attention: Ms(F)=σ(f7×7([AvgPoo1(F);MaxPoo1(F)])) Where F is the input feature map; σ is the Sigmoid activation function; The attention mechanism employs a dual-domain (spectral-spatial) attention approach. M f (F)=F -1 [σ(F(F)⊙W f )] Among them, M f (F) represents the frequency domain attention output, F(·) represents the Fourier transform operator, and F -1 (·) represents the inverse Fourier transform operator, ⊙ represents element-wise multiplication, and W f For a frequency domain bandpass filter, f is the frequency variable, f c The center frequency; Introducing potential gradient stabilization techniques to solve gradient explosion in potential field reconstruction: in, Let γ be the gradient of the loss function with respect to the parameters, ‖·‖2 be the L2 norm operator, γ be the gradient clipping threshold, ε be the minimum positive number, and min(1,·) be the threshold cutoff function. Gradient stabilization technology achieves dynamic control through a three-pronged mechanism: First, it monitors the gradient magnitude changes in backpropagation in real time and triggers an intervention mechanism immediately when an abnormal growth trend is detected; second, it uses adaptive scaling technology to precisely limit the range of gradient changes and constrain the update of runaway parameters within a physically reasonable threshold; and finally, it retains effective learning signals through an intelligent truncation strategy.

6. The ADSS optical cable potential measurement method based on CNN-PINN fusion network according to claim 1, characterized in that: The Poisson equation residual term introduced by the loss function of the PINN physical constraint module in S3: Among them, R phys For the residual term of the Poisson equation, ∈ corr The dielectric constant is dynamically corrected. Let p be the potential gradient and p be the space charge density; Field strength gradient constraint term: Among them, R curl For the field strength irrotational constraint term, Here, E is the curl operator, and E is the electric field intensity vector. Environmental parameter dielectric compensation item: Where, ∈ corr denoted as the dynamically corrected dielectric constant; ∈0 is the vacuum dielectric constant; χ0 is the intrinsic polarizability of the material; RH is the ambient relative humidity; τ is the humidity diffusion time constant; and γ is the hysteresis factor.

7. The ADSS optical cable potential measurement method based on CNN-PINN fusion network according to claim 1, characterized in that: The adaptive residual weight allocation strategy of the PINN physical constraint module in S3 is as follows: Residual weight adaptive algorithm: Where, λ k (t) represents the residual weight of the k-th term, λ k (0) represents the initial weight value, a is the time decay coefficient, t is the training time step, b is the transition gain, erf(.) is the Gaussian error function, and λ k (∞) represents the asymptotically stable weight.

8. The ADSS optical cable potential measurement method based on CNN-PINN fusion network according to claim 6, characterized in that: The dielectric compensation term for the environmental parameter is dynamically corrected by temperature and humidity: ∈ corr =1+0.02(RH-50%) Where, ∈ corr The dielectric constant is dynamically corrected, and RH is the ambient relative humidity.

9. The ADSS optical cable potential measurement method based on CNN-PINN fusion network according to claim 1, characterized in that: The potential field Φ(x,y,z,t) output in S4 is used for: Locate the surface potential distortion area of ​​the ADSS optical cable; Predicted risk level of electrical erosion: Among them, Risk represents the risk level of electrical erosion, ∫ s |·| represents the surface integral operator. Let ds be the potential gradient vector and ds be the area element.

10. The ADSS optical cable potential measurement method based on CNN-PINN fusion network according to claim 1, characterized in that: The quantitative fusion of environmental parameters includes establishing a composite environmental correction factor for humidity, temperature, and air pressure. Where k is the composite environmental correction factor, T is the ambient temperature, P is the atmospheric pressure, 0.7 is the humidity attenuation base, 65% is the humidity reference base, 1013 is the standard atmospheric pressure, and 0.01 is the air pressure compensation coefficient. Dynamic synergy between environmental perception and data augmentation: The composite environmental correction factor generates a quantitative environmental interference intensity index by fusing humidity, temperature and air pressure parameters in real time. This index directly affects the core noise generation module of the data augmentation strategy, dynamically adjusting the noise injection intensity of the training samples: enhancing the noise amplitude in high humidity and low pressure environments to simulate the real interference spectrum, and reducing the noise intensity in dry and stable conditions. Data augmentation strategies: τ = 100 μs Where σ is the noise amplitude, E peak t is the current peak value of the electric field, t is the current timestamp, t0 is the noise center time, τ is the pulse width constant, and 0.1 is the amplitude adjustment coefficient.

11. An electronic device comprising a processor, a memory, and a computer program stored in the memory, characterized in that, When the computer program is executed by the processor, it implements the ADSS optical cable potential measurement method based on the CNN-PINN fusion network as described in any one of claims 1-10.

12. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the ADSS optical cable potential measurement method based on the CNN-PINN fusion network as described in any one of claims 1-10.

Images