Method for detecting corrosion degree of mine double-loop power supply line

Abstract

The application relates to the technical field of power system monitoring and fault diagnosis, and discloses a corrosion degree detection method based on a mine double-loop power supply line. The method comprises the following steps: injecting a dynamic frequency modulated micro-amplitude excitation signal into the double loop; acquiring an interference characteristic spectrum and generating an anti-aliasing filter parameter through an environmental noise sensing module; separating high-frequency transient and low-frequency steady-state components by using wavelet packet decomposition and empirical mode decomposition dual-channel preprocessing; combining a deep time sequence encoder embedded with an electrochemical model prior to extract corrosion features, and reconstructing a pure corrosion signal through an interference suppression decoder driven by an attention mechanism; and finally outputting a quantitative index by a corrosion rate inversion model and triggering a graded early warning. The system comprises a signal injection unit, a noise sensing module, a dual-channel preprocessing unit, a deep encoding-decoding network and an early warning module. The application significantly improves the accuracy and robustness of corrosion detection through multi-scale feature decoupling, physical constraint modeling and noise adaptive suppression.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power system monitoring and fault diagnosis, and particularly relates to a corrosion degree detection method based on a mine double-loop power supply line. BACKGROUND

[0002] With the continuous expansion of the mining depth and scale, the safety and reliability of the power supply system has become a key link to ensure the continuity of production. As the core power infrastructure, the mine double-loop power supply line is long-term in a high-humidity, high-salt, strong-corrosive gas and complex electromagnetic environment, and the metal conductor and the connecting part are prone to electrochemical corrosion, which leads to an increase in contact resistance, a decrease in current-carrying capacity, and even causes power failure. In order to prevent such risks, the online detection technology of the corrosion state has attracted widespread attention, and the weak current signal monitoring method based on the electrochemical principle is widely used due to its high sensitivity and fast response.

[0003] The corrosion degree detection is usually performed by applying a micro excitation signal to the power supply line, collecting the corrosion current response generated thereby, and then inverting the corrosion rate and damage degree of the metal surface. This method relies on the accurate extraction and analysis of the weak electrochemical signal, and its effectiveness is highly dependent on the signal-to-noise ratio and the suppression ability of environmental interference.

[0004] In the actual mine scene, there are serious challenges: there are a large number of strong electromagnetic noise sources in the mine environment, such as large motor start-stop, frequency converter switching and arc discharge, etc. These transient interferences have a wide frequency spectrum and high energy, which are easily coupled into the detection circuit, causing the corrosion current signal to be severely distorted; the traditional detection system mostly uses fixed frequency excitation and single threshold filtering strategy, which cannot adapt to the signal characteristics of different corrosion stages, and it is also difficult to effectively distinguish the electrochemical response generated by the real corrosion from the pseudo signal caused by electromagnetic interference, resulting in a high false alarm rate. Especially in the double-loop alternating operation or load mutation working condition, the interference and corrosion signal are highly overlapped in time-frequency domain, the existing method lacks the deep recognition ability of the signal essential characteristics, and it is difficult to realize the corrosion state discrimination with high confidence, which limits the reliability and practicality of the intelligent operation and maintenance of the power supply line. SUMMARY

[0005] The application provides a corrosion degree detection method based on a mine double-loop power supply line, which realizes high-fidelity extraction of weak corrosion current signals and accurate discrimination of corrosion states in a strong electromagnetic interference environment by constructing an adaptive excitation-response signal coupling mechanism and a multi-scale time-frequency feature decoupling architecture.

[0006] The method first injects a micro excitation signal with dynamic frequency modulation characteristics into the double-loop power supply line, and synchronously collects the corresponding corrosion current response signal;

[0007] Subsequently, the interference characteristic spectrum under the current working condition is acquired in real time by using the environmental electromagnetic noise sensing module, and an anti-aliasing filter parameter set is generated accordingly;

[0008] Then, the original response signal is input into a double-channel signal preprocessing unit composed of wavelet packet decomposition and empirical mode decomposition, and high-frequency transient components and low-frequency steady-state components are extracted respectively; on this basis, a deep time encoder fusing prior knowledge of a physical electrochemical model is constructed, multi-scale feature embedding is performed on the preprocessed signal, and a potential representation of the corrosion state is formed; at the same time, an interference suppression decoder based on attention weight distribution is introduced, and the reconstruction weight of each frequency band channel is dynamically adjusted according to the environmental noise spectrum, so as to separate out a pure corrosion feature signal.

[0009] Finally, the pure signal is input into a corrosion rate inversion model, and a quantitative corrosion degree index is output, and a hierarchical early warning mechanism is triggered.

[0010] As an embodiment of the present application, the injected micro-excitation signal specifically includes: an isolation type signal injection unit is arranged between the grounding ends of the main circuit and the standby circuit of the double circuit power supply line, the unit is composed of a high-precision digital-to-analog converter, a power amplification circuit and an electrical isolation transformer; the high-precision digital-to-analog converter receives digital excitation instructions from the central controller to generate a sine sweep signal with an amplitude of 0.1 volts to 0.5 volts and a frequency range of 10 hertz to 5000 hertz; the power amplification circuit amplifies the signal to a level sufficient to overcome the distributed impedance of the line, and the output current does not exceed 5 milliamperes; the electrical isolation transformer ensures that the excitation signal only acts on the measured metal structure and does not affect the normal operation of the main power supply circuit; the frequency scanning strategy of the excitation signal adopts a nonlinear logarithmic stepping mode, the initial frequency is 10 hertz, the final frequency is 5000 hertz, the frequency point is updated every 10 milliseconds, and the single scanning period is two seconds.

[0011] As an embodiment of the present application, the environmental electromagnetic noise sensing module includes a wideband electromagnetic field probe array, a high-speed analog-to-digital converter and a noise feature extraction unit; the wideband electromagnetic field probe array is composed of three mutually orthogonal magnetic loop antennas and three mutually orthogonal electric dipole antennas, covering a frequency range from direct current to 100 megahertz; the sampling rate of the high-speed analog-to-digital converter is 250 megahertz, and the quantization bit number is 16, which is used for synchronous acquisition of electromagnetic field components of 6 channels; the noise feature extraction unit performs short-time Fourier transform, calculates the energy density distribution of each frequency band, and identifies transient interference events corresponding to energy mutation points; the noise characteristic spectrum is output in the form of a frequency domain energy vector, and the dimension is 1000, and the frequency resolution is 100 kilohertz.

[0012] In one embodiment of the present invention, the dual-channel signal preprocessing unit employs the db4 wavelet basis function in the wavelet packet decomposition channel, with six decomposition layers, generating 64 sub-band signals. The empirical mode decomposition channel adaptively decomposes the original response signal, generating several intrinsic mode function components until the standard deviation of the residual signal is less than a preset threshold of 0.001. The high-frequency transient components are reconstructed from the sub-band signals of the third to sixth layers in the wavelet packet decomposition channel, covering a frequency range of 625 Hz to 5000 Hz. The low-frequency steady-state components are superimposed from the first three intrinsic mode function components in the empirical mode decomposition channel, covering a frequency range of 10 Hz to 625 Hz.

[0013] In one embodiment of the present invention, the deep temporal encoder is composed of a three-layer stacked bidirectional long short-term memory network, with 256 hidden units in each layer; its input is a time-series concatenation vector of high-frequency transient components and low-frequency steady-state components, with a time window length of 4 seconds, a sampling interval of 10 milliseconds, and a total of 400 time steps; at the first layer input of the encoder, a physical constraint layer derived from electrochemical impedance spectroscopy theory is embedded, which forces the network output to satisfy the complex impedance relationship with a semi-circular arc feature in the Nyquist plot; the mathematical expression of the physical constraint layer is: the real impedance and the imaginary impedance satisfy the equation ,in For the resistance of the solution, Both are charge transfer resistors and are learnable parameters.

[0014] In one embodiment of the present invention, the interference suppression decoder includes a channel attention module and a temporal reconstruction module; the channel attention module receives the feature tensor output by the deep temporal encoder, which has a shape of 400 time steps × 512 feature dimensions; the module first performs global average pooling on each feature channel in the temporal dimension to obtain the channel description vector; then, it calculates the attention weight of each channel through a two-layer fully connected network, the hidden layer of which has 64 neurons and the activation function is a modified linear unit;

[0015] The input to the fully connected network is a concatenated vector of the channel description vector and the environmental noise feature spectrum; the temporal reconstruction module generates a pure corrosion current signal through a deconvolutional network based on the weighted feature tensor, and its output sampling rate is consistent with the original signal.

[0016] In one embodiment of the present invention, the corrosion rate inversion model is a multilayer perceptron network, whose inputs are the time-domain statistical features and frequency-domain spectral moment features of the pure corrosion current signal; the time-domain statistical features include mean, variance, skewness, kurtosis, and zero-crossing rate; the frequency-domain spectral moment features include first to fourth order central moments; the multilayer perceptron network contains three hidden layers with 128, 64, and 32 neurons respectively, and the output layer is a single neuron that directly outputs the corrosion rate in micrometers per year; the model is trained using a laboratory accelerated corrosion experiment dataset for supervised learning, and the labeled data is obtained through weightlessness calibration.

[0017] As one embodiment of the present invention, the graded early warning mechanism sets three threshold levels based on the value of the corrosion rate: when the corrosion rate is less than 10 micrometers per year, it is determined to be in a normal state and no early warning is triggered; when the corrosion rate is greater than or equal to 10 micrometers per year and less than 30 micrometers per year, it is determined to be mild corrosion, triggering a yellow warning, and it is recommended to check the section during the next planned maintenance.

[0018] When the corrosion rate is greater than or equal to 30 micrometers per year, it is judged as severe corrosion, triggering a red alert. The maintenance personnel are immediately notified to conduct on-site verification and prepare to switch to the backup circuit.

[0019] As one embodiment of the present invention, the method further includes a dual-loop cooperative detection mode; in this mode, when the main loop is in operation, the excitation signal is injected only into the main loop, and the backup loop is used as a reference ground;

[0020] When the system detects a sudden change in the load of the main circuit or a switching action, it automatically switches to the backup circuit to inject an excitation signal and uses the idle state of the main circuit as a noise reference channel; by comparing the response differences of the two circuits under the same excitation, common-mode interference is further eliminated.

[0021] As one embodiment of the present invention, the method further includes a historical data fusion and correction step; this step verifies the consistency between the corrosion rate output in the current detection cycle and the historical cumulative corrosion depth; the historical cumulative corrosion depth is obtained by integrating the current corrosion rate and the running time, and then summed with the cumulative value of the previous cycle.

[0022] If the corrosion rate in the current cycle causes the cumulative depth to increase non-monotonically or abruptly exceed 20%, the abnormality review process will be initiated, the signal of that section will be reacquired, and the signal acquisition window will be extended to 8 seconds to improve the signal-to-noise ratio.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] 1. This invention effectively solves the problem of corrosion signal distortion under strong electromagnetic interference by introducing a dynamic frequency modulated excitation signal and an adaptive filtering mechanism driven by environmental noise perception.

[0025] 2. By utilizing a dual-channel preprocessing architecture of wavelet packet decomposition and empirical mode decomposition, the high-frequency transient features and low-frequency steady-state features in corrosion signals are separated and extracted, overcoming the shortcomings of traditional single filtering methods that cannot take into account the fidelity of signals in different frequency bands.

[0026] 3. By embedding prior knowledge of the electrochemical physics model into the deep temporal encoder, the feature extraction process is ensured to conform to the basic laws of corrosion electrochemistry, thereby improving the model's generalization ability and physical interpretability. By using the channel attention mechanism to integrate the environmental noise feature spectrum into the interference suppression decoding process, dynamic attenuation of the interference frequency band and enhanced reconstruction of the corrosion feature frequency band are achieved, significantly improving the accuracy of signal separation.

[0027] 4. By integrating multi-dimensional features and verifying the consistency of historical data, a high-confidence quantitative assessment system for corrosion degree was constructed, which significantly reduced the false alarm rate and the missed alarm rate, providing reliable technical support for the safe operation and maintenance of dual-circuit power supply lines in mines. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall technical solution architecture of the corrosion degree detection method based on the dual-circuit power supply line in the mine proposed in this invention;

[0029] Figure 2 This is a schematic diagram of the core principle framework of the adaptive excitation-response signal coupling mechanism and the multi-scale time-frequency feature decoupling architecture in this invention;

[0030] Figure 3 This is a logic flowchart of the dual-channel signal preprocessing unit in this invention;

[0031] Figure 4 This is a schematic diagram of the joint feature extraction and reconstruction framework of the deep temporal encoder and the interference suppression decoder that integrates prior knowledge of the physical electrochemical model in this invention;

[0032] Figure 5 This is a schematic diagram of the interaction relationship between excitation injection and noise reference and the data flow in the dual-loop cooperative detection mode of this invention;

[0033] Figure 6 This is a flowchart illustrating the hierarchical early warning logic of corrosion rate inversion and historical data fusion correction in this invention. Detailed Implementation

[0034] Please refer to Figures 1 to 6This invention provides a corrosion degree detection method based on a dual-circuit power supply line in a mine, aiming to solve the technical challenge of high-fidelity extraction and accurate corrosion state determination of weak corrosion current signals under strong electromagnetic interference. The method achieves deep identification of the essential characteristics of corrosion signals and effective suppression of environmental noise by constructing an adaptive excitation-response signal coupling mechanism and a multi-scale time-frequency feature decoupling architecture.

[0035] The method includes the following steps:

[0036] S1, inject a micro-amplitude excitation signal with dynamic frequency modulation characteristics into the dual-circuit power supply line, and simultaneously collect the corresponding corrosion current response signal.

[0037] S2, using the environmental electromagnetic noise sensing module to obtain the interference characteristic spectrum under the current working condition in real time, and generating an anti-aliasing filter parameter set accordingly;

[0038] S3 inputs the original response signal into a dual-channel signal preprocessing unit consisting of wavelet packet decomposition and empirical mode decomposition, and extracts its high-frequency transient components and low-frequency steady-state components respectively.

[0039] S4. Construct a deep temporal encoder that integrates prior knowledge from a physical electrochemical model, and embed multi-scale features into the preprocessed signal to form a potential characterization of the corrosion state.

[0040] S5 introduces an interference suppression decoder based on attention weight allocation, which dynamically adjusts the reconstruction weights of each frequency band channel according to the environmental noise spectrum in order to separate out the pure corrosion feature signal.

[0041] S6. Input the clean signal into the corrosion rate inversion model, output a quantitative corrosion degree index, and trigger a graded early warning mechanism.

[0042] In the above method, step S1 involves injecting a micro-amplitude excitation signal with dynamic frequency modulation characteristics into the dual-circuit power supply line and simultaneously acquiring the corresponding corrosion current response signal. This step is accomplished by applying the excitation signal through an isolated signal injection unit between the grounding terminals of the main circuit and the backup circuit. The isolated signal injection unit consists of a high-precision digital-to-analog converter, a power amplifier circuit, and an electrical isolation transformer. The high-precision digital-to-analog converter receives digital excitation commands from the central controller and generates a sinusoidal sweep signal with an amplitude of 0.1V to 0.5V and a frequency range of 10Hz to 5000Hz. The power amplifier circuit amplifies this signal to a level where its driving capability is sufficient to overcome the distributed impedance of the line, and its output current does not exceed 5mA. The electrical isolation transformer ensures that the excitation signal only acts on the metal structure under test and does not affect the normal operation of the main power supply circuit.

[0043] The frequency scanning strategy for the excitation signal employs a nonlinear logarithmic stepping method, with an initial frequency of 10 Hz and a final frequency of 5000 Hz. The frequency point is updated every 10 milliseconds, and a single scan cycle is 2 seconds. Simultaneously with the excitation signal injection, the corrosion current response signal is synchronously acquired using a high-precision current sensor. The sampling rate is 100 kHz, the quantization bit depth is 16 bits, the time window length is 4 seconds, and a total of 40,000 sampling points are collected. This synchronous acquisition mechanism ensures strict time alignment between the excitation and response, providing an accurate input basis for subsequent signal analysis.

[0044] In the above method, step S2 involves using an environmental electromagnetic noise sensing module to acquire the interference characteristic spectrum under the current operating conditions in real time, and generating an anti-aliasing filter parameter set accordingly. The environmental electromagnetic noise sensing module includes a broadband electromagnetic field probe array, a high-speed analog-to-digital converter, and a noise feature extraction unit. The broadband electromagnetic field probe array consists of three mutually orthogonal magnetic loop antennas and three mutually orthogonal electric dipole antennas, covering a frequency range from DC to 100 MHz, used to capture vector information of the spatial electromagnetic field from all directions.

[0045] The high-speed analog-to-digital converter has a sampling rate of 250 MHz and a quantization bit depth of 16 bits, used for synchronous acquisition of electromagnetic field components from six channels. The noise feature extraction unit performs a short-time Fourier transform on the acquired raw data from the six channels with a window length of 4 milliseconds and an overlap rate of 50%, calculates the energy density distribution of each frequency band, and identifies transient interference events corresponding to energy abrupt change points.

[0046] The noise feature spectrum is output as a frequency domain energy vector with a dimension of 1000, corresponding to a frequency resolution of 100 kHz. This energy vector serves as the basis for generating subsequent anti-aliasing filter parameters. By identifying the region of maximum energy concentration, the interference frequency bands that need to be suppressed are determined, and corresponding band-stop filter parameters or wavelet threshold adjustment coefficients are generated. This step ensures that the system can dynamically adjust the signal processing strategy according to the real-time electromagnetic environment, avoiding the failure of fixed filter parameters under complex operating conditions.

[0047] In the above method, step S3 involves inputting the original response signal into a dual-channel signal preprocessing unit composed of wavelet packet decomposition and empirical mode decomposition, extracting its high-frequency transient components and low-frequency steady-state components, respectively. The dual-channel signal preprocessing unit contains two parallel processing paths. The first path is the wavelet packet decomposition channel, using the db4 wavelet basis function, with six decomposition layers, generating 64 sub-band signals, each covering a frequency width of 78.125 Hz. The second path is the empirical mode decomposition channel, which adaptively decomposes the original response signal, generating several intrinsic mode function components until the standard deviation of the residual signal is less than a preset threshold of 0.001.

[0048] The high-frequency transient component is reconstructed from the sub-band signals of layers three to six in the wavelet packet decomposition channel, covering a frequency range of 625 Hz to 5000 Hz. This band mainly contains the rapid electrochemical response and transient interference components induced by local corrosion points. The low-frequency steady-state component is superimposed from the first three eigenmode function components in the empirical mode decomposition channel, covering a frequency range of 10 Hz to 625 Hz. This band mainly reflects the slow evolution trend of the overall corrosion process and background noise. The time series length of both components is 4 seconds, with a sampling interval of 10 milliseconds, for a total of 400 time steps. This dual-channel architecture effectively separates the dynamic characteristics of different time scales in the corrosion signal, providing structured input for subsequent multi-scale feature embedding.

[0049] In the above method, step S4 involves constructing a deep temporal encoder that integrates prior knowledge from a physical electrochemical model. This encoder embeds multi-scale features into the preprocessed signal to form a potential characterization of the corrosion state. The deep temporal encoder consists of three stacked bidirectional long short-term memory networks, with 256 hidden units per layer. Its input is a time-series concatenation vector of high-frequency transient components and low-frequency steady-state components. The time window length is 4 seconds, the sampling interval is 10 milliseconds, and there are a total of 400 time steps, with each time step having an input dimension of 2. At the first input layer of the encoder, a physical constraint layer derived from electrochemical impedance spectroscopy theory is embedded. This physical constraint layer forces the network output to satisfy the complex impedance relationship with a semi-circular arc feature in the Nyquist plot. Specifically, the real impedance and imaginary impedance satisfy the following equation:

[0050]

[0051] in For the resistance of the solution, The charge transfer resistance and the charge transfer resistance are both learnable parameters, initially set to 0.5 ohms and 10 ohms respectively, and optimized during training using gradient descent. This physical constraint is introduced as an additional term in the loss function with a weight of 0.1, ensuring that the network learns data-driven features without deviating from fundamental electrochemical principles. The final output of the deep temporal encoder is a feature tensor with a shape of 400 time steps × 512 feature dimensions. This tensor fully encodes the multi-scale features of the corrosion signal in the time domain, frequency domain, and electrochemical physics, constituting a latent representation of the corrosion state.

[0052] In the above method, step S5 introduces an interference suppression decoder based on attention weight allocation. This decoder dynamically adjusts the reconstruction weights of each frequency band channel according to the environmental noise spectrum to separate the pure corrosion feature signal. The interference suppression decoder includes a channel attention module and a temporal reconstruction module. The channel attention module receives the feature tensor output by the deep temporal encoder, which has a shape of 400 time steps × 512 feature dimensions. This module first performs global average pooling on each feature channel in the temporal dimension to obtain a channel description vector of length 512.

[0053] Subsequently, the channel description vector is concatenated with the environmental noise feature spectrum obtained in step S2 to form a joint input vector of length 1512. This joint input vector is used to calculate the attention weights of each channel through a two-layer fully connected network. The hidden layer of this network has 64 neurons, the activation function is a modified linear unit, and the output layer is a 512-dimensional weight vector, which is normalized by the Sigmoid function and then used to weight the original feature tensor channel by channel.

[0054] The timing reconstruction module generates a clean corrosion current signal based on the weighted feature tensor through a deconvolutional network. Its output sampling rate is consistent with the original signal, at 100 kHz, and the duration is 4 seconds. This decoder explicitly models the correlation between environmental noise and feature channels, achieving dynamic attenuation of interference frequency band features and enhancement of corrosion feature frequency bands, thereby outputting a clean corrosion signal with a high signal-to-noise ratio.

[0055] In the above method, step S6 involves inputting the clean signal into the corrosion rate inversion model, outputting a quantitative corrosion degree index, and triggering a graded early warning mechanism. The corrosion rate inversion model is a multilayer perceptron network, whose inputs are the time-domain statistical characteristics and frequency-domain spectral moment characteristics of the clean corrosion current signal. The time-domain statistical characteristics include five features: mean, variance, skewness, kurtosis, and zero-crossing rate. The frequency-domain spectral moment characteristics are calculated by performing a fast Fourier transform on the clean signal, yielding first to fourth order central moments, totaling four features, for a nine-dimensional input feature vector.

[0056] The multilayer perceptron network consists of three hidden layers with 128, 64, and 32 neurons respectively. The activation function for all layers is a modified linear unit (MRU). The output layer is a single neuron that directly outputs the corrosion rate in micrometers per year. During the training phase, the model is trained using a laboratory accelerated corrosion experiment dataset under supervised learning. Label data is obtained through a weightlessness method to ensure the output results are physically interpretable and engineering-applicable.

[0057] The graded early warning mechanism sets three thresholds based on the corrosion rate: when the corrosion rate is less than 10 micrometers per year, it is considered a normal state and no warning is triggered; when the corrosion rate is greater than or equal to 10 micrometers per year but less than 30 micrometers per year, it is considered light corrosion and triggers a yellow warning, suggesting that the section be inspected during the next planned maintenance; when the corrosion rate is greater than or equal to 30 micrometers per year, it is considered severe corrosion and triggers a red warning, immediately notifying maintenance personnel to conduct on-site verification and prepare to switch to the backup circuit.

[0058] Furthermore, the method also includes a dual-loop cooperative detection mode. In this mode, when the main loop is in operation, the excitation signal is injected only into the main loop, with the backup loop serving as a reference ground. When the system detects a sudden change in the load on the main loop or a switching action, it automatically switches to injecting the excitation signal into the backup loop and utilizes the idle state of the main loop as a noise reference channel. By comparing the response differences of the two loops under the same excitation, common-mode interference is further eliminated.

[0059] In practice, the central controller monitors the rate of change of current in the main circuit in real time. If its absolute value exceeds a preset threshold of 50 amperes per second, it is determined to be a load surge event, triggering the circuit switching logic. After switching, the system executes the complete S1 to S6 process on the standby circuit, while simultaneously recording the induced voltage of the main circuit in the unexcited state as a noise reference. This reference signal is used to correct the common-mode component in the standby circuit response, improving the accuracy of corrosion signal extraction.

[0060] The method also includes a historical data fusion and correction step. This step verifies the consistency between the corrosion rate output in the current detection cycle and the historical cumulative corrosion depth. The historical cumulative corrosion depth is obtained by integrating the current corrosion rate and the running time, and then summed with the cumulative value of the previous cycle. If the corrosion rate in the current cycle causes the cumulative depth to show a non-monotonic increase or a sudden change exceeding 20%, an anomaly review process is initiated. The anomaly review process includes re-acquiring the signal for that segment and extending the signal acquisition window to 8 seconds to improve the signal-to-noise ratio. The extended signal processing flow remains unchanged, but the number of time steps within the time window is increased to 800, and the depth timing encoder adjusts its internal state buffer length accordingly. The corrected corrosion rate is used to update the historical database and serves as a benchmark reference for the next detection. This mechanism effectively prevents misjudgments caused by single detection anomalies, improving the long-term stability and reliability of the system.

[0061] In summary, this invention constructs a complete corrosion degree detection method system through the synergistic effect of multiple mechanisms, including dynamic excitation, environmental perception, dual-channel preprocessing, physical constraint encoding, attention decoding, and history correction. This method can achieve high-fidelity extraction of weak corrosion signals and accurate quantification of corrosion status even under strong electromagnetic interference environments, significantly outperforming existing fixed-frequency excitation and single-filtering strategies. This provides solid technical support for the safe and intelligent operation and maintenance of dual-circuit power supply lines in mines.

Claims

1. A method for detecting the degree of corrosion of a dual-circuit power supply line in a mine, characterized in that, include: A micro-amplitude excitation signal with dynamic frequency modulation characteristics is injected into the dual-circuit power supply line, and the corresponding corrosion current response signal is collected simultaneously. The interference characteristic spectrum under the current operating conditions is obtained in real time by using the environmental electromagnetic noise sensing module, and an anti-aliasing filter parameter set is generated accordingly. The corrosion current response signal is input to a dual-channel signal preprocessing unit consisting of wavelet packet decomposition and empirical mode decomposition, and its high-frequency transient component and low-frequency steady-state component are extracted respectively. A deep time-series encoder integrating prior knowledge from a physical-electrochemical model is constructed to embed multi-scale features into the high-frequency transient components and the low-frequency steady-state components to form a potential characterization of the corrosion state. An interference suppression decoder based on attention weight allocation is introduced, which dynamically adjusts the reconstruction weights of each frequency band channel according to the interference feature spectrum in order to separate out the pure corrosion feature signal. The purified corrosion characteristic signal is input into the corrosion rate inversion model, which outputs a quantitative corrosion degree index and triggers a graded early warning mechanism.

2. The corrosion detection method based on a dual-circuit power supply line in a mine according to claim 1, characterized in that, A micro-amplitude excitation signal with dynamic frequency modulation characteristics is injected into the dual-circuit power supply line, and the corresponding corrosion current response signal is acquired simultaneously, including: An isolated signal injection unit is provided between the grounding terminals of the main circuit and the backup circuit. The isolated signal injection unit consists of a high-precision digital-to-analog converter, a power amplifier circuit, and an electrical isolation transformer. The high-precision digital-to-analog converter receives digital excitation commands and generates a sinusoidal sweep signal with an amplitude of 0.1 volts to 0.5 volts and a frequency range of 10 Hz to 5000 Hz. The power amplifier circuit amplifies the sinusoidal sweep signal to an output current of no more than 5 mA. The electrical isolation transformer ensures that the excitation signal acts only on the metal structure under test; The sinusoidal sweep frequency signal adopts a nonlinear logarithmic stepping method, with an initial frequency of 10 Hz and an ending frequency of 5000 Hz. The frequency point is updated every 10 milliseconds, and the single scan period is 2 seconds. The corrosion current response signal is synchronously acquired using a high-precision current sensor with a sampling rate of 100 kHz and a time window length of 4 seconds.

3. The corrosion detection method based on a dual-circuit power supply line in a mine according to claim 2, characterized in that, The environmental electromagnetic noise sensing module is used to acquire the interference characteristic spectrum under the current operating conditions in real time, and an anti-aliasing filter parameter set is generated accordingly, including: Spatial electromagnetic field vector information is captured by a broadband electromagnetic field probe array, which consists of three mutually orthogonal magnetic loop antennas and three mutually orthogonal electric dipole antennas, covering a frequency range from DC to 100 MHz. The electromagnetic field components of 6 channels are simultaneously acquired by a high-speed analog-to-digital converter with a sampling rate of 250 MHz and a quantization bit depth of 16 bits. Perform a short-time Fourier transform on the collected data to calculate the energy density distribution of each frequency band and identify transient interference events corresponding to energy abrupt change points; The output frequency domain energy vector is used as the interference feature spectrum, with a dimension of 1000 and a corresponding frequency resolution of 100 kHz. Based on the frequency domain energy vector, the interference frequency bands to be suppressed are determined, and the corresponding anti-aliasing filter parameter set is generated.

4. The corrosion detection method based on a dual-circuit power supply line in a mine according to claim 3, characterized in that, The corrosion current response signal is input to a dual-channel signal preprocessing unit consisting of wavelet packet decomposition and empirical mode decomposition, to extract its high-frequency transient components and low-frequency steady-state components, including: The corrosion current response signal is input into the wavelet packet decomposition channel, and a six-level decomposition is performed using the db4 wavelet basis function to generate 64 sub-band signals. The corrosion current response signal is input into the empirical mode decomposition channel, and several intrinsic mode function components are adaptively decomposed until the standard deviation of the residual signal is less than 0.

001. The high-frequency transient component is reconstructed from the sub-band signals of the third to sixth layers of the wavelet packet decomposition channel, covering a frequency range of 625 Hz to 5000 Hz; The low-frequency steady-state component is formed by superimposing the first three intrinsic mode function components in the empirical mode decomposition channel, covering a frequency range of 10 Hz to 625 Hz. The time series lengths of both the high-frequency transient component and the low-frequency steady-state component are 4 seconds, and the sampling interval is 10 milliseconds.

5. The corrosion detection method based on a dual-circuit power supply line in a mine according to claim 4, characterized in that, A deep temporal encoder integrating prior knowledge from a physicochemical model is constructed to embed multi-scale features into the high-frequency transient components and the low-frequency steady-state components to form a potential characterization of the corrosion state, including: The time series of the high-frequency transient component and the low-frequency steady-state component are concatenated into an input vector. The time window length is 4 seconds, the sampling interval is 10 milliseconds, and there are a total of 400 time steps. The input vector is fed into a three-layer stacked bidirectional long short-term memory network, with 256 hidden units in each layer; A physical constraint layer is embedded at the first input layer to force the network output to satisfy the complex impedance relationship in electrochemical impedance spectroscopy theory: ,in For the resistance of the solution, Both of these are charge transfer resistance parameters and are learnable parameters. The output feature tensor with a shape of 400 time steps × 512 feature dimensions is used as a potential representation of the corrosion state.

6. The corrosion detection method based on a dual-circuit power supply line in a mine according to claim 5, characterized in that, An interference suppression decoder based on attention weight allocation is introduced, which dynamically adjusts the reconstruction weights of each frequency band channel according to the interference feature spectrum to separate the pure corrosion feature signal, including: Global average pooling is performed on each feature channel of the potential characterization of the corrosion state in the time dimension to obtain the channel description vector; The channel description vector is concatenated with the interference feature spectrum to form a joint input vector; The joint input vector is input into a two-layer fully connected network with 64 hidden layer neurons and a modified linear unit activation function. The output is an attention weight vector normalized by Sigmoid. The potential representation of the corrosion state is weighted channel-by-channel according to the attention weight vector; The weighted feature tensor is input into the deconvolutional network to generate a pure corrosion feature signal with a sampling rate of 100 kHz and a duration of 4 seconds.

7. The corrosion detection method based on a dual-circuit power supply line in a mine according to claim 6, characterized in that, The purified corrosion characteristic signal is input into the corrosion rate inversion model, which outputs a quantitative corrosion degree index and triggers a graded early warning mechanism, including: Time-domain statistical features and frequency-domain spectral moment features are extracted from the pure corrosion feature signal. The time-domain statistical features include mean, variance, skewness, kurtosis and zero crossing rate. The frequency-domain spectral moment features include first to fourth order central moments. The nine-dimensional feature vector is input into a multilayer perceptron network, which contains three hidden layers with 128, 64 and 32 neurons respectively, and the output layer is a single neuron. Output corrosion rate, in micrometers per year; When the corrosion rate is less than 10 micrometers per year, it is considered to be in a normal state; when the corrosion rate is greater than or equal to 10 micrometers per year but less than 30 micrometers per year, a yellow warning is triggered; when the corrosion rate is greater than or equal to 30 micrometers per year, a red warning is triggered.

8. The corrosion detection method based on a dual-circuit power supply line in a mine according to claim 7, characterized in that, It also includes a dual-loop collaborative detection mode, which includes: When the main circuit is in operation, the excitation signal is injected only into the main circuit, and the standby circuit is used as a reference ground; When the absolute value of the load change rate of the main circuit exceeds fifty amperes per second, switch to the standby circuit to inject an excitation signal and use the idle state of the main circuit as a noise reference channel. Common-mode interference is eliminated by comparing the response differences between the main circuit and the backup circuit under the same excitation.

9. The corrosion detection method based on a dual-circuit power supply line in a mine according to claim 8, characterized in that, It also includes a historical data fusion and correction step, which includes: Verify the consistency between the corrosion rate output in the current detection cycle and the historical cumulative corrosion depth. The historical cumulative corrosion depth is obtained by integrating the current corrosion rate and the running time, and then added to the cumulative value of the previous cycle. If the corrosion rate in the current cycle causes the cumulative depth to increase non-monotonically or abruptly exceed 20%, the abnormal review process will be initiated. The anomaly verification process includes re-acquiring the signal and extending the signal acquisition window to eight seconds to improve the signal-to-noise ratio.

10. The corrosion detection method based on a dual-circuit power supply line in a mine according to claim 9, characterized in that, The depth time encoder adjusts the length of its internal state cache during the anomaly review process to fit an 8-second time window, corresponding to 800 time steps.

Images