A power cable inspection robot system

Abstract

The application relates to the technical field of cable safety detection, and specifically discloses a power cable inspection robot system, which synchronously collects the self-generated sound emission waveform flow and the thermal elastic stress field distribution of a cable through an autonomous mobile platform; nonlinear modal decomposition and entropy evolution analysis are carried out by using an edge computing hub, and damage kinetic energy release rate and state phase change margin characteristic values are respectively extracted; the spatiotemporal correlation of the two is constructed through a comprehensive analysis module, a critical instability coefficient indicating the chain failure risk is calculated, a safety state grade is output through a risk prediction model, and finally a differentiated inspection decision is executed by an autonomous strategy generation module; the application realizes the fusion evaluation of the internal damage and thermal stability of the cable, can early and accurately predict the structure instability risk, and drives the autonomous adjustment of the inspection strategy of the robot, thereby improving the intelligentization, accuracy and response efficiency of the inspection.

Description

Technical Field

[0001] This invention relates to the field of cable safety inspection technology, specifically to a power cable inspection robot system. Background Technology

[0002] In modern power systems, power cables serve as the "main arteries" of power transmission, and their operational status directly affects the safety and stability of the entire power grid. Traditional manual inspection methods are inefficient, have limited coverage, and are difficult to operate in complex or hazardous environments, failing to meet the ever-increasing maintenance needs. Therefore, power cable inspection robot technology has emerged to achieve automated and intelligent line monitoring.

[0003] However, most existing inspection robot systems are limited to visible light or infrared imaging detection of cable surface defects. Their assessment methods are relatively simple, mainly relying on temperature anomalies or visual recognition, lacking in-depth analysis of the internal material damage evolution process and overall structural stability of the cable. These methods can usually only detect defects after they occur or when they are already quite obvious, resulting in insufficient early warning capabilities and an inability to effectively predict sudden, cascading failures caused by potential factors such as internal microcrack propagation, fatigue accumulation, or thermo-coupling instability. Summary of the Invention

[0004] The purpose of this invention is to provide a power cable inspection robot system to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A power cable inspection robot system includes an autonomous mobile platform module, which synchronously acquires the spontaneous acoustic emission waveform flow and the thermoelastic stress field distribution of the cable through a collaborative intelligent sensing system.

[0007] Edge computing hub module, the edge computing hub module is configured as follows:

[0008] Nonlinear mode decomposition was performed on the spontaneous acoustic emission waveform to extract the damage kinetic energy release rate characteristic value that characterizes the crack propagation rate inside the material;

[0009] Entropy evolution analysis was performed on the thermoelastic stress field distribution of the bulk material to extract the state phase transition margin characteristic value that characterizes the thermo-mechanical coupling stability under load fluctuations;

[0010] The comprehensive analysis module is used to construct the spatiotemporal correlation between the characteristic value of damage kinetic energy release rate and the characteristic value of phase transition margin, and to calculate the critical instability coefficient that predicts the risk of chain failure of the structure.

[0011] The risk prediction module inputs the critical instability coefficient into the cable fault prediction model and outputs the current safety status level of the cable.

[0012] An autonomous policy generation module generates and executes inspection decisions based on the security status level.

[0013] As a further aspect of the present invention: the nonlinear mode decomposition of the spontaneously emitted acoustic waveform specifically includes:

[0014] Empirical mode decomposition (EMD) is used to process the spontaneous acoustic emission waveform stream to obtain a set of intrinsic mode function (IMF) components. Hilbert transform is performed on each IMF component to calculate its instantaneous frequency and instantaneous amplitude. Target IMF components whose instantaneous frequencies match the typical frequency range of material crack propagation are selected, and the damage kinetic energy release rate is obtained by integrating the instantaneous amplitude of the target IMF component and dividing it by the total signal duration.

[0015] As a further aspect of the present invention: the calculation process of the damage kinetic energy release rate characteristic value is as follows:

[0016] Calculate the sample entropy of the target intrinsic mode function components within a set time window;

[0017] The sample entropy value is normalized and used as a stability weight coefficient.

[0018] The damage kinetic energy release rate is multiplied by the stability weighting coefficient, and the product is used as the final damage kinetic energy release rate characteristic value.

[0019] As a further aspect of the present invention: the entropy evolution analysis of the thermoelastic stress field distribution of the bulk body specifically includes:

[0020] The thermoelastic stress field distribution data is divided into several sub-regions in the spatial domain, and a time series of thermoelastic stress values ​​is constructed for each sub-region. For the time series of each sub-region, the permutation entropy is calculated using a sliding time window, thereby obtaining the sequence of permutation entropy changes with time for each sub-region. The permutation entropy sequences of all sub-regions are arranged according to spatial location to form a multidimensional entropy evolution field.

[0021] As a further aspect of the present invention: calculating the state phase transition margin characteristic value based on the entropy evolution field, specifically including:

[0022] The variance of the entropy evolution field is calculated over a set time period, and the reciprocal of the variance is used as the state phase transition margin characteristic value representing the stability of the thermodynamic coupling process within the corresponding time period.

[0023] As a further aspect of the present invention: the spatiotemporal correlation between the characteristic value of damage kinetic energy release rate and the characteristic value of state transition margin specifically includes:

[0024] The time series of damage kinetic energy release rate feature values ​​and the time series of state phase transition margin feature values ​​are subjected to dynamic time warping, and the cumulative distance between the two time series under the optimal warping path is calculated. The cumulative distance is input into a pre-trained radial basis function neural network. The input layer of the corresponding neural network contains three nodes corresponding to the cumulative distance and its first and second derivatives, respectively. The output layer generates a critical instability coefficient between 0 and 1 through the sigmoid activation function.

[0025] As a further aspect of the present invention: the construction process of the cable fault prediction model is as follows:

[0026] The cable fault prediction model includes multiple intermediate event nodes triggered by the critical instability coefficient; the weight allocation of each intermediate event node is determined by time series analysis; the critical instability coefficient is mapped to the basic event layer of the fault tree through fuzzy inference, and a nonlinear correspondence between the occurrence probability of basic events and the critical instability coefficient is established; finally, the system-level fault probability is obtained through uplink calculation, and the safety status level is divided according to the probability value range.

[0027] As a further aspect of the present invention: the division of the security status levels specifically includes:

[0028] The system-level failure probability values ​​are mapped to a percentage range of 0-100. Based on the range of the mapped values, four safety status levels are defined: 0-60 is a dangerous state, 60-80 is a warning state, 80-95 is a caution state, and 95-100 is a safe state. Different time series analysis window lengths and failure probability thresholds are set for dangerous and warning states. When the failure probability threshold of recent monitoring data exceeds the threshold of the corresponding state, the safety status level is automatically downgraded by one level.

[0029] As a further aspect of the present invention: the generation and execution of inspection decisions specifically includes:

[0030] When in a safe state, routine inspections are conducted along the preset basic path, collecting visible light images and infrared data. When in a state of alert, key re-inspections of historical defect points are added to the basic path to improve image acquisition resolution. When in a state of warning, high-frequency special inspections are initiated, the cycle between adjacent inspections is shortened, and extended inspections are conducted on related sections. When in a dangerous state, routine tasks are immediately terminated, and the system quickly navigates to the target location for continuous monitoring, while simultaneously sending real-time video streams to the backend platform and initiating emergency response procedures.

[0031] The beneficial effects of this invention are:

[0032] (1) This invention simultaneously acquires the spontaneous acoustic emission waveform flow and the thermoelastic stress field distribution of the cable, and extracts the "damage kinetic energy release rate characteristic value" which characterizes the crack propagation rate inside the material and the "state phase transition margin characteristic value" which characterizes the thermo-mechanical coupling stability under load fluctuations. By constructing the spatiotemporal correlation of these two key characteristic values, the "critical instability coefficient" that can predict the risk of chain failure of the structure is calculated and input into a specialized cable fault prediction model. This multi-physics field comprehensive analysis method, which integrates the internal damage dynamics of the material and the external thermo-mechanical coupling stability, can accurately identify the potential systemic fault risk of the cable and improve the foresight and reliability of power cable fault prediction.

[0033] (2) This invention not only performs risk assessments, but also automatically generates and executes differentiated inspection strategies based on the calculated safety status levels (safe, alert, warning, dangerous). For example, it performs routine inspections in a safe state; in an alert state, it increases the focus on reviewing historical defects; in a warning state, it initiates high-frequency special inspections; and in a dangerous state, it immediately terminates routine tasks, quickly navigates to the target location for continuous monitoring, and initiates emergency response procedures (such as sending real-time video streams, making phone calls to alarm, etc.). This closed-loop autonomous decision-making mechanism enables the inspection robot to intelligently allocate resources, concentrate limited inspection capabilities on high-risk areas, and improve the efficiency of inspection work and the timeliness of emergency response. Attached Figure Description

[0034] The invention will now be further described with reference to the accompanying drawings.

[0035] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0036] 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.

[0037] Please see Figure 1 As shown, the present invention is a power cable inspection robot system, comprising:

[0038] An autonomous mobile platform module is used to synchronously acquire the spontaneous acoustic emission waveform flow and the thermoelastic stress field distribution of the cable through a collaborative intelligent sensing system.

[0039] Edge computing hub module, the edge computing hub module is configured as follows:

[0040] Nonlinear mode decomposition was performed on the spontaneous acoustic emission waveform to extract the damage kinetic energy release rate characteristic value that characterizes the crack propagation rate inside the material;

[0041] Entropy evolution analysis was performed on the thermoelastic stress field distribution of the bulk material to extract the state phase transition margin characteristic value that characterizes the thermo-mechanical coupling stability under load fluctuations;

[0042] The comprehensive analysis module is used to construct the spatiotemporal correlation between the characteristic value of damage kinetic energy release rate and the characteristic value of phase transition margin, and to calculate the critical instability coefficient that predicts the risk of chain failure of the structure.

[0043] The risk prediction module inputs the critical instability coefficient into the cable fault prediction model and outputs the current safety status level of the cable.

[0044] An autonomous policy generation module generates and executes inspection decisions based on the security status level.

[0045] In the autonomous mobile platform module, the synchronous acquisition of cable data through a collaborative intelligent sensing system specifically includes:

[0046] The inspection device mounted on the autonomous mobile platform module synchronously collects acoustic signals and thermodynamic data of the cable through a collaborative sensing system. The acoustic emission sensing unit uses four piezoelectric sensors with a resonant frequency of 150kHz, which are fixed to the cable surface in a diamond pattern, with the sensor spacing strictly maintained at 1 / 4 of the cable circumference.

[0047] The infrared thermal imaging unit employs an uncooled infrared focal plane detector operating in the 8-14μm wavelength range. The detector array has a size of 640×512 pixels and a pixel pitch of 12μm. The thermal imager is mounted on a two-axis precision gimbal driven by a stepper motor, achieving an angle positioning accuracy of 0.05 degrees. During acoustic signal acquisition, a dual-trigger mechanism ensures signal integrity. The hardware trigger is set to a fixed threshold of 40dB; acquisition begins immediately when the signal amplitude exceeds this threshold. The software trigger uses a short-time averaging / long-time averaging ratio algorithm, with a ratio of 4.0 set as the trigger condition, and time window lengths of 1ms and 10ms respectively. Each trigger event records waveform data from 1024 sampling points, along with parameters such as rise time, duration, amplitude, and energy.

[0048] During thermal imaging data acquisition, the angle between the lens axis and the normal to the cable surface should not exceed 5 degrees, and the shooting distance should be controlled within 2.0 ± 0.2 meters. Environmental parameters are recorded synchronously with each image acquisition: ambient temperature is measured using a PT100 platinum resistance thermometer with an accuracy of ±0.1℃; relative humidity is measured using a capacitive humidity sensor with a range of 0-100%RH and an accuracy of ±2%RH; wind speed is measured using an ultrasonic anemometer with a range of 0-30 m / s and an accuracy of ±0.1 m / s; and solar radiation intensity is measured using a thermopile radiation sensor with a range of 0-2000 W / m² and an accuracy of ±5%.

[0049] The acoustic signal transmission channel employs a three-stage amplification structure: a built-in preamplifier in the sensor provides 20dB gain, a signal conditioning module provides adjustable gain from 0-60dB, and an ADC driver amplifier provides a fixed gain of 6dB. The signal is processed by a 100-400kHz bandpass filter with an attenuation slope set to 24dB / oct. A 16-bit ADC is used for analog-to-digital conversion, with a quantization error of less than 1.5LSB.

[0050] The thermal imager is calibrated regularly: non-uniformity correction is automatically performed every 30 minutes using a built-in blackbody reference source with blackbody temperature stability better than ±0.1℃; temperature drift compensation is performed every 2 hours by measuring a reference target with a known temperature for deviation correction; bad pixel repair uses a 3×3 neighborhood median filtering algorithm to replace abnormal pixel values ​​in real time.

[0051] The time-space synchronization system uses a GPS-disciplined cryogenic crystal oscillator as the time reference, generating a 10MHz reference clock and a 1PPS synchronization pulse. The acoustic acquisition system and the thermal imager receive synchronization signals via a BNC interface, with a time synchronization error of less than 1 microsecond. Each data packet is appended with a 64-bit timestamp, containing GPS time and sampling counter information.

[0052] The signal quality monitoring system evaluates data quality in real time: the acoustic signal-to-noise ratio is required to be greater than 20dB, the waveform distortion is less than 5%, the sensor coupling status is monitored by impedance measurement, and the impedance change range is controlled within ±10%; the thermal imaging data requires an image sharpness index greater than 0.5, temperature measurement fluctuation less than ±0.5℃, and environmental parameter measurement errors within the allowable range.

[0053] Data transmission employs a multi-channel redundancy design: the 4G network uses dual-band transmission (Band3 / Band38), the 5G network uses the n78 / n79 band, and Wi-Fi supports the 802.11ac standard. The transmission protocol uses a custom reliable data packet protocol, with each data packet containing a 16-bit CRC checksum, achieving a packet loss rate of less than 0.1%. Data encryption uses the AES-256 algorithm, with the key updated every 24 hours.

[0054] All acquired data is stored in a standardized format: acoustic emission data is stored in IEEE floating-point format, containing waveform data from 1024 sampling points and 12 bytes of header information; thermal imaging data is stored in 14-bit RAW format, with each frame accompanied by 20 bytes of metadata, including temperature calibration parameters and environmental data. Data storage employs a circular buffer mechanism with a buffer capacity of 8GB, supporting continuous 24-hour data recording.

[0055] The device's operating status is monitored through multiple indicators: CPU load rate remains below 60%, memory utilization does not exceed 70%, storage space utilization is below 80%, and battery level is maintained above 20%. When any indicator exceeds the normal range, the system automatically adjusts the collected parameters or activates the protection program.

[0056] In the edge computing hub module, the bandpass filter used in the acoustic emission signal preprocessing is a fourth-order Butterworth filter. This filter consists of two identical second-order filters cascaded together, each with a damping coefficient of 0.707. The filter allows frequency components between 100 kHz and 400 kHz to pass through, and for frequency components below 100 kHz and above 400 kHz, it achieves an attenuation of 40 dB for every tenfold increase in frequency.

[0057] Wavelet packet denoising uses the sym8 wavelet basis function for nine-level decomposition. Each level of decomposition employs the pyramid algorithm, which involves high-pass and low-pass filtering and downsampling of the approximate coefficients from the previous level. For soft thresholding, the absolute value of the wavelet coefficients is first calculated, then subtracted from a set threshold. If the difference is greater than zero, the difference is retained and multiplied by the original sign of the coefficient; if the difference is less than or equal to zero, the coefficient is set to 0. The threshold value is determined by multiplying the median of the first-level wavelet coefficients by a coefficient of 0.6475.

[0058] When performing adaptive noise-complete set empirical mode decomposition, independent white noise sequences are first generated. The length of each noise sequence is exactly the same as the input signal, and the amplitude fluctuation range of the noise is set to 0.2 times the amplitude fluctuation range of the input signal. Each noise sequence is added to the original signal, and empirical mode decomposition is performed on each noisy signal to obtain the corresponding set of intrinsic mode function components. Finally, the components of the same order in the independent white noise sequence decomposition results are arithmetically averaged to obtain the final intrinsic mode function components.

[0059] Each intrinsic mode function component must satisfy two conditions: the first condition is that the difference between the number of maxima and the number of minima in the component waveform does not exceed one; the second condition is that the arithmetic mean of the envelope formed by all maxima and the envelope formed by all minima at any time does not exceed 1 / 100 of the component amplitude at that time.

[0060] The Hilbert transform is implemented using a 91-tap digital filter, whose coefficients are obtained by multiplying the impulse response of the ideal Hilbert transformer by a Hanning window function. After performing the Hilbert transform on each eigenmode function component, an analytic signal is constructed. The real part of this signal is the original component, and the imaginary part is the result of the Hilbert transform. The instantaneous phase is obtained by calculating the arctangent of the ratio of the real to the imaginary parts of the analytic signal, and the instantaneous frequency is obtained by differential calculation of the instantaneous phase using the center-difference method.

[0061] When selecting target components, the arithmetic mean of the instantaneous frequency sequence of each component is first calculated. Components whose arithmetic mean falls within the range of 150 kHz to 250 kHz are selected. If multiple components meet this condition, the energy value of each component within that frequency range is calculated. The energy value is calculated by first bandpass filtering the component signal, and then calculating the sum of squares of all sampled values ​​of the filtered signal. The component with the highest energy value is selected as the target component. If no component falls within this frequency range, the component with an average value closest to 200 kHz is selected.

[0062] When calculating the damage kinetic energy release rate, the instantaneous amplitude of the target component is numerically integrated using the composite trapezoidal rule. The entire signal duration is divided into several equal time intervals, each one microsecond in length. The integral value is equal to half the sum of the products of the sum of all two adjacent instantaneous amplitudes and the length of the time interval. Dividing the integral result by the total signal duration yields the average damage kinetic energy release rate, which reflects the elastic wave energy released per unit time due to crack propagation.

[0063] In calculating sample entropy, the target component signal is divided into several time windows, each containing 1024 sampling points, with an overlap of 510 sampling points between adjacent windows. Within each window, the time series is reconstructed into a multi-dimensional vector with two dimensions and a time delay of one. The number of vector pairs that meet the condition is that the maximum difference between corresponding dimensions of the two vectors is less than a threshold, which is 0.2 times the standard deviation of the time series. The sample entropy value is equal to the negative of the natural logarithm of the number of matching vector pairs.

[0064] During normalization, the minimum and maximum values ​​of the sample entropy time series are first found. For each sample entropy value, the minimum value is subtracted, and then divided by the difference between the maximum and minimum values ​​to obtain the normalized value. If the maximum value equals the minimum value, all normalized values ​​are set to 0.5.

[0065] When calculating the final eigenvalue, the average damage kinetic energy release rate within each time window is multiplied by the corresponding normalized sample entropy value. The multiplication operation uses 32-bit floating-point precision, and the result is stored as a signed floating-point array. The sampling rate is consistent with the time window sliding rate, which is one thousand times per second.

[0066] When partitioning the thermoelastic stress field, the entire temperature field image is divided into several sub-regions, each containing 32 by 32 pixels. Adjacent sub-regions overlap by 16 pixels in both the horizontal and vertical directions. Each sub-region is assigned a unique identifier, recording its starting coordinates and region size in the original image.

[0067] When constructing the temperature time series, temperature values ​​are extracted from all pixels in each sub-region, and the arithmetic mean of these temperature values ​​is calculated as the representative temperature of the sub-region at that moment.

[0068] When calculating permutation entropy, a sliding time window is used to process the temperature sequence of each sub-region. Each window contains 250 data points, corresponding to a duration of 10 seconds, and the window slides 25 points at a time, corresponding to a 1-second interval. Within each window, the one-dimensional time series is reconstructed into a multi-dimensional phase space with an embedding dimension of three and a time delay of 5. Multiple multi-dimensional vectors are generated, and the values ​​in each vector are arranged in ascending order to obtain various possible permutation patterns.

[0069] Count the frequency of each permutation pattern and calculate the probability of each pattern. The permutation entropy is equal to the sum of the probabilities of all permutations multiplied by their natural logarithms, then taken as a negative value. The entropy reaches its maximum when all permutations have equal probabilities, and zero when there is only one permutation pattern. In actual calculations, the entropy value is normalized to a range between 0 and 1.

[0070] When constructing the entropy evolution field, the entropy time series of each sub-region is arranged according to its spatial location, forming a three-dimensional array. The first dimension of the array represents the horizontal position, the second dimension represents the vertical position, and the third dimension represents time. This three-dimensional array contains information on the spatial and temporal complexity evolution of the thermoelastic stress field.

[0071] When calculating the state transition margin, the time axis is divided into multiple analysis periods, each containing 250 time points, with 125 points overlapping between adjacent periods. Within each period, for a fixed time slice, the variance of the entropy values ​​for all spatial locations is calculated. The variance is calculated by first averaging the entropy values ​​for all spatial points, then averaging the squared differences between each entropy value and the average, and finally taking the average of these squared values. A small positive number is added to the variance to avoid division by zero errors, and the reciprocal is taken as the state transition margin characteristic value for that period.

[0072] The computation process employs a multi-core parallel processing approach, with two dedicated processor cores handling the acoustic emission signal and two dedicated processor cores handling the thermal data. The acoustic emission processing flow consists of four stages: the first stage performs signal filtering and noise reduction; the second stage performs mode decomposition; the third stage performs Hilbert transform; and the fourth stage calculates eigenvalues. The thermal data processing flow consists of three stages: the first stage partitions the data; the second stage calculates the permutation entropy; and the third stage calculates the state phase transition margin.

[0073] Memory management uses a circular buffer structure. The acoustic emission input buffer is 4 megabytes in size, divided into 16 data blocks; the acoustic emission output buffer is 2 megabytes in size, divided into 8 data blocks; the hot data input buffer is 8 megabytes in size, divided into 8 data blocks; and the hot data output buffer is 4 megabytes in size, divided into 4 data blocks. Each buffer is equipped with read / write pointers and status flags, and a memory barrier mechanism is used to ensure data consistency.

[0074] Error detection employs a cyclic redundancy check (CRC) method, using a specific polynomial for verification calculations. Each data block is appended with a four-byte checksum, which is verified after each stage of processing. Simultaneously, calculation deviations are monitored; if the instantaneous frequency calculation deviation exceeds 10 kHz, or the temperature calculation deviation exceeds 5 degrees Celsius, the current data block is automatically recalculated.

[0075] The output data uses a custom binary format. The file begins with a 16-byte header containing information such as sampling rate, range, and timestamp. Damage kinetic energy release rate features are stored at a rate of 1000 samples per second, with each sample occupying four bytes. State transition margin features are stored at a rate of 50 samples per second, with each sample occupying four bytes. All data is stored in big-endian byte order, and timestamps use UNIX time format.

[0076] In the comprehensive analysis module, data processing begins with unifying the sampling rate. Since the two sequences have different acquisition frequencies, a linear interpolation method is used to increase the sampling rate of the state transition margin eigenvalues ​​to 1,000 times per second. Specifically, 19 new data points are inserted between every two original data points. The value of each new point is calculated linearly based on the values ​​of the preceding and following original data points, proportional to time. For example, inserting a new data point at time t between two original data points at times t1 and t2 results in a value equal to v1 plus the difference between v2 and v1 multiplied by the time difference between t and t1, then divided by the time difference between t2 and t1. After this resampling process, both sequences have an acquisition frequency of 100 data points per second, a duration of ten minutes, and a total of 60,000 data points.

[0077] The standardization process employs a sliding window z-score method. The window duration is set to ten minutes, containing 60,000 data points. When processing each new data point, the arithmetic mean and standard deviation of all data within the window are recalculated. The arithmetic mean is calculated by summing the values ​​of all data points and dividing by the number of data points. The standard deviation is calculated as the square root of the sum of the squares of the differences between each data point and the mean. The standardized value of each data point is equal to the original value minus the mean, then divided by the standard deviation. After this process, the mean of the sequence becomes 0, and the standard deviation becomes 1. To optimize computational efficiency, a recursive calculation method is used to maintain the cumulative sum and sum of squares of the data points in real time, avoiding the need to recalculate all data each time.

[0078] Dynamic time warping first constructs a cumulative distance matrix. The matrix has 60,000 rows and 60,000 columns, and each matrix element stores the local distance between two sequences at their corresponding positions. The local distance is calculated as the square of the difference between the corresponding data points of the two sequences. Due to the large size of the matrix, a sparse storage method is used, storing only elements within a certain range near the diagonal. This range is set to 600 points, meaning that alignment cases with time offsets within 0.6 seconds are only considered.

[0079] The cumulative distance is calculated using dynamic programming. Starting from the top left corner of the matrix, the minimum cumulative distance is calculated row-wise and column-wise for each position. The minimum cumulative distance for each position is equal to the local distance of that position plus the minimum of the minimum cumulative distances of its three adjacent positions to its left, bottom, and bottom left. If an adjacent position is outside the matrix range, its cumulative distance is set to an extremely large value. During the calculation, only the cumulative distance values ​​of the current and previous rows are stored to save memory.

[0080] The optimal path is found using a backtracking method. Starting from the bottom right corner of the matrix, the movement proceeds towards the top left corner. At each position, the movement is made in the direction with the smallest cumulative distance among the three adjacent positions to the left, bottom, and bottom left. The coordinates of all positions along the movement path are recorded. During the path finding process, if a matrix boundary is encountered, the movement continues along the boundary until the top left corner is reached. The final sequence of path points represents the optimal correspondence between two sequences.

[0081] The average cumulative distance is calculated by summing all local distances along the path and dividing by the path length. The path length is equal to the number of points on the path. Since the path may not be a straight line, its length may be greater than the sequence length. The average cumulative distance represents the average degree of difference between two sequences after alignment.

[0082] The first derivative of the cumulative distance is calculated using the central difference method. For the derivative at point i, the distance value at the i-th plus one point is subtracted from the distance value at the i-th minus one point, and then divided by the time interval between the two points, which is 0.002 seconds. The second derivative is also calculated using the central difference method, by subtracting the derivative value at the i-th minus one point from the derivative value at the i-th plus one point. The derivatives at boundary points are calculated using either forward or backward difference.

[0083] The radial basis function neural network uses twenty nodes in its hidden layer. The center vector of each node is learned from the training data using the k-means clustering algorithm. The k-means clustering is iterated 100 times, with the initial centers randomly selected. The scaling factor for each node is set to half the distance from that node to its nearest neighbor. The hidden layer output is calculated as the negative of the exponential function of the square of the Euclidean distance between the input vector and the center vector, multiplied by the scaling factor. The Euclidean distance is calculated as the square root of the sum of the squares of the differences in each dimension of the two vectors.

[0084] The output layer calculates a weighted sum of the hidden layer outputs, adds a bias term, and then maps it using the sigmoid function. The weighted sum is calculated by multiplying each hidden layer output by its corresponding weight coefficient, summing the results, and then adding the bias value. The sigmoid function is calculated as one divided by one raised to the power of the negative weighted sum plus the natural constant. The final output value ranges from 0 to 1.

[0085] The neural network was trained using a dataset containing 10,000 samples. Each sample contained three features: the cumulative distance, its first derivative, and its second derivative, as well as a label indicating whether it was faulty. Training employed the backpropagation algorithm, iterating 1000 times. In each iteration, the cross-entropy loss between the predicted output and the true label for all samples was calculated, and then the weights and biases were updated. The initial learning rate was 0.1, multiplied by 0.5 every 100 iterations. Weight updates used stochastic gradient descent, with 32 samples used per update.

[0086] For real-time computation, multi-scale dynamic time warping is employed for acceleration. First, the sequence is downsampled to 100 samples per second for coarse alignment to obtain an approximate path. Then, fine-tuning is performed within a certain range near the coarsely aligned path at the original resolution. The range size is set to 100 points, representing a time offset of 0.1 seconds.

[0087] Quality monitoring includes three metrics: the signal-to-noise ratio (SNR) of the input data, the smoothness of the regularized path, and the confidence level of the neural network output. The SNR is calculated as the ratio of signal power to noise power, where signal power is the square of the sequence variance, and noise power is calculated from the residual after high-pass filtering. Path smoothness is calculated as the average of the distance differences between adjacent points on the path. The confidence level is calculated as twice the absolute value of the neural network output value's distance from 0.5.

[0088] Data storage uses a circular buffer. Each record contains a critical instability coefficient value (a four-byte floating-point number), a timestamp (an eight-byte integer), and a quality flag (one byte). The buffer write pointer advances one position every millisecond, and the read pointer reads one record per second. When the buffer is full, the oldest data is overwritten by the newest data.

[0089] The output interface generates one data packet per second. The serial interface uses the RS-485 protocol, with a data format of 8 data bits, one stop bit, and no parity bit. Each data packet contains a one-byte start flag, sixteen bytes of data content, and a two-byte checksum. The network interface uses the TCP protocol, with a destination port number of 5000.

[0090] During error handling, when an anomaly occurs during calculation, the error type and timestamp are first recorded. Then, an attempt is made to reinitialize the calculation parameters and recalculate the three most recent data points. If the retry fails after three attempts, the most recent reliable calculation result is used, and the estimation flag is set in the quality flags. Simultaneously, the error information, including the error code, timestamp, and relevant parameter values, is written to the error log.

[0091] Performance monitoring records key metrics every second. Computation latency is measured using a high-precision timer, representing the time difference between input data readiness and output result generation. Memory usage is obtained by querying the operating system interface to determine the current process's memory usage. CPU load is calculated as the percentage of CPU time used by the process in the previous second. When computation latency exceeds 50 milliseconds, memory usage exceeds 70% of total memory, or CPU load exceeds 0.7%, the processing precision is automatically reduced or the computation scope is decreased, and a warning message is sent.

[0092] In the risk prediction module, the critical instability coefficient time series is input at a rate of one sample per second, with each coefficient value being a 32-bit floating-point number ranging from 0 to 1. Preprocessing employs a moving average filter using a 60-second sliding window containing sixty data points. The weight distribution uses a Gaussian function, with a maximum weight of 1.0 at the center point, decreasing symmetrically towards both ends, with a weight of 0.1 at both ends. Upon arrival of each new data point, the weighted average of all data within the window is recalculated; the weight coefficients are pre-calculated and stored in an array of length 60.

[0093] The dynamic fault tree structure contains three intermediate event nodes and nine basic event nodes. The three intermediate events are identified as mechanical faults, electrical faults, and environmental faults, respectively. Each intermediate event is formed by combining the three basic events through a logic OR gate; that is, the occurrence of any basic event will trigger an intermediate event. The critical instability coefficient is assigned to different intermediate events according to its numerical range: when it is less than 0.3, mechanical faults are assigned a weight of 0.7, electrical faults 0.2, and environmental faults 0.1; when it is between 0.3 and 0.7, each of the three intermediate events is assigned a weight of 0.33; when it is greater than 0.7, electrical faults are assigned a weight of 0.7, mechanical faults 0.2, and environmental faults 0.1.

[0094] Weight allocation employs time series analysis. For each intermediate event's corresponding coefficient subsequence, its autocorrelation function is first calculated, followed by the autocorrelation coefficient at lags of 1 to 10 steps. The autocorrelation coefficient equals the covariance of the sequence and its lags divided by the variance of the sequence. Then, the partial autocorrelation function is calculated. The partial autocorrelation function measures the correlation between the sequence and its lags given intermediate lag values, obtained by solving the Yul-Walker equations. Based on the truncation or tailing characteristics of the autocorrelation and partial autocorrelation functions, an autoregressive model, a moving average model, or an autoregressive moving average model is selected. Model parameters are determined using maximum likelihood estimation, with parameter estimates obtained through iterative optimization of the likelihood function. Finally, the weights are determined based on the adjusted goodness of fit of the model, which is equal to one minus the ratio of the sum of squared residuals to the total sum of squares multiplied by an adjustment factor.

[0095] The fuzzy inference system contains 27 rules, with 3 rules corresponding to each basic event. The rule premises define three fuzzy sets for the critical instability coefficient: the low set (0-0.3), the medium set (0.2-0.8), and the high set (0.7-1). The membership function uses trigonometric functions: the vertices of the triangle for the low set are at 0, 0, and 0.3; for the medium set, at 0.2, 0.5, and 0.8; and for the high set, at 0.7, 1, and 1. The rule conclusions define three fuzzy sets for the probability of occurrence: the low probability set (0-0.3), the medium probability set (0.2-0.8), and the high probability set (0.7-1). Fuzzy inference uses the Mamdani method, employing a maximum-minimum composition operation. This involves taking the minimum value of the premise satisfaction as the rule activation strength, and then taking the maximum value of the output fuzzy sets of each rule as the composite result. Defuzzification uses the centroid method, calculating the centroid abscissa value of the area enclosed by the membership function curve of the composite fuzzy set and the abscissa.

[0096] After the basic event probabilities are calculated, the fault tree is then calculated upwards. The probability of each intermediate event is equal to 1 minus the product of the probabilities of all corresponding basic events not occurring. For example, the probability of a mechanical failure is equal to 1 minus the product of the probabilities of the first, second, and third basic events not occurring. The probability of the top event, i.e., the system-level failure probability, is equal to 1 minus the product of the probabilities of all intermediate events not occurring. All probability calculations employ precise arithmetic using 64-bit floating-point precision.

[0097] The system-level failure probability is mapped to a percentage score using a linear transformation: the score equals 1 minus the failure probability, then multiplied by 100. The result is rounded to the nearest integer, ranging from 0 to 100. Safety status levels are categorized as follows: 0-59 points indicate a hazardous state, 60-79 points indicate a warning state, 80-94 points indicate a caution state, and 95-100 points indicate a safe state. Each level corresponds to a specific color code: hazardous state uses RGB color values ​​255, 0, 0; warning state uses RGB color values ​​255, 165, 0; caution state uses RGB color values ​​255, 255, 0; and safe state uses RGB color values ​​0, 128, 0.

[0098] In the special handling mechanism, the hazardous state uses a 300-second analysis window containing 300 fault probability values. The threshold is set to 35%, meaning a downgrade check is triggered when the average value within the window exceeds 0.35. The warning state uses a 180-second analysis window with a threshold set to 25%. During real-time monitoring, the arithmetic mean of the fault probabilities within the most recent window is calculated every second. When the average value for three consecutive sampling periods exceeds the threshold, a downgrade operation is performed. The downgrade is by one level, such as downgrading from the warning state to the hazardous state.

[0099] After a downgrade, a 30-minute stabilization observation period begins. During this period, data is checked every second, and data values ​​are recorded without performing the downgrade operation. After the observation period, the downgrade mechanism is reactivated. All downgrade events are recorded, including: the timestamp of the downgrade (accurate to milliseconds), the level before the downgrade, the level after the downgrade, the average value of the window that triggered the downgrade, and the failure probability values ​​of the ten most recent sampling points at the time of the downgrade. Recorded data is saved to a dedicated downgrade event log.

[0100] Historical data is stored using a relational database structure. Each record contains eight fields: timestamp (datetime type), critical instability coefficient (float type), failure probability (float type), security level (integer type), level flag (integer type), data quality flag (integer type), update timestamp (datetime type), and version number (integer type). Composite indexes are created for the data tables, including timestamp and level indexes. The data cleansing strategy involves archiving data from three years ago every six months.

[0101] The verification mechanism includes three levels: a model self-verification is performed every ten minutes to check the rationality of the input coefficients and output probabilities, requiring the correlation coefficient between the failure probability and the critical instability coefficient to be greater than 0.7; an integrity check is performed daily to verify that the numerical range of all 900 computation nodes (three intermediate event nodes multiplied by 300 time points) is between 0 and 1; and parameter optimization is performed weekly, using the data from the most recent week to recalculate the parameters of the fuzzy inference rules, and using the gradient descent method to optimize the fuzzy set center value of the rule conclusion part.

[0102] Data output employs multiple formats: a real-time data stream outputs a JSON object per second, containing the current timestamp, failure probability, and security level; a statistical JSON object is output per minute, containing the minimum, maximum, and average values ​​of sixty sampling points; and a CSV file is output per hour. All output data is digitally signed using the RSA algorithm with SHA-256 hashing.

[0103] Monthly model updates utilize the most recent 30 days of running data. The training process employs 10-fold cross-validation, randomly dividing the data into ten parts, using nine parts for training and one part for validation in rotation. The optimization objective is to minimize the mean squared error between the predicted fault probability and the actual fault records. Parameter tuning uses mini-batch gradient descent, with a batch size of 120 records, a learning rate of 0.01, and 50 iterations.

[0104] Performance monitoring records key metrics every second: computation latency is measured using a high-precision timer, representing the time difference between input data readiness and output completion; memory usage is obtained by querying the process memory counter; and CPU utilization is calculated as the ratio of CPU time used by the process in the previous second to the total CPU time. When computation latency exceeds 100 milliseconds, data logging is automatically disabled; when memory usage exceeds 512 megabytes, cached data is automatically cleared; and when CPU utilization exceeds 70% for 30 consecutive seconds, the data processing frequency is automatically reduced to 0.5 samples per second. Simultaneously, alert messages are sent to the monitoring center via the SNMP protocol.

[0105] In the autonomous policy generation module, the system continuously receives security status level input, which is updated once per second. The level value ranges from 0 to 3, where 0 represents a safe state, 1 represents a state of alert, 2 represents a state of warning, and 3 represents a state of danger. The policy generation unit internally maintains a state machine, triggering the corresponding policy execution process based on the current level value. Each policy is predefined as an executable sequence of instructions, including specific operations such as movement control, data acquisition, and communication transmission.

[0106] Under safe conditions, a pre-set routine inspection plan is executed. The inspection path is defined by fifty latitude and longitude coordinate points, with an exact distance of 20 meters between adjacent points. Movement control uses a PID controller, with a speed set at 0.5 meters per second and a position error tolerance of ±0.1. Each inspection point is occupied for 10 seconds, with the first 0.5 seconds used for fine-tuning and the remaining 9.05 seconds for data acquisition. Visible light acquisition uses a two-megapixel CMOS sensor with a 50mm focal length and an aperture of f / 8, acquiring one frame of JPEG image per second. Infrared acquisition uses a 160x120 pixel microbolometer with a spectral response range of 8 to 14 micrometers, a temperature measurement accuracy of ±2 degrees Celsius, and acquires five frames of RAW data per second. After acquisition, the data is temporarily stored in local flash memory and packaged into a ZIP format every ten minutes, then transmitted to the backend server via TCP.

[0107] When the status changes to "Attention," a review task of ten historical defect points is added to the regular inspection. The coordinates of these defect points are stored in an SQLite database, and each point includes latitude, longitude, elevation, and defect type information. The movement speed is adjusted to 0.3 meters per second, and the position control accuracy is improved to ±0.05 meters. The system pauses at each historical defect point for 30 seconds to collect data from multiple angles: visible light acquisition is switched to a 5-megapixel mode, acquiring one frame each from 0 degrees, 45 degrees, and 90 degrees; infrared acquisition is switched to a 320x240 pixel mode, acquiring three frames from each angle. All acquired data is immediately transmitted to the backend platform via HTTPS, with a transmission bandwidth guaranteed to be no less than 10 megabits per second.

[0108] When in alert mode, a high-frequency inspection mode is activated. Path planning adds 20 monitoring points beyond the basic 50 points, determined based on historical fault probability distribution. The moving speed is increased to 0.8 meters per second, employing an adaptive cruise control algorithm. The dwell time at each detection point is reduced to 5.0 seconds, with 1 second for positioning and 4 seconds for data acquisition. Visible light acquisition uses an 8-megapixel mode, acquiring three frames per second; infrared acquisition uses a 640x480 pixel mode, acquiring ten frames per second; a three-axis MEMS vibration sensor is simultaneously activated, with a range of ±16g and a sampling rate of 100 Hz. All sensor data is streamed in real-time, while a 10GB circular buffer is allocated in local flash memory for data caching.

[0109] Upon detecting a dangerous situation, the current task is immediately terminated. First, an emergency stop command is sent to the motion controller, with a braking acceleration of -3 m / s². Then, the shortest path from the current location to the target point is calculated based on the A* algorithm, and the navigation speed is set to 1.5 m / s. Upon reaching the target point, continuous monitoring mode is entered; the infrared thermal imager switches to video mode. Acoustic emission sensors, vibration sensors, and temperature sensors are simultaneously activated. Data is transmitted through a dedicated QoS channel with a transmission priority of 7.

[0110] Upon activation of the emergency response, an emergency alarm data packet is first generated. This packet contains the current GPS coordinates, hazard level, on-site temperature, humidity, and wind speed data. It is sent to the backend platform via both UDP and TCP protocols. Then, it automatically dials three pre-set responsible personnel's phone numbers, playing a voice alarm; simultaneously, it sends SMS alarm messages to five designated mobile phone numbers. A 30-second timer is established to periodically send status update messages, including operating parameters such as battery level, signal strength, and equipment temperature. If no response is received after three consecutive attempts, it sequentially switches to the backup APN, microwave link, and satellite communication module.

[0111] All instruction execution processes are recorded in real time. Each instruction is assigned a unique sequence number, and the issuance time, execution start time, execution end time, and return code are recorded. Instructions that fail are automatically retried at intervals of 1 second, 3 seconds, and 5 seconds. An execution log file is generated every 5 minutes, containing instruction statistics, exception codes, resource usage information, etc. The log file is stored in binary format and includes a CRC32 checksum.

[0112] State transitions are gradual. When a new state level is received, the current detection point's data acquisition task is completed first, and then the various subtasks of the current strategy are gradually stopped. Motion control uses a smooth speed curve transition, with acceleration limited to ±1 m / s². Data acquisition tasks wait for the current frame to be acquired before switching modes. All switching events are recorded with millisecond-level timestamps and the reason for the switch is noted.

[0113] The communication link management and maintenance consists of three independent connections: a primary 4G / 5G link using the mobile network, a backup microwave link using the 60GHz band, and an emergency satellite link using the L-band. Each data packet is assigned a unique sequence number and is transmitted simultaneously through at least two links. The receiving end automatically deduplicates packets received. The transmission protocol implements a selective retransmission mechanism, triggering fast retransmission when the packet loss rate exceeds 5%.

[0114] The energy management unit monitors the lithium-ion battery pack status in real time. In a safe state, the lidar and backup computing unit are shut down; in a cautious state, all computing units are enabled but the CPU frequency is limited; in a warning state, all computing units run at full load; in a dangerous state, non-critical functions are disabled, prioritizing sensing and communication. When the battery voltage drops below 3.6 volts or the capacity drops below 20%, a path back to the charging point is automatically planned.

[0115] The working principle of this invention is as follows: A self-generated acoustic emission waveform and thermoelastic stress field distribution data of the cable are simultaneously collected via an autonomous mobile platform; a nonlinear modal decomposition of the acoustic emission data is performed using an edge computing center to extract the damage kinetic energy release rate characteristic value, and entropy evolution analysis is performed on the thermoelastic stress field to extract the state phase transition margin characteristic value; a comprehensive analysis module is used to construct the spatiotemporal correlation of the two characteristic values ​​and calculate the critical instability coefficient; based on this coefficient, a fault prediction model outputs the safety state level; finally, based on different state levels, corresponding inspection decisions are autonomously generated and executed, realizing intelligent cable inspection throughout the entire process from data collection, feature extraction, risk assessment to autonomous decision-making.

[0116] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A power cable inspection robot system, characterized in that, include: An autonomous mobile platform module, wherein the autonomous mobile platform module synchronously acquires the spontaneous acoustic emission waveform flow and the thermoelastic stress field distribution of the cable through a collaborative intelligent sensing system; Edge computing hub module, the edge computing hub module is configured as follows: Nonlinear mode decomposition is performed on the spontaneous acoustic emission waveform to extract the damage kinetic energy release rate characteristic value, which characterizes the crack propagation rate inside the material. The calculation process of the damage kinetic energy release rate characteristic value is as follows: Calculate the sample entropy of the target intrinsic mode function components within a set time window; The numerical value of the sample entropy is normalized and used as a stability weight coefficient. The damage kinetic energy release rate is multiplied by the stability weighting coefficient, and the product is used as the final characteristic value of the damage kinetic energy release rate. Entropy evolution analysis was performed on the thermoelastic stress field distribution of the bulk material to extract state phase transition margin characteristic values ​​that characterize the thermo-mechanical coupling stability under load fluctuations. Specifically, these include: The thermoelastic stress field distribution data is divided into several sub-regions in the spatial domain, and a time series of thermoelastic stress values ​​is constructed for each sub-region. For the time series of each sub-region, the permutation entropy is calculated using a sliding time window, thereby obtaining the sequence of permutation entropy changes with time for each sub-region. The permutation entropy sequences of all sub-regions are arranged according to spatial location to form a multidimensional entropy evolution field. The variance of the entropy evolution field is calculated over a set time period, and the reciprocal of the variance is used as the state phase transition margin characteristic value representing the stability of the thermodynamic coupling process within the corresponding time period; The comprehensive analysis module is used to construct the spatiotemporal correlation between the characteristic value of damage kinetic energy release rate and the characteristic value of phase transition margin, and to calculate the critical instability coefficient that predicts the risk of structural chain failure. Specifically, it includes: The time series of damage kinetic energy release rate feature value and the time series of state phase transition margin feature value are subjected to dynamic time warping, and the cumulative distance between the two time series under the optimal warping path is calculated. The cumulative distance is input into a pre-trained radial basis function neural network. The input layer of the corresponding neural network contains three nodes corresponding to the cumulative distance and its first and second derivatives, respectively. The output layer generates a critical instability coefficient between 0 and 1 through the sigmoid activation function. The risk prediction module inputs the critical instability coefficient into the cable fault prediction model and outputs the current safety status level of the cable. An autonomous policy generation module generates and executes inspection decisions based on the security status level.

2. The power cable inspection robot system according to claim 1, characterized in that, The nonlinear mode decomposition of the spontaneous acoustic emission waveform stream specifically includes: Empirical mode decomposition (EMD) is used to process the spontaneous acoustic emission waveform stream to obtain a set of intrinsic mode function (IMF) components. Hilbert transform is performed on each IMF component to calculate its instantaneous frequency and instantaneous amplitude. Target IMF components whose instantaneous frequencies match the typical frequency range of material crack propagation are selected, and the damage kinetic energy release rate is obtained by integrating the instantaneous amplitude of the target IMF component and dividing it by the total signal duration.

3. The power cable inspection robot system according to claim 1, characterized in that, The construction process of the cable fault prediction model is as follows: The cable fault prediction model includes multiple intermediate event nodes triggered by the critical instability coefficient; the weight allocation of each intermediate event node is determined by time series analysis; the critical instability coefficient is mapped to the basic event layer of the fault tree through fuzzy inference, and a nonlinear correspondence between the occurrence probability of basic events and the critical instability coefficient is established; finally, the system-level fault probability is obtained through uplink calculation, and the safety status level is divided according to the probability value range.

4. The power cable inspection robot system according to claim 1, characterized in that, The classification of security status levels specifically includes: The system-level failure probability values ​​are mapped to a percentage range of 0-100. Based on the range of the mapped values, four safety status levels are defined: 0-60 is a dangerous state, 60-80 is a warning state, 80-95 is a caution state, and 95-100 is a safe state. Different time series analysis window lengths and failure probability thresholds are set for dangerous and warning states. When the failure probability threshold of recent monitoring data exceeds the threshold of the corresponding state, the safety status level is automatically downgraded by one level.

5. The power cable inspection robot system according to claim 1, characterized in that, The generation and execution of inspection decisions specifically include: When in a safe state, routine inspections are conducted along the preset basic path, collecting visible light images and infrared data. When in a state of alert, key re-inspections of historical defect points are added to the basic path to improve image acquisition resolution. When in a state of warning, high-frequency special inspections are initiated, the cycle between adjacent inspections is shortened, and extended inspections are conducted on related sections. When in a dangerous state, routine tasks are immediately terminated, and the system quickly navigates to the target location for continuous monitoring, while simultaneously sending real-time video streams to the backend platform and initiating emergency response procedures.

Images