Tactile time sequence compensation system and method

Abstract

The invention discloses a touch time sequence compensation system and method, and relates to the technical field of power systems. The compensation system comprises a tactile signal processing module, a physical field modeling module, a wavelet phase alignment module, a space-time diagram network module and a reinforcement learning tuning module. The tactile signal processing module acquires and converts tactile signals into digital signals, so that the physical field modeling module constructs a wave equation and a delay matrix; a delay matrix output by the physical field modeling module is used as priori knowledge to assist the wavelet phase alignment module to perform phase alignment on the multi-band signals subjected to wavelet decomposition so as to generate time sequence alignment signals; the time sequence alignment signal is input into a space-time diagram network module for edge weight calculation; and the reinforcement learning tuning module dynamically compensates the tactile signal according to the edge weight and outputs tactile time sequence compensation data so as to feed back and correct the physical field modeling module and the wavelet phase alignment module, so that closed-loop optimization is formed, the mechanical fault recognition accuracy is high, and the accuracy and robustness of tactile time sequence compensation are improved.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a tactile timing compensation system and method. Background Technology

[0002] In the field of intelligent power equipment detection and high-precision signal processing in power systems, substation inspection robots are crucial for the rapid and accurate detection of mechanical faults. Mechanical faults such as circuit breaker jamming and disconnector wear, if not detected and addressed in a timely manner, can lead to serious power accidents and affect the safe and stable operation of the power grid. Tactile timing detection, as a key technology for identifying these mechanical faults, directly determines the accuracy of fault diagnosis through its detection precision and reliability.

[0003] Currently, traditional techniques have many shortcomings in sub-millisecond time-series fault detection in substations. For example, the traditional Kalman filter method, relying on linear assumptions, struggles to meet practical requirements when dealing with the nonlinear delays generated during elastic wave propagation, resulting in errors exceeding 2ms and significant deviations in time-series detection results, failing to provide accurate data support for fault diagnosis. While LSTM (Long Short-Term Memory) compensation networks offer certain advantages in time-series data processing, their lack of physical field modeling capabilities leads to significant phase drift in the complex and intense electromagnetic interference environment of substations, resulting in a false alarm rate exceeding 18%, greatly impacting the accuracy of fault identification. Although Dynamic Time Warping (DTW) can address the alignment problem of time-series data to some extent, its computational complexity is as high as O(n log n). The processing latency exceeds 50ms, which cannot meet the requirements of substation inspection robots for real-time detection of mechanical faults. Furthermore, the complex electromagnetic environment of substations can cause distortion of tactile data, further reducing the accuracy of mechanical fault identification. Summary of the Invention

[0004] This invention provides a tactile timing compensation system and method, which solves the technical problems of existing tactile timing detection methods being easily affected by electromagnetic interference and having long processing delays, resulting in low accuracy in mechanical fault identification.

[0005] The first aspect of the present invention provides a tactile timing compensation system, comprising: a tactile signal processing module, a physical field modeling module, a wavelet phase alignment module, a spatiotemporal graph network module, and a reinforcement learning optimization module;

[0006] The tactile signal processing module is used to acquire the tactile signals of the substation inspection robot and convert the tactile signals into digital signals.

[0007] The physical field modeling module is used to perform physical field modeling using the digital signal, and to construct wave equations and delay matrices;

[0008] The wavelet phase alignment module is used to use the delay matrix as prior knowledge to perform phase alignment on the multi-frequency signal after wavelet decomposition of the wave equation and generate a time-aligned signal.

[0009] The spatiotemporal graph network module is used to calculate the edge weights of the time-series alignment signal using a spatiotemporal graph network, and generate multiple edge weights.

[0010] The reinforcement learning optimization module is used to dynamically compensate the tactile signal according to the edge weights and generate tactile temporal compensation data.

[0011] Optionally, the tactile signal processing module includes a tactile sensor, a charge amplifier, a bandpass filter, and an analog-to-digital converter;

[0012] The tactile sensor is used to acquire tactile signals from the substation inspection robot;

[0013] The charge amplifier is used to convert the tactile signal into a voltage signal to generate an initial voltage signal;

[0014] The bandpass filter is used to filter out noise signals in the initial voltage signal and generate the target voltage signal;

[0015] The analog-to-digital converter is used to convert the target voltage signal into a digital signal.

[0016] Optionally, the physics modeling module performs the following steps:

[0017] The medium damping coefficient and electromagnetic mechanical coupling coefficient corresponding to the tactile signal were calibrated by finite element simulation.

[0018] A wave equation is constructed using the material-related wave velocity corresponding to the digital signal, the external excitation function, the medium damping coefficient, and the electromagnetic-mechanical coupling coefficient.

[0019] A delay matrix is ​​generated by constructing a matrix using the three-dimensional coordinates of the sensor node corresponding to the digital signal, a preset material correction coefficient, and the material-related wave velocity.

[0020] Optionally, the wavelet phase alignment module performs the following steps:

[0021] The tactile signal is subjected to complex Morlet wavelet transform to generate a time-frequency energy spectrum;

[0022] The instantaneous phase spectrum is obtained by calculating the instantaneous phase spectrum using the time-frequency energy spectrum.

[0023] With phase consistency as the objective, a cross-band phase alignment loss function is constructed using the wavelet coefficients and phase angles corresponding to the instantaneous phase spectrum;

[0024] The delay error corresponding to the delay matrix is ​​converted into a multi-band phase difference in the wavelet domain using the cross-band phase alignment loss function to generate a timing alignment signal.

[0025] Optionally, the spatiotemporal graph network module performs the following steps:

[0026] The contact pressure, temperature, and angular velocity corresponding to the digital signals, as well as the vibration spectrum vector corresponding to the time-aligned signal, are respectively used to calculate the feature vectors of the sensor nodes to generate multiple feature vectors.

[0027] The edge weights corresponding to the sensor nodes are generated by calculating the edge weights using the feature vector and the delay matrix respectively through the spatiotemporal graph attention mechanism.

[0028] Optionally, the reinforcement learning tuning module performs the following steps:

[0029] The state space is constructed using the error change rate corresponding to the time-series alignment signal, the edge weights, the wave velocity gradient corresponding to the wave equation, and the phase angle vector, to generate a state space vector.

[0030] The state space vector is used to dynamically compensate the policy network parameters corresponding to the tactile signal by employing the dual-delay deep deterministic policy gradient algorithm, thereby generating tactile temporal compensation data.

[0031] A second aspect of the present invention provides a tactile timing compensation method, comprising:

[0032] Acquire tactile signals from the substation inspection robot and convert the tactile signals into digital signals;

[0033] The digital signal is used to model the physical field, and wave equations and delay matrices are constructed.

[0034] Using the delay matrix as prior knowledge, phase alignment is performed on the multi-band signal after wavelet decomposition of the wave equation to generate a time-aligned signal.

[0035] A spatiotemporal graph network is used to calculate the edge weights of the time-series aligned signal, generating multiple edge weights.

[0036] The tactile signal is dynamically compensated based on the edge weights to generate tactile timing compensation data.

[0037] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the tactile timing compensation method as described in any of the preceding claims.

[0038] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the tactile timing compensation method as described in any of the preceding claims.

[0039] The fifth aspect of the present invention provides a computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer performs the tactile timing compensation method as described in any of the preceding claims.

[0040] As can be seen from the above technical solutions, the present invention has the following advantages:

[0041] This invention acquires tactile signals from a substation inspection robot through a tactile signal processing module and converts these signals from physical tactile signals to digital signals, providing high-quality input data for subsequent processing. A physics modeling module constructs a physical model of the tactile signal propagation in metal components, obtaining wave equations and delay matrices to provide theoretical constraints on signal delay and address the lack of physical mechanisms in traditional methods. A wavelet phase alignment module performs time-frequency decomposition and phase alignment of multi-sensor signals in the wavelet domain, generating time-aligned signals to solve nonlinear delay and phase drift problems, achieving sub-millisecond-level time-series compensation. A spatiotemporal graph network module fuses the spatiotemporal correlation and signal characteristics of sensor nodes, using a spatiotemporal graph network to calculate edge weights and generate edge weights corresponding to each sensor. By capturing abnormal propagation paths in the complex structure of the substation, it provides high-level feature input for reinforcement learning. A reinforcement learning optimization module dynamically adjusts the physical field and algorithm parameters to obtain tactile time-series compensation data, adapting to dynamic environments such as substation temperature changes and electromagnetic interference, forming a "modeling-compensation-optimization" closed loop to improve system robustness. This invention provides prior constraints through physical field modeling, solving the problem of uninterpretability in pure deep learning methods; it achieves decoupling of time-frequency features through wavelet decomposition, improving phase alignment accuracy; and it captures complex spatial correlations through spatiotemporal graph networks, reinforcing learning to adapt to dynamic environments, ultimately achieving a high accuracy rate in mechanical fault identification. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of a tactile timing compensation system provided in Embodiment 1 of the present invention;

[0044] Figure 2 This is a schematic diagram of the multimodal adaptive wavelet-graph enhancement compensation framework provided in Embodiment 1 of the present invention;

[0045] Figure 3 This is a flowchart of the steps of a tactile timing compensation method provided in Embodiment 2 of the present invention;

[0046] Figure 4 This is a structural block diagram of an electronic device provided in Embodiment 3 of the present invention. Detailed Implementation

[0047] This invention provides a tactile timing compensation system and method to solve the technical problem that existing tactile timing detection methods are easily affected by electromagnetic interference and have long processing delays, resulting in low accuracy in mechanical fault identification.

[0048] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0049] Please see Figure 1 , Figure 1 This is a schematic diagram of a tactile timing compensation system provided in Embodiment 1 of the present invention.

[0050] The present invention provides a tactile timing compensation system, comprising: a tactile signal processing module, a physical field modeling module, a wavelet phase alignment module, a spatiotemporal graph network module, and a reinforcement learning optimization module;

[0051] The tactile signal processing module is used to acquire tactile signals from the substation inspection robot and convert the tactile signals into digital signals.

[0052] The physics modeling module is used to model physical fields using digital signals and construct wave equations and delay matrices.

[0053] The wavelet phase alignment module is used to perform phase alignment on the multi-band signals after wavelet decomposition of the wave equation using the delay matrix as prior knowledge, and generate time-aligned signals.

[0054] The spatiotemporal graph network module is used to calculate edge weights for time-aligned signals using a spatiotemporal graph network, generating multiple edge weights.

[0055] The reinforcement learning optimization module is used to dynamically compensate tactile signals based on edge weights and generate tactile temporal compensation data.

[0056] In this embodiment of the invention, a tactile signal processing module acquires the tactile signals of a substation inspection robot and converts them from physical tactile signals to digital signals, providing high-quality input data for subsequent processing. A physical field modeling module constructs a physical model of the propagation of tactile signals in metal components, obtaining wave equations and delay matrices, providing theoretical constraints on signal delay and addressing the lack of physical mechanisms in traditional methods. A wavelet phase alignment module performs time-frequency decomposition and phase alignment of multi-sensor signals in the wavelet domain, generating a time-aligned signal to solve nonlinear delay and phase drift problems, achieving sub-millisecond-level time-series compensation. A spatiotemporal graph network module fuses the spatiotemporal correlation and signal features of sensor nodes, and uses a spatiotemporal graph network to calculate edge weights on the time-aligned signal, generating multiple edge weights. By capturing abnormal propagation paths in the complex structure of the substation, high-level feature input is provided for reinforcement learning. By dynamically adjusting the physical field and algorithm parameters through a reinforcement learning optimization module, tactile temporal compensation data is obtained and sent to the physical field modeling module and the spatiotemporal graph network module. This adapts to dynamic environments such as temperature variations and electromagnetic interference in substations, forming a "modeling-compensation-optimization" closed loop to improve system robustness. This invention provides prior constraints through physical field modeling, addressing the lack of interpretability in pure deep learning methods; it achieves time-frequency feature decoupling through wavelet processing, improving phase alignment accuracy; and it captures complex spatial correlations through a spatiotemporal graph network, enabling reinforcement learning to adapt to dynamic environments. Ultimately, this achieves a high accuracy rate in mechanical fault identification. It solves the technical problems of existing tactile temporal detection methods being susceptible to electromagnetic interference and having long processing delays, resulting in low accuracy in mechanical fault identification.

[0057] Furthermore, the tactile signal processing module includes a tactile sensor, a charge amplifier, a bandpass filter, and an analog-to-digital converter;

[0058] Tactile sensors are used to acquire tactile signals from the substation inspection robot;

[0059] A charge amplifier is used to convert tactile signals into voltage signals and generate an initial voltage signal.

[0060] A bandpass filter is used to filter out noise signals in the initial voltage signal and generate the target voltage signal;

[0061] An analog-to-digital converter is used to convert a target voltage signal into a digital signal.

[0062] In embodiments of the present invention, such as Figure 2 As shown, the tactile timing compensation system provided by this invention is a substation inspection robot tactile timing compensation system based on the MAWGRC (Multi-modal Adaptive Wavelet-Graph Reinforcement Compensation) framework. It includes a tactile sensor, a charge amplifier, a bandpass filter, an analog-to-digital converter (ADC), an FPGA preprocessing unit, and an FPGA wavelet phase alignment module. The FPGA preprocessing unit is the aforementioned physical field modeling module. The FPGA wavelet phase alignment module includes the aforementioned wavelet phase alignment module, a spatiotemporal graph network module, and a reinforcement learning tuning module. A multi-modal adaptive wavelet-graph reinforcement compensation framework is constructed using the FPGA preprocessing unit and the FPGA wavelet phase alignment module. Figure 3 As shown, the framework achieves high-precision compensation of tactile data through a four-level progressive structure of physical mechanism modeling → signal feature decoupling → spatiotemporal relationship modeling → dynamic closed-loop optimization.

[0063] like Figure 2 As shown, tactile signals are acquired by the tactile sensor of the substation inspection robot, then converted into voltage signals by a charge amplifier to generate an initial voltage signal. Next, noise signals are filtered out by a bandpass filter to generate the target voltage signal. Then, it is converted into a digital signal by an ADC (analog-to-digital converter), and the signal undergoes preliminary physical field modeling by an FPGA preprocessing unit. Subsequent signal processing is then performed by the FPGA wavelet phase alignment module, and finally, the results are sent to the backend via Ethernet. Specifically, the tactile sensor and charge amplifier bandwidth cover the entire tactile signal frequency band, and the FPGA implements 5-layer wavelet decomposition (delay <1ms).

[0064] like Figure 1As shown, the tactile signal processing module acquires the tactile signals of the substation inspection robot, converts the tactile signals into digital signals, and sends them to the physics modeling module. The physics modeling module then performs physics modeling, providing wave equations and delay matrices to the wavelet phase alignment module. The wavelet phase alignment module outputs a timing alignment signal to the spatiotemporal network module. The spatiotemporal graph network module extracts spatiotemporal correlation features and provides them to the reinforcement learning optimization module. The reinforcement learning optimization module outputs dynamic feedback parameter corrections to the physics modeling module for adjustment. Finally, the reinforcement learning optimization module continuously optimizes weight updates and provides them to the spatiotemporal graph network module for adjustment. Therefore, the tactile timing compensation system provided by this invention is a tactile timing compensation system that integrates physics modeling and deep learning, suitable for substation inspection robots to perform sub-millisecond timing detection of mechanical faults (such as circuit breaker jamming and disconnector wear).

[0065] Furthermore, the physics modeling module performs the following steps:

[0066] The medium damping coefficient and electromagnetic-mechanical coupling coefficient corresponding to the tactile signal were calibrated using finite element simulation.

[0067] A wave equation is constructed using the material-dependent wave velocity, external excitation function, medium damping coefficient, and electromagnetic-mechanical coupling coefficient corresponding to the digital signal.

[0068] A delay matrix is ​​generated by constructing a matrix using the three-dimensional coordinates of the sensor nodes corresponding to the digital signal, a preset material correction coefficient, and the material-related wave velocity.

[0069] In this embodiment of the invention, firstly, the medium damping coefficient and electromagnetic-mechanical coupling coefficient corresponding to the tactile signal are calibrated through finite element simulation. Then, using the material-dependent wave velocity, external excitation function, medium damping coefficient, and electromagnetic-mechanical coupling coefficient corresponding to the digital signal, a wave equation is constructed. The wave equation satisfied by the propagation of the tactile signal in the metal component is:

[0070] ;

[0071] in, denoted as displacement field (mm); c represents material-dependent wave velocity (5000 m / s for steel components, 3800 m / s for copper components). The medium damping coefficient (calibrated through finite element simulation); For external excitation functions (including electromagnetic interference terms); This refers to the electromagnetic-mechanical coupling coefficient (calibrated through finite element simulation). Specifically, it is calibrated through finite element simulation. and The delay inversion error is <0.5ms (60% improvement compared to traditional methods).

[0072] Secondly, a delay matrix is ​​constructed using the three-dimensional coordinates of the sensor nodes corresponding to the digital signal, a preset material correction coefficient, and the material-related wave velocity c.

[0073] ;

[0074] in, The theoretical time relationship of signal transmission between multiple sensors is provided for the theoretical propagation time of the signal from the sensor (calculated based on the elastic wave equation); Let be the three-dimensional coordinates of the sensor node of the i-th sensor; Let be the three-dimensional coordinates of the sensor node of the j-th sensor; c is the material-dependent wave velocity; The material correction factor is 1.0 for steel and 0.93 for copper.

[0075] Finally, a delay matrix is ​​constructed using the three-dimensional coordinates of the sensor nodes corresponding to the digital signal, a preset material correction coefficient, and the material-related wave velocity. The elastic wave equation, also known as the wave equation, describes the propagation law of tactile vibration waves in a metallic medium. The delay matrix calculation is based on the displacement field of the solution to the wave equation, quantifying the signal propagation delay between sensor nodes. The displacement field u is obtained by solving the wave equation using the finite element method, and then the time delay is calculated by combining it with the three-dimensional coordinates of the sensor. This provides physical constraints for subsequent modules.

[0076] Furthermore, the wavelet phase alignment module performs the following steps;

[0077] The tactile signal is subjected to complex Morlet wavelet transform to generate a time-frequency energy spectrum;

[0078] The instantaneous phase spectrum is obtained by calculating the instantaneous phase spectrum using the time-frequency energy spectrum.

[0079] To achieve phase consistency, a cross-band phase alignment loss function is constructed using the wavelet coefficients phase angle corresponding to the instantaneous phase spectrum.

[0080] A cross-band phase alignment loss function is used to perform wavelet domain multi-band phase difference transformation on the delay error corresponding to the delay matrix, generating a timing aligned signal.

[0081] In this embodiment of the invention, a complex Morlet wavelet transform is performed on the tactile signal to generate a time-frequency energy spectrum. Then, the instantaneous phase spectrum is calculated using the time-frequency energy spectrum to obtain the instantaneous phase spectrum. The corresponding transform formula is:

[0082] ;

[0083] in, This is a complex Morlet wavelet function used for time-frequency analysis of tactile signals; For bandwidth parameters (taken as 0.5-5kHz); The center frequency (selected according to the equipment vibration characteristics).

[0084] The following two types of core features were obtained:

[0085] Time-frequency energy spectrum: ,in, It is the time-frequency energy spectrum; The original tactile signal is related to time. The function records the tactile information collected by the sensor that changes over time; For complex Morlet wavelet functions; As the integration variable, it traverses the time range of the original signal during the integration process, and integrates the wavelet function with the original signal at different times to achieve energy analysis of different frequency components of the signal at different times. For the original tactile signals Complex Morlet wavelet function Convolution operations are performed to obtain the signal components at different frequencies and times. The time-frequency energy spectrum is obtained by squaring the modulus. The time-frequency energy spectrum characterizes the energy distribution of the signal at different time points t and different frequencies f.

[0086] Instantaneous phase spectrum: ,in, The instantaneous phase spectrum; To take complex numbers Phase angle manipulation. The instantaneous phase spectrum describes the phase evolution trajectory of each frequency band component of the signal and is used to capture timing deviations.

[0087] To achieve phase consistency, a cross-band phase alignment loss function is constructed using the wavelet coefficient phase angle corresponding to the instantaneous phase spectrum. The cross-band phase alignment loss function is as follows:

[0088] ;

[0089] in, The objective function value for phase consistency optimization is used to measure the degree of phase consistency between multi-sensor signals in different frequency bands; denoted by , where is the wavelet coefficient phase angle, reflecting the phase information of the signal at different frequencies and time points; K is the number of sensor nodes, representing the number of sensors participating in the detection; M is the number of time slices, obtained by discretizing time, used to analyze the signal phase at different time segments; and m is the time slice index, with a value range of . k is the sensor node index, with a value range of 1. ); For the t-th sensor in the time window The high-frequency phase angle (e.g., 5kHz) reflects the local temporal characteristics of high-frequency transient events in the robot (e.g., arc discharge, metal microcracks); For the k-th sensor in the time window The low-frequency phase angle (e.g., 200Hz) characterizes the macroscopic motion of a mechanical system (e.g., the vibration of a robot). Discretized time slices (e.g.) =0-10ms, =5-15ms), signal segmentation is achieved through a sliding window to balance time resolution and computational efficiency.

[0090] Data transfer from the physical field to the wavelet domain involves using a cross-band phase alignment loss function to transform the delay error corresponding to the delay matrix into a multi-band phase difference in the wavelet domain, generating a time-aligned signal. The delay matrix serves as prior knowledge, used to initialize the cross-band phase alignment loss function of the wavelet decomposition. The mechanism involves transforming the delay error from the physical field modeling into a multi-band phase difference in the wavelet domain, achieving dual-drive compensation of physical and data through optimized phase consistency. The phase alignment initialization method is as follows:

[0091] Time-phase mapping: converting time delay into phase offset.

[0092] ;

[0093] in, is the phase offset generated by the signal propagation from sensor i to j, used to describe the phase change of the signal as it propagates between different sensors. It is a key parameter for phase alignment initialization, and its unit is radians (rad); f is the center frequency of the target frequency band, which is a specific frequency value set according to the specific detection task, and its unit is Hertz (Hz). For the signal from the sensor arrive The theoretical propagation time is calculated based on the elastic wave equation and is expressed in seconds (s).

[0094] The initial expected value of the phase difference is: ,in The phase offset generated by the signal propagation from sensor i to j describes the phase change of the signal as it propagates between different sensors. It is a key parameter for phase alignment initialization and is measured in radians (rad). For the t-th sensor in the time window The high-frequency phase angle (e.g., 5kHz) reflects the local temporal characteristics of high-frequency transient events in the robot (e.g., arc discharge, metal microcracks); For the k-th sensor in the time window The low-frequency phase angle (e.g., 200Hz) characterizes the macroscopic motion of a mechanical system (e.g., the vibration of a robot).

[0095] Furthermore, the spatiotemporal graph network module performs the following steps:

[0096] The contact pressure, temperature, and angular velocity corresponding to the digital signals, as well as the vibration spectrum vectors corresponding to the time-aligned signals, are used to calculate the feature vectors of the sensor nodes, generating multiple feature vectors.

[0097] The spatiotemporal graph attention mechanism is used to calculate edge weights using feature vectors and delay matrices respectively, generating the edge weights corresponding to the sensor nodes.

[0098] In this embodiment of the invention, the contact pressure, temperature, and angular velocity corresponding to the digital signals, as well as the vibration spectrum vector corresponding to the timing alignment signal, are used to calculate the feature vectors of the sensor nodes, generating multiple feature vectors. Each feature vector includes pressure, vibration spectrum, temperature, and angular velocity. The calculation formula for the feature vectors is as follows:

[0099] ;

[0100] in, The node features, i.e., the feature vector, of the i-th sensor contain time-frequency features after wavelet decomposition (such as high-frequency phase angle). Time-frequency energy spectrum ); Contact pressure (unit: N); The vibration spectrum vector is (0-20kHz); T is time. ω is the angular velocity.

[0101] The spatiotemporal graph attention mechanism is used to calculate edge weights using feature vectors and delay matrices respectively, generating the edge weights corresponding to the sensor nodes.

[0102] In the MAWGRC framework, the edge weights of the Spatiotemporal Graph Attention Network (ST-GAT) are the attention weights. Instead of directly using the delay matrix Instead, they are dynamically linked through an attention mechanism constrained by physical fields. Their synergistic effect is as follows:

[0103] Delay matrix The theoretical delay of signal propagation between nodes is calculated based on the elastic wave equation (see formula). This reflects the spatial topology and physical propagation laws of sensor networks.

[0104] Function: To provide physical interpretability constraints for attention mechanisms and avoid weight bias caused by noise in purely data-driven models.

[0105] The formula for the edge weights used in the dynamic weighting of the attention mechanism is:

[0106] ;

[0107] in, The edge weight between nodes i and j is used to measure the degree of association between nodes and reflects the signal propagation relationship in the spatiotemporal graph network. The LeakyReLU activation function is used; W is the learnable weight matrix. This involves concatenating vectors. As the weight vector in the attention mechanism, it participates in the calculation of edge weights and determines the importance of different node features when calculating edge weights; This is a vector concatenation operation, which transforms the nodes after the weight matrix W. and The feature vectors are concatenated and used as input to the attention mechanism; For nodes The set of neighboring nodes is used to comprehensively consider the information of neighboring nodes when calculating edge weights; Let be the feature vector of the k-th sensor.

[0108] Dynamic association logic:

[0109] Node features , : Let be the node features, i.e., the feature vector, of the i-th sensor. The node features, i.e., the feature vector, of the j-th sensor are the node features. Includes time-frequency features after wavelet decomposition (such as high-frequency phase angle) Time-frequency energy spectrum ).

[0110] Implicit embedding delay matrix The attention weights are indirectly constrained by influencing the distribution of node features through physical field parameters. The generation of .

[0111] The multi-band signals (including pressure, vibration spectrum, etc.) after wavelet decomposition constitute the node attribute vectors of the graph neural network, while the delay matrix... Dynamically compute edge weights of graph structures This refers to attention weights, which are used to determine faults based on edge weights, and the fault results are then input into the Ethernet network.

[0112] Mathematical Integration: Spatiotemporal Graph Attention Mechanism (Formula) This method integrates physical delay with signal characteristics to capture abnormal propagation paths in the complex structure of substations.

[0113] Examples of mathematical-physical connections

[0114] In wear detection of robotic arms:

[0115] If the contact becomes loose, causing the propagation path to lengthen (increase), node characteristics... The phase difference increases synchronously, driving the attention network to assign higher weights to this path. This enhances the sensitivity of abnormal path detection.

[0116] Furthermore, the reinforcement learning optimization module performs the following steps:

[0117] The state space is constructed using the error change rate corresponding to the time-aligned signal, the edge weights, the wave velocity gradient corresponding to the wave equation, and the phase angle vector, and a state space vector is generated.

[0118] The dual-delay deep deterministic policy gradient algorithm uses state space vectors to dynamically compensate the policy network parameters corresponding to the tactile signal, generating tactile temporal compensation data.

[0119] In this embodiment of the invention, the state space vector includes the DTW error rate of change, wave velocity gradient, graph attention weights, and phase angle vector. The state space vector is constructed using the error rate of change corresponding to the time-aligned signal, edge weights, wave velocity gradient corresponding to the wave equation, and phase angle vector. The state space vector is:

[0120] ;

[0121] in, To reinforce learning, the state-space vector at time t contains multiple parameters reflecting the current state of the system; DTW is the error rate of change, and DTW (Dynamic Time Warping) is often used to measure the similarity between two time series. This describes how its error changes over time; The wave velocity gradient reflects the rate of change of wave velocity in space or time, and is used to represent the dynamic changes of wave velocity in a physical field. In spatiotemporal graph attention networks, the graph attention weights are weight vectors used to represent the connections between nodes, reflecting the strength of the association between nodes. It is a phase angle vector containing phase angle information of multiple sensor signals at different frequencies, used to describe the phase characteristics of the signal.

[0122] The TD3 algorithm update strategy uses a dual-delay, deep deterministic policy gradient algorithm to dynamically compensate the policy network parameters corresponding to the tactile signal using the state space vector, generating tactile temporal compensation data and sending it to the physics modeling module. The specific update process is as follows:

[0123] Policy network output actions:

[0124] ;

[0125] ;

[0126] in, For output actions; The parameters are the policy network parameters, i.e., the target network parameters (delayed-updating policy network parameters), and the neural network weight matrix. This is Ornstein-Uhlenbeck noise, or simply "O-Uhlenbeck noise"; The discount factor, also known as the policy gradient step size (learning rate, controlling the magnitude of parameter updates), ranges from 0.0001 to 0.001 (often 0.99). The target network update rate controls the update speed of the parameter, with a value range of 0.001~0.01 (usually 0.005). For displacement field; The policy gradient (loss function J with respect to policy parameters) The gradient ( ) is the value calculated during backpropagation; To explore noise (random perturbations generated by the Ornstein-Uhlenbeck process), the value range is: .

[0127] The closed-loop optimization of reinforcement learning involves using a dual-delay deep deterministic policy gradient algorithm to dynamically compensate the policy network parameters corresponding to the tactile signal using the state space vector, thereby generating tactile temporal compensation data.

[0128] The reinforcement learning agent uses a temporal consistency metric (such as DTW distance) as its reward function and adjusts dynamically.

[0129] Physical field parameters (material-related wave velocity c, damping coefficient) );

[0130] Wavelet basis function parameters (bandwidth parameters) Center frequency );

[0131] Graph attention weights are edge weights. .

[0132] Feedback mechanism: Forms a closed loop of physical field modeling → feature extraction → strategy update → parameter correction, adapting to dynamic environments such as substation temperature changes and electromagnetic interference.

[0133] Furthermore, the physical definition of the strategy

[0134] Policy Network Input state space vector (e.g., timing error, temperature, electromagnetic interference intensity), output action (such as wave speed correction amount) Phase compensation amount ).

[0135] Objective: To maximize long-term rewards (such as timing consistency and fault detection accuracy) by adjusting actions.

[0136] Role in the MAWGRC framework

[0137] Dynamic parameter tuning: Real-time optimization of physical field parameters (material-dependent wave velocity c, damping coefficient) ) and wavelet decomposition parameters (bandwidth parameters) Center frequency It adapts to the complex environment of substations.

[0138] Anti-overfitting: By updating the target network rate ( ) and exploration noise ( This helps avoid the strategy getting stuck in local optima.

[0139] The TD3 policy network updates the target policy network parameters with a delay. ) and exploration noise ( This was implemented in the MAWGRC framework:

[0140] 1. Dynamic optimization of physical field parameters (material-dependent wave velocity c, damping coefficient) );

[0141] 2. Robust compensation for environmental disturbances (temperature changes, electromagnetic noise);

[0142] 3. Significantly improved fault detection accuracy.

[0143] In this embodiment of the invention, the joint optimization of delay tracing and phase alignment is achieved by introducing an elastic wave backpropagation algorithm to optimize the delay matrix. The computational error is backpropagated to the wavelet decomposition layer to achieve:

[0144] ;

[0145] Where D is the theoretical delay matrix (calculated through the physical field model); The actual delay matrix (obtained through experimental measurement or high-precision calibration); The weighting factor is used to minimize the physical delay error and the signal phase error simultaneously; Optimize the objective function value for phase consistency.

[0146] Multimodal embedding of graph neural networks, i.e., designing heterogeneous graph structures:

[0147] Node embedding: ,in, The node features, i.e., the feature vector, of the i-th sensor contain time-frequency features after wavelet decomposition (such as high-frequency phase angle). Time-frequency energy spectrum ); Contact pressure (unit: N); The vibration spectrum vector is (0-20kHz); T is time.

[0148] Edge embedding: , where is a node and The embedding vector of the edge between them; This represents the delay matrix. MLP stands for Multi-Layer Perceptron, a feedforward neural network composed of multiple fully connected layers. It learns complex feature representations by performing nonlinear transformations on the input. Here, it is used to transform the concatenated node features to extract effective features related to the relationships between nodes. edge weight The gradient of the edge weights can be calculated during model training to understand the edge weights. The changing trend in the loss function is then used to adjust the edge weights and optimize the model's learning and representation of relationships between nodes.

[0149] Cross-modal interaction between physical field data and wavelet features is achieved through graph convolutional layers.

[0150] Define the policy network output action:

[0151] ;

[0152] in, To reinforce learning the output action at time t; Wave speed correction amount; Bandwidth parameter adjustment amount; Edge weight adjustment amount; Adjustment amount of medium damping coefficient

[0153] The overestimation problem in substations with strong noise is avoided by using the dual-Q network structure of the TD3 algorithm.

[0154] Furthermore, the substation scenario verification scheme is shown in the table below:

[0155]

[0156] Furthermore, a typical application example: Detection of loose transformer bushings:

[0157] 1. The vibration signal measured by the tactile array is compensated by MAWGRC, and the 3.2kHz characteristic frequency band is extracted.

[0158] 2. Spatiotemporal graph network identification of abnormal propagation paths: Sleeve → Support frame → Base

[0159] 3. The reinforcement learning module dynamically adjusts the detection threshold, reducing the false positive rate by 62%.

[0160] Disconnector switch contact wear assessment:

[0161] 1. Wavelet decomposition to separate contact pressure (0-200Hz) and arc interference (5-20kHz)

[0162] 2. By restoring the true contact timing through phase alignment, the wear estimation error is <8%.

[0163] This method solves the problem of tactile data distortion in the complex electromagnetic environment of substations by deeply integrating physical mechanisms and deep learning. Actual tests show that it can improve the accuracy of mechanical fault identification to 98.7%.

[0164] Please see Figure 3 , Figure 3 This is a flowchart of a tactile timing compensation method provided in Embodiment 2 of the present invention.

[0165] The present invention provides a tactile timing compensation method, comprising:

[0166] Step 301: Acquire the tactile signals of the substation inspection robot and convert the tactile signals into digital signals.

[0167] Step 302: Use digital signals to model the physical field and construct the wave equation and delay matrix.

[0168] Step 303: Using the delay matrix as prior knowledge, perform phase alignment on the multi-band signals after wavelet decomposition of the wave equation to generate a time-aligned signal.

[0169] Step 304: Use a spatiotemporal graph network to calculate the edge weights of the time-series aligned signal and generate multiple edge weights.

[0170] Step 305: Dynamically compensate the tactile signal according to the edge weights to generate tactile timing compensation data.

[0171] In this embodiment of the invention, the tactile sensors of the substation inspection robot are responsible for collecting tactile signals during equipment operation. These signals are typically analog signals. The tactile signals are then first converted into voltage signals by a charge amplifier, then filtered by a bandpass filter to remove noise, and finally converted into digital signals by an ADC (analog-to-digital converter) for subsequent digital processing. Physical field modeling provides physical constraints and theoretical support for subsequent processing, making the analysis of tactile signals more consistent with actual physical laws. The delay matrix is ​​used as prior knowledge to initialize the phase alignment target. The delay error from physical field modeling is transformed into multi-band phase difference in the wavelet domain. Phase consistency optimization achieves dual-drive compensation of physical and data, generating a time-aligned signal. Edge weight calculation reflects the correlation between sensor nodes, highlighting anomaly propagation paths and providing a more accurate basis for subsequent dynamic compensation. Based on edge weights and reinforcement learning, relevant parameters are dynamically adjusted to compensate for the tactile signals, improving time-series compensation accuracy, reducing false alarm rates, and increasing the accuracy of mechanical fault identification, meeting the high-precision requirements of substation equipment fault detection.

[0172] Please see Figure 4 , Figure 4 This is a structural block diagram of an electronic device provided in Embodiment 3 of the present invention.

[0173] An electronic device according to an embodiment of the present invention includes: a memory 401 and a processor 402. The memory 401 stores a computer program. When the computer program is executed by the processor 402, the processor 402 performs a tactile timing compensation method as described in any of the above embodiments.

[0174] Memory 401 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Memory 401 has storage space 403 for program code 413 for performing any of the method steps described above. For example, storage space 403 for program code may include individual program codes 413 for implementing the various steps in the methods described above. This program code may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above. This program code may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When this code is run by a computing device, it causes the computing device to perform the various steps in the haptic timing compensation method described above.

[0175] Embodiment 5 of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the tactile timing compensation method as described in any of the above embodiments.

[0176] Embodiment 6 of the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer performs the tactile timing compensation method as described in any of the above embodiments.

[0177] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0178] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0179] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0180] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0181] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0182] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A tactile timing compensation system, characterized in that, include: The module includes a tactile signal processing module, a physics field modeling module, a wavelet phase alignment module, a spatiotemporal graph network module, and a reinforcement learning optimization module. The tactile signal processing module is used to acquire the tactile signals of the substation inspection robot and convert the tactile signals into digital signals. The physical field modeling module is used to perform physical field modeling using the digital signal, and to construct wave equations and delay matrices; The wavelet phase alignment module is used to use the delay matrix as prior knowledge to perform phase alignment on the multi-frequency signal after wavelet decomposition of the wave equation and generate a time-aligned signal. The spatiotemporal graph network module is used to calculate the edge weights of the time-series alignment signal using a spatiotemporal graph network, and generate multiple edge weights. The reinforcement learning optimization module is used to dynamically compensate the tactile signal according to the edge weights and generate tactile temporal compensation data.

2. The tactile timing compensation system according to claim 1, characterized in that, The tactile signal processing module includes a tactile sensor, a charge amplifier, a bandpass filter, and an analog-to-digital converter; The tactile sensor is used to acquire tactile signals from the substation inspection robot; The charge amplifier is used to convert the tactile signal into a voltage signal to generate an initial voltage signal; The bandpass filter is used to filter out noise signals in the initial voltage signal and generate the target voltage signal; The analog-to-digital converter is used to convert the target voltage signal into a digital signal.

3. The tactile timing compensation system according to claim 1, characterized in that, The physics modeling module performs the following steps: The medium damping coefficient and electromagnetic mechanical coupling coefficient corresponding to the tactile signal were calibrated by finite element simulation. A wave equation is constructed using the material-related wave velocity corresponding to the digital signal, the external excitation function, the medium damping coefficient, and the electromagnetic-mechanical coupling coefficient. A delay matrix is ​​generated by constructing a matrix using the three-dimensional coordinates of the sensor node corresponding to the digital signal, a preset material correction coefficient, and the material-related wave velocity.

4. The tactile timing compensation system according to claim 1, characterized in that, The wavelet phase alignment module performs the following steps: The tactile signal is subjected to complex Morlet wavelet transform to generate a time-frequency energy spectrum; The instantaneous phase spectrum is obtained by calculating the instantaneous phase spectrum using the time-frequency energy spectrum. With phase consistency as the objective, a cross-band phase alignment loss function is constructed using the wavelet coefficients and phase angles corresponding to the instantaneous phase spectrum; The delay error corresponding to the delay matrix is ​​converted into a multi-band phase difference in the wavelet domain using the cross-band phase alignment loss function to generate a timing alignment signal.

5. The tactile timing compensation system according to claim 1, characterized in that, The spatiotemporal graph network module performs the following steps: The contact pressure, temperature, and angular velocity corresponding to the digital signals, as well as the vibration spectrum vector corresponding to the time-aligned signal, are respectively used to calculate the feature vectors of the sensor nodes to generate multiple feature vectors. The edge weights corresponding to the sensor nodes are generated by calculating the edge weights using the feature vector and the delay matrix respectively through the spatiotemporal graph attention mechanism.

6. The tactile timing compensation system according to any one of claims 1-5, characterized in that, The reinforcement learning tuning module performs the following steps: The state space is constructed using the error change rate corresponding to the time-series alignment signal, the edge weights, the wave velocity gradient corresponding to the wave equation, and the phase angle vector, to generate a state space vector. The state space vector is used to dynamically compensate the policy network parameters corresponding to the tactile signal by employing the dual-delay deep deterministic policy gradient algorithm, thereby generating tactile temporal compensation data.

7. A tactile timing compensation method, characterized in that, include: Acquire tactile signals from the substation inspection robot and convert the tactile signals into digital signals; The digital signal is used to model the physical field, and wave equations and delay matrices are constructed. Using the delay matrix as prior knowledge, phase alignment is performed on the multi-band signal after wavelet decomposition of the wave equation to generate a time-aligned signal. A spatiotemporal graph network is used to calculate the edge weights of the time-series aligned signal, generating multiple edge weights. The tactile signal is dynamically compensated based on the edge weights to generate tactile timing compensation data.

8. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the tactile timing compensation method as described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the tactile timing compensation method as described in claim 7.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the haptic timing compensation method as described in claim 7.

Images