An intelligent auxiliary winding method and device for high-voltage insulation self-fusion belt insulation treatment

Abstract

The application provides an intelligent auxiliary winding method and device for high-voltage insulation self-fusion belt insulation treatment, and belongs to the technical field of electric digital data processing and intelligent control. The method comprises the following steps: collecting multi-dimensional physical information of a winding process, and preprocessing to obtain a multi-source sensor fusion vector; an adaptive winding strategy network and a value network are constructed; the multi-source sensor fusion vector is input into the trained adaptive winding strategy network, and a winding parameter optimization vector is output; a quality evaluation prediction network is constructed in parallel based on a bidirectional long short-term memory network and a one-dimensional convolutional neural network, the multi-source sensor fusion vector and the winding parameter optimization vector are spliced and input into the quality evaluation prediction network, and a quality prediction confidence is output; and the winding parameter optimization vector and the quality prediction confidence are converted into a control instruction sequence, which is executed by an executing mechanism. The application realizes the collaborative optimization of winding tension and winding speed, and improves the adaptive ability of winding operation.

Description

Technical Field

[0001] This invention relates to an intelligent assisted winding method and device for insulation treatment of high-voltage self-fusion insulating tape, belonging to the field of electrical digital data processing and intelligent control technology. Background Technology

[0002] With the continuous advancement of intelligent operation and maintenance in power systems, the demand for precision and efficiency in insulation treatment on the 10kV side of main transformers is becoming increasingly prominent. High-voltage self-fusion insulation tape winding, as a core process in insulation treatment, directly affects the insulation protection effect, equipment heat dissipation performance, and operation and maintenance safety due to the coordinated precision of winding force, number of layers, and film peeling. At the same time, the operational difficulty of traditional manual winding methods within the confined space of the 10kV side of the main transformer is constantly increasing, posing a severe challenge to the digitalization and intelligentization of winding operations.

[0003] Currently, the main methods for controlling the winding of high-voltage insulating self-fusion tape include the following: (1) Manual winding control based on human experience: This method mainly relies on the operator holding the self-fusion insulating tape reel and performing the winding operation through visual observation and tactile feedback, relying on experience to judge the winding force and number of layers. The operation method is relatively direct and does not require special equipment. However, this method usually depends on the individual experience and physical condition of the operator, resulting in a large fluctuation range of winding force and a high error in the number of layers, which limits its application in the standardized management of the 10kV side of the main transformer (see: State Grid Corporation of China "Standardized Operation Guide for Power Equipment Operation and Maintenance", the measured range of manual winding tension fluctuation is ±2.5 Newtons, and the error in the number of layers is as high as ±3 layers).

[0004] (2) Open-loop assisted winding based on fixed program: Automatic conveying and winding of self-fusion tape is achieved by using a mechanical device with a preset fixed speed and pressure and a motor-driven winding wheel assembly. This method has the advantage of reducing the intensity of manual labor. However, due to the lack of a real-time feedback adjustment mechanism, it is impossible to make dynamic adjustments based on changes in the curvature of the insulation layer surface, fluctuations in the thickness of the tape, and changes in ambient temperature and humidity. It is difficult to meet the precise winding requirements of the complex morphology of the 10kV side of the main transformer (see: Chinese Patent CN110517878A, Automatic winding machine for coating transformer core and its tape winding method, which adopts a mechanical structure with a fixed speed and pressure and lacks real-time feedback adjustment capability).

[0005] (3) Winding control based on single sensor closed loop: Some research and engineering practices have attempted to collect process data in a single dimension using pressure or displacement sensors and combine them with PID controllers to achieve closed-loop adjustment of winding force or winding position. However, such methods rely on single sensor information and lack the ability to fully perceive the multi-physics coupling characteristics of the winding process. They cannot comprehensively evaluate the winding quality, which makes it difficult to guarantee the winding quality at key locations such as the edge of the insulation layer and the transition zone of the curved surface (see: ResearchGate paper "Design of Insulation Tape Tension Control System of Transformer Winding Machine Based on Fuzzy PID", which uses a single pressure sensor combined with a PID controller to achieve tension closed-loop adjustment, but only targets the single dimension of tension and does not integrate multi-source sensor information for comprehensive quality evaluation).

[0006] In summary, existing high-voltage self-fusion insulation tape winding control methods have certain limitations in terms of sensor information fusion, adaptive parameter optimization, and quality prediction and evaluation: manual methods rely on experience and have poor consistency, open-loop assisted methods lack feedback adjustment capabilities, and single-sensor closed-loop methods have limited information dimensions and cannot predict quality. Based on these shortcomings, it is necessary to develop an intelligent high-voltage self-fusion insulation tape winding control method and device that can integrate multi-source sensor information, adaptively optimize winding parameters, and predict winding quality in real time. This would improve the standardization level of winding operations and the quality of insulation treatment, thereby better meeting the actual needs of intelligent operation and maintenance of power equipment. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent assisted winding method and device for insulation treatment of high-voltage self-fusion insulating tape, which solves the problems of insufficient sensor information fusion, inability to adaptively optimize parameters, and difficulty in predicting and evaluating quality in existing high-voltage self-fusion insulating tape winding control methods.

[0008] To achieve the above objectives, the present invention employs the following technical solution: A smart assisted winding method for insulation treatment of high-voltage self-fusion insulating tape includes the following steps: Multidimensional physical information of the winding process is collected in real time by multiple types of sensors, and the multidimensional physical information is spatiotemporally aligned and feature fused to obtain a multi-source sensing fusion vector. An adaptive entanglement policy network and a value network are constructed based on deep reinforcement learning. The deep reinforcement learning algorithm adopts Markov decision and the proximal policy optimization algorithm is used to train the adaptive entanglement policy network. The multi-source sensing fusion vector is input into the trained adaptive winding strategy network, and the winding parameter optimization vector is output. A quality assessment and prediction network is constructed by paralleling a bidirectional long short-term memory network and a one-dimensional convolutional neural network. The multi-source sensor fusion vector and the winding parameter optimization vector are concatenated and input into the quality assessment and prediction network to output the quality prediction confidence. If it is lower than a preset threshold, an anomaly warning is triggered and the current state is marked as pending review. The quality prediction confidence is fed back to the reward function of deep reinforcement learning in real time as the input of the winding quality comprehensive score in the reward function. The winding parameter optimization vector and quality prediction confidence are converted into a sequence of control commands and sent to the actuator of the intelligent assisted winding device for execution.

[0009] Preferably, the multidimensional physical information includes: contact pressure data, axial displacement data, rotational encoding data, and spatial positioning data collected by pressure sensors, displacement sensors, counting encoders, and positioning sensors, respectively; The spatiotemporal alignment and feature fusion yield a multi-source sensing fusion vector, specifically including: The four types of sensor data collected were time-stamp aligned and dimension-normalized, and then spliced ​​together to form the original sensor matrix. A local temporal feature map is obtained by using a one-dimensional convolutional neural network to extract local temporal features from the original sensing matrix. The local feature map is input into a multi-head self-attention mechanism to calculate the correlation weight matrix between different sensing channels and perform weighted fusion to output the multi-source sensing fusion vector.

[0010] Preferably, the Markov decision uses a multi-source sensor fusion vector as the environmental state input; the action space is a two-dimensional continuous vector composed of the winding tension setting value and the winding speed setting value; The reward function is a weighted combination of insulation quality reward, edge coverage integrity reward, and material utilization penalty: , in, Rewards for the current moment. As a reward for the quality of the insulation layer, For edge coverage integrity rewards, Penalty for low material utilization; The insulation layer quality bonus is calculated based on the time-series statistical characteristics of the contact pressure data within the current time window. It takes the weighted sum of the standard deviation of the contact pressure and the deviation of the standard deviation of the contact pressure from the target winding tension, and maps it to the 0 to 1 interval through a Gaussian kernel function. The edge coverage integrity reward is calculated based on whether the positioning sensor detects the terminal end face boundary and whether the number of winding layers measured by the displacement sensor at the boundary meets the preset threshold. When the positioning sensor recognizes the end face and the number of layers meets the standard, a positive reward is given; otherwise, a negative penalty is given. The material utilization penalty is calculated based on the absolute value of the deviation between the actual winding length (converted from the cumulative pulse count of the counting encoder) and the target winding length.

[0011] Preferably, both the adaptive winding strategy network and the value network adopt a four-layer fully connected structure, the hidden layers of the adaptive winding strategy network adopt the ReLU activation function, and the output layer adopts the Tanh activation function. The winding force safety constraint layer is introduced after the output layer of the adaptive winding strategy network. The winding force safety constraint layer projects the winding tension setting value output by the adaptive winding strategy network to a safe range of 0.5 to 5.0 Newtons through a hard clipping function, and clips the winding speed setting value to a range of 10 to 100 millimeters per second. If the value exceeds the boundary value, the boundary value is directly taken. The adaptive winding strategy network total loss as follows: , , , in, For strategic losses, The value loss coefficient, The mean squared error loss of the value network. The entropy reward coefficient, For policy entropy, As a strategy, For the state value function, for The state at any given moment, Let be the cumulative reward at time t. For strategic losses, To calculate the expectation over time step t, The probability ratio between the old and new strategies. This is the estimate of the generalized advantage. For the clipping function, This is the cutting factor.

[0012] Preferably, the specific steps for the quality assessment prediction network to output the confidence level of the quality prediction include: After concatenating the multi-source sensing fusion vector and the winding parameter optimization vector, the concatenation vector is input in parallel to the bidirectional long short-term memory network branch and the one-dimensional convolutional neural network branch. The temporal context feature vector output by the bidirectional long short-term memory network branch is concatenated with the local mutation feature vector output by the one-dimensional convolutional neural network branch to obtain a parallel fused feature vector. The parallel fused feature vector is mapped to insulation layer thickness uniformity score, bonding tightness score and edge coverage integrity score through a fully connected layer. The quality prediction confidence score is obtained by weighted summation of the thickness uniformity score, bonding tightness score and edge coverage integrity score.

[0013] Preferably, the confidence level for quality prediction is calculated as follows: The Softmax function was used to normalize the scores for insulation layer thickness uniformity, bonding tightness, and edge coverage integrity. Calculate the relative weights of the scores for each dimension, and then calculate the confidence level of the quality prediction by weighted summation: , in, For quality prediction confidence level, To score the thickness uniformity, To score the tightness of the fit, Score the edge coverage integrity.

[0014] An anomaly detection threshold is set. If the quality prediction confidence level is below 0.85, the current winding state is determined to have a quality risk, triggering an audible and visual alarm and displaying an anomaly code and suggested handling measures on the touch operation panel. At the same time, the sensor data and parameter records within that time window are marked as an anomaly sample and stored in the local database. If the quality prediction confidence level is between 0.85 and 0.95, it is determined to be in a state of alert, and the control command update frequency is increased to 100 Hz. If the quality prediction confidence level is above 0.95, it is determined to be in a normal state, and standard control parameters are maintained.

[0015] Preferably, the step of converting the winding parameter optimization vector and the quality prediction confidence into a control command sequence specifically includes: The winding tension setting value in the winding parameter optimization vector is converted into the PWM duty cycle control value of the electromagnetic pressure regulating valve by the dynamic control generator, and the winding speed setting value is converted into the speed pulse frequency of the servo motor. The update frequency of control commands is dynamically adjusted based on the quality prediction confidence level. The control command sequence is sent to the PLC controller of the intelligent auxiliary winding device via the CAN bus, and the PLC controller drives the actuator to move.

[0016] A smart auxiliary winding device for high-voltage self-fusion insulating tape insulation treatment includes: The multi-source sensing layer includes pressure sensors, displacement sensors, counting encoders, and positioning sensors, which are used to collect multi-dimensional physical information of the winding process; The main control module includes an edge computing unit and a PLC controller. The edge computing unit is used to perform multi-source sensor data fusion, adaptive optimization of winding parameters, and real-time prediction of winding quality. The PLC controller is used to generate and issue control commands. The actuator layer includes a winding execution module, a force adjustment module, a layer number control module, a film peeling module, and a film recycling module, which are used to execute the control commands; The multi-source sensing layer, main control module, and actuator layer are connected through a data closed loop to form a complete control loop from perception and decision-making to execution.

[0017] Preferably, the pressure sensor is installed directly below the bearing seat of the clamping wheel of the passive winding wheel assembly to detect the radial clamping force of the self-fusion tape on the surface of the terminal block; the displacement sensor is installed on the side slide rail of the bearing seat of the active winding wheel assembly to monitor the reciprocating displacement of the wheel assembly along the axial direction of the terminal block; the counting encoder is installed on the output shaft of the drive motor to record the number of motor rotations; and the positioning sensor is embedded in the front end of the fixing fixture to identify the winding start and end positions.

[0018] The advantages of this invention are as follows: This invention constructs a multi-source sensing fusion and feature extraction module based on a multi-head self-attention mechanism, which performs spatiotemporal alignment and deep fusion of contact pressure, axial displacement, rotation encoding and spatial positioning data, realizes complementary enhancement and noise suppression of multi-dimensional information in the winding process, and improves the comprehensiveness and accuracy of state perception.

[0019] An adaptive winding strategy network based on a near-end strategy optimization algorithm was designed. Through deep reinforcement learning, it autonomously explores and generates the optimal winding strategy under complex working conditions, realizing the coordinated optimization of winding tension and winding speed, and improving the adaptability of winding operations.

[0020] A quality assessment and prediction network based on a bidirectional long short-term memory network and a one-dimensional convolutional neural network was designed. This network can predict and identify anomalies in real time the uniformity of insulation layer thickness, bonding tightness, and edge coverage integrity, thus realizing online assessment and closed-loop feedback of winding quality. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0022] Figure 1 A flowchart of the overall technology roadmap; Figure 2 This is a schematic diagram of the feature extraction structure for multi-source sensing fusion. Figure 3 This is a schematic diagram of the adaptive winding strategy network structure; Figure 4 This is a schematic diagram of the network structure for quality assessment prediction. Figure 5 Architecture diagram for generating and issuing intelligent control commands; Figure 6 This is a schematic diagram of the overall structure of the intelligent assisted winding device; Figure 7 Detailed diagram of sensor deployment in the winding wheel assembly area; Figure 8 For comparison curves of winding tension tracking; Figure 9 A cloud map showing the confidence distribution of quality predictions; Figure 10 A schematic diagram of the convergence trajectory to enhance learning and training. Detailed Implementation

[0023] 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.

[0024] Example 1 like Figures 1-5 As shown, an intelligent assisted winding method for insulation treatment of high-voltage self-fusion insulating tape is presented. This method uses multiple types of sensors deployed on an intelligent assisted winding device to collect contact pressure data, axial displacement data, rotational encoding data, and spatial positioning data in real time, forming a multi-source sensor fusion vector. Subsequently, a dynamic mapping relationship between winding parameters and winding quality is established based on an adaptive winding strategy network to achieve synergistic optimization of winding tension and winding speed. Furthermore, a quality assessment prediction network is used to predict the insulation layer thickness uniformity, bonding tightness, and edge coverage integrity of the current winding state in real time, outputting the quality prediction confidence level. Finally, a dynamic control generator converts the optimized winding parameters into a control command sequence, and dynamically sends the control command sequence to the actuator of the intelligent assisted winding device to complete closed-loop intelligent control.

[0025] Specifically, it includes the following steps: S1: Real-time acquisition of multi-dimensional physical information of the winding process by multiple types of sensors, and spatiotemporal alignment and feature fusion of the multi-dimensional physical information to obtain a multi-source sensing fusion vector; S2: An adaptive entanglement policy network and a value network are constructed based on deep reinforcement learning. The deep reinforcement learning algorithm adopts Markov decision and the proximal policy optimization algorithm is used to train the adaptive entanglement policy network. S3: Input the multi-source sensing fusion vector into the trained adaptive winding strategy network, and output the winding parameter optimization vector; S4: A quality assessment and prediction network is constructed by paralleling a bidirectional long short-term memory network and a one-dimensional convolutional neural network. The multi-source sensor fusion vector and the winding parameter optimization vector are concatenated and input into the quality assessment and prediction network. The output is the quality prediction confidence. If it is lower than the preset threshold, an anomaly warning is triggered and the current state is marked as pending review. The quality prediction confidence is fed back to the reward function of deep reinforcement learning in real time as the input of the winding quality comprehensive score in the reward function. S5: The winding parameter optimization vector and quality prediction confidence are converted into a sequence of control commands and sent to the actuator of the intelligent assisted winding device for execution.

[0026] As a refinement of the above embodiments, step S1 aims to comprehensively perceive and deeply fuse the multidimensional physical information of the winding process in order to ensure the completeness of information for subsequent winding parameter optimization. This step aims to acquire multidimensional physical information of the winding process in real time through multiple types of sensors deployed on the intelligent assisted winding device, and to perform spatiotemporal alignment and deep fusion of the multidimensional physical information to form a multi-source sensing fusion vector.

[0027] Specific examples Figure 2 As shown, it includes: S101: Multi-source sensor data acquisition and preprocessing: Contact pressure data is acquired using a pressure sensor installed on the force transmission path between the passive winding wheel assembly, i.e., the clamping wheel bearing housing and the electromagnetic pressure regulating valve. Specifically, it is located at the clamping force detection point directly below the contact surface between the self-fusion tape and the passive winding wheel assembly. The radial clamping force of the self-fusion tape on the terminal surface during the winding process is detected in real time at a sampling frequency of 1000 Hz. The range is 0 to 10 Newtons, and the resolution is 0.01 Newtons. It is used to characterize the tightness of the insulation layer adhesion.

[0028] Axial displacement data is collected using a displacement sensor. This displacement sensor is fixed to the side of the axial slide rail of the active winding wheel assembly, i.e., the active wheel bearing housing, via a magnetic base. The sensing probe faces the end face of the bearing housing and detects the reciprocating linear displacement of the wheel assembly along the terminal axis at a sampling frequency of 500 Hz. The range is 0 to 200 mm and the resolution is 0.01 mm. This data is used to calculate the winding stroke and the cumulative number of layers.

[0029] Rotational encoding data is acquired using a counting encoder, which is installed on the transmission chain between the output shaft of the drive motor and the reduction mechanism via a flexible coupling. The encoder records the number of motor rotations at a resolution of 4096 pulses per revolution and obtains the real-time rotational speed through first-order differential calculation, which is used to calculate the conveying length and winding speed of the self-fusion tape.

[0030] Spatial positioning data is collected using a positioning sensor. This positioning sensor is embedded on the edge recognition surface of the front end of the fixed fixture near the treatment area, with the sensing surface facing outward and directly opposite the boundary of the terminal end face. It identifies the start and end boundary positions of the treatment area at a sampling frequency of 100 Hz, which is used to determine the spatial range and edge coverage status of the winding operation. The above four types of sensor data are converted into digital signals by a 24-bit analog-to-digital converter and then buffered in the ring buffer of the edge computing unit.

[0031] S102: Spatiotemporal alignment and feature fusion: Based on the timestamp alignment algorithm, the sampling rate of axial displacement data, rotational encoding data and spatial positioning data is unified to 1000 Hz through linear interpolation using the time axis of contact pressure data, thereby eliminating the time delay difference caused by the asynchronous sampling between sensors.

[0032] The Z-score normalization method was used to normalize the dimensions of the four types of sensor data to eliminate the numerical scale differences between different physical quantities. Then, the normalized four types of sensor data were concatenated into a 4×T-dimensional original sensor matrix, where T is the time window length and is set to 128 sampling points.

[0033] A one-dimensional convolutional neural network (1DCNN) is used to extract local temporal features from the original sensor matrix. The kernel size is 3, the number of channels is 64, the stride is 1, and the padding method is same. The ReLU activation function is used to output a 64×T local feature map. Finally, the local feature map is input into a multi-head self-attention (MHSA) mechanism with 4 attention heads to calculate the correlation weight matrix between different sensor channels. After weighted fusion, a 256-dimensional multi-source sensor fusion vector is output through global average pooling.

[0034] As a refinement of the above embodiment, step S2 aims to establish a dynamic mapping relationship between the winding process state and the optimal control parameters. Through a deep reinforcement learning algorithm, it autonomously explores and converges to the optimal winding strategy under complex working conditions, outputting an optimized winding parameter vector. This achieves autonomous optimization and dynamic adaptation of winding parameters under complex working conditions, ensuring the optimal synergy between winding force and winding speed.

[0035] Specifically, such as Figure 3 The following are included: S201: Construction of Markov Decision Processes A 256-dimensional multi-source sensor fusion vector is used as the environmental state input to characterize the complete physical state of the current winding process. First, the action space is defined as a two-dimensional continuous vector. The first dimension represents the winding tension setting, ranging from 0.5 to 5.0 Newtons, and the second dimension represents the winding speed setting, ranging from 10 to 100 millimeters per second. Second, a three-dimensional reward function tightly coupled with the high-voltage insulation self-fusion tape winding process is constructed. This reward function is a weighted combination of three factors: insulation layer quality reward, edge coverage integrity reward, and material utilization penalty, specifically expressed as follows: , in, Rewards for the current moment. As a reward for the quality of the insulation layer, For edge coverage integrity rewards, Penalty for low material utilization.

[0036] The insulation layer quality bonus is calculated based on the time-series statistical characteristics of contact pressure data within the current time window. Specifically, it is the weighted sum of the pressure standard deviation and the deviation of the pressure mean from the target tension. The smaller the standard deviation and the closer the mean is to the target tension, the more uniform the insulation layer thickness, the tighter the self-fusion tape adheres to the terminal surface, and the lower the risk of bubbles and wrinkles. The bonus value is mapped to the 0-1 interval using a Gaussian kernel function, and its mathematical expression is: , in, As a reward for the quality of the insulation layer, The standard deviation of the contact pressure within the current time window. The average contact pressure. For the target winding tension, , These are weighting coefficients. The value range is [0,1]. Values ; Gaussian kernel bandwidth parameters.

[0037] The edge coverage integrity reward is calculated based on whether the positioning sensor detects the terminal block end face boundary and whether the number of winding layers measured by the displacement sensor at the boundary meets a preset threshold (6 in this embodiment). A positive reward is given when the positioning sensor identifies the end face and the number of layers meets the threshold. Add 1, otherwise give a negative penalty, that is Subtract one.

[0038] The material utilization penalty is calculated based on the absolute value of the deviation between the actual winding length (converted from the cumulative pulse count of the counting encoder) and the target winding length. The larger the deviation, the more serious the material waste or the more obvious the insufficient winding. The penalty value increases linearly with the deviation.

[0039] Material utilization penalty : , In the formula, The actual winding length converted from the cumulative pulse count of the counting encoder. The target winding length.

[0040] S202: Policy Network Initialization: Construct the policy network and value network. The policy network adopts a four-layer fully connected structure, with an input layer dimension of 256, hidden layer dimensions of 512, 256, and 128 respectively, and an output layer dimension of 2. All hidden layers use the ReLU activation function, and the output layer uses the Tanh activation function to constrain actions within legal boundaries. The value network adopts the same hidden layer structure as the policy network, with an output layer dimension of 1, used to estimate the state value function. An orthogonal initialization method is used to initialize the network weights, setting the learning rate to [missing information]. The discount factor is 0.99, the generalized advantage estimation parameter is 0.95, and the proximal pruning coefficient is 0.2.

[0041] S203: Safety Constraint Projection and Online Policy Inference: A winding force safety constraint layer is introduced after the policy network output layer. This constraint layer projects the winding tension setting value output by the policy network to a safe range of 0.5 to 5.0 Newtons through a hard clipping function, and clips the winding speed setting value to a range of 10 to 100 mm / s. Values ​​exceeding the boundary values ​​are directly taken. The experience playback buffer capacity is set to 2048 steps. A policy update is performed after collecting 2048 steps of experience, with each update iteration consisting of 10 rounds and a batch size of 64. Furthermore, in the policy update stage, the proximal clipping objective function is used to calculate the policy loss, specifically expressed as follows: , Simultaneously calculate the mean squared error loss of the value network: , The total loss function is: in, For strategic losses, The value loss coefficient, The mean squared error loss of the value network. The entropy reward coefficient, For policy entropy, As a strategy, For the state value function, for The state at any given moment, Let be the cumulative reward at time t. For strategic losses, To calculate the expectation over time step t, The probability ratio between the old and new strategies. This is the estimate of the generalized advantage. For the clipping function, =0.2 is the cutting factor.

[0042] The dominance estimation adopts the generalized dominance estimation method with parameter λ=0.95, which is obtained by weighted summation of multi-step temporal difference errors. In the online inference stage, the edge computing unit receives the latest multi-source sensor fusion vector in real time at a frequency of 50 Hz, inputs it into the adaptive winding strategy network after training convergence, and controls the forward computation time to within 5 milliseconds, outputting the winding parameter optimization vector. The winding parameter optimization vector is encapsulated into a control input packet and passed to the dynamic control generator. The dynamic control generator receives the winding parameter optimization vector and simultaneously receives the quality prediction confidence, and the two work together to complete the generation of control commands.

[0043] As a refinement of the above embodiment, step S4 aims to perform real-time prediction and anomaly identification of insulation layer thickness uniformity, bonding tightness, and edge coverage integrity based on historical winding data and the current process status, and output the quality prediction confidence level. This enables real-time assessment of winding quality and early warning of abnormal states, ensuring the reliability and traceability of insulation treatment quality.

[0044] Specifically, such as Figure 4 The following are included: S401: Parallel feature extraction network forward computation: The 256-dimensional multi-source sensing fusion vector and the 2-dimensional winding parameter optimization vector are concatenated into a 258-dimensional joint input vector, and stacked through a time window to form a 258×T-dimensional input tensor, where T is the historical time step, set to 64.

[0045] S402: Input tensors in parallel to a Bidirectional Long Short-Term Memory (BiLSTM) branch and a one-dimensional convolutional neural network branch. In the BiLSTM branch, two layers are set up, each with 128 hidden units and a dropout ratio of 0.2, outputting a 256-dimensional temporal context feature vector. In the one-dimensional convolutional neural network branch, three one-dimensional convolutional layers are set up, with kernel sizes of 7, 5, and 3 respectively, and channel numbers of 64, 128, and 256 respectively. Each layer is followed by... Following the normalization layer and ReLU activation function, a 256-dimensional local mutation feature vector is output through global max pooling. The temporal context feature vector and the local mutation feature vector are concatenated by channels to form a 512-dimensional parallel fusion feature vector. The parallel fusion feature vector is then reduced to 128-dimensional and 3-dimensional layers through two fully connected layers. Regularization with a ReLU activation and dropout ratio of 0.3 is used between the fully connected layers. The output layer uses the Sigmoid activation function to output the insulation layer thickness uniformity score, the bonding tightness score, and the edge coverage integrity score, respectively. The values ​​of the three scores are all in the range of 0 to 1.

[0046] S403: Quality Confidence Calculation and Anomaly Detection Feedback: The Softmax function is used to normalize the three-dimensional scores, calculate the relative weights of each dimension's scores, and calculate the quality prediction confidence score through weighted summation. The formula is expressed as follows: in, For quality prediction confidence level, To score the thickness uniformity, To score the tightness of the fit, Score the edge coverage integrity.

[0047] An anomaly detection threshold is set. If the quality prediction confidence level is below 0.85, the current winding state is determined to have a quality risk, triggering an audible and visual alarm and displaying an anomaly code and suggested handling measures on the touch operation panel. At the same time, the sensor data and parameter records within that time window are marked as anomalies and stored in the local database. If the quality prediction confidence level is between 0.85 and 0.95, it is determined to be in a state of attention, and the control command update frequency of S4 is increased to 100 Hz. If the quality prediction confidence level is above 0.95, it is determined to be in a normal state, and standard control parameters are maintained. Finally, the quality prediction confidence level is fed back to the reinforcement learning reward function calculation in real time, serving as the basic input for the comprehensive winding quality score in the reward function, forming a closed-loop linkage between quality prediction and strategy optimization.

[0048] As a refinement of the above embodiment, step S5 aims to convert the optimized winding parameters into specific electromechanical control commands, and send them in real time to the actuator of the intelligent auxiliary winding device via the industrial bus, thereby realizing closed-loop intelligent control of the winding process. This completes the electromechanical execution and closed-loop feedback control of the optimized winding parameters to ensure the real-time performance of the control commands and the accuracy of the coordinated action of the actuator.

[0049] like Figure 5 The specific steps shown are as follows: S501: Control Command Mapping and Dynamic Frequency Adjustment: The dynamic control generator receives the winding parameter optimization vector, parses it to obtain the winding tension setpoint and winding speed setpoint, and converts the winding tension setpoint into the PWM duty cycle (Pulse Width Modulation, PWM) control quantity of the electromagnetic pressure regulating valve through a calibrated mapping relationship. The PWM frequency is fixed at 10 kHz, the duty cycle range is 0 to 100%, corresponding to a winding tension of 0.5 to 5.0 Newtons. The winding speed setpoint is converted into the speed pulse frequency of the servo motor, with a pulse frequency range of 1000 to 10000 Hz, corresponding to a winding speed of 10 to 100 N. The speed is millimeters per second; it receives the quality prediction confidence level and dynamically adjusts the update frequency of the control command according to the confidence level. When the quality prediction confidence level is higher than 0.95, it maintains the standard update frequency of 50 Hz. When the quality prediction confidence level is between 0.85 and 0.95, it increases the update frequency to 100 Hz. When the quality prediction confidence level is lower than 0.85, it triggers emergency braking and suspends the winding operation. Then, it compares the current target number of layers with the axial displacement data collected in real time by the displacement sensor, calculates the remaining winding stroke, and generates a work completion flag and starts the deceleration and shutdown procedure when the actual number of winding layers reaches the preset number and the axial displacement is within the termination position threshold range.

[0050] S502: Bus Communication and Closed-Loop Execution: The PWM duty cycle control value, speed pulse frequency, and work completion flag are packaged into a CAN bus (Controller Area Network, CAN) message frame conforming to the ISO11898 standard. The message identifier is set to 0x210, and the data field length is 8 bytes. Bytes 0-1 represent the PWM duty cycle value, bytes 2-3 represent the speed pulse frequency value, byte 4 represents the update frequency level, byte 5 represents the work completion flag, and bytes 6-7 represent the checksum. The message frame is transmitted to the PLC controller of the intelligent auxiliary winding device via a CAN bus transceiver at a baud rate of 250 kilobits per second. The communication cycle is synchronized with the control command update frequency. The PLC controller parses the received message frame, drives the electromagnetic pressure regulating valve to adjust the winding wheel clamping force according to the PWM duty cycle control value, controls the servo motor driver to adjust the winding speed according to the speed pulse frequency, and controls the start / stop of the drive motor and the synchronous operation of the retraction motor according to the work completion flag. The PLC controller continuously collects real-time data from pressure sensors, displacement sensors, counting encoders, and positioning sensors, and transmits it back to the multi-source sensor fusion module of the edge computing unit via the CAN bus, forming a complete data closed loop from perception and decision-making to execution.

[0051] Experimental verification and analysis To verify the effectiveness of the intelligent winding control method for high-voltage self-fusion tape based on multi-source sensor fusion and deep reinforcement learning proposed in this invention, this experiment takes the insulation treatment of the 10kV side terminals of the main transformer as the research object. The experimental environment is based on the Ubuntu 20.04 operating system, using the PyTorch 2.0 deep learning framework, and model inference and training are performed on the NVIDIA Jetson Xavier NX edge computing unit. The intelligent auxiliary winding device participates in the closed-loop verification as a data acquisition device and actuator.

[0052] The experimental dataset was collected from the operation and maintenance site of Lanling County Power Supply Company of Shandong Electric Power Company. It covers the insulation winding operation process of 10kV side terminals of 50 different specifications of main transformers. The duration of each operation is 3 to 5 minutes. The acquisition frequency is 1000 Hz. After preprocessing and windowing, a total of about 12,000 valid samples were generated, of which the training set accounts for 80%, the validation set accounts for 10%, and the test set accounts for 10%.

[0053] analyze Figure 8It can be seen that the proposed method has better dynamic response characteristics and steady-state accuracy than the traditional PID method in winding tension tracking control. During the start-up phase, the proposed method reaches the target tension value within 0.5 seconds, while the PID method exhibits overshoot and oscillation of approximately 1.2 seconds. In the steady-state phase, the standard deviation of tension fluctuation for the proposed method is 0.08 Newtons, while that for the PID method is 0.23 Newtons. The proposed method improves tension control accuracy by 65.2%.

[0054] analyze Figure 9 As can be seen, the quality prediction confidence distribution of this method on the test set exhibits a clear trend of high-value concentration. 87.3% of the samples have a confidence level higher than 0.90, 9.5% have a confidence level between 0.85 and 0.90, and only 3.2% of the samples have a confidence level lower than 0.85 and are correctly identified as anomalous states. This distribution indicates that the quality assessment prediction network has good discriminative ability.

[0055] analyze Figure 10 It can be seen that the cumulative reward value of the adaptive entanglement policy network shows a stable upward trend during training, rising rapidly in the first 200 training rounds and basically converging to a stable value after 400 rounds, with a final convergence reward of 23.5, which is about 3.2 times higher than the baseline reward of the random policy. No obvious policy collapse phenomenon was observed during training, indicating that the pruning mechanism of the proximal policy optimization (PPO) algorithm effectively ensures the stability of policy updates.

[0056] Table 1 Summary and Comparison of Experimental Results This experiment quantitatively compares the proposed intelligent winding control method based on multi-source sensor fusion and deep reinforcement learning with traditional PID control and manual winding methods from multiple dimensions. The experimental results are shown in Table 1. In Table 1, the steady-state error of tension control shows that the proposed method is 0.05 N, the PID method is 0.23 N, and the manual method is 1.2 N; the quality prediction accuracy shows that the proposed method is 94.2%, the PID method cannot be scored due to the lack of a quality prediction module, and the manual method relies on visual inspection and cannot be quantified; the abnormal state recall rate shows that the proposed method is 91.7%, and neither the PID method nor the manual method has an abnormality detection mechanism; the single-piece operation time shows that the proposed method is 4.2 minutes, the PID method is 5.8 minutes, and the manual method is 8.0 minutes; the layer control error shows that the proposed method is 0.3 layers, the PID method is 1.5 layers, and the manual method is 2.0 layers.

[0057] Analysis of Table 1 shows that the proposed method achieves the expected goals in three dimensions: tension control accuracy, quality prediction performance, and operational efficiency. The steady-state error of tension control is 0.05 Newtons, meeting the process accuracy requirement of 0.1 Newtons; the quality prediction accuracy reaches 94.2%, and the abnormal state recall rate reaches 91.7%, effectively identifying winding quality risks; the single-piece operation time is shortened to 4.2 minutes, and the layer control error is only 0.3 layers. The single-piece operation time is reduced from approximately 8 minutes for manual winding to 4.2 minutes, and the layer control error is reduced from 1 to 2 layers for manual winding to 0.3 layers, improving both operational efficiency and consistency. Comprehensive experimental results demonstrate that the intelligent winding control method proposed in this invention can achieve efficient, accurate, and reliable intelligent operation in the insulation treatment scenario of the 10kV side of the main transformer.

[0058] Example 2 like Figures 6-7 As shown, an intelligent auxiliary winding device for high-voltage self-fusion insulation tape insulation treatment is presented. The device is divided into three functional layers according to the information flow direction: a multi-source sensing layer, a main control module, and an actuator layer. A main support module is located at the bottom. It includes: The multi-source sensing layer includes pressure sensors, displacement sensors, counting encoders, and positioning sensors, which are used to collect multi-dimensional physical information of the winding process; The main control module includes an edge computing unit and a PLC controller. The edge computing unit is used to perform multi-source sensor data fusion, adaptive optimization of winding parameters, and real-time prediction of winding quality. The PLC controller is used to generate and issue control commands. The actuator layer includes a winding execution module, a force adjustment module, a layer number control module, a film peeling module, and a film recycling module, which are used to execute the control commands; The multi-source sensing layer, main control module, and actuator layer are connected through a data closed loop to form a complete control loop from perception and decision-making to execution.

[0059] Figure 7 The sensor deployment details in the winding wheel area are shown. A pressure sensor is located directly below the bearing housing of the passive winding wheel's clamping wheel, directly detecting the radial clamping force of the self-fusion tape on the terminal surface. A displacement sensor is mounted on the side rail of the active winding wheel's bearing housing, monitoring the reciprocating displacement of the wheel along the terminal axis in real time. A counting encoder is mounted on the output shaft of the drive motor via a flexible coupling, recording the number of motor rotations and converting them into winding linear speed. A positioning sensor is embedded in the front end of the fixing fixture, facing the terminal end face boundary, used to identify the winding start and end positions.

[0060] The passive winding wheel assembly radially presses the self-fusion tape against the terminal surface to ensure a tight fit of the insulation layer. This passive winding wheel assembly is not directly driven by a drive motor; instead, it rotates passively due to the frictional force of the self-fusion tape conveyed by the drive wheel. A pressure sensor is installed directly below the bearing housing of the passive winding wheel assembly to detect the radial pressing force of the self-fusion tape on the terminal surface in real time. This pressing force is dynamically adjusted by an electromagnetic pressure regulating valve based on PWM control signals from the PLC. During operation, the passive winding wheel assembly and the drive winding wheel assembly clamp the terminal from both sides, forming a coordinated winding process of clamping, conveying, and pressing, ensuring that the self-fusion tape is evenly coated on the terminal surface under controlled tension and pressure.

[0061] The active winding wheel assembly is the core drive component of the winding execution module. Its function is to provide rotational power and control the winding speed for the self-fusion tape winding operation. The active wheel is directly driven to rotate by a drive motor through a mechanical transmission chain. The wheel surface maintains contact with the terminal surface, and friction drives the self-fusion tape to move continuously on the terminal surface. Simultaneously, the active wheel bearing housing is mounted on an axial slide rail, allowing it to reciprocate along the terminal axis, thereby driving the tape to complete layer-by-layer arrangement and winding stroke control. The drive motor output shaft is connected to a counting encoder via a flexible coupling. The encoder records the number of motor rotations and converts them into winding linear speed and cumulative winding length, providing speed feedback and layer calculation basis for the main control module.

[0062] Example 3 This disclosure also provides an intelligent assisted winding device for insulation treatment of high-voltage self-fusion insulating tape, including a processor and a memory. Optionally, the device may further include a communication interface and a bus. The processor, communication interface, and memory can communicate with each other via the bus. The communication interface can be used for information transmission. The processor can call logical instructions in the memory to execute the intelligent assisted winding method for insulation treatment of high-voltage self-fusion insulating tape described in the above embodiments.

[0063] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0064] As a computer-readable storage medium, the memory can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor executes the functional applications and data processing by running the program instructions / modules stored in the memory, thereby realizing the intelligent assisted winding method for high-voltage self-fusion tape insulation treatment in the above embodiments.

[0065] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory may include high-speed random access memory and may also include non-volatile memory.

[0066] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to perform the above-described intelligent assisted winding method for insulation treatment of high-voltage self-fusion tape.

[0067] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0068] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more 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 method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code. It can also be a transient storage medium.

[0069] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart assisted winding method for insulation treatment of high-voltage self-fusion insulating tape, characterized in that, Includes the following steps: Multidimensional physical information of the winding process is collected in real time by multiple types of sensors, and the multidimensional physical information is spatiotemporally aligned and feature fused to obtain a multi-source sensing fusion vector. An adaptive entanglement policy network and a value network are constructed based on deep reinforcement learning. The deep reinforcement learning algorithm adopts Markov decision and the proximal policy optimization algorithm is used to train the adaptive entanglement policy network. The multi-source sensing fusion vector is input into the trained adaptive winding strategy network, and the winding parameter optimization vector is output. A quality assessment and prediction network is constructed by paralleling a bidirectional long short-term memory network and a one-dimensional convolutional neural network. The multi-source sensor fusion vector and the winding parameter optimization vector are concatenated and input into the quality assessment and prediction network to output the quality prediction confidence. If it is lower than a preset threshold, an anomaly warning is triggered and the current state is marked as pending review. The quality prediction confidence is fed back to the reward function of deep reinforcement learning in real time as the input of the winding quality comprehensive score in the reward function. The winding parameter optimization vector and quality prediction confidence are converted into a sequence of control commands and sent to the actuator of the intelligent assisted winding device for execution.

2. The intelligent assisted winding method for high-voltage self-fusion insulation treatment according to claim 1, characterized in that, The multidimensional physical information includes: contact pressure data, axial displacement data, rotational encoding data, and spatial positioning data collected by pressure sensors, displacement sensors, counting encoders, and positioning sensors, respectively. The spatiotemporal alignment and feature fusion yield a multi-source sensing fusion vector, specifically including: The four types of sensor data collected were time-stamp aligned and dimension-normalized, and then spliced ​​together to form the original sensor matrix. A local temporal feature map is obtained by using a one-dimensional convolutional neural network to extract local temporal features from the original sensing matrix. The local feature map is input into a multi-head self-attention mechanism to calculate the correlation weight matrix between different sensing channels and perform weighted fusion to output the multi-source sensing fusion vector.

3. The intelligent assisted winding method for insulation treatment of high-voltage self-fusion insulating tape according to claim 1, characterized in that, The Markov decision uses a multi-source sensor fusion vector as the environmental state input; the action space is a two-dimensional continuous vector composed of the winding tension setting value and the winding speed setting value. The reward function is a weighted combination of insulation quality reward, edge coverage integrity reward, and material utilization penalty: ， in, Rewards for the current moment. As a reward for the quality of the insulation layer, For edge coverage integrity rewards, Penalty for low material utilization; The insulation layer quality bonus is calculated based on the time-series statistical characteristics of the contact pressure data within the current time window. It takes the weighted sum of the standard deviation of the contact pressure and the deviation of the standard deviation of the contact pressure from the target winding tension, and maps it to the 0 to 1 interval through a Gaussian kernel function. The edge coverage integrity reward is calculated based on whether the positioning sensor detects the terminal end face boundary and whether the number of winding layers measured by the displacement sensor at the boundary meets the preset threshold. When the positioning sensor recognizes the end face and the number of layers meets the standard, a positive reward is given; otherwise, a negative penalty is given. The material utilization penalty is calculated based on the absolute value of the deviation between the actual winding length (converted from the cumulative pulse count of the counting encoder) and the target winding length.

4. The intelligent assisted winding method for insulation treatment of high-voltage self-fusion tape according to claim 3, characterized in that, Both the adaptive winding strategy network and the value network adopt a four-layer fully connected structure. The hidden layers of the adaptive winding strategy network all use the ReLU activation function, and the output layer uses the Tanh activation function. The winding force safety constraint layer is introduced after the output layer of the adaptive winding strategy network. The winding force safety constraint layer projects the winding tension setting value output by the adaptive winding strategy network to a safe range of 0.5 to 5.0 Newtons through a hard clipping function, and clips the winding speed setting value to a range of 10 to 100 millimeters per second. If the value exceeds the boundary value, the boundary value is directly taken. The adaptive winding strategy network total loss as follows: ， ， ， in, For strategic losses, The value loss coefficient, The mean squared error loss of the value network. The entropy reward coefficient, For policy entropy, As a strategy, For the state value function, for The state at any given moment, Let be the cumulative reward at time t. For strategic losses, To calculate the expectation over time step t, The probability ratio between the old and new strategies. This is the estimate of the generalized advantage. For the clipping function, This is the cutting factor.

5. The intelligent assisted winding method for insulation treatment of high-voltage self-fusion tape according to claim 1, characterized in that, The specific steps for the quality assessment prediction network to output the quality prediction confidence level include: After concatenating the multi-source sensing fusion vector and the winding parameter optimization vector, the concatenation vector is input in parallel to the bidirectional long short-term memory network branch and the one-dimensional convolutional neural network branch. The temporal context feature vector output by the bidirectional long short-term memory network branch is concatenated with the local mutation feature vector output by the one-dimensional convolutional neural network branch to obtain a parallel fused feature vector. The parallel fused feature vector is mapped to insulation layer thickness uniformity score, bonding tightness score and edge coverage integrity score through a fully connected layer. The quality prediction confidence score is obtained by weighted summation of the thickness uniformity score, bonding tightness score and edge coverage integrity score.

6. The intelligent assisted winding method for insulation treatment of high-voltage self-fusion tape according to claim 5, characterized in that, The confidence level for quality prediction is calculated as follows: The Softmax function was used to normalize the scores for insulation layer thickness uniformity, bonding tightness, and edge coverage integrity. Calculate the relative weights of the scores for each dimension, and then calculate the confidence level of the quality prediction by weighted summation: ， in, For quality prediction confidence level, To score the thickness uniformity, To score the tightness of the fit, Score the edge coverage integrity. An anomaly detection threshold is set. If the quality prediction confidence level is below 0.85, the current winding state is determined to have a quality risk, triggering an audible and visual alarm and displaying an anomaly code and suggested handling measures on the touch operation panel. At the same time, the sensor data and parameter records within that time window are marked as an anomaly sample and stored in the local database. If the quality prediction confidence level is between 0.85 and 0.95, it is determined to be in a state of alert, and the control command update frequency is increased to 100 Hz. If the quality prediction confidence level is above 0.95, it is determined to be in a normal state, and standard control parameters are maintained.

7. The intelligent assisted winding method for insulation treatment of high-voltage self-fusion tape according to claim 1, characterized in that, The step of converting the winding parameter optimization vector and quality prediction confidence into a control command sequence specifically includes: The winding tension setting value in the winding parameter optimization vector is converted into the PWM duty cycle control value of the electromagnetic pressure regulating valve by the dynamic control generator, and the winding speed setting value is converted into the speed pulse frequency of the servo motor. The update frequency of control commands is dynamically adjusted based on the quality prediction confidence level. The control command sequence is sent to the PLC controller of the intelligent auxiliary winding device via the CAN bus, and the PLC controller drives the actuator to move.

8. An intelligent auxiliary winding device for insulation treatment of high-voltage self-fusion insulating tape, used to implement the method as described in any one of claims 1 to 7, characterized in that, include: The multi-source sensing layer includes pressure sensors, displacement sensors, counting encoders, and positioning sensors, which are used to collect multi-dimensional physical information of the winding process; The main control module includes an edge computing unit and a PLC controller. The edge computing unit is used to perform multi-source sensor data fusion, adaptive optimization of winding parameters, and real-time prediction of winding quality. The PLC controller is used to generate and issue control commands. The actuator layer includes a winding execution module, a force adjustment module, a layer number control module, a film peeling module, and a film recycling module, which are used to execute the control commands; The multi-source sensing layer, main control module, and actuator layer are connected through a data closed loop to form a complete control loop from perception and decision-making to execution.

9. The intelligent auxiliary winding device for high-voltage self-fusion tape insulation treatment according to claim 8, characterized in that, The pressure sensor is installed directly below the bearing seat of the clamping wheel of the passive winding wheel assembly to detect the radial clamping force of the self-fusion tape on the surface of the terminal block; the displacement sensor is installed on the side slide rail of the bearing seat of the active winding wheel assembly to monitor the reciprocating displacement of the wheel assembly along the axial direction of the terminal block; the counting encoder is installed on the output shaft of the drive motor to record the number of motor rotations; and the positioning sensor is embedded in the front end of the fixing fixture to identify the winding start and end positions.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the intelligent assisted winding method for insulation treatment of high-voltage self-fusion tape as described in any one of claims 1-7.

Images