A method and system for detecting an insulation fault of a transformer winding under voltage

Abstract

This invention discloses a method and system for live detection of transformer winding insulation faults. The method includes: acquiring transformer acoustic and vibration signals and constructing signal task segments, then inputting these segments into a pre-trained model to obtain a predicted computational complexity level and resource preference index; real-time monitoring of the underlying hardware resource status to calculate the system load level index and comparing it with a dynamic load window upper limit threshold; responding to system overload, comprehensively calculating segment priorities and allocating them to matching processing channels via a waiting queue; dynamically configuring the calling logic of feature extraction operators based on channel computing power limitations to extract deep features and obtain a joint feature vector; inputting the joint feature vector into a fault classification model with multiple inference exits, and triggering the corresponding exit to output diagnostic results based on computing power limitations. This invention achieves hardware awareness and adaptive scheduling of computing resources in an edge computing environment, balancing real-time detection and high accuracy under high load impacts, and effectively avoiding missed detection of critical faults.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, and in particular to a method and system for live detection of transformer winding insulation faults. Background Technology

[0002] With continuous technological advancements, acoustic and vibration signal analysis technology has gradually emerged, opening up new avenues for transformer fault detection. During operation, transformers generate abundant acoustic and vibration signals due to factors such as electromagnetic forces and mechanical vibrations. In-depth analysis of these signals allows for a more direct and comprehensive understanding of the winding insulation condition, thereby enabling effective fault diagnosis.

[0003] However, existing detection methods based on acoustic and vibration signals still face a trade-off between accuracy and computational power in practical applications. In the signal processing stage, to pursue higher detection accuracy, the technological trend is to deeply fuse multimodal acoustic and vibration signals and introduce complex deep learning models for feature extraction and pattern recognition. While this approach improves the accuracy of fault identification, it also leads to an exponential increase in computational complexity. For edge computing devices, typically deployed on the side of transformers with relatively limited computing and storage resources, running such highly complex algorithms in real time faces significant performance bottlenecks.

[0004] Furthermore, traditional signal processing workflows typically employ a static, one-size-fits-all approach. This means that regardless of whether the acquired signal is simple background noise or a complex, suspected fault waveform, the system allocates fixed computing resources and executes a standardized, end-to-end processing procedure. This inefficient resource allocation method has significant drawbacks. When the system is under low load, high-performance computing units may remain idle, resulting in wasted computing power. Conversely, when a sudden transformer failure (such as a short-circuit impact) causes massive concurrent data processing, the system cannot prioritize tasks, leading to instantaneous overload of the computing units.

[0005] When faced with extremely high load surges, existing scheduling mechanisms often employ simple first-in-first-out (FIFO) strategies or random packet loss strategies. This means that high-value signals containing critical fault characteristics are highly likely to be blocked outside the processing queue by a large amount of meaningless noise signals, or incorrectly discarded due to buffer overflow. This phenomenon of critical characteristics being overwhelmed by noise severely affects the real-time performance and reliability of transformer energization detection, failing to meet the urgent needs of power systems for rapid response to major faults.

[0006] Therefore, how to achieve high-precision fusion detection of acoustic and vibration signals and efficient adaptive scheduling of computing resources under the conditions of limited edge computing resources and dynamic load fluctuations is a technical problem that urgently needs to be solved. Summary of the Invention

[0007] To address the shortcomings of existing transformer winding insulation fault detection technologies, such as the lack of a hardware-aware adaptive computing resource scheduling mechanism leading to processing delays, key fault features being overwhelmed by noise queues, or packet loss when edge devices face sudden high loads, and the inability of traditional fixed-architecture deep learning models to balance detection accuracy and real-time performance under computing power constraints, this invention provides a method and system for live-line detection of transformer winding insulation faults. The method includes: acquiring transformer acoustic and vibration signals to construct signal task slices, inputting them into a pre-trained model to obtain a predicted computational complexity level and resource preference index; real-time monitoring of the underlying hardware resource status to calculate the system load level index and comparing it with a dynamic load window upper limit threshold; responding to system overload, comprehensively calculating slice priorities and allocating them to matching processing channels via a waiting queue; dynamically configuring the feature extraction operator's calling logic based on channel computing power limitations to extract deep features and obtain a joint feature vector; inputting the joint feature vector into a fault classification model with multiple inference exits, triggering corresponding exits to output diagnostic results based on computing power limitations. This invention achieves hardware awareness and adaptive scheduling of computing resources in an edge computing environment, balancing real-time detection and high accuracy under high load impacts, and effectively avoiding missed detections of key faults.

[0008] The present invention adopts the following technical solution.

[0009] In a first aspect, the present invention provides a method for live detection of transformer winding insulation faults, comprising: Acoustic and vibration sensor signals are acquired during transformer operation, initial feature extraction is performed, and signal task segments are constructed. Based on the signal task segmentation, a pre-trained resource demand prediction model is used to obtain a resource demand descriptor that includes the prediction computational complexity level and the sensor resource preference index. Real-time monitoring and acquisition of hardware resource status, and calculation of the current system load level index; The system load level index is compared with a preset upper limit threshold; in response to the system load level index exceeding the upper limit threshold, the priority value of the signal task slice is calculated by combining the prediction computation complexity level and the sensor resource preference index, and the signal task slice is dynamically allocated to the matching processing channel through the task slice waiting queue. For the signal task slices assigned to the matching processing channel, the calling logic of the feature extraction operator is dynamically configured according to the computing power limit of the matching processing channel, and deep feature extraction and fusion are performed to obtain a joint feature vector; Based on the joint feature vector, a fault classification model with multiple inference exits is used, and the corresponding inference exit is triggered according to the computing power limit of the processing channel to obtain the insulation fault mode identification result of the transformer winding.

[0010] Preferably, initial feature extraction is performed and constructed into signal task slices, and a resource demand descriptor is obtained using a pre-trained resource demand prediction model, including: A fixed-length sliding window is used to extract time-domain statistical features, frequency-domain energy features, and time-frequency domain joint features from the synchronously sampled acoustic sensor signals and vibration sensor signals, respectively. The extracted time-domain statistical features, frequency-domain energy features, and time-frequency domain joint features are standardized and concatenated into a fixed-dimensional multidimensional feature vector. The multidimensional feature vector is then encapsulated and bound to the original synchronous sampling signal segments of the acoustic sensor signal and the vibration sensor signal as the signal task slice. The multidimensional feature vectors from the signal task slices are input into the pre-trained resource demand prediction model, which is a multilayer perceptron model. The output layer of the multilayer perceptron model includes two branches: Computational complexity prediction branch: The Softmax function is used to output the probability that the signal task slice belongs to multiple discrete computational complexity levels, and the level with the highest probability is taken as the predicted computational complexity level; Resource preference index prediction branch: The Sigmoid function is used to output a multi-dimensional vector, which is used to quantify the dependence of the signal task slice on the acoustic sensor data channel, vibration sensor data channel and computing unit, respectively, as the sensor resource preference index.

[0011] Preferably, the system monitors and acquires hardware resource status in real time, and calculates the current system load level index, including: The system monitors and acquires the occupancy rate of the acoustic sensor data channel, the occupancy rate of the vibration sensor data channel, and the average utilization rate of the computing unit in real time. The occupancy rate of the sensor data channel is determined by acquiring the real-time data storage volume of the first-in-first-out buffer in the data acquisition card and calculating the ratio of the real-time data storage volume to the total capacity of the first-in-first-out buffer. Based on preset weighting coefficients, the occupancy rates of the acoustic sensor data channels, the vibration sensor data channels, and the average utilization rate of the computing unit are weighted and summed to obtain a weighted load sum. Based on preset no-load and heavy-load benchmark values, the weighted load is linearly transformed and truncated to obtain the system load level index constrained within a preset range.

[0012] Preferably, the processing channel includes a high-precision parallel processing channel, a standard processing channel, and a resource-optimized channel; for the signal task slices allocated to the processing channel, the calling logic of the feature extraction operator is dynamically configured according to the computing power limit of the processing channel, and deep feature extraction and fusion are performed, including: A pre-built feature extraction operator library is provided, which includes lightweight operators and heavyweight operators; When the signal task slice is allocated to the resource optimization channel, a logical masking instruction is sent to the preset feature extraction operator library to modify the control state of the feature extraction mask to block the triggering link of the heavyweight operator and call the lightweight operator to extract the minimum feature set to participate in feature fusion. In response to the dynamic adjustment of the mapping rules in the resource matching matrix, when the signal task slice is allocated to the resource optimization channel and the prediction computation complexity level is high, an exception interruption instruction is generated; the control state of the feature extraction mask is modified according to the exception interruption instruction, the masking of the heavyweight operator is removed, and the independent computing resources reserved by the system are called to execute the heavyweight operator to extract the complete feature set for feature fusion.

[0013] Preferably, the sensor resource preference index includes an acoustic channel preference index, a vibration channel preference index, and a computing unit preference index; The priority value of the signal task slice is calculated by combining the predicted computational complexity level and the sensor resource preference index, and the signal task slice is dynamically allocated to the matching processing channel via the task slice waiting queue, including: Convert the predicted computational complexity level into a numerical score; The acoustic channel preference index, the vibration channel preference index, and the computing unit preference index are weighted and summed using preset local weighting coefficients to obtain a scalar preference score. The priority value is obtained by weighting the numerical score and the scalar preference score using a preset global weight coefficient. The task fragment waiting queue is a multi-level hash table structure. Storing the signal task fragments into the task fragment waiting queue includes: Using the predicted computational complexity level as the first-level key, the signal task fragments are mapped to the first-level hash buckets corresponding to the first-level hash table; The value range of the sensor resource preference index is discretized into multiple level intervals. The combination of the level intervals is used as the second-level key. A secondary mapping is performed in the first-level hash bucket, and the signal task is stored in the corresponding sub-hash bucket.

[0014] Preferably, the upper limit threshold and the lower limit threshold corresponding to the upper limit threshold constitute a dynamic load window; the preset and update process of the upper limit threshold includes: Obtain time-series data of the system load level index under historical workloads, and extract the high percentile of the time-series data as the initial upper limit threshold; The upper limit threshold is adaptively updated using an exponentially weighted moving average algorithm. Extract the high-segment statistical mean of the system load level index within a preset time window as a reference benchmark; The updated upper limit threshold is obtained by weighting and summing the current upper limit threshold with the reference benchmark using preset weighting coefficients. And maintain a preset hysteresis interval between the upper limit threshold and the lower limit threshold.

[0015] Preferably, the pre-trained fault classification model is a hierarchical neural network model containing multiple network layers connected in sequence; the hierarchical neural network model has an emergency inference exit in the Nth network layer and a final inference exit in the last network layer; The joint feature vector is input into a fault classification model with multiple inference exits, and the corresponding inference exit is triggered according to the computing power limitation of the processing channel to obtain the insulation fault mode identification result of the transformer winding, including: When the signal task slice is allocated to the resource optimization channel, the joint feature vector is input into the hierarchical neural network model. After the calculation of the first N network layers is completed, the subsequent calculation is terminated, and the insulation fault mode identification result of the transformer winding is output by the emergency inference output. When the signal task slice is assigned to the high-precision parallel processing channel or the standard processing channel, the joint feature vector is input into the hierarchical neural network model and all the network layers are traversed and calculated. The insulation fault mode identification result of the transformer winding is output by the final inference output. The hierarchical neural network model includes a feature alignment layer between the Nth and N+1th network layers. The hierarchical neural network model is trained based on a joint loss function, which includes a first error calculated based on the prediction result of the emergency inference exit and a second error calculated based on the prediction result of the final inference exit.

[0016] Secondly, the present invention provides a live-line detection system for transformer winding insulation faults, comprising: The multimodal signal segmentation construction module is used to acquire acoustic sensor signals and vibration sensor signals during transformer operation, perform initial feature extraction, and construct signal task segments. The resource demand descriptor generation module is used to obtain a resource demand descriptor containing the prediction computational complexity level and the sensor resource preference index based on the signal task segmentation and a pre-trained resource demand prediction model. The system load status assessment module is used to monitor and obtain the hardware resource status in real time and calculate the current system load level index. The resource-coupled task sharding adaptive scheduling module is used to compare the system load level index with a preset upper limit threshold; in response to the system load level index exceeding the upper limit threshold, it calculates the priority value of the signal task shard by combining the predicted computational complexity level and the sensor resource preference index, and dynamically allocates the signal task shard to the matching processing channel through the task sharding waiting queue. The operator-level feature extraction path dynamic reconstruction module is used to slice the signal task allocated to the matching processing channel, dynamically configure the calling logic of the feature extraction operator according to the computing power limit of the matching processing channel, perform deep feature extraction and fusion, and obtain a joint feature vector. The multi-exit adaptive inference decision module is used to obtain the insulation fault mode identification result of the transformer winding by using a fault classification model with multiple inference exits based on the joint feature vector and triggering the corresponding inference exit according to the computing power limit of the processing channel.

[0017] Thirdly, the present invention provides a terminal, including a processor and a storage medium; the storage medium is used to store instructions. The processor is configured to operate according to the instructions to perform the steps according to the method.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of any of the methods described herein.

[0019] The beneficial effects of this invention are as follows: 1. Improved real-time performance and resource utilization of live-line detection: This invention introduces a resource demand prediction mechanism based on neural networks, enabling the prediction of the dependence and computational complexity of signal task slices on acoustic channels, vibration channels, and computing units before task processing. This allows the system to intelligently match resource demand descriptors rather than blindly processing sequentially. When facing system overload, the control state of the feature extraction mask is dynamically modified based on the computing power limitations of the processing channels to block heavyweight operators and trigger an emergency inference exit in the fault classification model. This dynamic truncation mechanism, combining hardware and software, eliminates resource contention bottlenecks, effectively avoiding overall detection delays caused by single resource overload, and ensuring the real-time requirements of live-line detection of transformer winding insulation faults.

[0020] 2. Adaptive Dynamic Balancing of System Load: This invention constructs an adaptive scheduling strategy based on a dynamic load window. By calculating a system load level index that includes data channel occupancy and computing unit utilization in real time, the system can flexibly switch between direct routing processing and buffered scheduling processing modes. During low load, direct processing is used to reduce latency, while during high load, a priority-based waiting queue is activated to ensure throughput and stability. The waiting queue employs a multi-level hash table data structure, using a combination of predicted computational complexity levels and sensor resource preference index levels as hash keys for two-dimensional retrieval, achieving accurate priority matching with low time complexity. This mechanism makes the system robust, capable of coping with sudden data surges or hardware performance fluctuations.

[0021] 3. Improved accuracy and reliability of fault diagnosis: This invention employs a feature-level fusion strategy of acoustic and vibration signals, combining time-domain, frequency-domain, and multi-dimensional feature extraction to overcome the limitations of monitoring a single physical quantity, enabling a more comprehensive capture of subtle signs of winding insulation faults. Simultaneously, the fault classification model of this invention adopts a hierarchical multi-intermediate-exit architecture, with a pre-set feature alignment network layer at the junction of shallow and deep networks. It uses a joint loss function containing both first and second errors for dual-objective iterative training. This ensures the consistency of feature logic across different inference depths for the joint feature vector, avoiding diagnostic result drift caused by dynamic switching of inference depth due to load fluctuations, and providing a high-quality data foundation for subsequent fault mode classification.

[0022] 4. Possesses continuous self-learning capabilities: This invention designs a dual-loop feedback optimization closed-loop mechanism combining short and long cycles. By recording the actual resource consumption of signal task slices in actual processing and calculating the deviation from the predicted values, the model parameters of the resource demand prediction neural network are updated periodically in the short cycle; while in the long cycle, the asynchronously generated physical inspection true values ​​are anchored with the previous diagnostic results to reconstruct the fault classification model. This enables the system to continuously adapt to changing environmental noise characteristics, hardware aging, or new fault characteristics over time, maintaining high accuracy in resource demand prediction and achieving full lifecycle performance optimization of the detection system.

[0023] 5. Optimized multi-sensor concurrent processing efficiency: By distinguishing between acoustic sensor data channel preferences and vibration sensor data channel preferences, this invention enables fine-grained resource isolation and parallel scheduling. When a certain type of channel, such as the vibration channel, is busy, the system can prioritize scheduling tasks that have a higher preference for another type of channel, such as the acoustic channel, maximizing the concurrent utilization of the multi-channel hardware interface and significantly improving the signal processing throughput per unit time. Attached Figure Description

[0024] Figure 1This is an overall flowchart of a method for live detection of transformer winding insulation faults provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0026] Please see Figure 1 This invention provides a method for live detection of transformer winding insulation faults, comprising: S1: Acquire acoustic and vibration sensor signals during transformer operation, perform initial feature extraction, and construct signal task segments.

[0027] S2: Based on the signal task segmentation, a pre-trained resource demand prediction model is used to obtain a resource demand descriptor that includes the prediction computational complexity level and the sensor resource preference index.

[0028] In a preferred embodiment of the present invention, initial feature extraction and construction into signal task slices, and obtaining a resource demand descriptor using a pre-trained resource demand prediction model, include: A fixed-length sliding window is used to extract time-domain statistical features, frequency-domain energy features, and time-frequency domain joint features from the synchronously sampled acoustic sensor signals and vibration sensor signals, respectively. The extracted time-domain statistical features, frequency-domain energy features, and time-frequency domain joint features are standardized and concatenated into a fixed-dimensional multidimensional feature vector. The multidimensional feature vector is then encapsulated and bound to the original synchronous sampling signal segments of the acoustic sensor signal and the vibration sensor signal as the signal task slice. The multidimensional feature vectors from the signal task slices are input into the pre-trained resource demand prediction model, which is a multilayer perceptron model. The output layer of the multilayer perceptron model includes two branches: Computational complexity prediction branch: The Softmax function is used to output the probability that the signal task slice belongs to multiple discrete computational complexity levels, and the level with the highest probability is taken as the predicted computational complexity level; Resource preference index prediction branch: The Sigmoid function is used to output a multi-dimensional vector, which is used to quantify the dependence of the signal task slice on the acoustic sensor data channel, vibration sensor data channel and computing unit, respectively, as the sensor resource preference index.

[0029] Optionally, the resource demand prediction model is pre-trained based on historical signal task slices and their corresponding real resource consumption data. The real resource consumption data includes computational unit resource consumption information and resource occupancy information for acoustic sensor data channels and vibration sensor data channels. The system constructs prediction computational complexity level labels and sensor resource preference index labels based on historical resource consumption data. The prediction computational complexity level labels are determined by the comprehensive resource consumption level during historical task processing, and the sensor resource preference index labels are obtained by normalizing the corresponding resource occupancy information. In this way, the resource demand prediction model can learn the mapping relationship between signal features and resource demands and output a resource demand descriptor.

[0030] It should be noted that the above-described scheme of using a multilayer perceptron model combined with the Softmax and Sigmoid functions for resource demand prediction is the preferred path designed in this invention to balance computational efficiency and prediction accuracy. In practical applications, those skilled in the art can also use machine learning or deep learning algorithms such as support vector machines, random forests, or lightweight one-dimensional convolutional neural networks to construct resource demand prediction models; at the same time, for the output function of the prediction branch, other normalized mapping functions can be used depending on the model architecture. These equivalent substitutions based on different algorithm architectures can also achieve the technical objective of extracting the computational complexity level and resource preference index from signal task slices.

[0031] S3: Monitor and obtain hardware resource status in real time, and calculate the current system load level index.

[0032] In a preferred embodiment of the present invention, real-time monitoring and acquisition of hardware resource status, and calculation of the current system load level index, include: The system monitors and acquires the occupancy rate of the acoustic sensor data channel, the occupancy rate of the vibration sensor data channel, and the average utilization rate of the computing unit in real time. The occupancy rate of the sensor data channel is determined by acquiring the real-time data storage volume of the first-in-first-out buffer in the data acquisition card and calculating the ratio of the real-time data storage volume to the total capacity of the first-in-first-out buffer. Based on preset weighting coefficients, the occupancy rates of the acoustic sensor data channels, the vibration sensor data channels, and the average utilization rate of the computing unit are weighted and summed to obtain a weighted load sum. Based on preset no-load and heavy-load benchmark values, the weighted load is linearly transformed and truncated to obtain the system load level index constrained within a preset range.

[0033] S4: Compare the system load level index with a preset upper limit threshold; when the system load level index does not exceed the upper limit threshold, directly route the signal task slice to the matching processing channel based on the predicted computational complexity level; in response to the system load level index exceeding the upper limit threshold, calculate the priority value of the signal task slice by combining the predicted computational complexity level and the sensor resource preference index, and dynamically allocate the signal task slice to the matching processing channel through the task slice waiting queue.

[0034] In a preferred embodiment of the present invention, the processing channel includes a high-precision parallel processing channel, a standard processing channel, and a resource-optimized channel; for the signal task slices allocated to the processing channel, the calling logic of the feature extraction operator is dynamically configured according to the computing power limit of the processing channel, and deep feature extraction and fusion are performed, including: A pre-built feature extraction operator library is provided, which includes lightweight operators and heavyweight operators; When the signal task slice is allocated to the resource optimization channel, a logical masking instruction is sent to the preset feature extraction operator library to modify the control state of the feature extraction mask to block the triggering link of the heavyweight operator and call the lightweight operator to extract the minimum feature set to participate in feature fusion. In response to the dynamic adjustment of the mapping rules in the resource matching matrix, when the signal task slice is allocated to the resource optimization channel and the prediction computation complexity level is high, an exception interruption instruction is generated; the control state of the feature extraction mask is modified according to the exception interruption instruction, the masking of the heavyweight operator is removed, and the independent computing resources reserved by the system are called to execute the heavyweight operator to extract the complete feature set for feature fusion.

[0035] Furthermore, the exception interrupt instruction is used to provide controlled feature extraction enhancement capabilities for high-complexity signal task slices based on the lightweight processing mechanism of the resource optimization channel. When the system is under high load, signal task slices may be allocated to the resource optimization channel for processing due to the dynamic adjustment of the mapping rules in the resource matching matrix. For signal task slices with high prediction computational complexity, the system generates the exception interrupt instruction to temporarily remove the call restrictions on heavyweight operators and call the independent computing resources reserved by the system to execute the heavyweight operators, so as to ensure the complete feature extraction capability of signals corresponding to difficult faults or potentially serious faults.

[0036] In a preferred embodiment of the present invention, the sensor resource preference index includes an acoustic channel preference index, a vibration channel preference index, and a computing unit preference index; The priority value of the signal task slice is calculated by combining the predicted computational complexity level and the sensor resource preference index, and the signal task slice is dynamically allocated to the matching processing channel via the task slice waiting queue, including: Convert the predicted computational complexity level into a numerical score; The acoustic channel preference index, the vibration channel preference index, and the computing unit preference index are weighted and summed using preset local weighting coefficients to obtain a scalar preference score. The priority value is obtained by weighting the numerical score and the scalar preference score using a preset global weight coefficient. The task fragment waiting queue is a multi-level hash table structure. Storing the signal task fragments into the task fragment waiting queue includes: Using the predicted computational complexity level as the first-level key, the signal task fragments are mapped to the first-level hash buckets corresponding to the first-level hash table; The value range of the sensor resource preference index is discretized into multiple level intervals. The combination of the level intervals is used as the second-level key. A secondary mapping is performed in the first-level hash bucket, and the signal task fragments are stored in the corresponding sub-hash buckets. The sub-hash buckets are implemented at the underlying level as a max-heap structure based on the priority value, and the signal task fragments are dequeued in order of priority value from high to low.

[0037] It should be noted that the above-described scheme, which uses a multi-level hash table structure combined with a combination of computational complexity level and discretization level interval keys for two-dimensional retrieval, is the preferred path for the present invention designed for the rapid scheduling and allocation of massive task fragments. In practical applications, those skilled in the art can also use data structures such as max-heap priority queues, red-black trees, or multi-dimensional skip lists to construct and maintain waiting queues; simultaneously, for the dynamic priority sorting mechanism, an aging compensation sorting algorithm based on attenuation coefficients can be introduced in conjunction with task waiting time. These equivalent substitutions based on different data structures and sorting logic can also achieve the technical objective of rapid addressing and dequeueing allocation of critical signal task fragments under high load conditions.

[0038] In a preferred embodiment of the present invention, the upper limit threshold and the lower limit threshold corresponding to the upper limit threshold constitute a dynamic load window; the preset and update process of the upper limit threshold includes: Obtain time-series data of the system load level index under historical workloads, and extract the high percentile of the time-series data as the initial upper limit threshold; The upper limit threshold is adaptively updated using an exponentially weighted moving average algorithm. Extract the high-segment statistical mean of the system load level index within a preset time window as a reference benchmark; The updated upper limit threshold is obtained by weighting and summing the current upper limit threshold with the reference benchmark using preset weighting coefficients. And maintain a preset hysteresis interval between the upper limit threshold and the lower limit threshold.

[0039] It should be noted that the above-described scheme, which uses an exponentially weighted moving average algorithm combined with a hysteresis interval for dynamic adaptive updating of the upper limit threshold of the load window, is the preferred path for the design of smooth system load transition and anti-oscillation control in this invention. In practical applications, those skilled in the art can also use Kalman filtering algorithms, autoregressive moving average models, or simple moving average algorithms based on fixed time windows to dynamically track and update the threshold; at the same time, the setting of the hysteresis interval can also be dynamically adjusted based on real-time ambient temperature or hardware heat dissipation status. These equivalent substitutions based on different time series analysis theories can also achieve the technical objective of smoothly switching the system scheduling mode under high and low load conditions.

[0040] S5: For the signal task slices allocated to the matching processing channel, dynamically configure the calling logic of the feature extraction operator according to the computing power limit of the matching processing channel, perform deep feature extraction and fusion, and obtain a joint feature vector.

[0041] S6: Based on the joint feature vector, a fault classification model with multiple inference exits is used, and the corresponding inference exit is triggered according to the computing power limit of the processing channel to change the inference path of the signal task segment and obtain the insulation fault mode identification result of the transformer winding.

[0042] In a preferred embodiment of the present invention, the pre-trained fault classification model is a hierarchical neural network model comprising multiple network layers connected in sequence; the hierarchical neural network model has an emergency inference exit in the Nth network layer and a final inference exit in the last network layer; In a preferred embodiment of the present invention, the joint feature vector is input into a fault classification model with multiple inference exits, and the corresponding inference exit is triggered according to the computing power limitation of the processing channel to obtain the insulation fault mode identification result of the transformer winding, including: When the signal task slice is allocated to the resource optimization channel, the joint feature vector is input into the hierarchical neural network model. After the calculation of the first N network layers is completed, the subsequent calculation is terminated, and the insulation fault mode identification result of the transformer winding is output by the emergency inference output. When the signal task slice is assigned to the high-precision parallel processing channel or the standard processing channel, the joint feature vector is input into the hierarchical neural network model and all the network layers are traversed and calculated. The insulation fault mode identification result of the transformer winding is output by the final inference output. The hierarchical neural network model includes a feature alignment layer between the Nth and N+1th network layers. The hierarchical neural network model is trained based on a joint loss function, which includes a first error calculated based on the prediction result of the emergency inference exit and a second error calculated based on the prediction result of the final inference exit.

[0043] It should be noted that the above-mentioned scheme for fault mode recognition, which employs a hierarchical neural network model with multiple intermediate inference exits combined with a feature alignment network layer and a joint loss function, is the preferred path for the design of dynamically constrained edge computing power and adaptive adjustment of diagnostic accuracy in this invention. In practical applications, those skilled in the art can also use architectures such as cascaded random forest models, multi-expert gated networks, or dynamic routing networks based on knowledge distillation for hierarchical inference; at the same time, for joint optimization of errors, weighted cross-entropy or focal loss functions can be used for differentiated alignment according to the business scenario. These equivalent substitutions based on different hierarchical inference architectures and loss evaluation dimensions can also achieve the technical objective of dynamically truncating the inference depth according to computing power limitations and outputting reliable diagnostic results.

[0044] Example 1: In specific implementation, the initial step of the method is to extract initial features from the signal and construct a multi-dimensional feature vector containing acoustic and vibration signal features as signal task slices. Each signal task slice corresponds to a strictly fixed-length sliding window of transformer operating data, for example, specifically set to 50ms. This fixed length is predetermined within an empirical range of 20ms to 100ms. The selection and configuration of this window length is to cover at least one complete power frequency cycle and the accompanying transient fault decay process.

[0045] Specifically, the transformer's operating signals are acquired first. Acoustic sensors are deployed on the transformer housing to collect sound signals generated during transformer operation, while vibration sensors collect mechanical vibration signals from the transformer tank surface. In one implementation, the two signals are pre-amplified and anti-aliasing filtered before being synchronously sampled at the microsecond level using a shared clock source or hardware triggering mechanism.

[0046] Subsequently, time-domain statistical features, frequency-domain energy features, and joint time-frequency features were extracted from the acoustic sensor signals and vibration sensor signals deployed on the transformer casing, respectively: Extraction of time-domain statistical features: This directly applies to the original sampled sequence, calculating parameters including the signal's mean, variance, root mean square (RMS), peak-to-peak value, skewness, kurtosis, and envelope waveform features based on the Hilbert transform. These parameters reflect the signal's distribution characteristics and fluctuations over time. Skewness measures the asymmetry of the signal waveform, while kurtosis is highly sensitive to early-stage impulse faults, such as pulses generated by partial discharge or mechanical loosening, reflecting the degree to which the signal deviates from a normal distribution.

[0047] Frequency domain energy feature extraction: Frequency domain transformation of the signal is required. A Fast Fourier Transform (FFT) algorithm is used to convert the time-domain signal into a frequency-domain representation. Subsequently, the energy values ​​within several pre-defined characteristic frequency bands are calculated. The division of these characteristic frequency bands is based on the inherent frequency characteristics of transformer winding vibration and the main frequency bands of electromagnetic noise. Specifically, the defined characteristic frequency bands include a fundamental frequency band (100Hz ± 10Hz) and a high-frequency band (1kHz-3kHz). The fundamental frequency band is used to capture the basic vibration characteristics of the winding caused by power frequency electromagnetic force, while the high-frequency band corresponds to the high-frequency resonance frequency excited by the rapidly increasing electrodynamic force when faults such as inter-turn short circuits occur in the winding. This band can also effectively cover high-frequency acoustic emission signals generated by insulation defects such as partial discharge.

[0048] Extraction of joint time-frequency features: The continuous wavelet transform method is adopted. Since the waveform of the Morlet wavelet is highly similar to the damped oscillation waveform of transformer partial discharge or mechanical transient impact, it can match the fault characteristics to the greatest extent. The Morlet wavelet is selected as the mother wavelet. By adjusting the scaling and translation parameters, the energy distribution spectrum of the signal in the joint time and frequency dimensions is obtained. The mean, variance, and amplitude corresponding to the maximum energy ridge of each characteristic frequency band are extracted from the spectrum as components of the feature vector.

[0049] In practice, after the initial feature extraction is completed, the acoustic signal features and vibration signal features need to be standardized. The purpose of standardization is to eliminate the influence of differences in the dimensions and numerical ranges of different feature parameters. For example, the difference between the dB value of sound pressure level and the g value of vibration acceleration is huge. This prevents features with larger values ​​from overshadowing features with smaller values, making the subsequent fusion and processing process more stable.

[0050] Specifically, the standardization process employs the Z-Score standardization method. For each feature dimension, the mean and standard deviation of that feature in the training dataset are calculated, and then a transformation formula is used to process the feature values. During the online detection phase, the feature values ​​extracted from newly acquired signals also undergo the same standardization process using the mean and standard deviation determined during the training phase. The standardized acoustic signal feature vector and vibration signal feature vector are concatenated into a longer, fixed-dimensional multi-dimensional feature vector. The concatenation operation is performed in the order of acoustic features first, followed by vibration features, or vice versa, but the order must remain consistent throughout the system. In practical implementation, the fixed-dimensional multi-dimensional feature vector is packaged and bound with the original synchronously sampled signal segments of the acoustic and vibration sensor signals, forming a complete signal task slice. The multi-dimensional feature vector serves as the basic unit for subsequent resource prediction, while the original synchronously sampled signal segments serve as the foundational data for subsequent deep processing and fault analysis.

[0051] In practical implementation, the specific steps for resource demand prediction analysis for each signal task segment are as follows: A fixed-dimensional multi-dimensional feature vector is input into a pre-trained resource demand prediction model. The pre-trained resource demand prediction model employs a multilayer perceptron structure, where the number of nodes in the input layer is strictly equal to the dimension of the multi-dimensional feature vector, ensuring that the feature vector can be completely input into the network. Hidden layers typically consist of two or three layers, each containing several neurons. The hidden layer neurons use a linear rectified function as the activation function to introduce nonlinear transformation capabilities. The output layer includes two branches: a computational complexity prediction branch and a resource preference index prediction branch. The computational complexity prediction branch uses a Softmax function to output the probability that the signal task segment belongs to multiple discrete computational complexity levels, and takes the level with the highest probability as the prediction result. The sensor resource preference index is a multi-dimensional vector used to quantify the dependence of the signal task segment on acoustic sensor data channels, vibration sensor data channels, and computing unit resources. The resource preference index prediction branch uses a Sigmoid function to constrain the output value between 0 and 1; the closer the value is to 1, the higher the dependence.

[0052] In some embodiments, the training process of the pre-trained resource demand prediction model relies on a historical dataset. The historical dataset contains a large number of historically acquired acoustic and vibration sensor signals, along with the actual resource consumption data recorded during the processing of the corresponding signal task segments. The actual resource consumption data includes the CPU time consumed in processing the task segments, memory usage, and the duration of occupancy of the acoustic and vibration sensor data channels. The training objective is to minimize the difference between the resource demand descriptors predicted by the resource demand prediction model and the actual resource consumption data. The loss function is typically chosen as the cross-entropy loss function for classification tasks with high computational complexity, and the mean squared error loss function for regression tasks involving sensor resource preference indices. Model parameter optimization employs a backpropagation algorithm combined with an adaptive moment estimation algorithm, iteratively updating the weights and bias parameters of the neural network until the model's performance on the validation set stabilizes. The trained resource demand prediction model is stored in the system's memory for online, real-time resource demand prediction of newly generated signal task segments. When constructing training labels, the system pre-sets a mapping threshold for physical resources. Specifically, a first threshold and a second threshold are set for CPU time and memory usage. Physical consumption below the first threshold is marked as a low-level label, consumption between the two is marked as a medium-level label, and consumption above the second threshold is marked as a high-level label, so that the cross-entropy loss function can be used to calculate the classification error. At the same time, the maximum-minimum normalization algorithm is used to map the historical channel occupancy time to a continuous interval of 0 to 1, which is used as the standard true value for the mean squared error loss function regression calculation.

[0053] It is understandable that the generation of resource demand descriptors is a key prerequisite for dynamic resource scheduling. After performing forward propagation calculations on a fixed-dimensional multi-dimensional feature vector, the resource demand prediction model outputs the prediction computational complexity level and sensor resource preference index corresponding to the signal task segment. These two outputs together constitute the resource demand descriptor.

[0054] The predicted computational complexity level directly indicates the amount of computational resources required to process the task slice. A higher level means more complex calculations are needed, potentially involving iterative algorithms or large-scale matrix operations. Based on the differences in predicted computational complexity levels, the system will match differentiated subsequent processing strategies in the subsequent task allocation and execution phases: For signal task slices predicted as low complexity, which typically correspond to background noise or well-defined regular operating signals, the system can be scheduled to use a lightweight, rule-based judgment method for rapid processing without invoking complex computational logic. For signal task slices predicted as high complexity, which often contain complex feature information suggesting potential faults, the system can be triggered to initiate corresponding deep processing mechanisms, mobilizing model resources adapted to high-complexity calculations for refined analysis. This ensures that fault-related features in the signal are fully extracted, creating a logical closed loop between resource requirement prediction and subsequent processing execution, preventing a disconnect between prediction results and actual processing.

[0055] The sensor resource preference index is a more refined guiding indicator. For example, a signal task slice might exhibit a high preference index for acoustic sensor data channels and a low preference index for vibration sensor data channels. This indicates that the processing of this task slice heavily relies on acoustic data, and the availability of acoustic channel resources should be prioritized during scheduling. The resource requirement descriptor, as metadata for the signal task slice, is encapsulated along with the feature vector data and passed to the subsequent scheduling module.

[0056] Optionally, the interpretation and application of the sensor resource preference index can be further refined. The sensor resource preference index can be a three-dimensional vector, corresponding to the acoustic sensor data channel preference index, vibration sensor data channel preference index, and computing unit resource preference index, respectively. Each preference index ranges from 0 to 1, and its value reflects the relative demand intensity of the signal task slice for that type of resource. In subsequent scheduling decisions, when multiple idle resources exist in the system, the scheduler will prioritize allocating the resource indicated by the high preference index to the corresponding task slice to optimize processing efficiency. For example, if a task slice has a high computing unit resource preference index and a low sensor channel preference index, it will tend to be scheduled to a processing node with sufficient computing resources, and will be less concerned about the bandwidth of the sensor data interface.

[0057] It is understandable that the accuracy of resource demand forecasting and analysis directly impacts the efficiency of resource scheduling throughout the entire system. Pre-trained resource demand forecasting models, by learning the complex mapping relationship between signal characteristic patterns and resource consumption patterns in historical data, can make reasonable inferences about the resource requirements of newly arrived, unseen signal task slices. This predictive capability allows the system to have a forward-looking understanding of the computational burden and resource allocation of tasks before actually allocating resources for processing. This enables adaptive load balancing, ensuring that when faced with massive concurrent data surges (such as external short-circuit oscillations in transformers), high-value fault signals can be accurately identified and reserved for them, avoiding the delay or discarding of critical fault features due to a one-size-fits-all queuing strategy. As the system's runtime accumulates, new resource consumption data can be continuously collected and used for periodic model updates, allowing the resource demand forecasting model to continuously adapt to changes in the operating environment and maintain predictive accuracy.

[0058] In practical implementation, the feature extraction and resource prediction stages need to ensure real-time processing. The algorithms for extracting acoustic and vibration signal features need to be highly efficient; the Fast Fourier Transform algorithm can be implemented using existing highly optimized mathematical libraries. The pre-trained resource demand prediction model also needs to meet time constraints during the forward inference stage. This is typically achieved by selecting a network structure with moderate computational cost or by using hardware acceleration techniques such as tensor processing units to run the resource demand prediction model. The entire process from signal acquisition to the generation of resource demand descriptors needs to be completed within a strict time window to ensure that subsequent scheduling and processing can be carried out in a timely manner, meeting the real-time requirements of live-line detection of transformer winding insulation faults. Once the signal task slices and their associated resource demand descriptors are generated, they are ready to accept pass-through allocation or queuing decisions based on the real-time system load level.

[0059] Example 2: In specific implementation, establishing a dynamic mapping relationship between the real-time load status of the system and the resource demand descriptor is the core of achieving adaptive task allocation. Monitoring the real-time load status of the system involves the continuous collection and comprehensive calculation of multiple hardware resource indicators.

[0060] Specifically, the system monitors in real time the occupancy rates of the acoustic sensor data channels, the vibration sensor data channels, and the average utilization rate of the computing unit. Data channel occupancy: The occupancy rate of the acoustic sensor data channel is obtained by real-time reading of the buffer status register of the data acquisition card or dedicated interface chip. The real-time monitoring of the sensor data channel occupancy rate does not refer to a constant physical transmission bandwidth, but rather to the dynamic fill level of the data buffer. Specifically, the occupancy rate of the acoustic sensor data channel is obtained by real-time reading of the ratio of the current number of data frames in the first-in-first-out (FIFO) buffer on the data acquisition card to the total buffer capacity. This indicator reflects the dynamic balance between the data production rate and the consumption rate of subsequent processing units. When the processing unit's consumption rate decreases due to excessive computational load, the buffer fill level increases, and the occupancy rate rises accordingly, indicating that the data channel faces the risk of congestion. The occupancy rate of the vibration sensor data channel uses a similar monitoring mechanism. Since the sampling rate of the vibration signal may differ from that of the acoustic signal, the occupancy rate of each channel needs to be calculated independently.

[0061] Average utilization of computing units: obtained through the performance monitoring interface provided by the operating system kernel. It queries the percentage of time the central processing unit was in a non-idle state within a set time window in the past. The length of the set time window is configured according to the system's response speed requirements to load changes, and is usually between hundreds of milliseconds and several seconds.

[0062] In practical implementation, calculating the system load level index requires a weighted summation and normalization mapping of the three monitoring indicators mentioned above. The weighting coefficients are pre-set based on the relative importance and bottleneck degree of various resources in the system. For example, if the computing unit is the bottleneck resource of the system, the weighting coefficient of the average load rate of the computing unit should be set higher; if the bandwidth limitation of the sensor data channel is more critical, the weighting coefficient of the channel occupancy rate of the corresponding channel should be increased. The weighted summation formula is:

[0063] in, For weighted load sum, Acoustic channel occupancy rate, For vibration channel occupancy rate, To calculate unit utilization, the weighting coefficients must meet the following requirements. This is to ensure the consistency of the calculations.

[0064] Subsequently, the summation result is mapped to the interval between 0 and 100. The mapping process uses a linear transformation, and the mapping formula is as follows:

[0065] Among them, the minimum benchmark value and maximum benchmark value It is determined based on historical monitoring data of the system under no-load and extreme heavy-load conditions. Through this normalization and truncation process, it is ensured that the output system load level index is always within the range of 0 to 100, thereby truly and stably reflecting the relative busyness of the system and providing a unified benchmark for subsequent task allocation.

[0066] In practical implementation, the system has a pre-set resource matching matrix as the basis for allocating signal task slices to matching processing channels. This resource matching matrix is ​​stored in the system's non-volatile memory in the form of a data structure. Processing channels include high-precision parallel processing channels, standard processing channels, and resource-optimized channels. High-precision parallel processing channels are typically configured with a large amount of parallel computing resources and high-speed data cache, designed to process as many signal task slices as possible per unit time, but with higher resource consumption. Standard processing channels employ a balanced configuration, achieving a balance between processing speed and resource consumption. Resource-optimized channels use simplified algorithms and resource reuse strategies, prioritizing task processing during resource-constrained periods, but with lower throughput.

[0067] The resource matching matrix defines the mapping relationship between the predicted computational complexity level and the processing channel. For example, in the pass-through mode when the system is not overloaded, or in the dequeueing stage after overload queuing, task fragments with high predicted computational complexity will be preferentially matched to the high-precision parallel processing channel; task fragments with medium predicted complexity will be matched to the standard processing channel; and task fragments with low predicted complexity will be matched to the resource optimization channel.

[0068] It's understandable that allocating signal task fragments to designated processing channels based on the current system load level is a continuous decision-making process. A load monitoring thread runs in the system background, refreshing the system load level index at a fixed interval (e.g., 10 times per second) and comparing it to a preset dynamic load window upper limit threshold. If the system load level index does not exceed the upper limit threshold, routing is performed directly according to the channel determined by the resource matching matrix; if it exceeds the upper limit threshold, the signal task fragments are stored in a priority-based task fragment waiting queue, to be allocated after dequeueing. The input queue of the high-precision parallel processing channel may be directly associated with a high-speed memory buffer, the input queue of the standard processing channel may be based on FIFO memory, and the input queue of the resource-optimized channel may have priority identifiers. This dynamic allocation mechanism ensures that the system can flexibly adjust its task processing strategy based on its real-time resource availability.

[0069] The mapping rules in the resource matching matrix can be dynamically adjusted based on actual operating conditions. This adjustment can be done manually or automatically optimized by the system based on long-term performance data. For example, if the system detects that tasks assigned to high-precision parallel processing channels are experiencing processing delays, the mapping rules can be dynamically modified to downgrade some tasks with high predicted computational complexity to standard processing channels, thereby avoiding congestion. The adjustment strategy can be based on statistical analysis of historical task processing delays, queue lengths, and other performance indicators. The system load level index and the resource matching matrix for different processing channels can also be designed as a multi-level or probabilistic mapping model with fuzzy boundaries to enhance the system's robustness against load fluctuations.

[0070] In practical implementation, the internal architecture and resource allocation strategies of different processing channels differ significantly to suit their design goals. High-precision parallel processing channels may employ a multi-core parallel processing architecture, breaking down signal tasks into multiple subtasks and distributing them across different processor cores for simultaneous execution, fully utilizing the parallel computing capabilities of the computing units. Standard processing channels may use a single-core or few-core pipelined processing mode, executing tasks sequentially according to fixed processing stages to ensure a stable processing rate. Resource-optimized channels may employ lightweight algorithms or approximate calculation methods to reduce computational complexity, while potentially using more aggressive data compression and reuse strategies to reduce the bandwidth and memory resource consumption of the sensor data channels. Each processing channel has an independent resource manager and task scheduler responsible for the task execution order and resource allocation within the channel.

[0071] To further improve the response efficiency of the resource optimization channel under extremely high loads, this embodiment introduces a variable-granularity feature extraction strategy. This strategy dynamically reconstructs the deep feature extraction and fusion process based on the predicted computational complexity level output by the pre-trained neural network model. Specifically, the system's pre-built feature extraction operator library adopts a structured modular architecture. Lightweight and heavyweight operators correspond to independent functional modules, execution code segments, and resource call interfaces, respectively. Key information such as the type identifier, execution logic path, resource call priority, and feature output dimension of each operator is recorded in detail in the operator management registry. This registry maintains real-time data synchronization with the system resource scheduling module to ensure fast querying and accurate matching during operator calls. The feature extraction mask is built around the execution logic path in the operator management registry. When the system load level index continuously exceeds the upper limit threshold of the dynamic load window and the task is assigned to the resource optimization channel, the system sends a preset logic masking instruction to the operator call interface to accurately block the function triggering link of all heavyweight operators, while retaining the normal call channel of lightweight operators. Only the minimum feature set composed of time-domain statistics such as mean and variance participates in the calculation, simplifying the calculation process to the greatest extent. In response to the dynamic adjustment of the mapping rules in the resource matching matrix, when the resource requirement descriptor of the signal task segment indicates that its prediction computational complexity level is high, the system immediately generates an exception interrupt instruction carrying the unique identifier of the task segment, resource requirement details, and operator unlocking permissions. According to the exception interrupt instruction, the system modifies the control state of the feature extraction mask, removes the masking of heavyweight operators, and calls the independent computing resources reserved by the system. Subsequently, it retrieves the complete execution code and resource call configuration of operators such as Fast Fourier Transform and Continuous Wavelet Transform from the operator management registry, restarts the feature extraction process, and comprehensively extracts the time-domain, frequency-domain, and time-frequency-domain features of acoustic and vibration signals, providing complete feature data support for subsequent fault diagnosis.

[0072] Furthermore, to prevent excessive exception interruptions under high load from causing a complete system crash or deadlock, a computing power isolation and hard circuit breaker mechanism is introduced within the resource optimization channel. Specifically, the system reserves an independent high-priority thread pool for such high-complexity exception tasks in memory and computing units. The resource quota of this thread pool is strictly limited by physical isolation (e.g., it is allowed to occupy a maximum of 10% of the emergency computing time slice) to avoid crowding out the survival space of standard scheduling threads. At the same time, a hard circuit breaker threshold is set, for example, within a one-second sliding time window, a maximum of a preset number of exception interrupts (e.g., 3 times) are allowed to be triggered. When the number of triggers reaches the circuit breaker threshold, subsequent high-complexity tasks will be forcibly downgraded and processed according to a lightweight process, or directly placed in a suspension queue. This mechanism safeguards the bottom line of critical fault diagnosis while ensuring the physical security of the edge computing device's operating system.

[0073] It is understandable that the effectiveness of the dynamic mapping and allocation mechanism depends on the accuracy of monitoring data and the timeliness of decision-making. The calculation cycle of the system load level index needs to be short enough to capture rapid changes in system load and avoid incorrect allocation decisions due to information lag. The acquisition of the instantaneous bandwidth utilization rates of the acoustic sensor data channels and the vibration sensor data channels requires hardware support for high-precision real-time monitoring. The statistical window for the average load rate of the computing unit needs to balance stability and sensitivity; a window that is too short will cause drastic fluctuations in the index, while a window that is too long will cause the system to respond slowly to load changes. The entire dynamic mapping relationship establishment and task allocation process constitutes a closed-loop control system. The real-time system load status serves as a feedback signal, and the task allocation decision serves as a control signal, working together to enable the system to automatically maintain a high-performance operating state.

[0074] In some embodiments, allocation decisions present a hierarchical decision-making mechanism that separates global load balancing from local resource adaptation. At the global scheduling level, allocation decisions are primarily based on the system load level index. For example, even if the resource requirement descriptor of a semaphore task fragment indicates that it requires high computing resources, it will still be globally routed to the resource optimization channel if the current system is severely overloaded.

[0075] However, at the local resource adaptation level within the channel, the resource optimization channel will perform privileged adjustments based on the resource requirement descriptor of the task slice: for exceptional tasks with high computational complexity, the channel will trigger an exception interrupt mechanism to forcibly call heavyweight operators (and execute them within the aforementioned circuit breaker threshold); while for regular tasks with medium to low complexity, the operators remain masked, and lightweight, fast computation is used. This hierarchical decision-making mechanism separates macro-level load balancing from micro-level diagnostic baselines, simplifying system design and improving decision-making efficiency. Once a signal task slice enters the designated processing channel, it will be processed according to the channel's dynamic reconstruction process until the insulation fault analysis is completed and the results are output.

[0076] Example 3: In specific implementation, when the system load level index continuously exceeds the upper limit threshold of the dynamic load window, the task fragmentation buffer and scheduler is activated. The core function of the task fragmentation buffer and scheduler is to manage signal task fragments that cannot be processed immediately due to temporary insufficient system resources. After activating the task fragmentation buffer and scheduler, the system first creates a priority-based task fragment waiting queue. The priority-based task fragment waiting queue is a dynamic data structure used to store all signal task fragments waiting to be processed and their associated resource requirement descriptors. The priority value of each signal task fragment in the queue is jointly determined by the predicted computational complexity level and the sensor resource preference index in the resource requirement descriptor. The calculation process involves converting the enumerated values ​​of the predicted computational complexity level (e.g., high=3, medium=2, low=1) into numerical scores, and simultaneously condensing the sensor resource preference index (a multi-dimensional vector) into a scalar preference score through weighted combination. The priority value P can be comprehensively evaluated using the following formula:

[0077] in: This represents a numerical score converted from the predicted computational complexity level. , and These represent the acoustic sensor data channel preference index, vibration sensor data channel preference index, and computing unit resource preference index in the resource demand descriptor, respectively. and It is a coefficient that balances the weights of the two types of factors, and . , and These are weights used to reconcile the influence of different resource preferences, satisfying These weighting coefficients need to be pre-configured based on the importance and scarcity of system resources. For example, in transformer fault detection scenarios, high computational complexity usually means that the signal contains rich abnormal features or potential serious fault modes. Therefore, it needs to be given a higher processing priority than low-complexity background noise signals to prevent critical faults from being missed. The higher the calculated priority value P, the higher the position of the signal task slice in the priority-based task slice waiting queue, and the greater the chance of obtaining resources. The queue management strategy usually combines priority and enqueue time, that is, high-priority signal task slices are given priority, and signal task slices of the same priority are arranged in a first-in-first-out order. Optionally, the queue management strategy also introduces a time aging compensation mechanism, that is, as the enqueue time increases, the priority P of the task will receive a certain gain compensation to ensure that low-priority tasks will eventually be processed as well.

[0078] To prevent low-priority semaphore task fragments from starving indefinitely under sustained high load due to being unable to compete with a continuous stream of high-priority new tasks, the time aging compensation mechanism is implemented by dynamically increasing the effective scheduling priority of tasks. Specific steps include: recording the enqueue time; when each semaphore task fragment enters the priority-based task fragment waiting queue, the system records a precise timestamp for it. The start time for queuing is used as the starting time. The waiting time is calculated for each task slice in the queue as the scheduler periodically scans the queue to select tasks to be processed. Calculate its current waiting time in real time. ,in Use the current system time. Calculate dynamic priorities and task sharding. Effective dynamic priority for scheduling Based on its initial static priority (i.e., calculated according to the original formula) The calculation formula is determined by both the number of passengers and the waiting time, as follows:

[0079] in, It is a preset aging gain coefficient ( This is used to control the strength of the impact of waiting time on priority. It is a non-negative, monotonically increasing function used to map waiting time to priority increments. A typical implementation is a linear function. or logarithmic function To smooth out initial growth and avoid premature effectiveness, dynamic priority scheduling means that the scheduler no longer relies on the initial static priority when selecting tasks from the waiting queue. Instead, it is based on the calculated effective dynamic priority. Sort in descending order. Select each time. The task with the highest priority value is dequeued and scheduled. An elevation cap can be set to prevent a task with an initially very low priority from having a dynamic priority that exceeds that of a truly urgent task with a high initial priority due to excessively long waiting times. ,when season This upper limit can be set to the static priority value corresponding to the highest emergency priority defined by the system.

[0080] In practice, the task fragmentation buffer and scheduler need to periodically scan the system resource status to find execution opportunities. The scanning operation is performed by a separate scheduling thread, and the scanning frequency is set according to the system's real-time requirements, typically between tens and hundreds of milliseconds. Scanning the system resource status includes checking the current idle bandwidth of the acoustic sensor data channels, the current idle bandwidth of the vibration sensor data channels, and the immediate idle computing power of the computing units. The scheduler matches the scanned idle resource information with the resource requirement descriptors of each signal task fragment in the priority-based task fragmentation waiting queue. The matching process mainly examines whether the idle resources meet the resource type requirements pointed to by the sensor resource preference index of the signal task fragment and the computational resource level implied by the predicted computational complexity level. Specifically, the computational resource level matching in the matching process relies on a pre-set quantitative resource admission mapping table. This mapping table strictly maps the qualitative label of the predicted computational complexity level to the minimum hardware immediate idle threshold required to execute the task. For example, if the prediction computational complexity level is high, the corresponding requirement in the mapping table is that the instantaneous idle utilization rate of the computing unit must not be less than 40% and the continuous idle memory must not be less than a preset megabyte; if the prediction level is medium, the corresponding requirement is that the idle utilization rate must not be less than 20%; if the prediction level is low, the corresponding requirement is that the idle utilization rate must not be less than 5%.

[0081] During the matching process, the scheduler first extracts the real-time idle hardware metrics from the operating system's underlying layer, and then compares them in absolute terms with the quantitative thresholds in the resource admission mapping table corresponding to the semaphore task slice. Only when the real-time idle utilization of the computing unit is strictly greater than or equal to the corresponding preset threshold in the mapping table is it determined that the matching condition of the computing resource scale is met, thereby triggering the subsequent allocation process. When it is detected that the current idle resource status is sufficient to meet the needs of one or more semaphore task slices in the priority-based task slice waiting queue, the scheduler initiates the allocation process.

[0082] In some embodiments, the scheduler selects signal task fragments from the priority-based task fragment waiting queue by choosing the highest priority signal task fragment that matches the currently idle resources. Matching is a crucial step; for example, if the currently idle resources are primarily acoustic sensor data channel bandwidth, then signal task fragments with higher acoustic sensor data channel preference indices in the sensor resource preference index will be considered to have a higher match. The scheduler traverses the priority-based task fragment waiting queue, finds the first signal task fragment that meets the matching criteria, removes it from the queue, and loads it onto the corresponding processing unit for execution. The loading process includes transferring the multidimensional feature vector data of the signal task fragment and the original synchronous sampling signal segment from the buffer to the processor's computation unit and configuring the corresponding sensor data channel interface. Once a signal task fragment begins processing, the resources it occupies are marked in the system resource status until processing is complete and the resources are released. To improve the management and retrieval efficiency of a large number of waiting tasks, the priority-based task fragment waiting queue uses a multi-level hash table structure for indexing. The first-level hash table in the multi-level hash table structure uses the predicted computational complexity level as the key to initially group signal task fragments into different buckets, such as high computational complexity buckets, medium computational complexity buckets, and low computational complexity buckets. A second-level hash table is then built within each bucket. The keys of the second-level hash table are generated based on the sensor resource preference index. A common practice is to discretize the continuous value range of the sensor resource preference index into several level intervals, and then use the combination of these level intervals as the key value. For example, [0, 0.3) is mapped to low preference, [0.3, 0.7) to medium preference, and [0.7, 1.0] to high preference. By using the combined key of the predicted computational complexity level and the sensor resource preference index level, signal task fragments with similar resource requirement descriptors can be grouped into the same final sub-hash bucket. To ensure that the highest priority signal task fragment can be retrieved quickly and stably within the same sub-hash bucket, each sub-hash bucket is specifically implemented as a max-heap based on effective dynamic priority in its underlying data structure. When a new semaphore task fragment is mapped to its corresponding sub-hash bucket after two levels of hash calculation, the scheduler performs a heap insertion operation, adjusting it upwards based on its priority value. When the scheduler locates a matching sub-hash bucket and is ready to retrieve the task, it directly pops the top element of the heap and adjusts the remaining elements downwards. This mechanism ensures that after locating a group with similar needs, the scheduler can reliably obtain the locally highest priority task within that group in logarithmic time complexity, completely eliminating the overhead of linear traversal within the bucket. This group management method allows the scheduler to quickly locate potentially matching semaphore task fragment groups when resource status changes, without needing to linearly scan the entire priority-based task fragment waiting queue.

[0083] In some embodiments, when the system resource status changes, such as the emergence of new idle resources or the release of resources upon completion of existing tasks, the scheduler can batch process multiple signal task fragments within the same group. If there are sufficient idle resources and the resource requirements of multiple signal task fragments within a certain sub-hash bucket are similar and can be satisfied, the scheduler can retrieve multiple signal task fragments from that sub-hash bucket at once and allocate them to available processing units for parallel processing. This batch processing mechanism significantly reduces the number of resource allocation decisions made by the scheduler, reduces scheduling overhead, and improves the overall throughput of the system under heavy load. The number of batch processing operations is limited by the total amount of available resources, and the scheduler needs to calculate the maximum number of task instances that available resources can support simultaneously. As a multi-level hash table structure serving as a task fragment waiting queue, the insertion and deletion operations of each sub-hash bucket are synchronized. Whenever a new signal task fragment needs to enter the buffer due to excessive system load, the signal task fragment is first inserted into the corresponding position of the multi-level hash table structure based on the hash key calculated according to its resource requirement descriptor, and is also inserted into the logical sorting position within that position according to its priority value. When a signal task fragment is retrieved and executed by the scheduler, the location and deletion of the target signal task fragment are performed within the sub-hash buckets corresponding to the multi-level hash table structure. Simultaneously, the locally ordered structure within each sub-hash bucket is updated to maintain priority consistency. During insertion or deletion operations, atomic operations or mutex lock mechanisms should be used to ensure the consistency of the composite structure and prevent data races caused by concurrent access. The multi-level hash table structure primarily accelerates resource-demand-based matching queries, while its sub-hash buckets implement priority scheduling of task fragments through a local priority maintenance mechanism, thereby achieving synergistic optimization of resource scheduling efficiency and sorting accuracy.

[0084] It is understandable that the activation and deactivation of the task fragmentation buffer and scheduler are directly determined by the threshold of the system load level index relative to the dynamic load window. When the system load level index falls below the lower limit threshold of the dynamic load window, it indicates that system resources have become relatively abundant and can directly process newly arriving tasks without complex scheduling. At this time, the task fragmentation buffer and scheduler are deactivated. The remaining signaled task fragments in the priority-based task fragmentation waiting queue, as well as newly arriving signaled task fragments, will no longer be sorted and matched by the scheduler. Instead, the signaled task fragments will be directly routed to the matching processing channel for processing based on the predicted computational complexity level. The deactivation process includes pausing the scheduling thread and emptying the priority-based task fragmentation waiting queue. The dynamic start-stop mechanism of the task fragmentation buffer and scheduler allows the system to adopt different processing strategies under high and normal load conditions, optimizing resource utilization efficiency.

[0085] Example 4: In practical implementation, the calculation of the system load level index is the quantitative basis for dynamic load management. Setting the upper and lower thresholds of the dynamic load window is a process that combines historical data analysis with system performance objectives.

[0086] In practice, setting the upper and lower thresholds of the dynamic load window requires comprehensive consideration of system performance goals and hardware capabilities, rather than simply using fixed values. The first step in threshold setting is to statistically analyze the distribution pattern of the system load level index based on historical load data. The system needs to collect time-series data of the system load level index generated over a long period of time under typical workloads. Statistical analysis is then performed on this historical data to calculate its percentiles, such as the 95th, 90th, 50th, and 10th percentiles. Higher percentiles (such as the 95th percentile) are often considered as candidates for the upper threshold because these points represent high load states that the system does not frequently experience. Lower percentiles (such as the 10th percentile) are often considered as candidates for the lower threshold, representing the system's basic load level. However, relying solely on the percentiles of historical data is insufficient to determine the optimal threshold; a comprehensive decision must be made considering system hardware performance indicators and the real-time requirements of the processing tasks. System hardware performance metrics include the maximum processing power of the computing unit, memory bandwidth, and the total bandwidth of the sensor data channels. High-performance hardware allows for higher thresholds because the system is capable of maintaining stable operation under high loads. The real-time requirement of the processing task refers to the maximum allowable delay for the system to process signal task slices and output results. High real-time requirements mean the system needs to be more sensitive to load increases, thus requiring relatively lower thresholds to initiate load regulation mechanisms (such as task slice buffers and schedulers) earlier. The final upper and lower thresholds are specific values ​​selected from candidate thresholds obtained based on historical data statistics, after balancing hardware performance and real-time requirements. For example, even if the 95th percentile of historical data is 85, if the real-time requirements are extremely stringent, the upper threshold might ultimately be set at 75 to provide a more sufficient safety margin.

[0087] Optionally, the upper and lower thresholds of the dynamic load window can be fixed values ​​or adaptively adjusted during system operation. Fixed thresholds are set during system deployment and remain unchanged, offering the advantage of simplicity and stability. Adaptive threshold adjustment allows the system to optimize performance based on long-term learning. For example, the system can periodically (e.g., every 24 hours) analyze the load distribution characteristics and task processing latency over a recent period. If it finds that the current threshold is causing the scheduler to activate too early or deactivate too late, thus affecting efficiency, the upper and lower thresholds can be automatically fine-tuned. Adaptive adjustment algorithms require careful design to avoid frequent threshold oscillations, typically employing a slow drift strategy. The slow drift strategy can be implemented using an exponentially weighted moving average algorithm. Taking the adaptive adjustment of the upper limit threshold as an example, the system first records the high-segment statistical mean of the system load level index within a preset time window (e.g., the past 24 hours) as a reference benchmark. The high-segment statistical mean refers to the arithmetic mean calculated from the top 10% of the system load level index sequences sorted by value within the preset time window. The calculation of the new upper limit threshold is based on the current upper limit threshold, introducing a weighting coefficient to control the adjustment range. The weighting coefficient is preset according to system stability requirements to ensure that each adjustment range is within a reasonable range and to avoid sudden threshold changes. Specifically:

[0088] in, The updated upper limit threshold. This is the current upper limit threshold. For reference, The weighting coefficients are preset. This formula ensures that each adjustment is within a reasonable range, avoiding sudden threshold changes. The adaptive adjustment of the lower threshold uses the same algorithm logic, only requiring the reference benchmark to be replaced with the low-segment statistical mean of the system load level index within the corresponding time period. The low-segment statistical mean refers to the arithmetic mean calculated by selecting the system load level index sequence with the smallest values ​​within a preset time window and sorting them by value.

[0089] This algorithm allows the threshold to be slowly adjusted according to the long-term load trend of the system, adapting to dynamic changes in the operating environment while effectively avoiding frequent threshold oscillations caused by short-term load fluctuations. Furthermore, a preset hysteresis interval is maintained between the updated upper and lower thresholds. Too small an interval will cause the task sharding buffer and scheduler to frequently activate and deactivate near the critical point of the load level index, generating unnecessary overhead. Too large an interval will result in an excessively large transition area between high and normal load states, potentially delaying the activation or deactivation of load control measures. A common practice is to ensure that the width of the dynamic load window (upper threshold minus lower threshold) is sufficient to accommodate normal fluctuations in the system load level, avoiding mode switching caused by short-term jitter. After the threshold is set, the system monitors the calculated system load level index in real time and compares it with the upper and lower thresholds of the dynamic load window to determine the state of the task sharding buffer and scheduler, achieving adaptive load management.

[0090] Example 5: In practical implementation, after the signal task is segmented and enters the matching processing channel, deep feature extraction and fusion, adapted to the computing power of that processing channel, will be performed, as well as subsequent fault mode identification. The core task of deep feature extraction and fusion is to generate a joint feature vector that can comprehensively reflect the device status.

[0091] First, based on the computational limitations of the processing channel (such as a resource optimization channel or a high-precision parallel processing channel) and the control state of the feature extraction mask, the original synchronous sampling signal segments of the acoustic and vibration sensor signals encapsulated in the signal task slice are extracted, and the corresponding operators in the feature extraction operator library are called to obtain the feature set. Then, a feature-level fusion strategy based on attention weighting is adopted. Specifically, the system has a pre-set learnable attention weight vector, the dimension of which is the same as the total dimension of the concatenated acoustic and vibration features. During the fusion process, the system does not assign equal importance to all features, but instead calculates a dynamic weight for each feature dimension in the concatenated joint feature vector using this attention weight vector. The weight is automatically learned by the pre-trained classification model during training based on the contribution of each feature to the classification of historical fault samples. Features with a higher correlation to typical insulation faults are assigned higher weights; while features with lower discriminative power or stronger correlation to background noise are assigned lower weights.

[0092] Fault mode recognition and classification rely on a pre-trained classification model. This model employs deep learning, such as convolutional neural networks or recurrent neural networks. Deep learning models can automatically learn the complex nonlinear relationships between fault features and possess multi-layered cascaded network topologies, offering potential advantages in processing high-dimensional, nonlinear acoustic-vibration fusion features. The classification model outputs the probability distribution of the transformer winding state corresponding to the signal task segment belonging to each predefined fault category. The final diagnostic result is the fault category with the highest probability. Training the classification model requires a large amount of historical data with accurate labels, covering acoustic and vibration signals under various normal operating conditions and typical insulation fault conditions.

[0093] In some embodiments, the feedback optimization step of the signal task fragmentation processing result is an important mechanism for continuously improving system performance. The core of the feedback optimization step is to use actual operating data to correct possible deviations in the resource demand prediction model.

[0094] The system records the actual resource data consumed by each signal task segment during processing. This actual resource data includes, but is not limited to, the CPU time consumed in processing the signal task segment, the peak memory usage, and the effective occupancy time of the acoustic sensor data channels and vibration sensor data channels. This actual resource data, along with the predicted values ​​in the resource demand descriptors generated by resource demand prediction analysis before the signal task segment enters the processing flow, are recorded side-by-side in the log database. Prediction bias data is generated by systematically comparing predicted and actual values. For prediction computational complexity levels, the bias reflects whether the predicted level matches the complexity level corresponding to the actual CPU time consumed. For the sensor resource preference index, the bias is the numerical difference between the predicted index value and the actual preference level reflected by the actual occupancy time of the acoustic sensor data channels during processing. Prediction bias data quantifies the prediction accuracy of the resource demand prediction analysis neural network model in the current operating environment. Continuous bias indicates that the model may not fully adapt to the mapping relationship between actual signal characteristics and resource consumption.

[0095] In some embodiments, periodically updating the neural network model parameters used for resource demand prediction analysis using prediction bias data is crucial for maintaining the long-term efficient operation of the system. The update process is typically performed offline or via low-priority online learning to avoid interfering with real-time fault detection tasks. The periodic update cycle can be set based on system operating intensity and data accumulation rate, for example, triggering a model update every 24 hours or after processing a certain number of signal task slices. During the update, the system extracts recently accumulated signal task slice data from the log database. Each data sample includes the multi-dimensional feature vector initially input to the neural network model, the resource demand descriptor initially output by the neural network model, and the "real" resource demand descriptor derived from the resource data consumed during the actual processing of that signal task slice. These samples constitute a new training dataset for incremental training or fine-tuning of the original resource demand prediction analysis neural network model. Model parameter updates can employ various strategies. One strategy is to directly retrain the original neural network model for a full cycle using the new training dataset. This strategy fully utilizes the new data but incurs significant computational overhead. Another strategy employs transfer learning or online learning algorithms with constraints, making minor adjustments to only some layers or parameters of the neural network model using new data. Before the new data is used in the model update, it undergoes outlier detection and filtering. Simultaneously, a parameter update magnitude threshold is set to limit the maximum range of a single adjustment, preventing parameter mutations. Furthermore, after the model parameters are updated, it first enters a background trial phase, continuously monitoring preset performance indicators, including resource demand prediction accuracy and prediction bias stability. Only after performance surpasses the old model for several consecutive monitoring periods without abnormal fluctuations is the old model officially replaced and put into use. This strategy is computationally efficient and can adapt to changes more quickly. After the neural network model parameters are updated, the new version of the model is deployed to the resource demand prediction and analysis module, replacing the old model for predicting subsequent newly arriving signal task segments. Through this continuous feedback optimization cycle, the resource demand prediction and analysis neural network model can gradually learn the dynamic relationship between signal characteristics and resource consumption that may change over time, thereby improving prediction accuracy and optimizing the overall system's task scheduling efficiency and resource utilization.

[0096] Optionally, the feedback optimization step can be extended to the fault classification model within the matched processing channel. Besides the resource prediction model, diagnostic deviation data can also be generated by comparing the signal task segmentation processing results with subsequent offline high-voltage tests or enclosure inspection results. This diagnostic deviation data can be used to periodically update the classification model parameters used for insulation fault mode identification and classification, enabling the classification model to adapt to possible drift in equipment status or the emergence of new fault modes, thereby improving the accuracy and reliability of fault diagnosis.

[0097] It is understandable that, unlike the aforementioned resource demand prediction model which is based on short-cycle automatic iteration of real-time computation logs, the fault classification model relies on long-cycle physical truth feedback for updates.

[0098] Specifically, the feedback optimization step extends to the fault classification model within the matched processing channel. Since the diagnosis of insulation faults typically relies on subsequent physical inspections such as offline high-voltage testing, oil chromatography analysis, or overhead inspections, these methods have significant time lags (ranging from hours to months). Therefore, the system establishes an asynchronous diagnostic bias database, anchoring historical diagnostic results from signal task segments with later-entered physical inspection results. Based on the generated diagnostic bias data, the system periodically (e.g., quarterly or annually) triggers retraining or parameter calibration of the classification model. This combined short-term and long-term mechanism ensures both real-time adaptive resource scheduling (corresponding to a fast loop) and rigorous medical-grade accuracy in fault diagnosis (corresponding to a slow loop).

[0099] The fault classification model employs a hierarchical neural network architecture with multiple intermediate exits. This architecture comprises multiple sequentially connected network layers, with an emergency inference exit at the Nth network layer and a final inference exit at the last network layer. The selection of the Nth layer is based on a balance between model inference efficiency and diagnostic accuracy, ensuring that shallow inference meets the real-time requirements of resource-constrained scenarios while retaining core fault discrimination capabilities. When a signal task slice is routed to the resource optimization channel, the system inputs the joint feature vector into the hierarchical neural network model. After completing the calculation of the Nth layer of the neural network, the inference process of the subsequent deep network is terminated, and the fault classification result is directly output from the emergency exit. This shallow inference process significantly reduces the processing time of a single task by simplifying network computation and adapts to resource constraints under high load. When a signal task slice is assigned to the high-precision parallel processing channel or the standard processing channel, the joint feature vector is input into the hierarchical neural network model, and all network layers are traversed and calculated. This fully utilizes the feature learning capabilities of the deep network to extract subtle fault correlation information from complex acoustic and vibration fusion features, and finally outputs a high-precision fault classification result from the final inference exit of the last network layer. To ensure diagnostic consistency across different inference modes, a feature alignment layer is incorporated at the junction of the Nth and N+1th layers of the neural network. The feature vector output from the Nth layer satisfies both the classification requirements of shallow inference and serves as an effective input basis for the deep network, ensuring that both inference paths are based on a unified core feature logic and avoiding deviations in fault identification due to differences in inference depth. Simultaneously, this neural network architecture employs a dual-objective optimization strategy during the training phase, constructing a joint loss function. This joint loss function includes a first error calculated based on the prediction result of the emergency inference exit and a second error calculated based on the prediction result of the final inference exit. Specifically, both the first and second errors are calculated using the cross-entropy loss function, and the joint loss function... The calculation formula is:

[0100] in, For the joint loss value, For the first error, This is the second error. and These are preset weight hyperparameters used to balance the gradient update magnitudes of shallow and deep networks.

[0101] Iterative training is performed using the joint loss function. The training data covers acoustic and vibration fusion feature samples under normal operating conditions and various typical insulation fault conditions. Through iterative training, both inference modes can be adapted to the discrimination scenario of transformer winding insulation faults, ensuring that reliable fault diagnosis results can be output regardless of whether the system is under high load or low load.

[0102] Example 6: A transformer winding insulation fault live-line detection system, comprising the above-mentioned transformer winding insulation fault live-line detection method, including: The multimodal signal segmentation construction module is used to acquire acoustic sensor signals and vibration sensor signals during transformer operation, perform initial feature extraction, and construct signal task segments. The resource demand descriptor generation module is used to obtain a resource demand descriptor containing the prediction computational complexity level and the sensor resource preference index based on the signal task segmentation and a pre-trained resource demand prediction model. The system load status assessment module is used to monitor and obtain the hardware resource status in real time and calculate the current system load level index. The resource-coupled task sharding adaptive scheduling module is used to compare the system load level index with a preset upper limit threshold; in response to the system load level index exceeding the upper limit threshold, it calculates the priority value of the signal task shard by combining the predicted computational complexity level and the sensor resource preference index, and dynamically allocates the signal task shard to the matching processing channel through the task sharding waiting queue. The operator-level feature extraction path dynamic reconstruction module is used to slice the signal task allocated to the matching processing channel, dynamically configure the calling logic of the feature extraction operator according to the computing power limit of the matching processing channel, perform deep feature extraction and fusion, and obtain a joint feature vector. The multi-exit adaptive inference decision module is used to obtain the insulation fault mode identification result of the transformer winding by using a fault classification model with multiple inference exits based on the joint feature vector and triggering the corresponding inference exit according to the computing power limit of the processing channel.

[0103] A terminal includes a processor and a storage medium; the storage medium is used to store instructions. The processor is configured to operate according to the instructions to perform the steps according to the method.

[0104] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of any of the methods described.

[0105] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0106] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0107] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0108] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for live-line detection of transformer winding insulation faults, characterized in that, include: Acoustic and vibration sensor signals are acquired during transformer operation, initial feature extraction is performed, and signal task segments are constructed. Based on the signal task segmentation, a pre-trained resource demand prediction model is used to obtain a resource demand descriptor that includes the prediction computational complexity level and the sensor resource preference index. Real-time monitoring and acquisition of hardware resource status, and calculation of the current system load level index; Compare the system load level index with a preset upper limit threshold; In response to the system load level index exceeding the upper limit threshold, the priority value of the signal task slice is calculated by combining the predicted computational complexity level and the sensor resource preference index, and the signal task slice is dynamically allocated to the matching processing channel via the task slice waiting queue. For the signal task slices assigned to the matching processing channel, the calling logic of the feature extraction operator is dynamically configured according to the computing power limit of the matching processing channel, and deep feature extraction and fusion are performed to obtain a joint feature vector; Based on the joint feature vector, a fault classification model with multiple inference exits is used, and the corresponding inference exit is triggered according to the computing power limit of the processing channel to obtain the insulation fault mode identification result of the transformer winding.

2. The method for live detection of transformer winding insulation faults according to claim 1, characterized in that, Initial feature extraction and signal task segmentation are performed, and a resource demand descriptor is obtained using a pre-trained resource demand prediction model, including: A fixed-length sliding window is used to extract time-domain statistical features, frequency-domain energy features, and time-frequency domain joint features from the synchronously sampled acoustic sensor signals and vibration sensor signals, respectively. The extracted time-domain statistical features, frequency-domain energy features, and time-frequency domain joint features are standardized and concatenated into a fixed-dimensional multidimensional feature vector. The multidimensional feature vector is then encapsulated and bound to the original synchronous sampling signal segments of the acoustic sensor signal and the vibration sensor signal as the signal task slice. The multidimensional feature vectors from the signal task slices are input into the pre-trained resource demand prediction model, which is a multilayer perceptron model. The output layer of the multilayer perceptron model includes two branches: Computational complexity prediction branch: The Softmax function is used to output the probability that the signal task slice belongs to multiple discrete computational complexity levels, and the level with the highest probability is taken as the predicted computational complexity level; Resource preference index prediction branch: The Sigmoid function is used to output a multi-dimensional vector, which is used to quantify the dependence of the signal task slice on the acoustic sensor data channel, vibration sensor data channel and computing unit, respectively, as the sensor resource preference index.

3. The method for live detection of transformer winding insulation faults according to claim 1, characterized in that, Real-time monitoring and acquisition of hardware resource status, calculation of the current system load level index, including: The system monitors and acquires the occupancy rate of the acoustic sensor data channel, the occupancy rate of the vibration sensor data channel, and the average utilization rate of the computing unit in real time. The occupancy rate of the sensor data channel is determined by acquiring the real-time data storage volume of the first-in-first-out buffer in the data acquisition card and calculating the ratio of the real-time data storage volume to the total capacity of the first-in-first-out buffer. Based on preset weighting coefficients, the occupancy rates of the acoustic sensor data channels, the vibration sensor data channels, and the average utilization rate of the computing unit are weighted and summed to obtain a weighted load sum. Based on preset no-load and heavy-load benchmark values, the weighted load is linearly transformed and truncated to obtain the system load level index constrained within a preset range.

4. The method for live detection of transformer winding insulation faults according to claim 1, characterized in that, The processing channels include a high-precision parallel processing channel, a standard processing channel, and a resource-optimized channel. For signal task slices allocated to the processing channels, the calling logic of feature extraction operators is dynamically configured according to the computing power limitations of the processing channels, and deep feature extraction and fusion are performed, including: A pre-built feature extraction operator library is provided, which includes lightweight operators and heavyweight operators; When the signal task slice is allocated to the resource optimization channel, a logical masking instruction is sent to the preset feature extraction operator library to modify the control state of the feature extraction mask to block the triggering link of the heavyweight operator and call the lightweight operator to extract the minimum feature set to participate in feature fusion. In response to the dynamic adjustment of the mapping rules in the resource matching matrix, when the signal task slice is allocated to the resource optimization channel and the prediction computation complexity level is high, an exception interruption instruction is generated; the control state of the feature extraction mask is modified according to the exception interruption instruction, the masking of the heavyweight operator is removed, and the independent computing resources reserved by the system are called to execute the heavyweight operator to extract the complete feature set for feature fusion.

5. The method for live detection of transformer winding insulation faults according to claim 1, characterized in that, The sensor resource preference index includes the acoustic channel preference index, the vibration channel preference index, and the computing unit preference index; The priority value of the signal task slice is calculated by combining the predicted computational complexity level and the sensor resource preference index, and the signal task slice is dynamically allocated to the matching processing channel via the task slice waiting queue, including: Convert the predicted computational complexity level into a numerical score; The acoustic channel preference index, the vibration channel preference index, and the computing unit preference index are weighted and summed using preset local weighting coefficients to obtain a scalar preference score. The priority value is obtained by weighting the numerical score and the scalar preference score using a preset global weight coefficient. The task fragment waiting queue is a multi-level hash table structure. Storing the signal task fragments into the task fragment waiting queue includes: Using the predicted computational complexity level as the first-level key, the signal task fragments are mapped to the first-level hash buckets corresponding to the first-level hash table; The value range of the sensor resource preference index is discretized into multiple level intervals. The combination of the level intervals is used as the second-level key. A secondary mapping is performed in the first-level hash bucket, and the signal task is stored in the corresponding sub-hash bucket.

6. The method for live detection of transformer winding insulation faults according to claim 1, characterized in that, The upper limit threshold and the corresponding lower limit threshold constitute a dynamic load window; the preset and update process of the upper limit threshold includes: Obtain time-series data of the system load level index under historical workloads, and extract the high percentile of the time-series data as the initial upper limit threshold; The upper limit threshold is adaptively updated using an exponentially weighted moving average algorithm. Extract the high-segment statistical mean of the system load level index within a preset time window as a reference benchmark; The updated upper limit threshold is obtained by weighting and summing the current upper limit threshold with the reference benchmark using preset weighting coefficients. And maintain a preset hysteresis interval between the upper limit threshold and the lower limit threshold.

7. The method for live detection of transformer winding insulation faults according to claim 4, characterized in that, The pre-trained fault classification model is a hierarchical neural network model containing multiple network layers connected in sequence; the hierarchical neural network model has an emergency inference exit in the Nth network layer and a final inference exit in the last network layer; The joint feature vector is input into a fault classification model with multiple inference exits, and the corresponding inference exit is triggered according to the computing power limitation of the processing channel to obtain the insulation fault mode identification result of the transformer winding, including: When the signal task slice is allocated to the resource optimization channel, the joint feature vector is input into the hierarchical neural network model. After the calculation of the first N network layers is completed, the subsequent calculation is terminated, and the insulation fault mode identification result of the transformer winding is output by the emergency inference output. When the signal task slice is assigned to the high-precision parallel processing channel or the standard processing channel, the joint feature vector is input into the hierarchical neural network model and all the network layers are traversed and calculated. The insulation fault mode identification result of the transformer winding is output by the final inference output. The hierarchical neural network model includes a feature alignment layer between the Nth and N+1th network layers. The hierarchical neural network model is trained based on a joint loss function, which includes a first error calculated based on the prediction result of the emergency inference exit and a second error calculated based on the prediction result of the final inference exit.

8. A transformer winding insulation fault live-line detection system, operating the transformer winding insulation fault live-line detection method as described in any one of claims 1-7, characterized in that, include: The multimodal signal segmentation construction module is used to acquire acoustic sensor signals and vibration sensor signals during transformer operation, perform initial feature extraction, and construct signal task segments. The resource demand descriptor generation module is used to obtain a resource demand descriptor containing the prediction computational complexity level and the sensor resource preference index based on the signal task segmentation and a pre-trained resource demand prediction model. The system load status assessment module is used to monitor and obtain the hardware resource status in real time and calculate the current system load level index. The resource-coupled task sharding adaptive scheduling module is used to compare the system load level index with a preset upper limit threshold. In response to the system load level index exceeding the upper limit threshold, the priority value of the signal task slice is calculated by combining the predicted computational complexity level and the sensor resource preference index, and the signal task slice is dynamically allocated to the matching processing channel via the task slice waiting queue. The operator-level feature extraction path dynamic reconstruction module is used to slice the signal task allocated to the matching processing channel, dynamically configure the calling logic of the feature extraction operator according to the computing power limit of the matching processing channel, perform deep feature extraction and fusion, and obtain a joint feature vector. The multi-exit adaptive inference decision module is used to obtain the insulation fault mode identification result of the transformer winding by using a fault classification model with multiple inference exits based on the joint feature vector and triggering the corresponding inference exit according to the computing power limit of the processing channel.

9. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that runs on the processor, and when the processor executes the program, it implements the steps of a method for live detection of transformer winding insulation faults as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when read and executed by a processor, implement the steps of a method for live detection of transformer winding insulation faults as described in any one of claims 1 to 7.

Images