An intelligent switch fault detection and self-diagnosis method based on deep learning

Abstract

This invention discloses a deep learning-based method for intelligent switch fault detection and self-diagnosis, comprising the following steps: collecting various state data of the intelligent switch during operation, such as voltage, current, temperature, and control signals, and constructing them into a time-series input sequence with a unified time step; introducing a neural differential modeling network to continuously model the input sequence and generate state change trajectories; processing using a recursive structure with key information extraction capabilities to extract time-series features; constructing a hybrid network structure containing continuous modeling and time-series processing, and jointly training it to form a mapping from input to fault features; setting a dynamic adjustment mechanism during the modeling process to automatically adjust the calculation step size according to input changes, improving model response efficiency; inputting the extracted key features into a classification model to output fault category and level; and adjusting modeling parameters based on classification results to achieve an adaptive detection and diagnosis process.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment monitoring and fault diagnosis technology, and in particular to an intelligent switch fault detection and self-diagnosis method based on deep learning. Background Technology

[0002] With the development of intelligent power systems, smart switches have been widely used in the monitoring and control of distribution networks. The safety and stability of their operation directly affect the overall reliability of the system. However, under complex operating conditions and dynamic load environments, smart switches are prone to intermittent, latent, or early-stage fault modes, posing challenges to operation and maintenance. In particular, achieving high-precision fault detection and effective self-diagnosis of smart switches within a data-driven monitoring framework has become an important research direction in the field of intelligent operation and maintenance.

[0003] In existing technologies, fault detection and diagnosis methods for intelligent switches mostly rely on rule-based models, statistical analysis methods, or traditional machine learning models based on expert knowledge. These methods have significant limitations when dealing with complex time-series characteristics and dynamic fault evolution processes, mainly in the following aspects: 1. Insufficient ability to model continuous state evolution: Existing methods are generally based on discrete time modeling, which makes it difficult to capture the continuous dynamic characteristics in the fault development process, resulting in detection lag or misjudgment.

[0004] 2. Limited feature extraction capability: Traditional methods mostly rely on predefined features or fixed window processing, which makes it difficult to automatically extract deep-seated key fault modes from time series data, thus limiting the model's adaptability to diverse fault types.

[0005] 3. Weak real-time performance and edge deployment capabilities of models: In edge computing scenarios, traditional models have high computational complexity and lack dynamic control mechanisms for computational efficiency, making it difficult to meet the needs of real-time fault diagnosis.

[0006] 4. Lack of end-to-end adaptive mechanism: Existing methods are mostly phased processing flows, with breakpoints in information transmission. They cannot dynamically adjust sampling parameters and model inference configuration based on detection results, which limits the diagnostic closed-loop capability.

[0007] Therefore, how to provide a deep learning-based intelligent switch fault detection and self-diagnosis method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a deep learning-based intelligent switch fault detection and self-diagnosis method. This invention employs a neural differential modeling network to perform continuous state modeling on multi-dimensional time-series input data, and combines a gated recursive structure to extract key time-series features. It constructs a hybrid neural network architecture to achieve end-to-end joint training, and introduces an inference rate adjustment strategy to dynamically control the numerical solution step size. Based on the classification output results, it adjusts the modeling configuration parameters, which has the advantages of high modeling accuracy, fast response speed, and strong adaptability.

[0009] A method for fault detection and self-diagnosis of intelligent switches based on deep learning according to an embodiment of the present invention includes the following steps: S1. Collect multi-dimensional state data of the smart switch during operation. The multi-dimensional state data includes voltage data, current data, temperature data and control signal data, and construct it into a multivariable time-series input sequence with a unified time step. S2. Input the multivariate time-series input sequence into the neural differential modeling network, and use the improved neural ordinary differential equation to perform continuous-time modeling of the multivariate time-series input sequence to generate continuous state trajectories; S3. Input the continuous state trajectory into the gated recursive unit structure, and use the update gate mechanism and reset gate mechanism to perform time-series processing on the continuous state trajectory and extract key time-series features. S4. Construct a hybrid neural network architecture that includes a neural differential modeling network and a gated recursive unit structure, and perform joint training to establish a mapping relationship from input to features; S5. Set an inference rate adjustment strategy in the neural differential modeling network to dynamically adjust the numerical solution step size of the neural ordinary differential equation based on the difference between adjacent time step data in the multidimensional time series input data. S6. Input the key time-series features into the multi-label fault classification network and output the fault category, fault confidence and fault level label of the smart switch. S7. Adjust the solution parameters of the neural differential modeling network based on the feedback from the output of the multi-label fault classification network.

[0010] Optionally, S1 specifically includes: periodically collecting voltage data, current data, temperature data, and control signal data during operation through voltage sensors, current transformers, temperature sensors, and control signal acquisition interfaces installed at key parts of the smart switch, with the sampling interval set to a fixed time step; aligning the above different types of raw data based on a unified timestamp; generating multidimensional state data arranged in a time sequence after synchronization processing; cleaning the synchronized multidimensional state data, including missing data completion, outlier removal, and data standardization transformation, so that the state data of each dimension has a unified dimension and numerical range; segmenting the continuously collected data according to a sliding window method based on a set fixed time window length; and combining each segment of data into a multivariable time-series input sequence according to the time sequence.

[0011] Optionally, S2 specifically includes: S21. Construct a neural differential modeling network. The neural differential modeling network includes an input encoding layer, an improved neural ordinary differential equation modeling unit, and a state decoding layer. The input encoding layer is used to receive multivariable time-series input sequences and encode them into initial state vectors. The solution configuration parameters of the neural differential modeling network are set. The solution configuration parameters include the initial numerical integration step size, the time resolution range, and the dynamic step size adjustment enable state. S22. Input the initial state vector into the improved neural ordinary differential equation modeling unit. The improved neural ordinary differential equation modeling unit is used to process the initial state vector based on the continuous-time modeling mechanism and establish the continuous-time evolution path of the state changing with time. S23. Discrete-time solution of the improved neural ordinary differential equation modeling unit is performed by using a discrete solution strategy. An adjustable time step parameter is set and time-step integration is performed to obtain the state nodes corresponding to multiple consecutive time points and form a continuous state trajectory. S24. Input the continuous state trajectory into the state decoding layer. The state decoding layer is used to restore the continuous state trajectory into an output sequence with the same time step as the multivariable time series input sequence. S25. During the training process, the parameters of the input encoding layer, the improved neural ordinary differential equation modeling unit, and the state decoding layer in the neural differential modeling network are jointly optimized and trained so that the continuous state trajectory can express the dynamic evolution process of the multivariable temporal input sequence.

[0012] Optionally, the improved neural ordinary differential equation specifically includes: Based on the data difference between adjacent time steps in the multivariate time series input sequence, the state change amplitude in each continuous time interval is calculated, and the state change amplitude is used to construct a step size adjustment factor that corresponds one-to-one with the time interval. Using the step size adjustment factor, the current state vector, and the current time variable as inputs, a state change function is constructed, which is defined as follows: ; in, For state vectors, For time variables, This is the step size adjustment factor. For neural network parameters; During numerical integration, the numerical solution step size for state solution is dynamically set based on the value of the step size adjustment factor, and the integration calculation is performed in intervals to generate state solutions at multiple time points. The state solutions obtained by solving the step size adjustment factor are combined into a continuous state trajectory and input into the state decoding layer to form a time step output sequence consistent with the multivariate time series input sequence.

[0013] Optionally, S3 specifically includes: S31. Input the continuous state trajectory into the gated recursive unit structure in time sequence. The gated recursive unit structure consists of multiple time-step recursive processing units. Each recursive processing unit includes an update gate mechanism and a reset gate mechanism. S32. At each time step, extract the state vector corresponding to the current time step and the hidden state vector of the previous time step, and select two state vectors that are not corresponding to the current time step from the continuous state trajectory, and use them as the forward auxiliary state vector and the backward auxiliary state vector respectively. S33. Take the state vector corresponding to the current time step, the hidden state vector of the previous time step, the forward auxiliary state vector, and the backward auxiliary state vector as joint inputs and input them into the update gate mechanism to calculate the update gate control parameters. S34. Input the state vector corresponding to the current time step and the hidden state vector of the previous time step into the reset gate mechanism, calculate the reset gate control parameters, and generate candidate state representations based on the reset gate control parameters; S35. The control parameters based on the update gate mechanism are used to perform weighted fusion of the candidate state representation and the hidden state vector of the previous time step to generate the hidden state vector output of the current time step. S36. Combine the hidden state vector outputs of all time steps in chronological order to form key temporal features.

[0014] Optionally, S4 specifically includes: S41. Construct a hybrid neural network architecture, which includes a neural differential modeling network and a gated recursive unit structure. The output of the neural differential modeling network and the input of the gated recursive unit structure are sequentially connected to establish an information transmission path. S42. The continuous state trajectory output by the neural differential modeling network is used as an intermediate feature representation and input into the gated recursive unit structure. At the same time, the multivariate time-series input sequence is subjected to dimensionality compression and the compressed multivariate time-series input sequence is input into the gated recursive unit structure in parallel. S43. In the gated recursive unit structure, the continuous state trajectory and the compressed multivariate time-series input sequence are aligned according to the time step dimension, and joint fusion processing is performed in each time step to generate a fused feature representation. S44. Set a joint training objective function, which includes an output error loss term for fused feature representation, a regularization term for the neural differential modeling network, and a regularization term for the gated recursive unit structure. S45. During the training process, the multivariate temporal input sequence is sequentially fed into the neural differential modeling network and the gated recursive unit structure through forward propagation to generate the output and calculate the loss value of the joint training objective function. S46. Based on the loss value of the joint training objective function, the trainable parameters in the neural differential modeling network and the trainable parameters in the gated recursive unit structure are updated simultaneously using backpropagation to complete the joint training process of the hybrid neural network architecture.

[0015] Optionally, S5 specifically includes: S51. Set an inference rate adjustment strategy, which includes a state difference calculation step and a step size adjustment rule. The inference rate adjustment strategy is embedded in the numerical integration process of the neural differential modeling network. S52. Perform a sliding window traversal on the multivariate time series input data, calculate the state vector difference between the data of each adjacent time step, and obtain the state change amplitude index corresponding to each time interval. S53. Based on the state change amplitude index, call the step size adjustment rule to generate the numerical solution step size for each time interval. The step size adjustment rule is a mapping function, with the state change amplitude index as the input and the corresponding numerical solution step size as the output. S54. Distribute the generated numerical solution step size to the solution nodes of the neural differential modeling network in the corresponding time interval, and control the numerical integration granularity of the neural ordinary differential equation in the corresponding time interval. S55. During continuous-time modeling, the above numerical integration steps are dynamically executed in chronological order, so that the neural differential modeling network has an adjustable numerical solution step size in different state change intervals, thereby completing the inference rate adjustment process based on the characteristics of input data changes.

[0016] Optionally, S6 specifically includes: constructing a multi-label fault classification network, wherein the multi-label fault classification network includes an input layer for receiving key temporal features, a shared feature extraction layer for extracting general discrimination information, and multiple output branches for predicting fault category, fault confidence, and fault level labels, respectively. The key temporal features are expanded by time steps and then input into the multi-label fault classification network. The shared feature extraction layer extracts intermediate feature representations, and the intermediate feature representations are input into each output branch to obtain the corresponding fault category label, fault confidence value, and fault level label output results.

[0017] Optionally, S7 specifically includes: using the fault category label, fault confidence value, and fault level label output by the multi-label fault classification network as feedback input, constructing a feedback mapping strategy, mapping the fault level label and fault confidence value in the feedback input as adjustment factors for the solution configuration parameters of the neural differential modeling network, and applying the adjustment factors to the solution configuration parameters of the neural differential modeling network, such as the inference step size, the upper and lower limits of the inference rate, and the differential solution tolerance.

[0018] The beneficial effects of this invention are: (1) This invention constructs a hybrid neural network architecture that integrates neural differential modeling network and gated recursive unit structure, introduces an improved neural ordinary differential equation modeling mechanism, realizes continuous time modeling and dynamic evolution process description of the operating state of intelligent switch, can capture small temporal changes, improve the ability to identify intermittent faults and early faults, and optimize the mapping efficiency and accuracy from input sequence to key temporal features through joint training strategy.

[0019] (2) By setting an inference rate adjustment strategy, the present invention dynamically adjusts the solution step size of the neural differential modeling network based on the state differences of adjacent time steps in the multivariate time-series input data, thereby enhancing the real-time inference and resource adaptability of the model in the edge device environment. At the same time, the solution configuration parameters of the neural differential modeling network are adjusted based on the feedback of the output of the multi-label fault classification network, thereby realizing the closed-loop adaptive modeling and diagnosis process, and effectively supporting the continuous detection and self-diagnosis optimization of multiple types of intelligent switch faults during operation.

[0020] (3) The present invention constructs a deep learning framework for the entire process from operation data acquisition, multivariate modeling, feature extraction, fault classification to model self-adjustment, which improves the system’s ability to understand and predict the entire process of fault development and significantly enhances the autonomous operation and maintenance capability and operational safety of the intelligent switch system under complex working conditions. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based intelligent switch fault detection and self-diagnosis method proposed in this invention; Figure 2 This is a schematic diagram of the hybrid modeling process based on neural differential modeling network and gated recursive unit structure proposed in this invention; Figure 3 This is a flowchart of a dynamic step size adjustment mechanism for inference rate regulation proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figures 1-3 A deep learning-based intelligent switch fault detection and self-diagnosis method includes the following steps: S1. Collect multi-dimensional state data of the smart switch during operation. The multi-dimensional state data includes voltage data, current data, temperature data and control signal data, and construct it into a multivariable time-series input sequence with a unified time step. S2. Input the multivariate time-series input sequence into the neural differential modeling network, and use the improved neural ordinary differential equation to perform continuous-time modeling of the multivariate time-series input sequence to generate continuous state trajectories; S3. Input the continuous state trajectory into the gated recursive unit structure, and use the update gate mechanism and reset gate mechanism to perform time-series processing on the continuous state trajectory and extract key time-series features. S4. Construct a hybrid neural network architecture that includes a neural differential modeling network and a gated recursive unit structure, and perform joint training to establish a mapping relationship from input to features; S5. Set an inference rate adjustment strategy in the neural differential modeling network to dynamically adjust the numerical solution step size of the neural ordinary differential equation based on the difference between adjacent time step data in the multidimensional time series input data. S6. Input the key time-series features into the multi-label fault classification network and output the fault category, fault confidence and fault level label of the smart switch. S7. Adjust the solution parameters of the neural differential modeling network based on the feedback from the output of the multi-label fault classification network.

[0024] In this embodiment, S1 specifically includes: periodically collecting voltage data, current data, temperature data, and control signal data during operation through voltage sensors, current transformers, temperature sensors, and control signal acquisition interfaces installed at key parts of the smart switch, with the sampling interval set to a fixed time step; aligning the above-mentioned different types of raw data based on a unified timestamp; forming multidimensional state data arranged in a time sequence after synchronization processing; cleaning the synchronized multidimensional state data, including missing data completion, outlier removal, and data standardization transformation, so that the state data of each dimension has a unified dimension and numerical range; segmenting the continuously collected data according to a sliding window method based on a set fixed time window length; and combining each segment of data into a multivariable time-series input sequence according to the time sequence.

[0025] This implementation deploys voltage sensors, current transformers, temperature sensors, and control signal acquisition interfaces at key components of the smart switch to periodically collect various state data during operation. Synchronization and alignment based on a unified time step ensure consistency across all data types over time. Subsequently, missing value completion, outlier removal, and standardization improve the quality and usability of the multidimensional state data. Finally, a fixed-length sliding window mechanism is used to segment the continuously cleaned data into standardized multivariate time-series input sequences, providing consistent, comparable, and structured input for subsequent models. This enhances the model's ability to perceive state evolution characteristics and provides a high-quality raw data foundation for smart switch fault detection and self-diagnosis.

[0026] In this embodiment, S2 specifically includes: S21. Construct a neural differential modeling network. The neural differential modeling network includes an input encoding layer, an improved neural ordinary differential equation modeling unit, and a state decoding layer. The input encoding layer is used to receive multivariable time-series input sequences and encode them into initial state vectors. The solution configuration parameters of the neural differential modeling network are set. The solution configuration parameters include the initial numerical integration step size, the time resolution range, and the dynamic step size adjustment enable state. S22. Input the initial state vector into the improved neural ordinary differential equation modeling unit. The improved neural ordinary differential equation modeling unit is used to process the initial state vector based on the continuous-time modeling mechanism and establish the continuous-time evolution path of the state changing with time. S23. Discrete-time solution of the improved neural ordinary differential equation modeling unit is performed by using a discrete solution strategy. An adjustable time step parameter is set and time-step integration is performed to obtain the state nodes corresponding to multiple consecutive time points and form a continuous state trajectory. S24. Input the continuous state trajectory into the state decoding layer. The state decoding layer is used to restore the continuous state trajectory into an output sequence with the same time step as the multivariable time series input sequence. S25. During the training process, the parameters of the input encoding layer, the improved neural ordinary differential equation modeling unit, and the state decoding layer in the neural differential modeling network are jointly optimized and trained so that the continuous state trajectory can express the dynamic evolution process of the multivariable temporal input sequence.

[0027] This implementation constructs a neural differential modeling network comprising an input encoding layer, an improved neural ordinary differential equation modeling unit, and a state decoding layer. First, it receives a multivariate temporal input sequence and encodes it into an initial state vector, initializing it with predefined solution configuration parameters. Then, it generates an evolution path of the state over time based on a continuous-time modeling mechanism. A discrete solution strategy is used to perform time-step integration, obtaining a series of state nodes forming a continuous state trajectory. The state decoding layer then restores this trajectory to a temporal output sequence consistent with the original input. During training, the network effectively establishes a mapping relationship between multivariate temporal data and its dynamic evolution process by jointly optimizing the parameters of the encoding layer, modeling unit, and decoding layer. This achieves continuous modeling and efficient representation of the state changes of intelligent switches, laying the foundation for subsequent temporal feature extraction and fault classification.

[0028] In this embodiment, the improved neural ordinary differential equation specifically includes: Based on the data difference between adjacent time steps in the multivariate time series input sequence, the state change amplitude in each continuous time interval is calculated, and the state change amplitude is used to construct a step size adjustment factor that corresponds one-to-one with the time interval. Using the step size adjustment factor, the current state vector, and the current time variable as inputs, a state change function is constructed, which is defined as follows: ; in, For state vectors, For time variables, This is the step size adjustment factor. For neural network parameters; During numerical integration, the numerical solution step size for state solution is dynamically set based on the value of the step size adjustment factor, and the integration calculation is performed in intervals to generate state solutions at multiple time points. The state solutions obtained by solving the step size adjustment factor are combined into a continuous state trajectory and input into the state decoding layer to form a time step output sequence consistent with the multivariate time series input sequence.

[0029] This implementation constructs a step-size adjustment factor based on the data difference between adjacent time steps in a multivariate time-series input sequence and introduces it into the state change function to achieve dynamic adjustment of the step size for solving neural ordinary differential equations. By constructing the state change function using the step-size adjustment factor, state vector, and time variable as inputs, and dynamically setting the solution step size according to the numerical integration process based on the value of the step-size adjustment factor, the state evolution path has flexible resolution capabilities in different time intervals, thereby forming a continuous state trajectory that more closely matches the actual state change patterns. This mechanism effectively enhances the adaptability of the neural differential modeling network to modeling different dynamic modes, and by outputting a modeling result consistent with the time structure of the original input sequence through the state decoding layer, it achieves efficient modeling and representation of complex time-series data.

[0030] In this embodiment, S3 specifically includes: S31. Input the continuous state trajectory into the gated recursive unit structure in time sequence. The gated recursive unit structure consists of multiple time-step recursive processing units. Each recursive processing unit includes an update gate mechanism and a reset gate mechanism. S32. At each time step, extract the state vector corresponding to the current time step and the hidden state vector of the previous time step, and select two state vectors that are not corresponding to the current time step from the continuous state trajectory, and use them as the forward auxiliary state vector and the backward auxiliary state vector respectively. S33. Take the state vector corresponding to the current time step, the hidden state vector of the previous time step, the forward auxiliary state vector, and the backward auxiliary state vector as joint inputs and input them into the update gate mechanism to calculate the update gate control parameters. S34. Input the state vector corresponding to the current time step and the hidden state vector of the previous time step into the reset gate mechanism, calculate the reset gate control parameters, and generate candidate state representations based on the reset gate control parameters; S35. The control parameters based on the update gate mechanism are used to perform weighted fusion of the candidate state representation and the hidden state vector of the previous time step to generate the hidden state vector output of the current time step. S36. Combine the hidden state vector outputs of all time steps in chronological order to form key temporal features.

[0031] In this embodiment, deep processing of time-series data is achieved by inputting continuous state trajectories sequentially into a gated recursive unit structure composed of update and reset gate mechanisms. Forward and backward auxiliary state vectors are introduced at each time step to participate in the calculation of update gate control parameters, enabling the gate mechanism to dynamically update not only based on local states but also by integrating the global temporal context. Candidate state representations are calculated through the reset gate mechanism and combined with the update gate mechanism to complete the fusion generation of the hidden state at the current time step. Finally, the hidden state vectors output from all time steps are combined into key temporal features. This method effectively enhances the modeling capability of the state evolution process, making the extracted temporal features more discriminative and expressive, providing solid foundational data support for subsequent intelligent switch fault identification and diagnosis.

[0032] In this embodiment, S4 specifically includes: S41. Construct a hybrid neural network architecture, which includes a neural differential modeling network and a gated recursive unit structure. The output of the neural differential modeling network and the input of the gated recursive unit structure are sequentially connected to establish an information transmission path. S42. The continuous state trajectory output by the neural differential modeling network is used as an intermediate feature representation and input into the gated recursive unit structure. At the same time, the multivariate time-series input sequence is subjected to dimensionality compression and the compressed multivariate time-series input sequence is input into the gated recursive unit structure in parallel. S43. In the gated recursive unit structure, the continuous state trajectory and the compressed multivariate time-series input sequence are aligned according to the time step dimension, and joint fusion processing is performed in each time step to generate a fused feature representation. S44. Set a joint training objective function, which includes an output error loss term for fused feature representation, a regularization term for the neural differential modeling network, and a regularization term for the gated recursive unit structure. S45. During the training process, the multivariate temporal input sequence is sequentially fed into the neural differential modeling network and the gated recursive unit structure through forward propagation to generate the output and calculate the loss value of the joint training objective function. S46. Based on the loss value of the joint training objective function, the trainable parameters in the neural differential modeling network and the trainable parameters in the gated recursive unit structure are updated simultaneously using backpropagation to complete the joint training process of the hybrid neural network architecture.

[0033] This implementation constructs a hybrid neural network architecture combining a neural differential modeling network and a gated recursive unit structure. It jointly fuses the continuous state trajectory generated by the former with the compressed multivariate temporal input sequence, fully preserving the characteristics of continuous modeling and temporal memory. It extracts features by fusing them step-by-step using a feature alignment mechanism. Based on a joint training objective function, it generates a predicted output and calculates the loss value through forward propagation. Combined with the backpropagation process, it simultaneously optimizes the trainable parameters of the two types of sub-networks, thereby achieving joint modeling capabilities with structural synergy and feature complementarity. This helps improve the model's ability to recognize complex state evolution patterns and its classification accuracy.

[0034] In this embodiment, S5 specifically includes: S51. Set an inference rate adjustment strategy, which includes a state difference calculation step and a step size adjustment rule. The inference rate adjustment strategy is embedded in the numerical integration process of the neural differential modeling network. S52. Perform a sliding window traversal on the multivariate time series input data, calculate the state vector difference between the data of each adjacent time step, and obtain the state change amplitude index corresponding to each time interval. S53. Based on the state change amplitude index, call the step size adjustment rule to generate the numerical solution step size for each time interval. The step size adjustment rule is a mapping function, with the state change amplitude index as the input and the corresponding numerical solution step size as the output. S54. Distribute the generated numerical solution step size to the solution nodes of the neural differential modeling network in the corresponding time interval, and control the numerical integration granularity of the neural ordinary differential equation in the corresponding time interval. S55. During continuous-time modeling, the above numerical integration steps are dynamically executed in chronological order, so that the neural differential modeling network has an adjustable numerical solution step size in different state change intervals, thereby completing the inference rate adjustment process based on the characteristics of input data changes.

[0035] This implementation introduces an inference rate adjustment strategy into the neural differential modeling network, combining state difference calculation and step size adjustment rules to achieve dynamic adjustment of the numerical integration step size. Specifically, it involves: using a sliding window to traverse data from adjacent time steps, calculating the state vector difference and generating a state change magnitude index; based on this index, generating the solution step size for the corresponding time interval through a mapping function; and mapping the step size to the solution nodes of the neural differential modeling network to achieve dynamic adjustment of the integration granularity during continuous-time modeling. This mechanism enables the model to flexibly adjust the modeling analytical accuracy according to the rate of change of the input state, enhancing the model's adaptability to dynamic sequence features without introducing additional complex structures, and contributing to improved stability and representational ability in fault evolution modeling.

[0036] In this embodiment, S6 specifically includes: constructing a multi-label fault classification network, which includes an input layer for receiving key temporal features, a shared feature extraction layer for extracting general discrimination information, and multiple output branches for predicting fault category, fault confidence, and fault level labels, respectively. The key temporal features are expanded by time steps and then input into the multi-label fault classification network. The shared feature extraction layer extracts intermediate feature representations, and the intermediate feature representations are input into each output branch to obtain the corresponding fault category label, fault confidence value, and fault level label output results.

[0037] This implementation constructs a multi-label fault classification network comprising an input layer, a shared feature extraction layer, and multiple output branches. Key temporal features are expanded into the input model step by step, and intermediate feature representations are uniformly extracted using the shared feature extraction layer. These representations are then fed into the output branches for fault category, fault confidence, and fault level labels, respectively. This enables multi-dimensional discrimination and classification of smart switch faults, effectively separating diagnostic tasks of different dimensions. Consequently, the clarity and reliability of fault detection results are improved, enhancing the diagnostic system's ability to characterize and identify complex fault states.

[0038] In this embodiment, S7 specifically includes: using the fault category label, fault confidence value, and fault level label output by the multi-label fault classification network as feedback input, constructing a feedback mapping strategy, mapping the fault level label and fault confidence value in the feedback input as adjustment factors for the solution configuration parameters of the neural differential modeling network, and applying the adjustment factors to the solution configuration parameters of the neural differential modeling network, such as the inference step size, the upper and lower limits of the inference rate, and the differential solution tolerance.

[0039] This implementation constructs a feedback mapping strategy by using the fault category label, fault confidence value, and fault level label output by the multi-label fault classification network as feedback inputs. This strategy is then used to dynamically adjust the solution configuration parameters of the neural differential modeling network, specifically including the inference step size, the upper and lower limits of the inference rate, and the differential solution tolerance. This feedback mechanism enables the modeling network to adaptively adjust the computational accuracy and inference rate during subsequent modeling processes based on fault results of different levels and confidence levels. This achieves a more flexible continuous-time modeling scheduling strategy, effectively enhancing the model's sensitivity and responsiveness to changes in fault characteristics.

[0040] Example 1: To verify the feasibility of this invention in practice, it was applied to the low-voltage switch monitoring system of a 220kV substation under a State Grid Corporation of China. This substation is equipped with 60 intelligent switches, operating in a complex environment with frequent load fluctuations, demanding extremely high equipment stability and maintenance response speed. Traditional fault detection methods mainly rely on periodic inspections and threshold alarm mechanisms, which suffer from problems such as detection delays, frequent false alarms and missed alarms, and long response times, making it difficult to meet the requirements for real-time, high-precision fault diagnosis.

[0041] The system collected multi-dimensional state data samples over a period of 30 days, covering four key monitoring dimensions: voltage, current, temperature, and control signals. The initial data sampling frequency was set to 10Hz. The collected data was input into a hybrid neural network architecture composed of a neural differential modeling network and a gated recursive unit structure constructed in this invention. This method enables dynamic modeling of the operating state of switching equipment, extraction of key temporal features, and multi-label fault identification. The system automatically adjusts the model inference rate in real time, dynamically adjusting the solution step size and tolerance threshold of the neural differential modeling network based on temporal data fluctuations. In the fault classification output stage, the system feeds back the diagnostic results to adjust the inference configuration parameters, thereby improving modeling accuracy and system stability.

[0042] During operation, the system detected 38 instances of abnormal switching behavior, including five typical fault types such as poor contact, abnormal electromagnetic coil drive, and control signal failure. All diagnostic events were alerted within 5 seconds of fault occurrence, and the system automatically output fault confidence and fault level labels, while updating the differential solution granularity configuration within the model. This mechanism significantly improved the system's ability to identify transient abnormal states and its response rate, reducing the diagnostic inference latency from an average of 7.8 seconds to 2.1 seconds.

[0043] To verify the performance advantages of this method, the project team conducted a parallel comparison with the traditional diagnostic mechanism based on fixed threshold + time window clustering, and collected key performance indicators such as diagnostic accuracy, response time, false alarm rate, and model stability, which are summarized in the following table: Table 1 Performance Comparison of Intelligent Switch Fault Diagnosis Systems

[0044] The data shows that, compared with traditional methods, the method described in this invention improves fault identification accuracy by 11.4 percentage points, reduces response time by 72.9%, and lowers false alarm rate and false negative rate by 2.7% and 4.3%, respectively. Particularly in terms of inference rate adjustment capability, traditional methods have an average daily adjustment frequency of 0, while this method can achieve more than 9 dynamic step-size adjustments per day, effectively adapting to the fluctuations of the operating environment, and improving overall system stability by 22.4%.

[0045] In summary, this embodiment fully demonstrates that the method proposed in this invention can efficiently and stably realize fault detection and self-diagnosis of intelligent switches, and has advantages such as high accuracy, fast response, and adaptive adjustment. It is suitable for intelligent maintenance scenarios of large-scale power distribution automation systems and can be extended to the status monitoring and health assessment of other power intelligent equipment.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent switch fault detection and self-diagnosis based on deep learning, characterized in that, Includes the following steps: S1. Collect multi-dimensional state data of the smart switch during operation. The multi-dimensional state data includes voltage data, current data, temperature data and control signal data, and construct it into a multivariable time-series input sequence with a unified time step. S2. Input the multivariate time-series input sequence into the neural differential modeling network, and use the improved neural ordinary differential equation to perform continuous-time modeling of the multivariate time-series input sequence to generate continuous state trajectories; S3. Input the continuous state trajectory into the gated recursive unit structure, and use the update gate mechanism and reset gate mechanism to perform time-series processing on the continuous state trajectory and extract key time-series features. S4. Construct a hybrid neural network architecture that includes a neural differential modeling network and a gated recursive unit structure, and perform joint training to establish a mapping relationship from input to features; S5. Set an inference rate adjustment strategy in the neural differential modeling network to dynamically adjust the numerical solution step size of the neural ordinary differential equation based on the difference between adjacent time step data in the multidimensional time series input data. S6. Input the key time-series features into the multi-label fault classification network and output the fault category, fault confidence and fault level label of the smart switch. S7. Adjust the solution parameters of the neural differential modeling network based on the feedback from the output of the multi-label fault classification network.

2. The method for intelligent switch fault detection and self-diagnosis based on deep learning according to claim 1, characterized in that, S1 specifically includes: periodically collecting voltage data, current data, temperature data, and control signal data during operation through voltage sensors, current transformers, temperature sensors, and control signal acquisition interfaces installed at key parts of the smart switch, with the sampling interval set to a fixed time step; aligning the above different types of raw data based on a unified timestamp; generating multi-dimensional state data arranged in a time sequence after synchronization processing; cleaning the synchronized multi-dimensional state data, including missing data completion, outlier removal, and data standardization transformation, so that the state data of each dimension has a unified dimension and numerical range; segmenting the continuously collected data according to a sliding window method based on a set fixed time window length; and combining each segment of data into a multivariable time-series input sequence according to the time order.

3. The method for intelligent switch fault detection and self-diagnosis based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Construct a neural differential modeling network. The neural differential modeling network includes an input encoding layer, an improved neural ordinary differential equation modeling unit, and a state decoding layer. The input encoding layer is used to receive multivariable time-series input sequences and encode them into initial state vectors. The solution configuration parameters of the neural differential modeling network are set. The solution configuration parameters include the initial numerical integration step size, the time resolution range, and the dynamic step size adjustment enable state. S22. Input the initial state vector into the improved neural ordinary differential equation modeling unit. The improved neural ordinary differential equation modeling unit is used to process the initial state vector based on the continuous-time modeling mechanism and establish the continuous-time evolution path of the state changing with time. S23. The improved neural ordinary differential equation modeling unit is solved in discrete time by using a discrete solution strategy. An adjustable time step parameter is set and time step integration is performed to obtain the state nodes corresponding to multiple consecutive time points and form a continuous state trajectory. S24. Input the continuous state trajectory into the state decoding layer. The state decoding layer is used to restore the continuous state trajectory into an output sequence with the same time step as the multivariable time series input sequence. S25. During the training process, the parameters of the input encoding layer, the improved neural ordinary differential equation modeling unit, and the state decoding layer in the neural differential modeling network are jointly optimized and trained.

4. The method for intelligent switch fault detection and self-diagnosis based on deep learning according to claim 3, characterized in that, The improved neural ordinary differential equation specifically includes: Based on the data difference between adjacent time steps in the multivariate time series input sequence, the state change amplitude in each continuous time interval is calculated, and the state change amplitude is used to construct a step size adjustment factor that corresponds one-to-one with the time interval. Using the step size adjustment factor, the current state vector, and the current time variable as inputs, a state change function is constructed, which is defined as follows: ； in, For state vectors, For time variables, This is the step size adjustment factor. For neural network parameters; During numerical integration, the numerical solution step size for state solution is dynamically set based on the value of the step size adjustment factor, and the integration calculation is performed in intervals to generate state solutions at multiple time points. The state solutions obtained by solving the step size adjustment factor are combined into a continuous state trajectory and input into the state decoding layer to form a time step output sequence consistent with the multivariate time series input sequence.

5. The method for intelligent switch fault detection and self-diagnosis based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Input the continuous state trajectory into the gated recursive unit structure in time sequence. The gated recursive unit structure consists of multiple time-step recursive processing units. Each recursive processing unit includes an update gate mechanism and a reset gate mechanism. S32. At each time step, extract the state vector corresponding to the current time step and the hidden state vector of the previous time step, and select two state vectors that are not corresponding to the current time step from the continuous state trajectory, and use them as the forward auxiliary state vector and the backward auxiliary state vector respectively. S33. Take the state vector corresponding to the current time step, the hidden state vector of the previous time step, the forward auxiliary state vector, and the backward auxiliary state vector as joint inputs and input them into the update gate mechanism to calculate the update gate control parameters. S34. Input the state vector corresponding to the current time step and the hidden state vector of the previous time step into the reset gate mechanism, calculate the reset gate control parameters, and generate candidate state representations based on the reset gate control parameters; S35. The control parameters based on the update gate mechanism are used to perform weighted fusion of the candidate state representation and the hidden state vector of the previous time step to generate the hidden state vector output of the current time step. S36. Combine the hidden state vector outputs of all time steps in chronological order to form key temporal features.

6. The method for intelligent switch fault detection and self-diagnosis based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Construct a hybrid neural network architecture, which includes a neural differential modeling network and a gated recursive unit structure. The output of the neural differential modeling network and the input of the gated recursive unit structure are sequentially connected to establish an information transmission path. S42. The continuous state trajectory output by the neural differential modeling network is used as an intermediate feature representation and input into the gated recursive unit structure. At the same time, the multivariate time-series input sequence is subjected to dimensionality compression and the compressed multivariate time-series input sequence is input into the gated recursive unit structure in parallel. S43. In the gated recursive unit structure, the continuous state trajectory and the compressed multivariate time-series input sequence are aligned according to the time step dimension, and joint fusion processing is performed in each time step to generate a fused feature representation. S44. Set a joint training objective function, which includes an output error loss term for fused feature representation, a regularization term for the neural differential modeling network, and a regularization term for the gated recursive unit structure. S45. During the training process, the multivariate temporal input sequence is sequentially fed into the neural differential modeling network and the gated recursive unit structure through forward propagation to generate the output and calculate the loss value of the joint training objective function. S46. Based on the loss value of the joint training objective function, the trainable parameters in the neural differential modeling network and the trainable parameters in the gated recursive unit structure are updated simultaneously using backpropagation to complete the joint training process of the hybrid neural network architecture.

7. The method for intelligent switch fault detection and self-diagnosis based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Set an inference rate adjustment strategy, which includes a state difference calculation step and a step size adjustment rule. The inference rate adjustment strategy is embedded in the numerical integration process of the neural differential modeling network. S52. Perform a sliding window traversal on the multivariate time series input data, calculate the state vector difference between the data of each adjacent time step, and obtain the state change amplitude index corresponding to each time interval. S53. Based on the state change amplitude index, call the step size adjustment rule to generate the numerical solution step size for each time interval. The step size adjustment rule is a mapping function, with the state change amplitude index as the input and the corresponding numerical solution step size as the output. S54. Distribute the generated numerical solution step size to the solution nodes of the neural differential modeling network in the corresponding time interval, and control the numerical integration granularity of the neural ordinary differential equation in the corresponding time interval. S55. During continuous-time modeling, the above numerical integration steps are dynamically executed in chronological order, so that the neural differential modeling network has an adjustable numerical solution step size in different state change intervals.

8. The method for intelligent switch fault detection and self-diagnosis based on deep learning according to claim 1, characterized in that, S6 specifically includes: constructing a multi-label fault classification network, which includes an input layer for receiving key temporal features, a shared feature extraction layer for extracting general discrimination information, and multiple output branches for predicting fault category, fault confidence, and fault level labels, respectively. The key temporal features are expanded by time steps and then input into the multi-label fault classification network. The shared feature extraction layer extracts intermediate feature representations, and the intermediate feature representations are input into each output branch to obtain the corresponding fault category label, fault confidence value, and fault level label output results.

9. The method for intelligent switch fault detection and self-diagnosis based on deep learning according to claim 1, characterized in that, S7 specifically includes: using the fault category label, fault confidence value, and fault level label output by the multi-label fault classification network as feedback input, constructing a feedback mapping strategy, mapping the fault level label and fault confidence value in the feedback input as adjustment factors for the solution configuration parameters of the neural differential modeling network, and applying the adjustment factors to the solution configuration parameters of the neural differential modeling network, such as the inference step size, the upper and lower limits of the inference rate, and the differential solution tolerance.

Images