A solid-state transformer intelligent monitoring and diagnosis method and system based on multi-source sensor fusion

Abstract

This invention provides a method and system for intelligent monitoring and diagnosis of solid-state transformers based on multi-source sensor fusion. The method includes: Step 1, collecting multi-dimensional operating parameters, converting the collected analog signals into digital signals, and buffering them; Step 2, classifying the digital signals into two categories: high-frequency electrical fast-changing signals and thermal and mechanical slow-changing signals; Step 3, extracting fault-sensitive features from the high-frequency electrical fast-changing signals to form a first input feature set; Step 4, purifying the fault-sensitive features from the thermal and mechanical slow-changing signals to form a second input feature set; Step 5, obtaining a standardized feature set; Step 6, inputting the standardized feature set into an end-to-end fusion model to complete the fusion of multi-source fault-sensitive features and fault diagnosis. This invention enables full-dimensional online monitoring and accurate fault diagnosis of solid-state transformers, ensuring the safe and stable operation of new power systems.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring and fault diagnosis technology, specifically relating to a method and system for intelligent monitoring and diagnosis of solid-state transformers based on multi-source sensor fusion. Background Technology

[0002] Solid-state transformers (SSTs) are crucial energy routing equipment for building new power systems, and their operational reliability is paramount. SSTs integrate a large number of power semiconductor devices, high-frequency magnetic components, and passive components, and simultaneously withstand high temperatures during operation. , The electrical stress, along with the coupling effects of thermal, mechanical, and other physical field stresses, makes the fault mechanism complex.

[0003] Existing SST condition monitoring methods are mostly based on single or a few signals (such as temperature and current), failing to fully utilize the multi-source heterogeneous information generated during operation. Furthermore, SST fault signals exhibit significant differences: faults in power devices such as insulated-gate bipolar transistors (IGBTs) and partial discharge generate high-frequency transient electrical signals; while faults such as core vibration and junction temperature drift manifest as slowly changing mechanical and thermal signals. Traditional methods employ a uniform processing strategy, making it difficult to simultaneously extract transient features from rapidly changing signals and refine trend information from slowly changing signals, resulting in insufficient fault feature extraction and difficulty in identifying early, weak faults.

[0004] At the level of feature fusion and diagnosis, existing methods usually simply stitch together features extracted from different sensors, ignoring the inherent topological relationships and spatiotemporal scale differences between different physical quantities. This fails to achieve deep heterogeneous feature fusion and limits the accuracy and generalization ability of diagnostic models.

[0005] Therefore, there is an urgent need for a method that can differentiate and intelligently integrate fault signals from multiple sources in SST to achieve comprehensive perception and accurate early warning of its operating status. Summary of the Invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method and system for intelligent monitoring and diagnosis of solid-state transformers based on multi-source sensor fusion, which aims to achieve synchronous perception of the multi-dimensional operating status of SST, accurate processing and feature extraction of fast and slow differentiated signals, and high-precision fault diagnosis and location through a deep feature fusion model.

[0007] The method includes the following steps:

[0008] Step 1: Construct a distributed multi-source sensor array to simultaneously collect multi-dimensional operating parameters such as temperature, vibration, partial discharge, current, voltage, and heat dissipation from the core module of the Solid State Transformer (SST). The collected analog signals are converted into digital signals and then buffered. The core module includes a power module, a rectifier module, an inverter module, a filter module, a core module, and a heat dissipation module.

[0009] Step 2: Based on the signal acquisition sampling frequency, the digital signals obtained in Step 1 are divided into two categories: high-frequency electrical fast-changing signals and thermal and mechanical slow-changing signals. Among them, the high-frequency electrical fast-changing signals are transient electrical signals corresponding to the switching transients and high-frequency partial discharges of the Insulated Gate Bipolar Transistor (IGBT) generated during the operation of the solid-state transformer (SST), while the thermal and mechanical slow-changing signals are thermal and mechanical steady-state signals corresponding to the core vibration and temperature drift generated during the operation of the solid-state transformer (SST).

[0010] Step 3: The core parameters of Multi-Scale Adaptive Variational Singular Spectrum Decomposition (MS-AVSSD) are optimized using the Chaos Adaptive Cauchy Variation-Eagle Optimization Algorithm (CACV-EO). The optimized MS-AVSSD is then used to decompose and denoise the high-frequency electrical fast-changing signal, extracting fault-sensitive features from the signal to form the first input feature set.

[0011] Step 4: Combining Improved Adaptive Frequency Focusing Wavelet Filtering (IAFF-WF) and Adaptive Synergistic Geometric Mode Decomposition (ASGMD), the thermal and mechanical slowly varying signals are filtered and decomposed to purify the fault-sensitive features in the thermal and mechanical slowly varying signals, forming the second input feature set;

[0012] Step 5: Perform feature normalization on the first input feature set extracted in step 3 and the second input feature set purified in step 4 to obtain a standardized feature set;

[0013] Step 6: Input the standardized feature set obtained in Step 5 into the Feature Projection-Gated Channel Attention Module-Gated Modal Attention Module-Conformer (FP-GCMA-GMNA-Conformer) end-to-end fusion model. The model is used to complete the fusion of multi-source fault-sensitive features and fault diagnosis, and output the fault type, fault location and fault severity of the solid-state transformer SST.

[0014] In step 1, the distributed multi-source sensor array includes temperature sensors, vibration sensors, partial discharge sensors, current sensors, voltage sensors, and flow sensors. Each type of sensor is deployed at a corresponding characteristic monitoring point of each core module of the solid-state transformer (SST). The sampling frequency of the sensors is adapted and set according to the type of monitoring parameter. For high-frequency electrical parameters, the sampling frequency is no less than 10kHz, while for thermal and mechanical parameters, the sampling frequency is 1Hz to 10Hz. The analog signal is converted to a digital signal using an analog-to-digital converter (ADC). The conversion formula is as follows:

[0015] ,

[0016] in, The converted digital signal value, The analog signal value collected by the sensor. This represents the minimum range of the analog signal. This represents the maximum value of the analog signal range. This represents the number of bits used in the analog-to-digital conversion.

[0017] In step 2, the classification criterion for digital signals is the signal acquisition sampling frequency, and the determination formula is as follows:

[0018] ,

[0019] in, The sampling frequency for signal acquisition; This indicates high-frequency electrical fast-changing signals, corresponding to transient electrical signals generated during the operation of the solid-state transformer (SST) and the switching transients and high-frequency partial discharges of the insulated gate bipolar transistor (IGBT). This represents the thermal and mechanical slow-changing signals, corresponding to the thermal and mechanical steady-state signals generated during the operation of the solid-state transformer (SST) and the core vibration and temperature drift. The classified high-frequency electrical fast-changing signals... Let be the input signal for the subsequent feature extraction stage. Slowly varying thermal and mechanical signals This is denoted as the input signal for the subsequent feature purification stage. ,in This represents the sequence number of the discrete sampling point.

[0020] Step 3, the implementation process of the core parameters includes:

[0021] Step 3-1: Construct a multi-scale adaptive variational singular spectral decomposition (MS-AVSSD) parameter optimization sample set; determine the core parameter to be optimized as the number of decomposition scales. Punishment factor and number of iterations The objective function for optimization is the signal-to-noise ratio of the processed high-frequency electrical fast-changing signal. ;

[0022] ,

[0023] in, The original high-frequency electrical fast conversion signal Each sample value, The signal after MS-AVSSD processing for multi-scale adaptive variational singular spectrum decomposition is the first... Each sample value, This represents the total number of signal sampling points.

[0024] Step 3-2, obtain the signal-to-noise ratio. The MS-AVSSD parameter combination at maximum; defining the fitness function of the chaotic adaptive Cauchy mutation-based eagle optimization algorithm CACV-EO. :

[0025] ;

[0026] Step 3-3: Initialize the location of the vulture population using Logistic chaotic mapping, and set the population size to [value missing]. The maximum number of iterations is ;

[0027] Steps 3-4, during the vulture's search space selection phase, the position update formula is:

[0028] ,

[0029] in, For the updated position, This is the current globally optimal position. The average position of the population. Current position A random number within the range [0,1]. It is a contraction factor;

[0030] Steps 3-5, during the vulture's swooping down to capture prey, introduce adaptive Cauchy mutation, with the position update formula as follows:

[0031] ,

[0032] in, A random number within the range [0,1]. , As a learning factor, , For varying asynchronous lengths following a Cauchy distribution, the scale parameter With the number of iterations Adaptive Decreasing: ,in This indicates the maximum value of the scale parameter. This represents the minimum value of the scale parameter. Indicates the maximum number of iterations;

[0033] Steps 3-6 define the optimization decision relationship of the CACV-EO algorithm as follows:

[0034] ,

[0035] in, This represents the current iteration number of the algorithm. Let g be the fitness value of the g-th iteration. This is the convergence accuracy threshold. This indicates a logical AND operation.

[0036] Steps 3-7 involve iterative optimization in the parameter space. When the optimization criterion is satisfied, the optimal parameter combination of the multi-scale adaptive variational singular spectral decomposition MS-AVSSD is output. ,in This is the optimal decomposition scale. As the optimal penalty factor, This represents the optimal number of iterations.

[0037] In step 3, the optimal parameter combination is used. The configured multi-scale adaptive variational singular spectral decomposition MS-AVSSD for processing high-frequency electrical fast-changing signals includes the following steps:

[0038] Steps 3-8, for the fast-changing input signal Constructing the Hankel matrix ;

[0039] Steps 3-9, for Perform Singular Value Decomposition (SVD):

[0040] ,

[0041] in and It is a unitary matrix. It is a singular value matrix. The constructed Hankel matrix;

[0042] Steps 3-10, using and Construct and solve the constrained variational model of variational mode decomposition (VMD) to obtain Individual eigenmode function IMF components ;

[0043] Step 3-11: Calculate the IMF components of each intrinsic mode function. With the original signal correlation coefficient Select Greater than the threshold Component reconstruction to remove noise signal ;

[0044] Steps 3-12, for Hilbert envelope demodulation is performed, and the amplitude, variance, and kurtosis features of the fault feature frequencies in the envelope spectrum are extracted as the first input feature set.

[0045] Step 4 includes:

[0046] Step 4-1, for the slowly changing input signal The pre-filtered signal is obtained by performing improved adaptive frequency focusing wavelet filtering (IAFF-WF). :

[0047] ,

[0048] in, For the selected wavelet basis function type, The filtering threshold is adaptively calculated based on the signal characteristics;

[0049] Step 4-2, pre-filtering the signal Adaptive Co-Geometric Mode Decomposition (ASGMD) is performed to obtain a series of geometric mode components. :

[0050] ,

[0051] in, As an adaptive collaborative factor, To decompose the precision control parameters, The index of the geometric modal component;

[0052] Step 4-3: The geometric modal components are selected according to preset screening criteria. Purification: Calculate the energy percentage of each component. and With the original signal correlation coefficient Select the one that satisfies and The conditional components are reconstructed to obtain the feature-enhanced signal. ;in, This represents the energy percentage threshold. The threshold for the correlation coefficient;

[0053] Step 4-4, from the feature enhancement signal Extract time-domain statistical features to form the second input feature set.

[0054] In step 5, the first input feature set and the second input feature set are standardized using the min-max normalization method, as shown in the formula:

[0055] ,

[0056] in, These are the normalized, standardized eigenvalues. These are the original fault-sensitive characteristic values. It is the minimum value of the original feature of all samples in the same feature dimension as the original feature value. This is the maximum value of the original features of all samples within the same feature dimension as the original feature value;

[0057] In step 6, the FP-GCMA-GMNA-Conformer end-to-end fusion model performs fusion and diagnosis on the standardized feature set, specifically including the following steps:

[0058] Step 6-1: Map the normalized feature set to a high-dimensional space through the feature projection layer FP:

[0059] ,

[0060] in, To input a standardized feature set, , FC represents the number of neurons in the fully connected layer. This is the high-dimensional feature representation obtained after feature projection layer mapping;

[0061] Step 6-2: Model the topological relationships and dependencies between features using the Graph Convolutional Memory Attention (GCMA) module:

[0062] ,

[0063] in, Let l be the feature matrix of the l-th layer. For the normalized adjacency matrix, For trainable weight matrix, As a multi-head attention mechanism, For activation functions;

[0064] Step 6-3: Fuse multi-scale neighborhood information using the gated multi-scale neighborhood aggregation module GMNA:

[0065] ,

[0066] in, For nodes The final expression, This is a vector concatenation operation. The maximum number of aggregate hops, For nodes of Jump to neighbor set, Attention coefficient For trainable weight matrix, Features of neighboring nodes;

[0067] Step 6-4: The feature transformation and classification are completed through the Conformer convolutional feedforward network module, and the fault diagnosis results are output.

[0068] In step 6, the output fault types include open circuit fault of insulated gate bipolar transistor (IGBT), short circuit fault of freewheeling diode (FWD), inter-turn short circuit fault of high frequency transformer, capacitance decay fault of filter capacitor, and stall fault of cooling fan.

[0069] Based on the deviation of fault characteristics Classifying the severity of the fault :

[0070] ,

[0071] in, The percentage of fault characteristics that deviate from the normal range. This is a minor fault. This is a moderate fault. This is a severe fault.

[0072] The present invention also provides a solid-state transformer intelligent monitoring and diagnosis system based on multi-source sensor fusion implemented using the method described above, comprising a unit device layer, an embedded control layer, and a decision application layer;

[0073] The unit device layer includes a temperature sensor, a vibration sensor, a partial discharge sensor, a current sensor, a voltage sensor, a flow sensor, and a signal conversion module, which are used to synchronously acquire signals and upload them via the Modbus protocol.

[0074] The embedded control layer is used to implement signal buffering, classification, feature extraction, and fusion diagnosis;

[0075] The decision application layer communicates with the embedded control layer via 5G / WIFI for result display, risk assessment, and data management.

[0076] The present invention has the following beneficial effects:

[0077] 1. Multi-dimensional synchronous perception: Through a distributed multi-source sensor array, the synchronous acquisition of multiple physical quantities such as electrical, thermal, and mechanical properties of SST is realized, providing a data foundation for comprehensive condition assessment.

[0078] 2. Differentiated and precise processing: Based on the physical characteristics of fast and slow changing signals, optimized MS-AVSSD and IAFF-WF+ASGMD combination strategies are used for feature extraction, which significantly improves the extraction quality and noise resistance of various fault-sensitive features.

[0079] 3. Deep heterogeneous feature fusion: The proposed FP-GCMA-GMNA-Conformer model can effectively mine and fuse deep correlations between multi-source heterogeneous features through feature projection, graph attention and multi-scale aggregation mechanism, which greatly improves the diagnostic model's ability to recognize patterns of complex faults.

[0080] 4. Intelligent diagnosis and early warning: The method can output specific fault types, module locations and severity levels, providing a reliable technical means for realizing predictive maintenance and intelligent operation and maintenance of SST, which helps to ensure the safe and stable operation of the power system. Attached Figure Description

[0081] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention as described above or otherwise will become clearer.

[0082] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0083] Figure 2 This is a flowchart of the Condor Optimization Algorithm (CACV-EO) that incorporates chaotic adaptive Cauchy mutation.

[0084] Figure 3 This is a flowchart of the process for purifying thermal and mechanical slow-changing signals (IAFF-WF combined with ASGMD).

[0085] Figure 4 This is a schematic diagram of the FP-GCMA-GMNA-Conformer end-to-end fusion model structure. Detailed Implementation

[0086] like Figure 1 As shown, this embodiment of the invention provides an intelligent monitoring and diagnostic method for solid-state transformers based on multi-source sensor fusion, comprising the following steps:

[0087] Step 1: Multi-source sensor data acquisition and analog-to-digital conversion;

[0088] like Figure 1 As shown, the distributed multi-source sensor array includes temperature sensors, vibration sensors, partial discharge sensors, current sensors, voltage sensors, and flow sensors. Each type of sensor is deployed at a corresponding characteristic monitoring point of each core module of the solid-state transformer (SST). The sampling frequency of the sensors is adapted and set according to the type of monitoring parameter. For high-frequency electrical parameters, the sampling frequency is no less than 10kHz, while for thermal and mechanical parameters, the sampling frequency is 1Hz to 10Hz. The analog signal is converted to a digital signal using an analog-to-digital converter (ADC), and the conversion formula is:

[0089] ,

[0090] In the formula, The converted digital signal value, The analog signal value collected by the sensor. The minimum value of the analog signal range is 0V. The maximum value of the analog signal range is 5V. The number of bits for analog-to-digital conversion is 16 to 24 bits, preferably 24 bits.

[0091] Step 2: Classify digital signals according to sampling frequency;

[0092] like Figure 1 As shown, based on the signal acquisition sampling frequency, the digital signals obtained in step 1 are divided into two categories: high-frequency electrical fast-changing signals and thermal and mechanical slow-changing signals. The high-frequency electrical fast-changing signals are transient electrical signals corresponding to the switching transients and high-frequency partial discharges of the Insulated Gate Bipolar Transistor (IGBT) generated during the operation of the solid-state transformer (SST). The thermal and mechanical slow-changing signals are the thermal and mechanical steady-state signals corresponding to the core vibration and temperature drift generated during the operation of the SST. The determination formula is:

[0093] ,

[0094] in, The sampling frequency for signal acquisition; This indicates high-frequency electrical fast-changing signals, corresponding to transient electrical signals generated during the operation of the solid-state transformer (SST) and the switching transients and high-frequency partial discharges of the insulated gate bipolar transistor (IGBT). This represents the thermal and mechanical slow-changing signals, corresponding to the thermal and mechanical steady-state signals generated during the operation of the solid-state transformer (SST) and the core vibration and temperature drift. The classified high-frequency electrical fast-changing signals... Let be the input signal for the subsequent feature extraction stage. Slowly varying thermal and mechanical signals This is denoted as the input signal for the subsequent feature purification stage. ,in This refers to the discrete sampling point number. Step 3: CACV-EO optimizes the MS-AVSSD parameters and extracts fault features of rapidly changing signals;

[0095] like Figure 2 As shown, the core parameters of the multi-scale adaptive variational singular spectral decomposition (MS-AVSSD) are optimized using the Condor Optimization Algorithm with Chaotic Adaptive Cauchy Mutation (CACV-EO). The specific implementation process is as follows:

[0096] (31) Construct a sample set for MS-AVSSD parameter optimization; determine the core parameter to be optimized as the number of decomposition scales. Punishment factor Number of iterations The objective function for optimization is the signal-to-noise ratio of the processed high-frequency electrical fast-changing signal. ;

[0097] ,

[0098] In the formula, The original high-frequency electrical fast conversion signal Each sample value, The signal after MS-AVSSD processing Each sample value, This represents the total number of signal sampling points.

[0099] (32) Calculate the signal-to-noise ratio The MS-AVSSD parameter combination at maximum; defining the fitness function of the CACV-EO algorithm:

[0100] ,

[0101] In the formula, The fitness value of the CACV-EO algorithm. Signal-to-noise ratio of high-frequency electrical fast-changing signals after MS-AVSSD processing;

[0102] (33) The location of the vulture population is initialized using Logistic chaotic mapping, and the population size is set to be... The maximum number of iterations is ;

[0103] (34) During the vulture's search space selection phase, the position update formula is:

[0104] ,

[0105] in, For the updated position, This is the current globally optimal position. The average position of the population. Current position A random number within the range [0,1]. It is a contraction factor;

[0106] (35) During the vulture's dive to capture prey, an adaptive Cauchy mutation is introduced, and the position update formula is:

[0107] ,

[0108] in, A random number within the range [0,1]. , As a learning factor, , For a variable-length asynchrony that follows a Cauchy distribution, its scale parameter With the number of iterations Adaptive Decreasing: ,in This indicates the maximum value of the scale parameter. This represents the minimum value of the scale parameter. This indicates the maximum number of iterations.

[0109] (36) Define the optimization decision relation of the CACV-EO algorithm as follows:

[0110] ,

[0111] In the formula, This represents the current iteration number of the algorithm. This represents the maximum number of iterations for the algorithm. The fitness value (signal-to-noise ratio) of the g-th iteration. ), Indicates the first The fitness value of each iteration. The convergence accuracy threshold (usually taken as...) ), The "AND" operator represents the logical AND condition, indicating that both conditions must be met simultaneously.

[0112] (37) Iteratively optimize in the parameter space, and when the optimization judgment relationship is satisfied, output the optimal parameter combination of the multi-scale adaptive variational singular spectral decomposition MS-AVSSD. .in This is the optimal decomposition scale. As the optimal penalty factor, This represents the optimal number of iterations.

[0113] like Figure 1 As shown, using the optimal parameter combination The specific steps for processing high-frequency electrical fast-changing signals using the configured multi-scale adaptive variational singular spectrum decomposition MS-AVSSD include:

[0114] (5.1) For fast-changing input signals Constructing the Hankel matrix ;

[0115] (5.2) To Perform Singular Value Decomposition (SVD):

[0116] ,

[0117] in and It is a unitary matrix. It is a singular value matrix. The constructed Hankel matrix;

[0118] (5.3) Utilization and Construct and solve the constrained variational model of variational mode decomposition (VMD) to obtain Individual eigenmode function IMF components ;

[0119] (5.4) Calculate the IMF components of each intrinsic mode function. With the original signal correlation coefficient Select Greater than the threshold Component reconstruction to remove noise signal ;

[0120] (5.5) Hilbert envelope demodulation is performed, and the amplitude, variance, and kurtosis features of the fault characteristic frequencies in the envelope spectrum are extracted as the first input feature set.

[0121] Step 4: IAFF-WF combined with ASGMD to purify the fault characteristics of slow-changing signals;

[0122] like Figure 3As shown, by combining Improved Adaptive Frequency Focusing Wavelet Filtering (IAFF-WF) and Adaptive Synergistic Geometric Mode Decomposition (ASGMD), the thermal and mechanical slowly varying signals are filtered and decomposed to purify the fault-sensitive features in the signals, forming a second input feature set. The specific processing steps include:

[0123] (6.1) For slowly varying input signals The pre-filtered signal is obtained by performing improved adaptive frequency focusing wavelet filtering (IAFF-WF). :

[0124] ,

[0125] in, For the selected wavelet basis function type, The filtering threshold is adaptively calculated based on the signal characteristics;

[0126] (6.2) For the pre-filtered signal Adaptive Co-Geometric Mode Decomposition (ASGMD) is performed to obtain a series of geometric mode components. :

[0127] ,

[0128] in, As an adaptive collaborative factor, To decompose the precision control parameters, The index of the geometric modal component;

[0129] (6.3) The geometric modal components are selected according to the preset screening criteria. Purification: Calculate the energy percentage of each component. and its relationship with the original signal correlation coefficient Select the one that satisfies and The conditional components are reconstructed to obtain the feature-enhanced signal. ;in, This represents the energy percentage threshold. This is the threshold for the correlation coefficient.

[0130] (6.4) Enhancement signal from the features Extract time-domain statistical features to form the second input feature set.

[0131] Step 5: Multi-source feature normalization processing;

[0132] like Figure 1 As shown, the first input feature set extracted in step 3 and the second input feature set purified in step 4 are normalized to obtain a standardized feature set. The min-max normalization method is used to standardize the two types of fault-sensitive features, and the formula is:

[0133] ,

[0134] in, These are the normalized, standardized eigenvalues. These are the original fault-sensitive characteristic values. It is the minimum value of the original feature of all samples in the same feature dimension as the original feature value. This is the maximum value of the original features of all samples within the same feature dimension as the original feature value;

[0135] Step 6: Fusion of model features and fault diagnosis;

[0136] like Figure 4 As shown, the standardized feature set obtained in step 5 is input into the end-to-end fusion model of Feature Projection-Gated Channel Attention Module-Gated Modal Attention Module-Conformer (FP-GCMA-GMNA-Conformer). This model completes the fusion of multi-source fault-sensitive features and fault diagnosis, and outputs the fault type, fault location, and fault severity of the solid-state transformer (SST). The specific implementation process of the hierarchical structure and data processing flow of this end-to-end fusion model is as follows:

[0137] (8.1) The normalized feature set is mapped to a high-dimensional space through the feature projection layer FP:

[0138] ,

[0139] in, To input a standardized feature set, , FC represents the number of neurons in the fully connected layer. This is the high-dimensional feature representation obtained after feature projection layer mapping;

[0140] (8.2) Modeling the topological relationships and dependencies between features using the Graph Convolutional Memory Attention Module (GCMA):

[0141] ,

[0142] in, Let l be the feature matrix of the l-th layer. For the normalized adjacency matrix, For trainable weight matrix, As a multi-head attention mechanism, For activation functions;

[0143] (8.3) Fusing multi-scale neighborhood information through the gated multi-scale neighborhood aggregation module GMNA:

[0144] ,

[0145] in, For nodes The final expression, This is a vector concatenation operation. The maximum number of aggregate hops, For nodes of Jump to neighbor set, Attention coefficient For trainable weight matrix, Features of neighboring nodes;

[0146] (8.4) The feature transformation and classification are completed through the Conformer convolutional feedforward network module, and the fault diagnosis results are output.

[0147] The fault types output in step 6 include open-circuit faults in Insulated Gate Bipolar Transistors (IGBTs), short-circuit faults in Freewheeling Diodes (FWDs), inter-turn short-circuit faults in High-Frequency Transformers, capacitance decay faults in Filter Capacitors, and stall faults in Cooling Fans; the severity of the faults is also listed. Based on the deviation of fault characteristics Divide into:

[0148] ,

[0149] in, The severity level of the fault. The percentage of fault characteristics that deviate from the normal range. This is a minor fault. This is a moderate fault. This is a severe fault;

[0150] This embodiment also provides a solid-state transformer SST differentiated signal processing intelligent monitoring and diagnosis system based on multi-source sensor fusion implemented by the method, including a unit device layer, an embedded control layer and a decision application layer;

[0151] The unit device layer includes a temperature sensor, a vibration sensor, a partial discharge sensor, a current sensor, a voltage sensor, a flow sensor, and a signal conversion module, which are used to synchronously collect multi-dimensional operating parameters of the solid-state transformer (SST) and upload the converted digital signals to the embedded control layer via the Modbus communication protocol.

[0152] The embedded control layer includes a data caching module, a signal classification module, an intelligent algorithm optimization module, a multi-source feature extraction module, and a fault diagnosis module. The data caching module caches and stores the acquired digital signals. The signal classification module classifies the digital signals into high-frequency electrical fast-changing signals and thermal / mechanical slow-changing signals according to the sampling frequency. The intelligent algorithm optimization module optimizes the core parameters of the feature extraction model using the CACV-EO algorithm. The multi-source feature extraction module extracts fault-sensitive features of high-frequency electrical fast-changing signals and purifies fault-sensitive features of thermal / mechanical slow-changing signals based on the optimized model. The fault diagnosis module fuses multi-source features using the FP-GCMA-GMNA-Conformer model and outputs the fault type, fault location, and fault severity of the solid-state transformer (SST).

[0153] The decision application layer achieves bidirectional communication with the embedded control layer via 5G or WIFI, and includes a fault visualization system, a hierarchical early warning system, and a data management server. The fault visualization system displays fault diagnosis results and characteristic trends in real time. The hierarchical early warning system makes risk assessments based on the severity of the fault and executes corresponding hierarchical early warning schemes. The data management server effectively manages and stores monitoring and diagnostic data, and monitors the server's operating status, network connection, and load in real time to complete fault troubleshooting and recovery.

[0154] In one specific embodiment of the present invention, the following contents are included:

[0155] 1. Experimental Data Collection

[0156] 1.1 Experimental Data Acquisition

[0157] The experimental data comes from a solid-state transformer hardware-in-the-loop simulation test platform and a power electronic equipment reliability accelerated aging test platform, covering typical normal operating conditions and various fault conditions of solid-state transformers, including steady-state rated operation, load surges, AC / DC side voltage sags, power device degradation, winding insulation aging, and abnormal heat dissipation systems. A distributed multi-source sensor array synchronously collects multi-dimensional operating parameters of the solid-state transformer's power module, rectifier module, inverter module, filter module, core module, and heat dissipation module. Core parameters include module temperature, core and structural component vibration signals, partial discharge signals, AC / DC side current and voltage, and cooling medium flow rate and pressure. The collected dataset is divided into training and test sets in a 7:3 ratio. After signal classification and preprocessing, it is used for training, validation, and performance testing of the feature extraction model and fault diagnosis model.

[0158] 1.2 Signal Classification and Preprocessing Parameters

[0159] The acquired digital signals are divided into two categories based on their sampling frequency: high-frequency electrical fast-changing signals (such as IGBT switching transients and partial discharge signals, with a sampling frequency of not less than 10kHz) and thermal / mechanical slow-changing signals (such as temperature and vibration signals, with a sampling frequency of 1-10Hz). The signals are buffered after 24-bit high-precision analog-to-digital conversion and their validity is verified, laying the foundation for subsequent differential processing.

[0160] 2. Setting core model parameters

[0161] 2.1 Core Parameters of High-Frequency Electrical Fast-Conversion Signal Processing

[0162] (1) The core parameters of the multi-scale adaptive variational singular spectral decomposition (MS-AVSSD) were optimized using the bald eagle optimization algorithm with chaotic adaptive Cauchy mutation (CACV-EO). The algorithm was set with a population size of 30, a maximum number of iterations of 200, and a convergence accuracy threshold of . The core parameters to be optimized include the number of decomposition scales, the penalty factor, and the number of iterations in MS-AVSSD. The parameter search range is set as: number of decomposition scales. Punishment factor Number of iterations The optimization objective is to maximize the signal-to-noise ratio of the processed signal.

[0163] (2) In the MS-AVSSD signal processing stage, the multi-scale decomposition, singular value decomposition and variational mode decomposition of the signal are completed based on the optimal parameters obtained by optimization. The effective intrinsic mode components are screened by using a correlation coefficient threshold of 0.3. After removing the noise-dominant components, the denoised signal is reconstructed.

[0164] (3) Fault-sensitive feature extraction step: For the denoised high-frequency fast-changing signal, extract its multi-dimensional fault-sensitive features in time domain, frequency domain and time-frequency domain to form the first input feature set.

[0165] 2.2 Core Parameters for Purification of Slowly Changing Thermal and Mechanical Signals

[0166] (1) Improved adaptive frequency focusing wavelet filtering (IAFF-WF) is used to pre-filter and denoise the slowly changing signal. The db6 wavelet basis function is selected, and the thresholding of the wavelet coefficients is completed by adaptively calculating the filtering threshold. The pre-filtered signal is obtained by wavelet reconstruction.

[0167] (2) Modal decomposition of the pre-filtered signal is performed using Adaptive Cooperative Geometric Mode Decomposition (ASGMD). The initial value of the adaptive cooperative factor is set to 0.1, and the decomposition accuracy threshold is... Adaptive decomposition yields multiple geometric modal components;

[0168] (3) Based on the energy ratio threshold of 0.05 and the correlation coefficient threshold of 0.2, the geometric modal components are screened and purified, the effective components dominated by fault information are retained and the feature enhancement signal is reconstructed, and the fault-sensitive statistical features such as the trend features and mutation features of the signal are extracted to form the second input feature set.

[0169] 2.3 Multi-source feature fusion and fault diagnosis model parameters

[0170] (1) The min-max normalization method is used to standardize the two types of fault-sensitive features, mapping all feature values ​​to the [0,1] interval, eliminating the influence of features of different dimensions on the model, and obtaining a standardized feature set;

[0171] (2) Network structure parameters of the FP-GCMA-GMNA-Conformer end-to-end fusion model: The number of neurons in the first fully connected layer of the feature projection layer is 128, and the number of neurons in the second fully connected layer is 256; the number of heads in the attention module is 8, the number of Conformer blocks is 2, and the number of neurons in the classification layer matches the number of fault types of solid-state transformers.

[0172] (3) In the fault diagnosis stage, the matching rules between the fault type and the monitoring points of the core module of the solid transformer are set. The severity of the fault is classified based on the fault characteristic deviation. The judgment thresholds of the three levels of faults (mild, moderate and severe) are clarified, and finally the complete fault diagnosis results are output.

[0173] 3. Experimental Results and Analysis

[0174] 3.1 Comparison Model and Evaluation Indicators:

[0175] (1) Traditional Model 1: General wavelet denoising + traditional BP neural network fault diagnosis, which does not differentiate between high-frequency fast-changing signals and thermal and mechanical slow-changing signals, adopts a unified general feature extraction method, and has no targeted multi-source feature fusion mechanism;

[0176] (2) Traditional Model 2: Variational Mode Decomposition (VMD) + CNN-LSTM fault diagnosis only performs feature extraction for a single type of signal, does not consider the characteristic differences of multi-source heterogeneous signals of solid transformers, and lacks differentiated weight allocation for different modal features;

[0177] (3) The model of this invention: multi-source sensor synchronous acquisition + fast and slow signal differential processing (CACV-EO optimized MS-AVSSD, IAFF-WF combined with ASGMD) + FP-GCMA-GMNA-Conformer multi-source feature deep fusion diagnosis, and completes comprehensive verification on datasets of various typical working conditions and fault types of solid-state transformers.

[0178] The evaluation metrics used are the fault diagnosis accuracy (Acc) and fault type identification precision (Precision), which are commonly used in the field of fault diagnosis, to comprehensively evaluate the overall performance of the model. The formula is as follows:

[0179] ,

[0180] ,

[0181] In the formula, The number of samples that correctly diagnose the fault. The total number of samples in the test set; The number of fault samples to correctly identify the type, This represents the number of fault samples that are false alarms.

[0182] 3.2 Diagnostic Performance Comparison Analysis

[0183] Table 1 Comparison of Fault Diagnosis Performance of Different Models

[0184]

[0185] As shown in Table 1, the FP-GCMA-GMNA-Conformer end-to-end fusion diagnostic model proposed in this invention achieves a fault diagnosis accuracy of 99.26% under various typical fault scenarios of solid-state transformers. Compared with the two traditional comparison models, the diagnostic accuracy is improved by 16.91 percentage points and 7.79 percentage points respectively, which is far superior to the traditional comparison scheme. It has strong anti-interference ability and generalization performance for fluctuations under different operating conditions.

[0186] This solution effectively addresses the core issues in existing technologies, such as poor adaptability to differentiated signal processing of solid-state transformers, insufficient extraction of early weak fault features, and inadequate fusion of multi-source features. The integrated design of the entire process balances diagnostic accuracy and inference efficiency, enabling full-dimensional online monitoring and accurate fault diagnosis of solid-state transformers, and fully adapting to the intelligent operation and maintenance needs of solid-state transformers throughout their entire lifecycle in new power systems.

[0187] This invention provides a method and system for intelligent monitoring and diagnosis of solid-state transformers based on multi-source sensor fusion. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for intelligent monitoring and diagnosis of solid-state transformers based on multi-source sensor fusion, characterized in that, Includes the following steps: Step 1: Construct a distributed multi-source sensor array to simultaneously collect multi-dimensional operating parameters such as temperature, vibration, partial discharge, current, voltage, and heat dissipation from the core module of the solid-state transformer (SST). The collected analog signals are converted into digital signals and then buffered. The core module includes a power module, a rectifier module, an inverter module, a filter module, a core module, and a heat dissipation module. Step 2: Based on the signal acquisition sampling frequency, the digital signals obtained in Step 1 are divided into two categories: high-frequency electrical fast-changing signals and thermal and mechanical slow-changing signals. Among them, the high-frequency electrical fast-changing signals are transient electrical signals corresponding to the switching transients and high-frequency partial discharges of the Insulated Gate Bipolar Transistor (IGBT) generated during the operation of the solid-state transformer (SST), while the thermal and mechanical slow-changing signals are thermal and mechanical steady-state signals corresponding to the core vibration and temperature drift generated during the operation of the solid-state transformer (SST). Step 3: The core parameters of the multi-scale adaptive variational singular spectrum decomposition (MS-AVSSD) are optimized using the bald eagle optimization algorithm CACV-EO, which incorporates chaotic adaptive Cauchy mutation. The optimized MS-AVSSD is then used to decompose and denoise the high-frequency electrical fast-changing signal, extracting fault-sensitive features from the high-frequency electrical fast-changing signal to form the first input feature set. Step 4: Combine the improved adaptive frequency focusing wavelet filter IAFF-WF with the adaptive cooperative geometric mode decomposition ASGMD to filter and decompose the thermal and mechanical slowly varying signals, purify the fault-sensitive features in the thermal and mechanical slowly varying signals, and form the second input feature set. Step 5: Perform feature normalization on the first input feature set extracted in step 3 and the second input feature set purified in step 4 to obtain a standardized feature set; Step 6: Input the standardized feature set obtained in Step 5 into the end-to-end fusion model of feature projection-gated channel attention-gated modal attention-convolutional transformer FP-GCMA-GMNA-Conformer. The model is used to complete the fusion of multi-source fault-sensitive features and fault diagnosis, and output the fault type, fault location and fault severity of the solid-state transformer SST.

2. The method according to claim 1, characterized in that, In step 1, the distributed multi-source sensor array includes temperature sensors, vibration sensors, partial discharge sensors, current sensors, voltage sensors, and flow sensors. Each type of sensor is deployed at a corresponding characteristic monitoring point of each core module of the solid-state transformer (SST). The sampling frequency of the sensors is adapted and set according to the type of monitoring parameter. For high-frequency electrical parameters, the sampling frequency is no less than 10kHz, while for thermal and mechanical parameters, the sampling frequency is 1Hz to 10Hz. The analog signal is converted to a digital signal using an analog-to-digital converter (ADC). The conversion formula is as follows: ， in, The converted digital signal value, The analog signal value collected by the sensor. This represents the minimum range of the analog signal. This represents the maximum value of the analog signal range. This represents the number of bits used in the analog-to-digital conversion.

3. The method according to claim 2, characterized in that, In step 2, the classification criterion for digital signals is the signal acquisition sampling frequency, and the determination formula is as follows: ， in, The sampling frequency for signal acquisition; This indicates high-frequency electrical fast-changing signals, corresponding to transient electrical signals generated during the operation of the solid-state transformer (SST) and the switching transients and high-frequency partial discharges of the insulated gate bipolar transistor (IGBT). This represents the thermal and mechanical slow-changing signals, corresponding to the thermal and mechanical steady-state signals generated during the operation of the solid-state transformer (SST) and the core vibration and temperature drift. The classified high-frequency electrical fast-changing signals... Let be the input signal for the subsequent feature extraction stage. Slowly varying thermal and mechanical signals This is denoted as the input signal for the subsequent feature purification stage. ,in This represents the sequence number of the discrete sampling point.

4. The method according to claim 3, characterized in that, Step 3, the implementation process of the core parameters includes: Step 3-1: Construct a multi-scale adaptive variational singular spectral decomposition (MS-AVSSD) parameter optimization sample set; determine the core parameter to be optimized as the number of decomposition scales. Punishment factor and number of iterations The objective function for optimization is the signal-to-noise ratio of the processed high-frequency electrical fast-changing signal. ; ， in, The original high-frequency electrical fast conversion signal Each sample value, The signal after MS-AVSSD processing for multi-scale adaptive variational singular spectrum decomposition is the first... Each sample value, This represents the total number of signal sampling points. Step 3-2, obtain the signal-to-noise ratio. The MS-AVSSD parameter combination at maximum; defining the fitness function of the chaotic adaptive Cauchy mutation-based eagle optimization algorithm CACV-EO. : ； Step 3-3: Initialize the location of the vulture population using Logistic chaotic mapping, and set the population size to [value missing]. The maximum number of iterations is ; Steps 3-4, during the vulture's search space selection phase, the position update formula is: ， in, For the updated position, This is the current globally optimal position. The average position of the population. Current position A random number within the range [0,1]. It is a contraction factor; Steps 3-5, during the vulture's swooping down to capture prey, introduce adaptive Cauchy mutation, with the position update formula as follows: ， in, A random number within the range [0,1]. , As a learning factor, , For varying asynchronous lengths following a Cauchy distribution, the scale parameter With the number of iterations Adaptive Decreasing: ,in This indicates the maximum value of the scale parameter. This represents the minimum value of the scale parameter. Indicates the maximum number of iterations; Steps 3-6 define the optimization decision relationship of the CACV-EO algorithm as follows: ， in, This represents the current iteration number of the algorithm. Let g be the fitness value of the g-th iteration. This is the convergence accuracy threshold. This indicates a logical AND operation. Steps 3-7 involve iterative optimization in the parameter space. When the optimization criterion is satisfied, the optimal parameter combination of the multi-scale adaptive variational singular spectral decomposition MS-AVSSD is output. ,in This is the optimal decomposition scale. As the optimal penalty factor, This represents the optimal number of iterations.

5. The method according to claim 4, characterized in that, In step 3, the optimal parameter combination is used. The configured multi-scale adaptive variational singular spectral decomposition MS-AVSSD for processing high-frequency electrical fast-changing signals includes the following steps: Steps 3-8, for the fast-changing input signal Constructing the Hankel matrix ; Steps 3-9, for Perform Singular Value Decomposition (SVD): ， in and It is a unitary matrix. It is a singular value matrix. The constructed Hankel matrix; Steps 3-10, using and Construct and solve the constrained variational model of variational mode decomposition (VMD) to obtain Individual eigenmode function IMF components ; Step 3-11: Calculate the IMF components of each intrinsic mode function. With the original signal correlation coefficient Select Greater than the threshold Component reconstruction to remove noise signal ; Steps 3-12, for Hilbert envelope demodulation is performed, and the amplitude, variance, and kurtosis features of the fault feature frequencies in the envelope spectrum are extracted as the first input feature set.

6. The method according to claim 5, characterized in that, Step 4 includes: Step 4-1, for the slowly changing input signal The pre-filtered signal is obtained by performing improved adaptive frequency focusing wavelet filtering (IAFF-WF). : ， in, For the selected wavelet basis function type, The filtering threshold is adaptively calculated based on the signal characteristics; Step 4-2, pre-filtering the signal Adaptive Co-Geometric Mode Decomposition (ASGMD) is performed to obtain a series of geometric mode components. : ， in, As an adaptive collaborative factor, To decompose the precision control parameters, The index of the geometric modal component; Step 4-3: The geometric modal components are selected according to preset screening criteria. Purification: Calculate the energy percentage of each component. and With the original signal correlation coefficient Select the one that satisfies and The conditional components are reconstructed to obtain the feature-enhanced signal. ;in, This represents the energy percentage threshold. The threshold for the correlation coefficient; Step 4-4, from the feature enhancement signal Extract time-domain statistical features to form the second input feature set.

7. The method according to claim 6, characterized in that, In step 5, the first input feature set and the second input feature set are standardized using the min-max normalization method, as shown in the formula: ， in, These are the normalized, standardized eigenvalues. These are the original fault-sensitive characteristic values. It is the minimum value of the original feature among all samples belonging to the same feature dimension as the original feature value. It is the maximum value of the original features of all samples in the same feature dimension as the original feature value.

8. The method according to claim 7, characterized in that, In step 6, the FP-GCMA-GMNA-Conformer end-to-end fusion model performs fusion and diagnosis on the standardized feature set, specifically including the following steps: Step 6-1: Map the normalized feature set to a high-dimensional space through the feature projection layer FP: ， in, To input a standardized feature set, , FC represents the number of neurons in the fully connected layer. This is the high-dimensional feature representation obtained after feature projection layer mapping; Step 6-2: Model the topological relationships and dependencies between features using the Graph Convolutional Memory Attention (GCMA) module: ， in, Let l be the feature matrix of the l-th layer. For the normalized adjacency matrix, For trainable weight matrix, As a multi-head attention mechanism, For activation functions; Step 6-3: Fuse multi-scale neighborhood information using the gated multi-scale neighborhood aggregation module GMNA: ， in, For nodes The final expression, This is a vector concatenation operation. The maximum number of aggregate hops, For nodes of Jump to neighbor set, Attention coefficient For trainable weight matrix, Features of neighboring nodes; Step 6-4: The feature transformation and classification are completed through the Conformer convolutional feedforward network module, and the fault diagnosis results are output.

9. The method according to claim 8, characterized in that, In step 6, the output fault types include open circuit fault of insulated gate bipolar transistor (IGBT), short circuit fault of freewheeling diode (FWD), inter-turn short circuit fault of high frequency transformer, capacitance decay fault of filter capacitor, and stall fault of cooling fan. Based on the deviation of fault characteristics Classifying the severity of the fault : ， in, The percentage of fault characteristics that deviate from the normal range. This is a minor fault. This is a moderate fault. This is a severe fault.

10. A solid-state transformer intelligent monitoring and diagnostic system based on multi-source sensor fusion, implemented using the method described in any one of claims 1-9, characterized in that, It includes a unit device layer, an embedded control layer, and a decision application layer; The unit device layer includes a temperature sensor, a vibration sensor, a partial discharge sensor, a current sensor, a voltage sensor, a flow sensor, and a signal conversion module, which are used to synchronously acquire signals and upload them via the Modbus protocol. The embedded control layer is used to implement signal buffering, classification, feature extraction, and fusion diagnosis; The decision application layer communicates with the embedded control layer via 5G / WIFI for result display, risk assessment, and data management.

Images