Power supply detection method and system for power supply equipment

Abstract

The invention discloses a power supply detection method and system for power supply equipment, and belongs to the technical field of power supply equipment detection, and the method specifically comprises the steps: carrying out the synchronous collection of dissolved gas data in oil, partial discharge signal data and infrared thermal imaging data through employing a clock synchronization protocol, and obtaining an enhanced feature vector through combining with a tensor decomposition method; inputting the enhanced feature vector into a pre-trained gradient boosting decision tree model, and outputting fault types such as partial discharge, overheating, insulation aging and the like and corresponding probabilities; according to the output probability, a three-level probability threshold value obtained through calculation of a Weibull distribution-Gaussian mixture model is utilized, differential early warning is started, and an inspection work order is automatically generated; through synchronous acquisition and fusion analysis of the multi-source data, the accuracy and timeliness of fault detection of the power supply equipment are improved, differential early warning and work order generation can be realized according to the fault risk level, and safe and stable operation of the power supply system is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of power supply equipment detection, and in particular relates to a power supply detection method and system for power supply equipment. Background Art

[0002] The dam power supply system is critical infrastructure for ensuring the proper operation and safe management of the dam. The main transformer, as a core component, has a significant impact on the reliability and stability of the power supply. However, existing inspection methods for dam main transformers present numerous challenges. Manual inspections struggle to detect potential internal faults, such as winding deformation and insulation aging. Traditional inspection methods are inefficient and unable to meet the dam power supply system's demand for real-time equipment monitoring. Furthermore, insufficient data processing and analysis capabilities hinder accurate assessment of equipment status and prediction of faults. Therefore, a more efficient and accurate inspection method is urgently needed to ensure the safe and stable operation of dam main transformers.

[0003] For example, a Chinese patent with authorization announcement number CN104426707B discloses a method and power supply device for detecting Ethernet power supply, including: a PSE obtains a signature resistance value and a signature capacitance value of a powered device (PD), the PSE obtains a signature attribute of the PD based on the signature resistance value and the signature capacitance value of the PD, and when the signature attribute of the PD is Gray Zone, the PSE obtains a graded current based on a measuring device, and obtains a graded attribute of the PD based on the graded current. When the graded attribute of the PD is a grade in a PD grade table, the PSE supplies power to the PD, thereby being able to supply power to the PD even when the signature resistance value and the signature capacitance value deviate due to signal interference.

[0004] For example, Chinese patent application CN113640582B discloses a method and device for detecting the legitimacy of an Ethernet power supply system, including: providing a first voltage signal at a power supply port and detecting the port voltage; when the port voltage is greater than a first preset voltage signal, providing the first preset voltage signal at the power supply port and detecting the port voltage and port current until they stabilize, thereby obtaining a first steady-state voltage and a first steady-state current; providing a second power supply signal at the power supply port, the second power supply signal being different from the first voltage signal, detecting the port voltage and port current of the power supply port until the port voltage and port current of the power supply port reach a stable state, thereby obtaining a second steady-state voltage, a second steady-state current, multiple discharge voltage values, and a discharge time; and obtaining the signature impedance of a powered device based on the first steady-state voltage, the first steady-state current, the second steady-state voltage, the second steady current, the multiple discharge voltage values, and the discharge time. This technical solution also provides a power supply device that can quickly and accurately obtain the signature impedance of a powered device.

[0005] The above existing technologies have the following problems: they only detect the signature impedance of the powered device in the Ethernet power supply system, focusing on the electrical matching legitimacy between the power supply system and the powered device; they mainly obtain the signature impedance based on electrical parameters such as voltage and current of the power supply port, and the data type is single; they can only determine whether the powered device complies with the electrical specifications of the power supply system, and cannot directly diagnose faults in the power supply device itself. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention proposes a power supply detection method and system for power supply equipment, which adopts a clock synchronization protocol to synchronously collect dissolved gas data in oil, partial discharge signal data and infrared thermal imaging data, and combines the tensor decomposition method to obtain enhanced feature vectors; the enhanced feature vectors are then input into a pre-trained gradient boosting decision tree model to output fault types such as partial discharge, overheating, insulation aging and the corresponding probabilities; according to the output probability, the three-level probability threshold calculated by the Weibull distribution-Gaussian mixture model is used to start differentiated early warning and automatically generate inspection work orders; the present invention improves the accuracy and timeliness of power supply equipment fault detection through synchronous acquisition and fusion analysis of multi-source data, and can realize differentiated early warning and work order generation according to the fault risk level, thereby ensuring the safe and stable operation of the power supply system.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A power supply detection method for a power supply device, comprising:

[0009] A clock synchronization protocol is used to synchronously collect dissolved gas data in oil, partial discharge signal data, and infrared thermal imaging data. Combined with the tensor decomposition method, enhanced feature vectors are obtained.

[0010] Input the enhanced feature vector into a pre-trained gradient boosting decision tree model to output the fault type and probability; the fault type includes partial discharge, overheating, and insulation aging;

[0011] Based on the output probability and combined with the three-level probability threshold, differentiated warnings are initiated and inspection work orders are automatically generated; the three-level probability threshold is calculated using the Weibull distribution-Gaussian mixture model.

[0012] Specifically, the clock synchronization protocol is used to synchronously collect dissolved gas data in oil, partial discharge signal data, and infrared thermal imaging data, and combined with the tensor decomposition method to obtain enhanced feature vectors, including:

[0013] A1: Deploy a GPS / NTP server in the substation as the master clock source to generate the UTC time reference. At the same time, slave clock devices synchronize with the master clock source using the IEEE 1588v2 protocol and perform timestamp calibration.

[0014] The slave clock device includes a DGA sensor, a PD detector, and an infrared camera;

[0015] The timestamp calibration method is to add a UTC timestamp to each data point from a clock device during acquisition and correct network delays using a two-way time transfer algorithm;

[0016] A2: Based on the calibrated unified time base, the dissolved gas data in oil, partial discharge signal data, and infrared thermal imaging data are synchronously collected from the clock device and preprocessed. The preprocessing process includes:

[0017] Normalizing the dissolved gas in oil data to obtain normalized dissolved gas in oil data;

[0018] Perform noise reduction filtering on partial discharge signal data to extract discharge characteristic parameters;

[0019] Perform temperature field calibration on infrared thermal imaging data, mark hot spots, and obtain temperature characteristic parameters;

[0020] A3: The normalized dissolved gas data, discharge characteristic parameters, and temperature characteristic parameters are constructed into a three-dimensional tensor of time-characteristic-mode according to the time axis alignment method. , where I represents the time dimension, J represents the feature dimension, K represents the modal dimension, and R represents the real number domain;

[0021] A4: For three-dimensional tensors Decompose to get the core tensor and factor matrices, and generates enhanced feature vectors v by tensor reconstruction, where 、 、 Represent the core tensor rank respectively.

[0022] Specifically, the specific steps of A4 include:

[0023] A4.1: Obtain the three-dimensional tensor of time-feature-mode in A3 ;

[0024] A4.2: For each dimension ,pass Get the matrix expansion of X in dimension n ;

[0025] A4.3: For each implement , we get the left singular matrix and the singular value matrix , where T represents transpose;

[0026] A4.4: Determining the core tensor rank based on the singular value decay property 、 、 , and satisfy: ,in, represents the i-th singular value;

[0027] A4.5: Based on core tensor rank pairs Before Column interception to obtain the factor matrix .

[0028] Specifically, the specific steps of A4 further include:

[0029] A4.6: Pass Get the core tensor ,in, 、 、 Respectively represent 1-module product, 2-module product, 3-module product, 、 、 Respectively represent the time mode factor matrix, characteristic mode factor matrix, and modal mode factor matrix;

[0030] A4.7: Pass The core tensor Expand into matrix form to obtain the expanded core tensor ;

[0031] A4.8: Perform feature concatenation on the factor matrix to obtain the concatenated factor matrix ;

[0032] A4.9: Expand the core tensor And the concatenated factor matrix Perform secondary splicing to obtain enhanced feature vector .

[0033] Specifically, the enhanced feature vector is input into the pre-trained gradient boosting decision tree model to output the fault type and probability, including:

[0034] B1: Get the enhanced feature vector v and perform normalization. Divide into three subsequences , get the dissolved gas data subsequence , partial discharge signal data subsequence , infrared thermal imaging data subsequence ;

[0035] B2: Use the pre-trained gradient boosting decision tree model to calculate 、 、 The feature importance score of

[0036] B3: Assign importance weights to each subsequence by calculating the ratio of the feature importance score of the jth subsequence to the total feature importance score of the three subsequences .

[0037] Specifically, the step of inputting the enhanced feature vector into a pre-trained gradient boosting decision tree model and outputting the fault type and probability further includes:

[0038] B4: For the t-th decision tree in the gradient boosting decision tree model, the enhanced feature vector after standardization Input into the gradient boosting decision tree model, finally fall on leaf node d through the decision tree path, and output the score of the leaf node ,in, represents the preset score value of the leaf node d of the t-th decision tree for the k-th type of fault, where k represents the fault category index;

[0039] B5: By formula Integrate the prediction results of H trees to obtain the single model integration result ,in, represents the k-th category bias term;

[0040] B6: Use the three subsequence data to train the gradient boosting decision tree sub-model to obtain the dissolved gas sub-model , local discharge submodel , infrared thermal imaging sub-model , and use the formula Dissolved gas submodel , local discharge submodel , infrared thermal imaging sub-model Make predictions and get the prediction results of different gradient boosting decision tree sub-models , where j represents the gradient boosting decision tree sub-model index, and j=1 is , j=2 is , j=3 is ;

[0041] B7: Use the formula Perform multimodal integration on the prediction results of the gradient boosting decision tree sub-model trained for each sub-sequence to obtain the multimodal integration result ,in, Represents the weight of the j-th gradient boosting decision tree sub-model.

[0042] Specifically, the step of inputting the enhanced feature vector into a pre-trained gradient boosting decision tree model and outputting the fault type and probability further includes:

[0043] B8: Fusion of single model ensemble results Integration results with multimodal , and obtain the final model prediction results ,in, represents the equilibrium parameter;

[0044] B9: Convert to probability distribution ;

[0045] B10: According to probability distribution , get the fault type and failure probability .

[0046] Specifically, the output probability is combined with the three-level probability threshold to initiate differentiated early warning and automatically generate inspection work orders, including:

[0047] C1: Obtain failure probability P and probability distribution ;

[0048] C2: Use the preloaded Weibull distribution-Gaussian mixture model to perform probability density fitting on the failure probability P and obtain the probability density function ;

[0049] C3: Divide the probability distribution into There are three intervals;

[0050] C4: Probability density function Integrate to obtain the cumulative distribution function, and determine the three-level probability threshold by finding the quantiles where the cumulative probability reaches the three intervals 、 、 , and satisfies ;

[0051] like , it is judged as low risk, no warning is triggered, and no inspection work order is generated;

[0052] like , it is determined to be medium risk, triggering a yellow warning and automatically generating a regular inspection work order; the regular inspection work order contains the equipment name, possible fault type, and recommended inspection time;

[0053] like , it is judged as high risk, triggering an orange warning and automatically generating an expedited inspection work order; the expedited inspection work order includes the content of the regular inspection work order and the detection method recommendation;

[0054] like , it is determined to be the highest risk, triggering a red alert, automatically generating an emergency repair work order, and notifying the operation and maintenance person in charge; the emergency repair work order includes the content of the expedited inspection work order, emergency fault handling measures and on-site arrival time requirements.

[0055] C5: Retrain the Weibull distribution-Gaussian mixture model based on the newly generated fault data.

[0056] A power supply detection system for power supply equipment, comprising: a data acquisition module, a feature engineering module, a fault diagnosis module, and an early warning and work order generation module;

[0057] The data acquisition module is used to synchronously collect dissolved gas data in oil, partial discharge signal data and infrared thermal imaging data through a clock synchronization protocol, and perform preliminary time stamp calibration and preprocessing;

[0058] The feature engineering module is used to construct the preprocessed data into a three-dimensional tensor and generate an enhanced feature vector through tensor decomposition and reconstruction;

[0059] The fault diagnosis module is used to input the enhanced feature vector into a pre-trained gradient boosting decision tree model and output the fault type and probability through processing;

[0060] The warning and work order generation module is used to start differentiated warnings and automatically generate inspection work orders based on the probability of fault prediction output in combination with three-level probability thresholds.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. The present invention proposes a power supply detection system for power supply equipment, and optimizes and improves the architecture, operation steps and processes. The system has the advantages of simple process, low investment and operation costs, and low production work costs.

[0063] 2. The present invention proposes a power supply detection method for power supply equipment, which realizes the synchronous acquisition of multi-source data through a clock synchronization protocol, and combines tensor decomposition to generate enhanced feature vectors, effectively integrating the complementary information of dissolved gas in oil, partial discharge signals and infrared thermal imaging data, thereby improving the characterization capability of equipment status characteristics; when using the gradient boosting decision tree model for fault diagnosis, the dual strategy of single model integration and multimodal sub-model fusion is adopted to improve the accuracy and reliability of fault type identification, especially for the precise classification of different fault types, such as partial discharge, overheating, and insulation aging.

[0064] 3. The present invention proposes a power supply detection method for power supply equipment, which determines the three-level probability threshold based on the Weibull distribution-Gaussian mixture model, realizes differentiated early warning of fault risks and intelligent work order generation. This hierarchical early warning mechanism can automatically match the corresponding operation and maintenance strategy according to the fault probability, forming a closed-loop management process from routine inspections to emergency repairs, greatly improving the operation and maintenance efficiency; the work order content automatically generated by the system contains key information such as equipment name, fault type, and detection method recommendations, which effectively shortens the fault handling cycle, reduces the risk of equipment damage, and ensures the safe and stable operation of the power supply system. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a schematic diagram of a power supply detection method for a power supply device according to the present invention;

[0066] Figure 2 This is a flow chart showing the output of fault types and probabilities of a power supply detection method for power supply equipment according to the present invention;

[0067] Figure 3 This is an architecture diagram of a power supply detection system for power supply equipment of the present invention. DETAILED DESCRIPTION

[0068] Example 1

[0069] See also Figure 1 An embodiment of the present invention provides a method for detecting power supply of a power supply device, the method comprising steps S1 to S3, including the following steps:

[0070] S1: A clock synchronization protocol is used to synchronously collect dissolved gas data in oil, partial discharge signal data, and infrared thermal imaging data, and the enhanced feature vector is obtained by combining the tensor decomposition method;

[0071] S2: Input the enhanced feature vector into the pre-trained gradient boosting decision tree model to output the fault type and probability; the fault type includes partial discharge, overheating, and insulation aging;

[0072] S3: Based on the output probability and the three-level probability threshold, a differentiated warning is initiated and an inspection work order is automatically generated; the three-level probability threshold is calculated by the Weibull distribution-Gaussian mixture model.

[0073] The specific steps of S1 include:

[0074] A1: Deploy a GPS / NTP server in the substation as the master clock source to generate the UTC time reference. At the same time, slave clock devices synchronize with the master clock source using the IEEE 1588v2 protocol and perform timestamp calibration.

[0075] The slave clock device includes a DGA sensor, a PD detector, and an infrared camera;

[0076] The timestamp calibration method is to add a UTC timestamp to each data point from the clock device during acquisition and use a two-way time transfer algorithm to Correction of network delays, where Indicates the time offset between the slave clock source and the master clock device. represents the one-way propagation delay of the message in the network, Indicates the time when the master clock source sends the synchronization message. Indicates the time when the synchronization message is received from the clock device. Indicates that the delay request time is returned from the clock device. Indicates the delay time for receiving the request from the main clock source;

[0077] Furthermore, the specific steps of A1 include:

[0078] (1) Deploy an NTP / PTP server with a GPS receiver in the substation control room. Configure the server to support both NTP and IEEE 1588v2.

[0079] (2) Select an intelligent terminal that supports IEEE 1588v2 hardware timestamps and configure the slave clock working mode to Boundary Clock or Ordinary Clock;

[0080] (3) Perform master clock configuration and calibration, including:

[0081] GPS receiver configuration: Ensure the GPS antenna is installed in an unobstructed location, receives signals from at least four satellites, and configures the receiver to output 1PPS (pulse per second) signals and IRIG-B time codes.

[0082] It should be noted that 1PPS stands for 1 Pulse Per Second, which means one pulse per second. It is a commonly used time synchronization signal, usually generated by a high-precision clock source such as a rubidium clock or an atomic clock. It is used to synchronize the time bases of various devices and ensure a high level of time synchronization accuracy between different devices.

[0083] PTP server configuration: Set the server to master clock mode, configure the clock class to 1, and set the time reference source priority to: GPS > external rubidium clock > internal crystal oscillator;

[0084] It should be noted that the external rubidium clock is a high-precision clock device, which is usually used as an external reference clock source to provide an accurate time base for other devices. It uses the energy level transition characteristics of rubidium atoms to generate a stable oscillation signal with extremely high frequency stability and accuracy.

[0085] UTC time base generation: The server obtains UTC time through a GPS receiver, and its internal timekeeping algorithm maintains time accuracy when the GPS signal is lost;

[0086] (4) Perform slave clock synchronization and timestamp calibration, including:

[0087] Slave clock initialization: When the slave clock device starts, it uses the internal clock by default and obtains the PTP master clock IP address and configuration parameters through DHCP;

[0088] PTP synchronization process: The slave clock receives the Sync message sent by the master clock and records it. Then, the slave clock device sends a Delay_Req message and records , at the same time, receiving the master clock return Finally, according to the two-way time transfer algorithm formula in A1, and , adjust the local clock;

[0089] Timestamp generation: At the moment of data collection, the hardware directly generates a PTP timestamp, where the timestamp format is: UTC seconds + sub-second offset;

[0090] (5) Enable the PTP transparent clock function and boundary clock mode on the network switch. The boundary clock mode records the forwarding delay and corrects the timestamp for the switch;

[0091] (6) The master clock periodically sends Announce messages containing its own clock quality information. Then, the slave clocks select the best master clock through the Best Master Clock Algorithm (BMCA).

[0092] (7) Real-time monitoring Fluctuations and synchronization status change logs are recorded.

[0093] For example, assume that the master clock's actual time is 10:00:00, and the slave clock displays 10:00:05 locally, which is an offset of +5 seconds and a network delay of 2ms.

[0094] When the master clock is at 10:00:00 ( ) sends a Sync message, and the slave clock is at 10:00:05.002 ( )receive;

[0095] From the clock at 10:00:10 ( ) sends Delay_Req, the master clock is at 10:00:10.002 ( )receive;

[0096] calculate: , that is, the 2.5-second offset and 2.502-second delay are correctly captured. Through this two-way measurement, the time offset and network delay are separated to achieve high-precision synchronization.

[0097] A2: Based on the calibrated unified time base, the dissolved gas data in oil, partial discharge signal data, and infrared thermal imaging data are synchronously collected from the clock device and preprocessed. The preprocessing process includes:

[0098] The dissolved gas in oil data is normalized to obtain normalized dissolved gas in oil data. In the present invention, the normalization adopts the Z-score method, which is a prior art in the field and does not constitute an inventive solution of the present application, and is not described in detail here.

[0099] Perform noise reduction filtering on partial discharge signal data to extract discharge characteristic parameters;

[0100] Furthermore, the partial discharge signal data is subjected to noise reduction filtering to extract discharge characteristic parameters, including:

[0101] (1) Obtaining partial discharge signal data;

[0102] (2) performing anti-aliasing filtering on the partial discharge signal data according to the Nyquist theorem to obtain anti-aliasing filtered data, wherein the sampling rate is greater than or equal to twice the maximum frequency. For example, a sampling rate of 100 MHz is used for 30 MHz high-frequency partial discharge signal data;

[0103] (3) performing wavelet transform on the anti-aliasing filtered data and reconstructing the signal through inverse wavelet transform to obtain partial discharge signal data after time domain noise reduction, wherein wavelet transform and its inverse transform are prior art contents in this field and are not the inventive solution of this application, and are not described in detail here;

[0104] (4) performing a fast Fourier transform on the local discharge signal data after time domain noise reduction, and returning the data to the time domain through an inverse transform to obtain the local discharge signal data after frequency domain noise reduction. The fast Fourier transform and its inverse transform are prior art contents in this field and are not the inventive solution of this application, and are not described in detail here;

[0105] (5) Feature extraction is performed on the local discharge signal data after frequency domain noise reduction, including time domain features, frequency domain features and phase features.

[0106] Perform temperature field calibration on infrared thermal imaging data, mark hot spots, and obtain temperature characteristic parameters;

[0107] Furthermore, a two-point calibration method is used when calibrating the temperature field of infrared thermal imaging data. The voltage is measured at low temperature and high temperature respectively, and then the voltage difference is compared with the difference between low temperature and high temperature to obtain a first calibration coefficient; then the product of the first calibration coefficient and the low temperature value is subtracted from the voltage measured at low temperature to obtain a second calibration coefficient.

[0108] Furthermore, the hotspot area annotation process includes:

[0109] (1) performing median filtering on the infrared thermal imaging data after temperature field calibration to obtain infrared thermal imaging data after median filtering, wherein median filtering is a prior art in this field and is not an inventive solution of the present application and is not described in detail here;

[0110] (2) Using a threshold segmentation method combined with a region growing method to identify hotspots in the infrared thermal imaging data after median filtering to obtain hotspot areas, wherein the threshold segmentation method and the region growing method are prior art contents in this field and are not the inventive solutions of this application, and are not described in detail here;

[0111] (3) Based on the hotspot area, extract hotspot parameters, including hotspot temperature, hotspot area and relative temperature difference;

[0112] (4) Based on the hotspot parameters, the hotspot is marked using a rectangular frame.

[0113] A3: The normalized dissolved gas data, discharge characteristic parameters, and temperature characteristic parameters are constructed into a three-dimensional tensor of time-characteristic-mode according to the time axis alignment method. , where I represents the time dimension, J represents the feature dimension, K represents the modal dimension, and R represents the real number domain;

[0114] Among them, the time dimension refers to the number of sampling points; the feature dimension refers to the number of normalized dissolved gas data in oil, discharge characteristic parameters and temperature characteristic parameters. For example, if there are 8 kinds of dissolved gases in oil after normalization, 5 discharge characteristic parameters and 3 temperature characteristic parameters, then the feature dimension is 16; in the modal dimension, the normalized dissolved gas data in oil is the first dimension, the discharge characteristic parameters are the second dimension, and the temperature characteristic parameters are the third dimension.

[0115] A4: For three-dimensional tensors Decompose to get the core tensor and factor matrices, and generates enhanced feature vectors v by tensor reconstruction, where 、 、 They represent the core tensor rank respectively.

[0116] The specific steps of A4 include:

[0117] A4.1: Obtain the three-dimensional tensor of time-feature-mode in A3 ;

[0118] A4.2: For each dimension ,pass Get the matrix expansion of X in dimension n ;

[0119] Furthermore, the expansion rules include:

[0120] , indicating expansion by time dimension;

[0121] , means expanding by feature dimension;

[0122] , indicating expansion by modal dimension.

[0123] A4.3: For each implement , we get the left singular matrix and the singular value matrix , where T represents transpose;

[0124] A4.4: Determining the core tensor rank based on the singular value decay property 、 、 , and satisfy: ,in, represents the i-th singular value;

[0125] In the present invention, the specific formula of singular value decomposition is the existing technical content in this field, which is not the inventive solution of this application and will not be described in detail here;

[0126] A4.5: Based on core tensor rank pairs Before Column interception to obtain the factor matrix ;

[0127] A4.6: Pass Get the core tensor ,in, 、 、 Respectively represent 1-module product, 2-module product, 3-module product, 、 、 Respectively represent the time mode factor matrix, characteristic mode factor matrix, and modal mode factor matrix;

[0128] A4.7: Pass The core tensor Expand into matrix form to obtain the expanded core tensor ;

[0129] A4.8: Perform feature concatenation on the factor matrix to obtain the concatenated factor matrix ,in, Indicates that 、 、 Splice by row;

[0130] A4.9: Expand the core tensor And the concatenated factor matrix Perform secondary splicing to obtain enhanced feature vector ,in, Indicates that and Splice by column.

[0131] Example 2

[0132] See also Figure 2 , the specific steps of S2 in this embodiment include:

[0133] B1: Get the enhanced feature vector v and perform normalization. Divide into three subsequences , get the dissolved gas data subsequence , partial discharge signal data subsequence , infrared thermal imaging data subsequence ;

[0134] B2: Use the pre-trained gradient boosting decision tree model to calculate 、 、 The feature importance score of

[0135] Furthermore, the specific steps of B2 include:

[0136] (1) Obtaining dissolved gas data subsequence , partial discharge signal data subsequence , infrared thermal imaging data subsequence ;

[0137] (2) For the lth feature in the jth subsequence, its importance score for: ,in, Indicates the number of times feature l is used to split in the t-th tree, represents the total number of splits of the t-th tree, represents the information gain brought by the splitting of feature l in the t-th tree;

[0138] (3) For the jth subsequence, its overall importance score for: , where M represents all features in the j-th subsequence.

[0139] B3: Assign importance weights to each subsequence by calculating the ratio of the feature importance score of the jth subsequence to the total feature importance score of the three subsequences ;

[0140] B4: For the t-th decision tree in the gradient boosting decision tree model, the enhanced feature vector after standardization Input into the gradient boosting decision tree model, finally fall on leaf node d through the decision tree path, and output the score of the leaf node ,in, The preset score value of the leaf node d of the t-th decision tree for the k-th type of fault is determined during the training phase, where k represents the fault category index;

[0141] Furthermore, in the gradient boosting decision tree, the process of processing the standardized enhanced feature vector by the t-th decision tree can be decomposed into the following steps:

[0142] (1) For the normalized enhanced feature vector , the prediction process of the t-th decision tree is:

[0143] Starting from the root node, the node is split into the left or right subtree according to the node's splitting condition, and traversed until it reaches a leaf node; the splitting condition is a preset feature threshold;

[0144] Among them, for internal nodes, if the normalized enhanced feature vector of the u-th node is If the value is less than or equal to the preset feature threshold of the u-th node, go to the left subtree; otherwise, go to the right subtree;

[0145] (2) When the normalized enhanced feature vector When reaching the leaf node of the t-th decision tree, output the score of the leaf node.

[0146] B5: By formula Integrate the prediction results of H trees to obtain the single model integration result ,in, represents the k-th category bias term;

[0147] B6: Use the three subsequence data to train the gradient boosting decision tree sub-model to obtain the dissolved gas sub-model , local discharge submodel , infrared thermal imaging sub-model , and use the formula Dissolved gas submodel , local discharge submodel , infrared thermal imaging sub-model Make predictions and get the prediction results of different gradient boosting decision tree sub-models , where j represents the gradient boosting decision tree sub-model index, and j=1 is , j=2 is , j=3 is ;

[0148] B7: Use the formula Perform multimodal integration on the prediction results of the gradient boosting decision tree sub-model trained for each sub-sequence to obtain the multimodal integration result ,in, represents the weight of the j-th gradient boosting decision tree sub-model;

[0149] B8: Fusion of single model ensemble results Integration results with multimodal , and obtain the final model prediction results ,in, represents the equilibrium parameter;

[0150] B9: Convert to probability distribution ;

[0151] B10: According to probability distribution , get the fault type and failure probability .

[0152] The specific steps of S3 include:

[0153] C1: Obtain failure probability P and probability distribution ;

[0154] C2: Use the Weibull distribution-Gaussian mixture model to perform probability density fitting on the failure probability P and obtain the probability density function ;

[0155] It should be understood that the probability density function of the Weibull distribution-Gaussian mixture model is a weighted combination of the Weibull distribution and the Gaussian mixture model. The specific formula is:

[0156] ;

[0157] in, represents the weight of the Weibull distribution, represents the probability density function of the Weibull distribution, represents the scale parameter, satisfying ,satisfy , m represents the shape parameter, and satisfies m>0, g represents the number of Gaussian components, Represents the weight of the qth Gaussian component, and the cumulative sum is 1, Represents the probability density function of the qth Gaussian distribution, and the probability density function of the Gaussian distribution satisfies the mean , the variance is distributed, represents the mean of the qth Gaussian component, represents the variance of the qth Gaussian component, Represents the exponential function.

[0158] C3: Divide the probability distribution into There are three intervals;

[0159] C4: Probability density function Integrate to get the cumulative distribution function , and determine the three-level probability threshold by finding the quantiles where the cumulative probability reaches the three intervals 、 、 , and satisfies , where z represents the integration variable;

[0160] In the present invention, the quantiles of the three intervals are set to 0.2, 0.6, and 0.9 respectively.

[0161] like , it is judged as low risk, no warning is triggered, and no inspection work order is generated;

[0162] like , it is determined to be medium risk, triggering a yellow warning and automatically generating a regular inspection work order; the regular inspection work order contains the equipment name, possible fault type, and recommended inspection time;

[0163] like , it is judged as high risk, triggering an orange warning and automatically generating an expedited inspection work order; the expedited inspection work order includes the content of the regular inspection work order and the detection method recommendation;

[0164] like , it is determined to be the highest risk, triggering a red alert, automatically generating an emergency repair work order, and notifying the operation and maintenance person in charge; the emergency repair work order includes the content of the expedited inspection work order, emergency fault handling measures and on-site arrival time requirements.

[0165] C5: Retrain the Weibull distribution-Gaussian mixture model based on the newly generated fault data.

[0166] Example 3

[0167] See also Figure 3 Another embodiment of the present invention provides a power supply detection system for a power supply device, comprising:

[0168] Data acquisition module, feature engineering module, fault diagnosis module, early warning and work order generation module;

[0169] The data acquisition module is used to synchronously collect dissolved gas data in oil, partial discharge signal data, and infrared thermal imaging data through a clock synchronization protocol, and perform preliminary time stamp calibration and preprocessing;

[0170] The feature engineering module is used to construct the preprocessed data into a three-dimensional tensor and generate enhanced feature vectors through tensor decomposition and reconstruction to improve the data feature expression capability;

[0171] The fault diagnosis module is used to input the enhanced feature vector into the pre-trained gradient boosting decision tree model and output the fault type and probability through processing;

[0172] The early warning and work order generation module is used to initiate differentiated early warnings and automatically generate inspection work orders based on the probability of fault prediction output and combined with the three-level probability threshold.

[0173] The feature engineering module includes: tensor construction unit and tensor decomposition unit;

[0174] A tensor construction unit is used to construct the characteristic parameters of different types of data into a three-dimensional tensor of time-feature-modality;

[0175] The tensor decomposition unit is used to decompose and reconstruct the three-dimensional tensor to generate enhanced feature vectors.

[0176] The fault diagnosis module includes: data partition unit, importance assessment unit, and model integration unit;

[0177] a data division unit, for obtaining enhanced feature vectors and performing normalization processing, and dividing them into dissolved gas data subsequences, partial discharge electron data subsequences, and infrared thermal imaging data subsequences, so as to perform separate analysis and processing on different types of data;

[0178] The importance evaluation unit is used to calculate the feature importance score of each subsequence using a pre-trained gradient boosting decision tree model, and assign an importance weight to each subsequence by calculating the score ratio to understand the contribution of different data types to fault diagnosis.

[0179] The model integration unit uses the gradient boosting decision tree model to make predictions and integrates the results of single models and multimodal sub-models.

[0180] The early warning and work order generation module includes: probability processing unit, threshold determination unit, and work order generation unit;

[0181] The probability processing unit is used to obtain the fault probability and probability distribution, and use the Weibull distribution-Gaussian mixture model to perform probability density fitting on the fault probability to obtain the probability density function, which provides a basis for determining the probability threshold;

[0182] A threshold determination unit is used to divide the probability distribution into three intervals according to the fault risk level and determine the three-level probability threshold as a standard for judging the fault risk level;

[0183] The work order generation unit is used to determine the fault risk level based on the comparison results of the fault probability with the three-level probability threshold, trigger the corresponding warning and automatically generate different types of inspection work orders. For example, low risk does not trigger a warning work order, medium risk triggers a yellow warning and generates a regular inspection work order, high risk triggers an orange warning and generates an expedited inspection work order, and the highest risk triggers a red warning and generates an emergency repair work order. At the same time, the operation and maintenance person in charge is notified to clarify the response measures and work order content under different risk levels.

[0184] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention. These are all protected by the present invention.

Claims

1. A power supply detection method for a power supply device, characterized in that: include: A clock synchronization protocol is used to synchronously collect dissolved gas data in oil, partial discharge signal data, and infrared thermal imaging data. Combined with the tensor decomposition method, enhanced feature vectors are obtained. Input the enhanced feature vector into a pre-trained gradient boosting decision tree model to output the fault type and probability; the fault type includes partial discharge, overheating, and insulation aging; Based on the output probability and combined with the three-level probability threshold, differentiated warnings are initiated and inspection work orders are automatically generated; the three-level probability threshold is calculated using the Weibull distribution-Gaussian mixture model.

2. A power supply detection method for power supply equipment according to claim 1, characterized in that: The clock synchronization protocol is used to synchronously collect dissolved gas data in oil, partial discharge signal data, and infrared thermal imaging data, and combined with the tensor decomposition method to obtain enhanced feature vectors, including: A1: Deploy a GPS / NTP server in the substation as the master clock source to generate the UTC time reference. At the same time, slave clock devices synchronize with the master clock source using the IEEE 1588v2 protocol and perform timestamp calibration. The slave clock device includes a DGA sensor, a PD detector, and an infrared camera; The timestamp calibration method is to add a UTC timestamp to each data point from a clock device during acquisition and correct network delays using a two-way time transfer algorithm; A2: Based on the calibrated unified time base, the dissolved gas data in oil, partial discharge signal data, and infrared thermal imaging data are synchronously collected from the clock device and preprocessed. The preprocessing process includes: Normalizing the dissolved gas in oil data to obtain normalized dissolved gas in oil data; Perform noise reduction filtering on partial discharge signal data to extract discharge characteristic parameters; Perform temperature field calibration on infrared thermal imaging data, mark hot spots, and obtain temperature characteristic parameters; A3: The normalized dissolved gas data, discharge characteristic parameters, and temperature characteristic parameters are constructed into a three-dimensional tensor of time-characteristic-mode according to the time axis alignment method. , where I represents the time dimension, J represents the feature dimension, K represents the modal dimension, and R represents the real number domain; A4: For three-dimensional tensors Decompose to get the core tensor and factor matrices, and generates enhanced feature vectors v by tensor reconstruction, where 、 、 Represent the core tensor rank respectively.

3. A power supply detection method for power supply equipment according to claim 2, characterized in that: The specific steps of A4 include: A4.1: Obtain the three-dimensional tensor of time-feature-mode in A3 ; A4.2: For each dimension ,pass Get the matrix expansion of X in dimension n ; A4.3: For each implement , we get the left singular matrix and the singular value matrix , where T represents transpose; A4.4: Determining the core tensor rank based on the singular value decay property 、 、 , and satisfy: ,in, represents the i-th singular value; A4.5: Based on core tensor rank pairs Before Column interception to obtain the factor matrix .

4. A power supply detection method for power supply equipment according to claim 3, characterized in that: The specific steps of A4 also include: A4.6: Pass Get the core tensor ,in, 、 、 Respectively represent 1-module product, 2-module product, 3-module product, 、 、 Respectively represent the time mode factor matrix, characteristic mode factor matrix, and modal mode factor matrix; A4.7: Pass The core tensor Expand into matrix form to obtain the expanded core tensor ; A4.8: Perform feature concatenation on the factor matrix to obtain the concatenated factor matrix ; A4.9: Expand the core tensor And the concatenated factor matrix Perform secondary splicing to obtain enhanced feature vector .

5. A power supply detection method for power supply equipment according to claim 4, characterized in that: The enhanced feature vector is input into the pre-trained gradient boosting decision tree model to output the fault type and probability, including: B1: Get the enhanced feature vector v and perform normalization. Divide into three subsequences , get the dissolved gas data subsequence , partial discharge signal data subsequence , infrared thermal imaging data subsequence ; B2: Use the pre-trained gradient boosting decision tree model to calculate 、 、 The feature importance score of B3: Assign importance weights to each subsequence by calculating the ratio of the feature importance score of the jth subsequence to the total feature importance score of the three subsequences .

6. A power supply detection method for power supply equipment according to claim 5, characterized in that: The step of inputting the enhanced feature vector into a pre-trained gradient boosting decision tree model and outputting the fault type and probability further includes: B4: For the t-th decision tree in the gradient boosting decision tree model, the enhanced feature vector after standardization Input into the gradient boosting decision tree model, finally fall on leaf node d through the decision tree path, and output the score of the leaf node ,in, represents the preset score value of the leaf node d of the t-th decision tree for the k-th type of fault, where k represents the fault category index; B5: By formula Integrate the prediction results of H trees to obtain the single model integration result ,in, represents the k-th category bias term; B6: Use the three subsequence data to train the gradient boosting decision tree sub-model to obtain the dissolved gas sub-model , local discharge submodel , infrared thermal imaging sub-model , and use the formula Dissolved gas submodel , local discharge submodel , infrared thermal imaging sub-model Make predictions and get the prediction results of different gradient boosting decision tree sub-models , where j represents the gradient boosting decision tree sub-model index, and j=1 is , j=2 is , j=3 is ; B7: Use the formula Perform multimodal integration on the prediction results of the gradient boosting decision tree sub-model trained for each sub-sequence to obtain the multimodal integration result ,in, Represents the weight of the j-th gradient boosting decision tree sub-model.

7. A power supply detection method for power supply equipment according to claim 6, characterized in that: The step of inputting the enhanced feature vector into a pre-trained gradient boosting decision tree model and outputting the fault type and probability further includes: B8: Fusion of single model ensemble results Integration results with multimodal , and obtain the final model prediction results ,in, represents the equilibrium parameter; B9: Convert to probability distribution ; B10: According to probability distribution , get the fault type and failure probability .

8. A power supply detection method for power supply equipment according to claim 7, characterized in that: Based on the output probability and the three-level probability threshold, differentiated warnings are initiated and inspection work orders are automatically generated, including: C1: Obtain failure probability P and probability distribution ; C2: Use the preloaded Weibull distribution-Gaussian mixture model to perform probability density fitting on the failure probability P and obtain the probability density function ; C3: Divide the probability distribution into There are three intervals; C4: Probability density function Integrate to obtain the cumulative distribution function, and determine the three-level probability threshold by finding the quantiles where the cumulative probability reaches the three intervals 、 、 , and satisfies ; like , it is judged as low risk, no warning is triggered, and no inspection work order is generated; like , it is determined to be medium risk, triggering a yellow warning and automatically generating a regular inspection work order; the regular inspection work order contains the equipment name, possible fault type, and recommended inspection time; like , it is judged as high risk, triggering an orange warning and automatically generating an expedited inspection work order; the expedited inspection work order includes the content of the regular inspection work order and the detection method recommendation; like , it is determined to be the highest risk, triggering a red alert, automatically generating an emergency repair work order, and notifying the operation and maintenance person in charge; the emergency repair work order includes the content of the expedited inspection work order, emergency fault handling measures, and on-site arrival time requirements; C5: Retrain the Weibull distribution-Gaussian mixture model based on the newly generated fault data.

9. A power supply detection system for power supply equipment, used to implement a power supply detection method for power supply equipment according to any one of claims 1 to 8, characterized in that: include: Data acquisition module, feature engineering module, fault diagnosis module, early warning and work order generation module; The data acquisition module is used to synchronously collect dissolved gas data in oil, partial discharge signal data and infrared thermal imaging data through a clock synchronization protocol, and perform preliminary time stamp calibration and preprocessing; The feature engineering module is used to construct the preprocessed data into a three-dimensional tensor and generate an enhanced feature vector through tensor decomposition and reconstruction; The fault diagnosis module is used to input the enhanced feature vector into a pre-trained gradient boosting decision tree model and output the fault type and probability through processing; The warning and work order generation module is used to start differentiated warnings and automatically generate inspection work orders based on the probability of fault prediction output in combination with three-level probability thresholds.

Images