Power equipment inspection device fault detection and diagnosis method based on deep learning

Abstract

The application discloses a power equipment inspection device fault detection and diagnosis method based on deep learning, comprising the following steps: collecting multi-modal data to generate an original data set; performing time calibration normalization on the original data set to generate a preprocessed data package; performing abnormality detection on the preprocessed data package to output an abnormal candidate region set and generate a feature slice set; performing group convolution operation on the feature slice set to output a structure preserving representation; performing classification and calibration on the structure preserving representation to output a fault type and a confidence vector; performing risk assessment on the fault type to output a risk level code; performing time series processing on the structure preserving representation to output a fault trend curve and a deterioration rate; organizing the abnormal candidate region set, archiving the risk level code, labeling the fault trend curve and the deterioration rate, and generating a diagnosis report data. The application realizes power equipment inspection device fault detection and diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system operation and maintenance, and particularly relates to a power equipment inspection device fault detection and diagnosis method based on deep learning. BACKGROUND

[0002] In the process of power system operation and maintenance, the state of power equipment is directly related to the safety and stability of the power grid. In order to ensure the long-term reliable operation of the equipment, inspection devices are widely used in the detection and diagnosis of power transmission and transformation equipment. The existing inspection devices mostly rely on single or multi-modal sensing means such as image acquisition, infrared temperature measurement, sound detection or partial discharge monitoring, and identify faults through traditional feature extraction algorithms or shallow machine learning methods. Although this kind of method can find surface abnormalities to some extent, due to the limited expression ability of the algorithm, it often relies on manually set features or experience thresholds, resulting in insufficient accuracy and robustness of the diagnosis results. At the same time, the operating environment of power equipment is complex, and the data types present high dimensionality, multi-modality and dynamic evolution, etc. The existing technology is difficult to effectively model the feature consistency under geometric changes such as rotation and translation, and the fault detection results are easily affected by environmental interference and equipment posture changes.

[0003] In recent years, deep learning technology has been gradually introduced into the inspection and diagnosis of power equipment, and convolutional neural networks or time series models are used to realize automatic feature extraction and classification judgment. Although these methods have improved the recognition accuracy compared with traditional algorithms, conventional convolutional networks have inherent shortcomings in modeling rotation and spatial translation. When the power equipment has different shooting angles, changes in viewing angle or environmental disturbances during inspection, the feature expression often distorts, resulting in a decline in the generalization performance of the classification model. In addition, the existing deep learning methods lack a systematic overall design in multi-modal data fusion, time series modeling and risk assessment, and often only stay at the level of classification and recognition of static data, and cannot give the trend and deterioration rate of the fault over time, making it difficult to provide forward-looking decision basis for equipment operation and maintenance.

[0004] Therefore, how to provide a power equipment inspection device fault detection and diagnosis method based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a power equipment inspection device fault detection and diagnosis method based on deep learning. Through multi-modal data acquisition, preprocessing, anomaly detection, group convolution feature modeling, classification calibration, risk assessment and time series analysis, the present application generates fault types, risk levels and development trends, and finally forms a diagnosis report, realizing accurate detection and intelligent diagnosis of power equipment faults.

[0006] The deep learning-based power equipment inspection device fault detection and diagnosis method according to the embodiment of the application comprises the following steps:

[0007] Collecting multi-modal perception data of the power equipment to generate an original data set with time identifiers and sensor identifiers;

[0008] Performing time synchronization calibration and feature scale normalization on the original data set to generate a preprocessed data package;

[0009] Performing abnormality detection on the preprocessed data package to output an abnormal candidate region set containing position indexes, boundary information and confidence scores, and to generate a feature slice set corresponding to the candidate regions;

[0010] Inputting the abnormal candidate region set and the corresponding feature slice set into a quaternion-SE(2) equivariant group convolution network to perform quaternion channel mapping and SE(2) group convolution operation, and outputting a structure-preserving representation tensor with rotation and translation equivariance, wherein the SE(2) represents a two-dimensional special Euclidean group;

[0011] Performing classification operation and calibration operation on the structure-preserving representation tensor to output a fault type label and a corresponding confidence vector;

[0012] Performing risk assessment on the fault type label and the confidence vector, performing level determination and threshold comparison, and outputting a risk level code;

[0013] Performing time series processing on the structure-preserving representation tensor and the fault type label, performing state evolution analysis and dynamic parameter extraction, and outputting a fault development trend curve and a degradation rate scalar;

[0014] Organizing the abnormal candidate region set, archiving the risk level code, labeling the fault development trend curve and the degradation rate scalar, generating a diagnosis report data, and outputting the diagnosis report data in real time by an edge deployment engine of the inspection device.

[0015] Optionally, the generation process of the original data set specifically comprises:

[0016] Establishing a collection configuration, creating a multi-modal list and a sensor list, establishing a one-to-one correspondence relationship between each mode and the corresponding sensor and recording the relationship, setting a unified time reference moment and a unified sampling interval to generate a sampling time list, setting a sampling rate, a range upper and lower limit, an allowed time deviation range and a device identifier code for each sensor, and outputting the sampling time list and the sensor parameter table;

[0017] The synchronization acquisition and frame construction are performed, the multi-modal sensing data of each sensor is synchronously acquired within the allowed time deviation range according to the sampling time list, a time identifier and a sensor identifier are written into each acquisition sample to generate a record unit, each record unit is spliced into a frame at the same sampling time according to the preset modal order and written into a time index and a modal order index, a complete integrity check code is generated for each frame and attached to the frame tail, a quality mark is generated for each sample according to the upper and lower limits of the sensor parameter table and marked as qualified or unqualified, and a record unit set, a frame set with integrity check and a quality mark set are outputted;

[0018] The data archiving and output are performed, the record unit set, the frame set and the quality mark set are grouped and sorted according to the time index and the sensor identifier, and the grouping result is written into a raw data set storage structure to generate a retrievable directory item, and a raw data set with a time identifier and a sensor identifier is formed.

[0019] Optionally, the generation process of the preprocessed data package specifically includes:

[0020] The time synchronization calibration is performed on all record units of the raw data set according to a unified time reference and a sampling interval, the time identifiers of the record units are aligned to the same time index sequence, the time correction is performed on the record units not within the allowed range, and a time index table is generated, and the record unit set after time alignment is outputted;

[0021] The integrity check and frame missing completion are performed on the record unit set, the time index and the modal order are compared one by one, the missing sample is found, the adjacent valid sample is used for completion and a completion mark is recorded, a frame missing log is generated, and a complete time sequence sample set is outputted;

[0022] The amplitude normalization and feature scale unification processing are performed on the complete time sequence sample set, the average value, the standard deviation, the minimum value and the maximum value of each modal are respectively calculated, the scaling and boundary truncation are performed on each sample, a scale-unified feature set is formed, and the time index table, the complete time sequence sample set and the feature set are merged into a preprocessed data package.

[0023] Optionally, the generation process of the feature slice set corresponding to the candidate region specifically includes:

[0024] Based on the preprocessed data package, a processing batch list is established according to the time index and the modal order, the sliding block and the fixed step clipping are performed on each processing batch, a window queue is generated, the time index, the window number and the modal number are written into each window, and a to-be-inspected window set is outputted;

[0025] An anomaly detection process is executed on each of the windows to be inspected to generate an initial candidate list containing location index, boundary information and confidence score. Threshold filtering and overlap suppression are performed on the initial candidate list. Candidates that pass the filtering are retained and written with candidate number, time index and window number. An anomaly candidate region set is output.

[0026] Based on the boundary information of the abnormal candidate region set, the corresponding region is cropped from the preprocessed data packet, and size alignment and coordinate archiving are performed. For each cropping result, a candidate number, time index and position index are written to form a feature slice set corresponding to the candidate region.

[0027] Optionally, the structure that preserves the output representing the tensor specifically includes:

[0028] The system receives a set of abnormal candidate regions and a set of feature slices. It performs geometric normalization and resampling operations on each candidate region, completes the scaling, rotation and translation uniformity processing, generates aligned candidate feature blocks and establishes an alignment index table. According to the quaternion channel mapping configuration, it distributes multi-channel real-valued features to four quaternion components according to the Hamiltonian multiplication operation rules. Each quaternion component includes one real part and three imaginary parts. It performs channel aggregation and quaternion encoding at each pixel position and outputs a set of candidate quaternion fields.

[0029] Define a set of angle sampling sets and assign a weight value to each angle. Simultaneously, based on the candidate quaternion field set, establish a lifting operator from the image plane domain to the rotation-translation group domain. Calculate the group domain representation for each candidate number and any position and angle pair. Perform angle rotation on the relative coordinates according to the two-dimensional rotation operator. Weight the rotated neighborhood according to the direction-selective window function and accumulate the quaternion features in the neighborhood using Hamiltonian multiplication as the rule to obtain the group domain representation entry for the current position and angle pair. After completing the calculation for all position and angle pairs, write the group domain representation entry into the group domain representation cache and output the group domain representation cache and the position and angle index list.

[0030] Based on the convolution kernel configuration, a set of rotation-translation group convolution kernels with quaternion values ​​is generated. Angle sampling registration, kernel rotation registration, and angle association of quaternion rotation factors are completed. Sliding calculation is performed on the group domain representation cache according to the position and angle dimensions. According to the set step size and boundary processing strategy, the group domain representation entries and the corresponding rotated convolution kernel entries are read position by position and angle by angle. Quaternion combination and accumulation are performed according to Hamiltonian multiplication to obtain angle and position response values. Each calculated angle and position response value is written into the angle and position response set, and the convolution kernel identifier, angle index, position index and calculation timestamp are recorded. The angle and position response set and calculation log are output.

[0031] According to the angle sampling set and the corresponding weight, and in combination with a preset isometric aggregation rule, a weighted aggregation is performed on the angle-dimension angle and position response set to obtain a position-dimension structure-preserving response, the isometric aggregation rule is set to ensure that the convolution output is still isometric under rotation and translation, a subsampling is performed on the position dimension according to a set spatial step, and a candidate number and a position index are added to the sampling result, and a structure-preserving representation tensor with rotation and translation isovariance is output.

[0032] Optionally, the output of the fault type label and the corresponding confidence vector specifically includes:

[0033] An index table is established according to the candidate numbers of the structure-preserving representation tensor set, each candidate tensor is traversed according to the position, the response of the tensor at each position point is decomposed into real value components according to the real part and three imaginary parts of the quaternion, all components in the same partition are point-by-point accumulated and arithmetically averaged to form a statistical vector of the current partition, and statistical vector sets of all partitions are sequentially spliced according to the candidate number order to output a candidate representation vector set;

[0034] A preset classification parameter matrix and a bias vector are loaded, the classification parameter matrix and the bias vector are pre-trained and stored by the system, and are loaded when the inspection device is initialized, a linear transformation is performed on each candidate representation vector, the representation vector is multiplied by the classification parameter matrix, and then the corresponding components of the bias vector are added to obtain an original score vector of each candidate on all fault type categories, the original score vector and corresponding parameter calling information are output together to generate an original score vector set;

[0035] A probability normalization operation is performed on each candidate original score vector, the original scores of each category are respectively subjected to exponential transformation, and then the sum of the exponential values of all categories is taken as a normalization factor, the exponential value of each category is divided by the normalization factor to obtain an uncalibrated probability distribution of the candidate on all categories, a temperature scaling calibration process is performed on the uncalibrated probability distribution, that is, the original score vector is scaled according to a preset temperature parameter, the exponential transformation and normalization are repeated to obtain a calibrated confidence vector of the candidate, the category with the maximum confidence in each calibrated confidence vector is selected as the fault type label, and a diagnosis result item is formed by pairing the fault type label with the corresponding confidence component, and the fault type label and the corresponding confidence vector are output.

[0036] Optionally, the output process of the risk level code specifically includes:

[0037] For each candidate record, a confidence vector with the same name as the fault type label is located, a label-confidence pair list is generated, and a candidate number index is output.

[0038] The risk weight table saved when the system is initialized is called, and the risk evaluation operation is performed on each tag-confidence pair in the list of tag-confidence pairs, the risk weight corresponding to the fault type tag is read, and the risk weight is weighted and superimposed with the corresponding confidence component to obtain the risk score of the current candidate. After the calculation of all candidates is completed, a risk score list is generated;

[0039] The power equipment inspection industry standard grid safety risk level is called, and the risk score list is compared with the threshold value in order from low to high. When the risk score of a candidate first falls into a certain threshold interval, the level corresponding to the interval is identified as the risk level code of the current candidate.

[0040] Optionally, the output of the fault development trend curve and the degradation rate scalar specifically includes:

[0041] The structure-preserving representation tensor set and the fault type tag set are received, the structure-preserving representation tensors of the same candidate number are time-ordered according to the time index table, a time series tensor stream is established for each time, the fault type tag corresponding to the time is attached to the time series tensor stream, the one-to-one correspondence mapping of time-tensor-tag is completed, and a time series data set is output;

[0042] The state evolution analysis is performed on the time series data set, the structure-preserving representation tensors of two adjacent times are taken in turn along the time index, the element difference operation is performed, the difference value is squared element by element and then summed and squared to obtain the overall change amplitude value of the current time interval, and the change amplitude values of all intervals are arranged in time order to form a change sequence. The change sequence is smoothed according to a preset sliding window width and step, and the unequal interval time periods are weighted according to the actual interval length. The smoothed change sequence and the corresponding fault type tag sequence are time-aligned and labeled, and a state evolution curve set is output.

[0043] The dynamic parameter extraction is performed on the state evolution curve. In each adjacent time interval, the instantaneous change rate is obtained by dividing the change amplitude value of the current interval by the actual time length of the interval. The instantaneous change rates are weighted and averaged according to the time length within the observation window to obtain a degradation rate scalar. The state evolution curve is attached with a time axis and a candidate number identifier, and is paired with the corresponding degradation rate scalar. After being integrated according to the candidate number, a fault development trend curve and a degradation rate scalar are output.

[0044] Optionally, the generation of the diagnostic report data specifically includes:

[0045] The abnormal candidate region set, the risk level code set, the fault development trend curve and the degradation rate scalar are received, and are aligned according to the time index table and the candidate number. A candidate data alignment list is generated.

[0046] The abnormal candidate region is sorted according to time index and position, the boundary information is corrected according to unified coordinate specification, the risk level code is written into the level mapping table according to the candidate number and time index, and the sorted abnormal item table and risk archive table are generated;

[0047] The labeling operation is performed on the fault development trend curve and the degradation rate scalar in the candidate data alignment list, the integration is performed according to the candidate number, the trend rate labeling table is generated, and the abnormal item table, the risk archive table and the trend rate labeling table are combined and loaded to generate and output the diagnosis report data.

[0048] The beneficial effects of the present application are:

[0049] 1、The present application realizes the unification and standardization of data among different sensors by collecting and time synchronously calibrating the multi-modal perception data of the power equipment, avoids the feature deviation problem caused by the collection time error and the inconsistent scale, and provides a stable data foundation for subsequent fault detection.

[0050] 2、The present application realizes the structure preserving representation of rotation and translation invariance in the feature slice processing of the abnormal candidate region by introducing the quaternion-SE(2) equivariant group convolution network, overcomes the feature distortion defect of the existing deep learning method under geometric change, and thus improves the fault recognition ability of the model under different angles and different postures.

[0051] 3、After the fault type determination, the present application combines the classification calibration and the risk level evaluation, can not only output the accurate fault type, but also can give the risk level code according to the confidence and the risk weight, realizes the whole process closed loop from recognition to risk management.

[0052] 4、The present application extracts the development trend curve and the degradation rate of the fault by time sequence processing on the structure preserving representation and the fault label, integrates the development trend curve and the degradation rate with the abnormal region and the risk level to generate the diagnosis report, so that the inspection device not only has the static detection ability, but also can predict the dynamic evolution of the fault, provides a reliable basis for the preventive maintenance and decision of the power equipment. BRIEF DESCRIPTION OF DRAWINGS

[0053] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation on the present application. In the drawings:

[0054] Fig. 1 The flow chart of the power equipment inspection device fault detection and diagnosis method based on deep learning proposed in the present application;

[0055] Fig. 2The algorithm structure diagram of the power equipment inspection device fault detection and diagnosis method based on deep learning is provided for the present application. DETAILED DESCRIPTION

[0056] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only show the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0057] REFERENCE Figs. 1-2 The power equipment inspection device fault detection and diagnosis method based on deep learning comprises the following steps:

[0058] Collecting multi-modal perception data of the power equipment to generate an original data set with time and sensor identifiers;

[0059] Performing time synchronization calibration and feature scale normalization on the original data set to generate a preprocessed data package;

[0060] Performing anomaly detection on the preprocessed data package to output an abnormal candidate region set containing position index, boundary information and confidence score, and generating a feature slice set corresponding to the candidate region;

[0061] Inputting the abnormal candidate region set and the corresponding feature slice set into a quaternion-SE(2) equivariant group convolution network to perform quaternion channel mapping and SE(2) group convolution operation, and outputting a structure-preserving representation tensor with rotation and translation equivariance, wherein SE(2) represents a two-dimensional special Euclidean group;

[0062] Performing classification operation and calibration operation on the structure-preserving representation tensor to output fault type label and corresponding confidence vector;

[0063] Performing risk assessment on the fault type label and the confidence vector, performing level determination and threshold comparison, and outputting a risk level code;

[0064] Performing time series processing on the structure-preserving representation tensor and the fault type label, performing state evolution analysis and dynamic parameter extraction, and outputting a fault development trend curve and a degradation rate scalar;

[0065] Organizing the abnormal candidate region set, archiving the risk level code, labeling the fault development trend curve and the degradation rate scalar, generating a diagnosis report data, and outputting the diagnosis report data in real time by the edge deployment engine of the inspection device.

[0066] The application realizes high-precision detection and diagnosis of power equipment faults by collecting and preprocessing multi-modal data, combining with quaternion-SE(2) equivariant group convolution network. Through the generation and classification calibration of structure-preserving representation tensors, the robustness of the system under rotation and translation equivariance is enhanced, ensuring accurate identification and risk assessment of fault types. Further time series analysis and dynamic parameter extraction not only provide real-time state evolution of equipment faults, but also predict their development trend and degradation rate.

[0067] In the embodiment, the generation process of the original data set specifically includes:

[0068] The acquisition configuration is established, a multi-modal list and a sensor list are created, a one-to-one correspondence relationship between each mode and the corresponding sensor is established and recorded, a unified time reference moment and a unified sampling interval are set to generate a sampling time list, a sampling rate, a range upper and lower limit, an allowed time deviation range and a device identification code are set for each sensor, and the sampling time list and the sensor parameter table are output;

[0069] Synchronous acquisition and frame construction are performed, multi-modal sensing data are synchronously acquired by each sensor within the allowed time deviation range according to the sampling time list, a time identifier and a sensor identifier are written into each acquisition sample to generate a record unit, each record unit is spliced into a frame at the same sampling time according to the preset mode order and is written into a time index and a mode order index, a complete check code is generated for each frame and is attached to the frame tail, a quality mark is generated for each sample according to the range upper and lower limit of the sensor parameter table and is marked as qualified or unqualified, and the record unit set, the frame set with integrity check and the quality mark set are output;

[0070] Data archiving and output are performed, the record unit set, the frame set and the quality mark set are grouped and sorted according to the time index and the sensor identifier, the grouping results are written into the original data set storage structure to generate a retrievable directory item, and the original data set with the time identifier and the sensor identifier is formed.

[0071] The application effectively avoids the influence of time error on data quality by accurately setting the sampling time, the sampling rate and the allowed deviation range. The construction and integrity check of the data frame ensure the accuracy and integrity of the data, and at the same time, the generation of the quality mark ensures the reliability of the acquisition data, and through data archiving and sorting, an efficient and structured data storage method is provided for subsequent data analysis and retrieval, improving the precision and efficiency of multi-modal data processing.

[0072] In the embodiment, the generation process of the original data set specifically includes:

[0073] According to a unified time reference and sampling interval, time synchronization calibration is performed on all record units of the original data set, time identifiers of the record units are aligned to the same time index sequence, time correction is performed on the record units not within the allowed range, and a time index table is generated, and a record unit set after time alignment is output;

[0074] Integrity checking and frame missing filling are performed on the record unit set, the time index and the modal order are compared one by one, when missing samples are found, adjacent valid samples are used for estimation and filling, and a filling mark is recorded, a frame missing log is generated, and a complete time sequence sample set is output;

[0075] Amplitude normalization and feature scale unification processing are performed on the complete time sequence sample set, scaling and boundary truncation are performed on each sample according to modal average value, standard deviation, minimum value and maximum value, a scale-unified feature set is formed, and the time index table, the complete time sequence sample set and the feature set are merged into a pre-processing data package.

[0076] The present application effectively ensures the time sequence consistency and data integrity of data collected by different sensors through time synchronization calibration, integrity checking and frame missing filling, avoids the influence of data loss on fault diagnosis through estimation and filling of missing samples, improves the reliability of data, and further ensures the unity of multi-modal data in the same scale through amplitude normalization and feature scale unification processing.

[0077] In the embodiment, the generation process of the feature slice set corresponding to the candidate region specifically includes:

[0078] Based on the pre-processing data package, a processing batch list is established according to the time index and the modal order, sliding block and fixed step clipping are performed on each processing batch, a window queue is generated, and time index, window number and modal number are written for each window, and a to-be-inspected window set is output;

[0079] An abnormality detection process is performed on the to-be-inspected window set one by one, an initial candidate list containing position index, boundary information and confidence score is generated, threshold screening and overlap suppression are performed on the initial candidate list, the candidates passing the screening are retained, and candidate number, time index and window number are written, and an abnormal candidate region set is output;

[0080] According to the boundary information of the abnormal candidate region set, the corresponding region is clipped from the pre-processing data package, size alignment and coordinate archiving are performed, candidate number, time index and position index are written for each clipping result, and a feature slice set corresponding to the candidate region is formed.

[0081] The application establishes a processing list by batches, performs sliding block and cropping operations, generates a window queue containing time index, window number and modal number, ensures the orderliness and efficiency of processing data, and generates a preliminary candidate region list by performing abnormality detection on each window to be detected, retains effective candidate regions by threshold screening and overlap suppression algorithm, and further extracts the feature slice corresponding to the abnormal region by cropping and size alignment operations.

[0082] In the embodiment, the output of the structure maintaining representation tensor specifically includes:

[0083] The set of abnormal candidate regions and the set of feature slices are received, geometric normalization and resampling operations are performed on each candidate region to complete the consistent processing of scale, rotation and translation, the aligned candidate feature blocks are generated and the alignment index table is established, and according to the configuration of the quaternion channel mapping, the multi-channel real value features are distributed to four quaternion components according to the operation rules of Hamilton multiplication, the quaternion components include one real part and three imaginary parts, channel aggregation and quaternion encoding are completed at each pixel position, and the set of candidate quaternion fields is output;

[0084] A set of angle sampling sets is defined, a weight value is assigned to each angle, and based on the set of candidate quaternion fields, a lifting operator from the image plane domain to the rotation-translation group domain is established, and the group domain representation is calculated for each candidate number, any position and angle pair:

[0085] ;

[0086] wherein, represents the group domain representation result of the abnormal region with the candidate number at the position and the angle , represents the traversal and weighted accumulation performed on all pixel positions in the pixel grid of the candidate region, represents a direction-selective window function, represents a two-dimensional rotation operator, represents a pixel position index in the candidate region, represents the quaternion feature value of the candidate quaternion field with the candidate number at the pixel position , Hamilton multiplication representing quaternion; performing angle rotation on relative coordinates according to a two-dimensional rotation operator, weighting the rotated neighborhood according to a direction-selective window function, and weighting and accumulating the quaternion features in the neighborhood according to the Hamilton multiplication rule to obtain a group domain representation entry of the current position and angle pair, after calculating all position and angle pairs, writing the group domain representation entry into the group domain representation cache, and outputting the group domain representation cache and the position and angle index list;

[0087] According to the convolution kernel configuration, a set of rotation-translation group convolution kernels of quaternion values is generated, angle sampling registration, kernel rotation registration and angle association of quaternion rotation factors are completed, sliding calculation is performed on the group domain representation cache according to the position dimension and the angle dimension, the group domain representation entry and the corresponding rotated convolution kernel entry are read position by position and angle by angle according to the set step and the boundary processing strategy, the quaternion combination and accumulation are performed according to the Hamilton multiplication, the angle and position response value is obtained, and the angle and position response value of each calculation is written into the angle and position response set, and the convolution kernel identifier, angle index, position index and calculation timestamp are recorded, and the angle and position response set and the calculation log are output.

[0088] According to the angle sampling set and the corresponding weight, and combining the preset isovar aggregation rule, the angle and position response set is weighted and aggregated in the angle dimension to obtain a structure-preserving response in the position dimension, the isovar aggregation rule is set to ensure that the convolution output is still isovar under rotation and translation, and the position dimension is subsampled according to the set spatial step, and the sampling result is appended with a candidate number and a position index, and a structure-preserving representation tensor with rotation and translation isovariance is output.

[0089] The application effectively solves the rotation and translation isovariance problem in power equipment fault detection by introducing a quaternion channel mapping and a rotation-translation group convolution network. Through the operation rules of geometric normalization, resampling and Hamilton multiplication, the feature consistency is accurately maintained, the fault feature recognition under different perspectives and postures is ensured, the processing capability of multi-modal data is improved, the structure-preserving property under rotation and translation is ensured, and the accuracy and robustness of fault diagnosis are significantly enhanced.

[0090] In the embodiment, the output of the fault type label and the corresponding confidence vector specifically includes:

[0091] An index table is established according to the candidate number of the structure-preserving representation tensor set, each candidate tensor is traversed according to the position, the response of the tensor at each position point is decomposed into real value components according to the real part and three imaginary parts of the quaternion, all components in the same partition are point-by-point accumulated and the arithmetic mean is taken to form a statistical vector of the current partition, and the statistical vectors of all partitions are sequentially spliced according to the candidate number order to output a candidate representation vector set.

[0092] Load the preset classification parameter matrix and bias vector, which are pre-trained and stored by the system, and load them when the inspection device is initialized, perform linear transformation on each candidate feature vector, multiply the feature vector by the classification parameter matrix, and then add the corresponding component of the bias vector to obtain the original score vector of each candidate in all fault type categories, output the original score vector and the corresponding parameter call information together to generate the original score vector set;

[0093] Perform probability normalization operation on the original score vector of each candidate, respectively perform exponential transformation on the original scores of each category, and then sum the exponential values of all categories as a normalization factor, divide the exponential value of each category by the normalization factor to obtain the uncalibrated probability distribution of the candidate in all categories, perform temperature scaling calibration processing on the uncalibrated probability distribution, that is, scale the original score vector according to the preset temperature parameter, repeat the exponential transformation and normalization to obtain the calibration confidence vector of the candidate, select the category with the highest confidence in each calibration confidence vector as the fault type label, and pair it with the corresponding confidence component to form a diagnostic result item, and output the fault type label and the corresponding confidence vector.

[0094] The present application successfully extracts the multi-dimensional fault features of the power equipment by indexing the candidate number of the structure-preserving representation tensor and generating the feature vector, and accurately classifies the equipment fault type by loading the preset classification parameter matrix and bias vector. Combined with temperature scaling calibration and probability normalization, the accuracy of fault diagnosis is further optimized, and the confidence calculation of each candidate is more stable and reliable.

[0095] In the embodiment, the output process of the risk level code specifically includes:

[0096] For each candidate record, locate the confidence vector with the same name as the fault type label, generate a label-confidence pair list, and output the candidate number index;

[0097] Call the risk weight table and evaluation configuration saved during system initialization, perform risk evaluation operation on the label-confidence pair list, read the risk weight corresponding to the fault type label, and then weight and superimpose the risk weight and the corresponding confidence component to obtain the risk score of the current candidate, and after the calculation of all candidate numbers is completed, generate a risk score list;

[0098] Call the power grid safety risk level of the power equipment inspection industry standard, and perform threshold comparison on the risk score list in order from low to high, and when the risk score of a candidate first falls into a certain threshold interval, confirm the level corresponding to the interval as the risk level code of the current candidate.

[0099] The application combines the fault type label with the confidence vector, accurately evaluates the risk of the power equipment, uses the preset risk weight table to weight and calculate the confidence of each candidate, obtains the corresponding risk score, ensures the accuracy of the fault diagnosis result, and divides the risk level according to the power equipment inspection industry standard, so that the risk evaluation is more authoritative and operable.

[0100] In the embodiment, the output of the fault development trend curve and the degradation rate scalar specifically includes:

[0101] The structure-preserving representation tensor set and the fault type label set are received, the structure-preserving representation tensors of the same candidate number are time-sequenced according to the time index table, the time sequence tensor flow of each time point is established, the fault type label corresponding to the time point is attached to the time sequence tensor flow, the one-to-one correspondence relationship mapping of the time point-tensor-label is completed, and the time sequence data set is output;

[0102] The state evolution analysis is performed on the time sequence data set, the structure-preserving representation tensors of two adjacent time points are sequentially taken along the time index, the element difference operation is performed, the difference value is squared element by element and then summed and squared to obtain the overall change amplitude value of the current time interval, the change amplitude values of all intervals are sequentially arranged according to time to form a change sequence, the change sequence is smoothed according to a preset sliding window width and step, and the smoothed change sequence and the corresponding fault type label sequence are time-aligned and labeled, and a set of state evolution curves is output;

[0103] The dynamic parameter extraction is performed on the state evolution curve, the change amplitude value of the current interval is divided by the actual time length of the interval to obtain the instantaneous change rate, the instantaneous change rates in the observation window are weighted and averaged according to the time length to obtain the degradation rate scalar, the time axis and the candidate number identifier are attached to the state evolution curve, and the fault development trend curve and the degradation rate scalar are output after being paired and integrated according to the candidate number.

[0104] The application successfully extracts the evolution trend and the degradation rate of the power equipment fault by using the time sequence analysis technology and combining the structure-preserving representation tensor and the fault type label. Through the state evolution analysis, the limitation of single detection is eliminated, and the dynamic process of the change of the equipment state with time can be reflected.

[0105] In the embodiment, the generation of the diagnosis report data specifically includes:

[0106] The abnormal candidate region set, the risk level code set, the fault development trend curve and the degradation rate scalar are received, and are aligned according to the time index table and the candidate number, and a candidate data alignment list is generated;

[0107] The abnormal candidate area is sorted according to time index and position, the boundary information is corrected according to unified coordinate specification, the risk level code is written into the level mapping table according to the candidate number and time index, and the sorted abnormal item table and risk archive table are generated;

[0108] The labeling operation is performed on the fault development trend curve and the degradation rate scalar in the candidate data alignment list, the integration is performed according to the candidate number, the trend rate labeling table is generated, and the abnormal item table, the risk archive table and the trend rate labeling table are combined and loaded to generate and output the diagnosis report data.

[0109] The present application effectively integrates and labels the abnormal candidate area, the risk level code and the fault development trend, ensures the comprehensiveness and accuracy of fault diagnosis, eliminates the influence of time synchronization and spatial deviation through unified alignment and sorting of data, thereby improves the consistency and reliability of the diagnosis result, helps the accurate maintenance and long-period operation and maintenance management of the power equipment, and improves the safety and efficiency of the system

[0110] Embodiment 1:

[0111] In order to verify the feasibility of the present application in implementation, the present application is applied to the daily detection scene of a city power equipment inspection device. The traditional inspection method often relies on manual experience or simple threshold algorithm, and is easily disturbed by environmental factors such as light, angle, background noise, etc., resulting in missed detection or false alarm. For example, when the wire joint is slightly loose, the traditional method may not be able to find it in time, and an alarm will not be given until the temperature rises significantly, which is often too late and has a great safety hazard. The present application collects and time-synchronizes the data of different sensors through multi-modal data acquisition and time synchronization calibration, and avoids the information missing problem caused by inconsistent sampling, fundamentally solving the shortcomings of traditional methods in data quality and consistency.

[0112] In the specific application process, the visible light camera, infrared temperature measurement, partial discharge sensor and vibration sensor carried by the inspection device work simultaneously, and the collected data is preprocessed and detected after the abnormality, and the quaternion-SE(2) group convolution network is used to extract features, which ensures the stable recognition effect under different angles and attitudes. The system will automatically give the fault type, risk level and future development trend, for example, when detecting a slight discharge signal of the insulator, it can not only judge that it belongs to the insulator discharge fault, but also output the risk level as medium, and predict the upward trend of the degradation rate, providing early warning for maintenance personnel.

[0113] In the effect comparison, the application is evaluated with the traditional method. From the detection accuracy, false positive rate and trend prediction ability, the application is significantly better than the traditional scheme. Although the traditional method can find obvious abnormalities, it performs poorly in slight degradation and complex background; while the application greatly reduces false positives while maintaining high accuracy, and can provide predictive information through trend analysis, and table 1 gives typical comparison results.

[0114] Table 1 Performance comparison of the application method and the traditional method in power equipment inspection

[0115] ;

[0116] As can be seen from table 1, the application method is better than the traditional method in various indicators. The detection accuracy is improved from 85% to 95%, and the detection rate is increased from 83% to 96%, which shows that the identification of equipment failure is more comprehensive; the false positive rate is reduced from 12% to 4%, which significantly reduces invalid alarm and reduces repeated checking of operation and maintenance personnel; the fault positioning deviation is reduced from ± 6 pixels to ± 3 pixels, which makes the fault position determination more accurate; the risk classification accuracy is improved from 80% to 93%, which ensures that the risk assessment result is more in line with the actual situation; at the same time, the application can predict the fault trend 2 cycles in advance on average, while the traditional method does not have this function; the fault report completeness rate is increased from 82% to 94%, and the artificial review consistency rate is increased from 84% to 92%, which shows that the diagnosis report is more perfect and highly consistent with artificial judgment.

[0117] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited to this, any person skilled in the art in the technical range disclosed by the application, according to the technical scheme and the invention concept of the application, equivalent replacement or change, should be covered in the protection scope of the application.

Claims

1. A method for fault detection and diagnosis of a power equipment inspection device based on deep learning, characterized in that, The method comprises the following steps: Collecting multi-modal sensing data of the power equipment to generate an original data set with time identifiers and sensor identifiers; Performing time synchronization calibration and feature scale normalization on the original data set to generate a preprocessed data package; Performing anomaly detection on the preprocessed data package to output an abnormal candidate region set and generate a feature slice set corresponding to the candidate region; Inputting the abnormal candidate region set and the corresponding feature slice set into a quaternion-two-dimensional special Euclidean group equivariant group convolution network to perform quaternion channel mapping and two-dimensional special Euclidean group convolution operation, and output a structure-preserving representation tensor with rotation and translation equivariance; Performing classification operation and calibration operation on the structure-preserving representation tensor to output a fault type label and a corresponding confidence vector; Performing risk assessment on the fault type label and the confidence vector, performing level determination and threshold comparison, and outputting a risk level code; Performing time series processing on the structure-preserving representation tensor and the fault type label, performing state evolution analysis and dynamic parameter extraction, and outputting a fault development trend curve and a degradation rate scalar; Organizing the abnormal candidate region set, archiving the risk level code, and labeling the fault development trend curve and the degradation rate scalar to generate a diagnosis report data. 2.The deep learning based power equipment inspection device fault detection and diagnosis method of claim 1, wherein, The generation process of the original data set specifically comprises: Establishing a collection configuration, setting a unified time reference moment and a unified sampling interval to generate a sampling time list; Performing synchronous collection and frame construction, synchronously collecting multi-modal sensing data according to the sampling time list, writing time identifiers and sensor identifiers into each collection sample to generate a record unit, concatenating record units at the same sampling time into a frame and writing time index and modality order index, generating quality marks for each sample and labeling, outputting a frame set and a quality mark set; Performing data archiving and outputting, grouping and sorting the record unit set, the frame set and the quality mark set, and writing them into an original data set storage structure to form an original data set with time identifiers and sensor identifiers. 3.The deep learning based power equipment inspection device fault detection and diagnosis method of claim 1, wherein, The generation process of the preprocessed data package specifically comprises: Performing time synchronization calibration on all record units of the original data set, aligning time identifiers to the same time index sequence, performing time correction and generating a time index table, and outputting a time-aligned record unit set; Performing integrity check and frame missing completion on the record unit set, simultaneously generating a frame missing log, and outputting a complete time sequence sample set; Performing amplitude normalization and feature scale unification processing on the complete time sequence sample set, performing scaling and boundary truncation on each sample to form a scale-unified feature set, and merging the time index table, the complete time sequence sample set and the feature set into a preprocessed data package. 4.The deep learning based power equipment inspection device fault detection and diagnosis method of claim 1, wherein, The generation process of the feature slice set corresponding to the candidate region specifically comprises: Based on the preprocessed data package, a processing batch list is established according to time index and modality order, sliding block and fixed step clipping are performed on each processing batch, and a window set to be detected is outputted; The abnormality detection procedure is performed on each of the to-be-detected window set to generate an initial candidate list, threshold screening and overlap suppression are performed on the initial candidate list, the candidates passing the screening are retained, and the candidates are numbered, time-indexed and window-numbered, and an abnormal candidate region set is output; According to the boundary information of the abnormal candidate region set, the corresponding region is cropped from the preprocessed data packet to form a feature slice set corresponding to the candidate region. 5.The deep learning based power equipment inspection device fault detection and diagnosis method of claim 1, wherein, The output of the structure-preserving representation tensor specifically includes: The abnormal candidate region set and the feature slice set are received, geometric normalization and resampling operations are performed on each candidate region to complete the scale, rotation and translation consistency processing, aligned candidate feature blocks are generated and an alignment index table is established, and according to the quaternion channel mapping configuration, the multi-channel real value features are distributed to four quaternion components according to the operation rules of Hamilton multiplication, channel aggregation and quaternion encoding are completed at each pixel position, and a candidate quaternion field set is output; A set of angle sampling sets are defined, and each angle is assigned a weight value, and based on the candidate quaternion field set, a lifting operator from the image plane domain to the rotation-translation group domain is established, the group domain representation is calculated for each candidate number and any position and angle pair, the relative coordinates are rotated according to the two-dimensional rotation operator, the rotated neighborhood is weighted according to the direction-selective window function, and the quaternion features in the neighborhood are weighted and accumulated according to the Hamilton multiplication rule, to obtain the group domain representation entry of the current position and angle pair. After the calculation of all positions and angle pairs is completed, the group domain representation entry is written into the group domain representation cache, and the group domain representation cache and the position and angle index list are output. The sliding calculation is performed on the group domain representation cache according to the position dimension and the angle dimension, the group domain representation entry and the corresponding rotated convolution kernel entry are read position by position and angle by angle according to the set step distance and the boundary processing strategy, the quaternion combination and accumulation are performed according to the Hamilton multiplication, the angle and position response value is obtained, and the angle and position response value of each calculation is written into the angle and position response set, and the angle and position response set and the calculation log are output. According to the angle sampling set and the corresponding weight, and combined with the preset isometric aggregation rule, the angle and position response set is weighted and aggregated in the angle dimension to obtain the position dimension structure-preserving response, and the position dimension is subsampled according to the set spatial step, the candidate number and the position index are appended to the sampling result, and the structure-preserving representation tensor with rotation and translation isometry is output. 6.The deep learning based power equipment inspection device fault detection and diagnosis method of claim 1, wherein, The output of the fault type label and the corresponding confidence vector specifically includes: An index table is established according to the candidate number of the structure-preserving representation tensor set, all components in the same partition are point-by-point accumulated and the arithmetic mean is taken according to the position for each candidate tensor, and the candidate representation vector set is sequentially spliced and output according to the candidate number order; Linear transformation is performed on each candidate representation vector to obtain an original score vector of each candidate, and an original score vector set is generated. The probability normalization operation is performed on each candidate original score vector to obtain an uncalibrated probability distribution of the candidate in all categories, a temperature scaling calibration process is performed on the uncalibrated probability distribution to obtain a calibrated confidence vector of the candidate, the category with the highest confidence in each calibrated confidence vector is selected as a fault type label, and the fault type label and the corresponding confidence vector are output. 7.The deep learning based power equipment inspection device fault detection and diagnosis method of claim 1, wherein, The output process of the risk level code specifically includes: For each candidate record, locate the confidence vector with the same name as the fault type label, generate a label-confidence pair list, and output the candidate number index; Call the risk weight table and evaluation configuration saved when the system is initialized, and perform risk evaluation operation on the label-confidence pair list one by one, and after the calculation of all candidate numbers is completed, generate a risk score list; Call the power equipment inspection industry standard grid safety risk level, and perform threshold comparison on the risk score list one by one in the order from low to high, and when the risk score of a certain candidate first falls into a certain threshold interval, the level corresponding to the interval is confirmed as the risk level code of the current candidate. 8.The deep learning based power equipment inspection device fault detection and diagnosis method of claim 1, wherein, The output of the fault development trend curve and the degradation rate scalar specifically includes: Receive a set of structure retention representation tensors and a set of fault type labels, time-sort the structure retention representation tensors of the same candidate number according to the time index table, and append the fault type label corresponding to the time to the time series tensor stream, and output a time series dataset; Perform state evolution analysis on the time series dataset, group the change amplitude values of all intervals in chronological order to form a change sequence, perform smoothing processing on the change sequence according to a preset sliding window width and step, time-align and label the smoothed change sequence and the corresponding fault type label sequence, and output a set of state evolution curves; Perform dynamic parameter extraction on the state evolution curve, and append the time axis and the candidate number identifier, and pair with the corresponding degradation rate scalar, and output the fault development trend curve and the degradation rate scalar after being integrated according to the candidate number. 9.The deep learning based power equipment inspection device fault detection and diagnosis method of claim 1, wherein, The generation of the diagnostic report data specifically includes: Receive the abnormal candidate region set, the risk level code set, the fault development trend curve and the degradation rate scalar, align them according to the time index table and the candidate number, and generate a candidate data alignment list; Sort the abnormal candidate regions according to the time index and the position, correct the boundary information according to the unified coordinate specification, write the risk level code into the level mapping table according to the candidate number and the time index, generate the sorted abnormal item table and the risk archive table; Perform labeling operation on the fault development trend curve and the degradation rate scalar in the candidate data alignment list, integrate them according to the candidate number, generate a trend rate labeling table, and merge the abnormal item table, the risk archive table and the trend rate labeling table to generate and output the diagnostic report data.

Images