Substation fault prediction method based on multi-stage equipment collaboration

Abstract

The invention relates to the technical field of transformer substation operation and maintenance, in particular to a transformer substation fault prediction method based on multi-stage equipment collaboration, which comprises the following steps of: in the operation process of a transformer substation, acquiring multi-source real-time monitoring data of target equipment in a current period to obtain an equipment real-time state feature vector; acquiring a fault feature range and a feature weight corresponding to each feature under each preset fault type in the current period, matching each feature data in the real-time state feature vector of the equipment, and if at least two preset fault types are matched, recording the fault type as a target fault type; for any target fault type, in combination with the feature conformity and the feature weight of each feature in any target fault type, obtaining the fault matching confidence of any target fault type; and according to the fault matching confidence of each target fault type, the final fault type of the target equipment in the current period is determined, so that the fault false alarm rate can be reduced to a great extent, and the fault prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of substation operation and maintenance, and particularly to a substation fault prediction method based on multi-level device collaboration. Background Art

[0002] With the advancement of the intelligent transformation of the power system, the substation, as the core hub of the power grid, faces increasingly complex challenges in its safe operation. Among them, the fault prediction of substation equipment is a key link to ensure the safe and stable operation of the power grid. Currently, the existing technologies generally adopt multi-source perception technologies (such as infrared thermal imaging, partial discharge detection, vibration acoustic fingerprint monitoring, etc.) to collect the status data of substation equipment and use artificial intelligence models for fault early warning. However, in actual deployment, the core problem faced by multi-source perception technologies is the low efficiency of multi-source data collaboration, resulting in a high false alarm rate of faults, which seriously restricts the improvement of fault prediction accuracy. The false alarms in substations not only cause the consumption of ineffective operation and maintenance resources, but also may lead to the concealment of real potential hazards in false warnings.

[0003] Therefore, in the existing substation fault prediction methods, most of them improve the collaboration effect of multi-source data through data alignment algorithms or feature-level fusion, but there are essential defects that make the proportion of false alarms caused by the failure of multi-source data collaboration still relatively high, restricting the improvement of fault prediction accuracy. Among them, the essential defects are specifically manifested as follows: static fault prediction rules cannot adapt to changes in equipment status and environment (such as equipment aging, climate change), and only raw data or low-level features are transmitted between edge terminals (such as unmanned aerial vehicles, inspection robots, fixed sensors) and the central system, without realizing fault semantic-level collaborative reasoning (such as the mapping verification between the combined feature of "vibration + temperature + current" and the "mechanical jamming" fault mode). Summary of the Invention

[0004] In view of this, an embodiment of the present invention provides a substation fault prediction method based on multi-level device collaboration to solve the problem that the low efficiency of multi-source data collaboration restricts the accuracy of substation fault prediction.

[0005] An embodiment of the present invention provides a substation fault prediction method based on multi-level device collaboration, and the method includes the following steps: During the operation of the substation, multi-source real-time monitoring data of a target device in the current cycle is collected by multi-level devices, and feature extraction is performed on the multi-source real-time monitoring data to obtain a device real-time status feature vector composed of at least two feature data; Obtain the fault feature range and feature weight corresponding to each feature in the current cycle under each preset fault type respectively. Use the fault feature range corresponding to each feature under each preset fault type to match each feature data in the real-time state feature vector of the device. If at least two preset fault types are matched, record the matched preset fault types as the target fault types; For any target fault type, according to the relationship between each feature data in the real-time state feature vector of the device and the fault feature range corresponding to each feature under the any target fault type, obtain the feature compliance of each feature under the any target fault type respectively. Combine the feature compliance and feature weight of each feature under the any target fault type to obtain the fault matching confidence of the any target fault type; Obtain the fault matching confidence of each target fault type. According to the fault matching confidence of each target fault type, determine the final fault type of the target device in the current cycle.

[0006] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows: In the present invention, by converting the correlation of core physical parameters (such as discharge is necessarily accompanied by acoustic-optical-thermal effects) into a device state feature vector that can represent the operating state of the device, and then dynamically adjusting the initial feature hyperparameter set obtained based on simulation, the fault feature range and feature weight corresponding to each feature in the current cycle under each preset fault type are obtained, so as to avoid the accuracy attenuation of fault prediction using static fault thresholds and feature parameters due to device aging and environmental changes. Therefore, use the fault feature range and feature weight corresponding to each feature in the current cycle under each preset fault type to match the feature data in the real-time state feature vector of the device in the current cycle. And when at least two preset fault types are matched, by analyzing the fault matching confidence of each preset fault type to determine the final fault type, the false alarm rate of faults can be greatly reduced and the accuracy of fault prediction can be improved. Description of the Drawings

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0008] Figure 1 It is a flowchart of a substation fault prediction method based on multi-level device collaboration provided in Embodiment 1 of the present invention. Detailed Embodiment

[0009] Embodiments of the present disclosure will be described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present disclosure, and should not be construed as a limitation to the present disclosure.

[0010] It should be noted that the terms "first", "second", etc. in the specification of the present disclosure and the above-mentioned accompanying drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure.

[0011] In order to illustrate the technical solution of the present invention, specific embodiments will be used for illustration below.

[0012] See Figure 1 , which is a method flow chart of a substation fault prediction method based on multi-level device collaboration provided in the first embodiment of the present invention. As Figure 1 shown, the method may include: Step S101, during the operation of the substation, use multi-level devices to collect multi-source real-time monitoring data of the target device in the current cycle, extract features from the multi-source real-time monitoring data, and obtain a device real-time state feature vector composed of at least two feature data.

[0013] The main purpose of the present invention is to predict the fault type and the authenticity of the occurrence of this type of fault for each device in the substation through a method of multi-level device collaboration, so as to improve the prediction accuracy, reduce the false alarm rate of faults, thereby enhancing the emergency protection ability of the substation and reducing the operation and maintenance costs of the substation. Among them, the devices include, but are not limited to, heat dissipation-dependent devices (such as transformers, reactors) and mechanical operation devices (such as circuit breakers, disconnectors) in the substation, as well as other key devices with multi-physical quantity monitoring characteristics. Therefore, in the embodiments of the present invention, any device in the substation is taken as an example, and any device is used as the target device. By collecting multi-source monitoring data of the target device, predicting its fault type, and transmitting the prediction result to the early warning center, the early warning center can take corresponding operation and maintenance measures according to the prediction result.

[0014] Among them, the multi-source monitoring data includes: infrared thermal imaging of the surface of the target device obtained by using an unmanned aerial vehicle to scan the surface of the target device at high altitude; vibration signals and acoustic emission signals of the target device obtained by using a quadruped robot to approach the base of the target device; hydrogen concentration in the oil valve obtained by using a patrol robot to connect to the oil valve of the target device. The specific acquisition method of the multi-source monitoring data is: (1)Time synchronization Adopt a clock synchronization protocol to unify the timestamps of all enabled multi-level devices (unmanned aerial vehicles, quadruped robot dogs, patrol robots), thereby unifying the timestamps of all collected parameters. There is asynchrony in the sensor data acquisition of multi-level devices, and timestamp unification can eliminate the timing deviation caused by sampling frequency differences (such as 10fps for thermal imaging of unmanned aerial vehicles and 44.1kHz sampling for acoustic fingerprints of robot dogs). For example, when an unmanned aerial vehicle detects abnormal temperature rise of a device, it needs to be precisely aligned with the acoustic emission signal collected by the robot dog on the time axis to verify whether it is the same fault event.

[0015] (2)Spatial synchronization Unmanned aerial vehicle: The local coordinate system is the local point cloud coordinate system constructed by SLAM (Simultaneous Localization and Mapping). The global UWB (Ultra-Wideband) point cloud is matched through the ICP (Iterative Closest Point) algorithm, and the 6-degree-of-freedom rigid body transformation matrix is solved. Based on the rigid body transformation matrix, the local coordinates of the unmanned aerial vehicle are converted into global coordinates.

[0016] Quadruped robot dog: The local coordinate system is a two-dimensional plane coordinate system based on odometer + IMU (Inertial Measurement Unit), and the global coordinates are directly obtained through the UWB tag on the back.

[0017] Patrol robot: The local coordinate system is a preset track coordinate system. When installed, the starting point of the track is laser-calibrated, the displacement is calculated through the encoder, and based on the displacement, the local coordinates of the track patrol robot are mapped into global coordinates.

[0018] In this way, even if multi-level devices collect data based on different coordinate systems (such as the local coordinate system of SLAM of unmanned aerial vehicles and the global coordinate system of UWB), the coordinates of the collected data can ultimately be converted to the global coordinate system to achieve spatial synchronization.

[0019] Since the occurrence of equipment failures does not form in a short period of time, the set fault prediction period should not be too short. However, if the fault prediction period is too long, it will cause delays in fault prediction and it is easy to have untimely operation and maintenance. Therefore, in the embodiment of the present invention, the period of equipment fault prediction is set to one day, which is not limited here and can be set according to the implementation scenario.

[0020] Based on the above multi-source monitoring data acquisition method, multi-source monitoring data of the target device in each period can be collected, and then multi-source real-time monitoring data of the target device in the current period can be obtained. The multi-source real-time monitoring data includes an infrared thermal imaging image sequence, vibration signals, acoustic emission signals, and a hydrogen concentration sequence in the oil valve.

[0021] Due to the dimensional differences and uneven numerical scales of multi-source monitoring data, in order to avoid large-order features dominating the decision-making, it is necessary to extract features from the multi-source real-time monitoring data to obtain a device real-time status feature vector composed of at least two feature data. The specific method is as follows: Use the Sobel operator to calculate the gradient value of each pixel point in each image of the infrared thermal imaging image sequence respectively, obtain the absolute value of the maximum gradient value, and normalize the absolute value of the maximum gradient value to obtain the maximum temperature gradient value; Perform Fourier transform on the vibration signal to obtain a spectrogram, obtain the entropy value of the spectrogram, and normalize the entropy value to obtain the vibration spectrum entropy; Obtain the energy value corresponding to the acoustic emission signal, and normalize the energy value to obtain the acoustic emission energy; Perform moving average filtering on the hydrogen concentration sequence to obtain a filtered sequence, calculate the average value of the filtered sequence, and normalize the average value to obtain the average hydrogen concentration; Take the maximum temperature gradient value, the vibration spectrum entropy, the acoustic emission energy, and the average hydrogen concentration as feature data to form a device real-time status feature vector.

[0022] It should be noted that the absolute value of the maximum gradient value, the entropy value of the spectrogram, and the average value of the filtered sequence belong to continuous features, and the Z-score (standard score) method is used for normalization; the energy value corresponding to the acoustic emission signal belongs to a count feature, and the Min-Max (range normalization) method is used for normalization. The Sobel operator and normalization both belong to existing technologies and will not be elaborated here.

[0023] Similarly, the device status feature vectors in each cycle before the current cycle of the target device can also be obtained.

[0024] Step S102: Respectively obtain the fault feature range and feature weight corresponding to each feature in each preset fault type in the current cycle. Use the fault feature range corresponding to each feature in each preset fault type to match each feature data in the device real-time status feature vector. If at least two preset fault types are matched, record the matched preset fault types as the target fault types.

[0025] Before determining the fault type of the target device in the current period, it is necessary to obtain the fault thresholds corresponding to each characteristic data involved in each fault type, so as to determine which fault type the target device belongs to according to the real-time state characteristic vector of the target device. Therefore, the accuracy of the fault thresholds corresponding to each characteristic data involved in each fault type will directly affect the accuracy of fault type judgment. Since in the past, the fault thresholds corresponding to each characteristic data involved in each fault type were mostly set through statistical experience, which could not match the current real situation of the device, resulting in false alarms of fault types. Therefore, the embodiments of the present invention establish a digital twin model of the target device, match the current real situation of the target device, and simulate multiple fault types to obtain accurate fault thresholds.

[0026] Specifically, first construct a digital twin model of the target device. When constructing the digital twin model, it is necessary to first determine the geometric model of the target device (modeled based on the actual size of the target device), physical models (including electromagnetic field, thermal field, structural mechanics field, and fluid dynamics field), and boundary conditions (including environmental temperature, load current, voltage level, cooling conditions, etc.) to ensure the consistency between the digital twin model and the actual situation of the target device, so as to ensure the accuracy of the simulation results obtained during subsequent simulations.

[0027] Then, due to the uncertainty of the target device state, such as device size deviation, heat conduction deviation of material properties, operation conditions and fault degree differences, a single simulation cannot cover most working conditions, which will lead to the inability of the fault thresholds obtained based on a single simulation to adapt to actual fluctuations, and the accuracy of setting the fault boundary conditions is not high. Therefore, the embodiments of the present invention use the Monte Carlo method to perform multiple fault simulations on the digital twin model to solve the uncertainty of the fault thresholds and boundary conditions through probability sampling.

[0028] The embodiments of the present invention set 4 preset fault types: partial discharge, mechanical looseness, insulation dampness, and overheating fault. There is no limitation here and can be set according to the implementation scenario. Then, 100 Monte Carlo simulations are respectively run for each preset fault type to obtain a set of device simulation state characteristic vectors under each preset fault type. One Monte Carlo simulation outputs one device state characteristic vector.

[0029] Finally, according to the set of device simulation state feature vectors under each preset fault type, obtain the initial fault feature range corresponding to each feature under each preset fault type. Specifically, in Monte Carlo simulation, since the input parameter (fault severity) is perturbed with a normal distribution, according to the central limit theorem, the output device state feature vectors approximately follow a normal distribution. For a normal distribution, by querying the standard normal distribution table, it can be obtained that the interval taking 1.96 times the standard deviation can cover 95% of the data, which means that in 100 simulations, the eigenvalue of about 95 samples will fall into this interval. Therefore, this interval reflects the natural fluctuation range of the device state feature vectors under a specific fault type. Device state feature vectors outside this range are very likely to be other faults or abnormal states. In the embodiments of the present invention, for the j-th feature under the i-th preset fault type, according to the feature simulation data corresponding to the j-th feature in the set of device simulation state feature vectors under the i-th preset fault type, obtain the average feature simulation data and the standard deviation of the feature simulation data, and set as the initial fault feature range corresponding to the j-th feature under the i-th preset fault type, where represents the average feature simulation data, represents the standard deviation of the feature simulation data.

[0030] Illustrative example: When the fault type is mechanical looseness, after performing Monte Carlo simulation, the vibration spectrum entropy in the device simulation state feature vector is significantly higher than the normal state, indicating that the device vibration is significantly enhanced. Therefore, the lower and upper limits of the fault feature threshold of the vibration spectrum entropy for the fault type of mechanical looseness will be significantly higher than the normal state; similarly, when the fault type is insulation dampness, after performing Monte Carlo simulation, the average hydrogen concentration in the device simulation state feature vector is significantly increased, indicating a sudden increase in hydrogen, and at the same time, the maximum temperature gradient value decreases significantly, indicating abnormal cooling of the device. Therefore, the lower and upper limits of the fault feature threshold of the average hydrogen concentration for the fault type of insulation dampness will be significantly higher than the normal state, while the lower and upper limits of the fault feature threshold of the maximum temperature gradient value will be lower than the normal state; when the fault type is overheating fault, after performing Monte Carlo simulation, the maximum temperature gradient value exceeds twice the normal value, indicating extreme overheating of the device. Then, the lower and upper limits of the fault feature threshold of the maximum temperature gradient value for the fault type of overheating fault will be higher than the normal state and other fault types.

[0031] Meanwhile, since there may be overlapping features for different fault types, relying solely on fault thresholds to determine fault types may lead to confusion in fault types. For example, multiple fault type determinations may be satisfied simultaneously, resulting in false alarms. Therefore, it is necessary to calculate the confidence level of each fault type occurrence, and the feature weight is the key basis for calculating the confidence level, which represents the ability of each feature corresponding to each fault type to distinguish this fault from other faults. Thus, the fault confidence level is calculated by assigning the feature weights of each feature corresponding to each fault type, improving the ability to determine fault types based on confidence levels, and thereby enhancing the accuracy of fault prediction.

[0032] In the embodiments of the present invention, for the j-th feature under the i-th preset fault type, first, according to the difference in feature simulation data of the j-th feature under the i-th preset fault type and under other preset fault types, the discrimination ability index of the j-th feature is obtained. The obtaining method is as follows: Since after performing 100 Monte Carlo simulations for each preset fault type, a set of feature simulation data for each feature of each preset fault type is obtained. Each set of feature simulation data is independently generated and is a one-dimensional array containing 100 feature simulation data. Therefore, the Wasserstein distance can be used to quantify the distribution difference between the set of feature simulation data of the j-th feature under the i-th preset fault type and the set of feature simulation data of the j-th feature under the y-th preset fault type. The Wasserstein distance can quantify the cost of transporting one distribution to another. If the distribution difference (such as the mean difference and standard deviation difference) is large, the transportation cost will be large, that is, the Wasserstein distance will be large. Then, the Wasserstein distance between the sets of feature simulation data of the j-th feature corresponding to two different preset fault types is used as the feature separation degree of the j-th feature for these two different preset fault types.

[0033] The set of feature simulation data of the j-th feature of the i-th preset fault type is used as the reference group, and the set of feature simulation data of the j-th feature under the y-th preset fault type is used as the comparison group. Since the two sets of data need to be aligned according to their respective quantiles (i.e., from smallest to largest) for one-to-one comparison, the reference group is sorted in ascending order to obtain the target sequence; the comparison group is sorted in ascending order to obtain the reference sequence.

[0034] Calculate the absolute value of the difference between the two elements at each same position between the target sequence and the reference sequence, and obtain the mean value of the absolute values of the differences as the feature separation degree of the j-th feature between the i-th preset fault type and the y-th preset fault type. The formula for the feature separation degree is as follows:

[0035] where represents the feature separation degree of the j-th feature between the i-th preset fault type and the y-th preset fault type, n represents the element data in the target sequence or the reference sequence, represents the a-th data in the target sequence, represents the a-th data in the reference sequence, and | | represents the absolute value symbol.

[0036] It should be noted that the larger the W, that is, the larger the feature separation degree, the greater the distribution difference between the two groups of data, and the easier it is to distinguish the two groups of data; the smaller the W, that is, the smaller the feature separation degree, the smaller the distribution difference between the two groups of data, and the more difficult it is to distinguish the two groups of data. Then the j-th feature is more important for judging the i-th preset fault type and the y-th preset fault type, and the feature weight assigned to this feature should be larger to avoid misjudgment of the fault type.

[0037] Similarly, obtain the separation degree of the j-th feature between the i-th preset fault type and each reference type. The feature separation degree reflects the distribution difference of a certain feature between different fault types. After Monte Carlo simulation, if the distribution difference of the feature simulation data of the same feature of two fault types is smaller, it means that the similarity of the two fault types in this feature is higher and it is more difficult to distinguish, so the feature separation degree is smaller. At the same time, it also means that this feature is more important for distinguishing the two fault types. Therefore, the separation degree of the j-th feature between the i-th preset fault type and each reference type is obtained, and the minimum separation degree is taken as the discrimination ability index of the j-th feature under the i-th preset fault type.

[0038] Finally, obtain the discrimination ability index of each feature under the i-th preset fault type, use the norm function to normalize the discrimination ability index of all features under the i-th preset fault type to obtain the corresponding normalized value, and take the difference between the constant 1 and each normalized value as the initial feature weight corresponding to each feature under the i-th preset fault type.

[0039] At this time, the initial fault feature range and the initial feature weight corresponding to each feature under the i-th preset fault type are obtained. Similarly, the initial fault feature range and the initial feature weight corresponding to each feature under each preset fault type are obtained, and the initial fault feature ranges and the initial feature weights corresponding to each feature under all preset fault types are combined into an initial feature parameter set, which is used to determine the fault type of the target device according to the actually obtained device state feature vector.

[0040] After determining the initial feature parameter set, each feature data in the device state feature vector obtained by the target device in any cycle can be compared with the initial fault feature range corresponding to each feature under each preset fault type in the initial feature parameter set, so as to match the corresponding fault type. If no match is found, it is determined that the target device is operating normally; if at least one fault type is matched, the analysis of the fault confidence level needs to be carried out to facilitate the equipment operation and maintenance based on the fault confidence level.

[0041] Since the initial feature parameter set is obtained based on simulation, with factors such as equipment aging and environmental changes, static fault thresholds and feature weights will affect the accuracy of fault prediction. Therefore, the embodiment of the present invention proposes a method for updating the fault threshold and the feature weight based on the initial feature parameter set, that is, after determining the final fault type of the target device in each cycle, the initial feature parameter set is dynamically updated according to the cumulative detection results of the final fault type, and by adjusting the fault feature range and the feature weight corresponding to each feature under the preset fault type, the problem of prediction accuracy attenuation caused by equipment aging and environmental changes is avoided.

[0042] Specifically, first, according to the initial feature parameter set and the device state feature vector of the target device in each cycle, the final fault type of the target device in each cycle and the corresponding fault matching confidence level are respectively obtained. The specific obtaining method is referred to below. Then, the number of times that each preset fault type is the final fault type is cumulatively recorded, and at least one preset fault type corresponding to the cumulative number meeting the preset number threshold is obtained, which is recorded as the optimized fault type. It should be noted that if no optimized fault type is obtained, the initial feature parameter set is always used to predict the final fault type of the target device in each cycle.

[0043] Preferably, the preset number threshold of the present invention is set to 10, that is, when the number of times the final fault type is the same preset fault type exceeds 10 times, the adjustment of the fault feature range and feature weight can be triggered. It should be noted that the preset number threshold cannot be too small, otherwise there will be insufficient sample quantity and poor data reliability, resulting in a decrease in the updated prediction accuracy; if the preset number threshold is too high, there will be an update gap and the prediction accuracy will be unsatisfactory for a period of time. Specifically, the preset number threshold corresponding to each preset fault type can be adjusted according to the climate of the area where the substation is located and the actual situation of the substation. For example, if the area where the substation is located has continuous rain and humid climate all year round, the possibility of insulation dampness failure is relatively high and the sample quantity is relatively sufficient, and the preset number threshold can be appropriately increased so that when updating the fault threshold and feature weight for insulation dampness failure, the data basis is more sufficient. There is no limitation here.

[0044] Then, for any optimized fault type, assuming that any optimized fault type is obtained after the final fault type in the b-th cycle of the target device is determined, then all fault matching confidence levels corresponding to taking any optimized fault type as the final fault type before the b-th cycle are obtained, as well as the device state feature vectors of the target device in all corresponding cycles. And according to all the device state feature vectors and all the fault matching confidence levels corresponding to any optimized fault type, the initial fault feature range and initial feature weight corresponding to each feature under any optimized fault type are updated to obtain the new fault feature range and new feature weight corresponding to each feature under any optimized fault type. The specific update method is as follows: (1) According to all the device state feature vectors corresponding to any optimized fault type, the mean value of the feature data of each feature is obtained , and according to all the fault matching confidence levels corresponding to any optimized fault type, the mean value of the fault matching confidence levels is calculated .

[0045] (2) For any feature under any optimized fault type, obtain the intermediate value in the initial fault feature range corresponding to any feature, calculate the absolute value of the difference between the mean value of the feature data of any feature and the intermediate value, and take the ratio of the absolute value of the difference to the intermediate value as the weight gradient of any feature; take the product of the weight gradient, the mean value of the fault matching confidence levels and the preset weight learning rate as the weight adjustment value, and take the sum of the initial feature weight corresponding to any feature and the weight adjustment value as the new feature weight corresponding to any feature.

[0046] In an embodiment, taking the j-th feature under any optimized fault type as an example, the calculation formula for the new feature weight corresponding to the j-th feature is:

[0047] Among them, represents the new feature weight corresponding to the j-th feature, represents the initial feature weight corresponding to the j-th feature, represents the preset weight learning rate, represents the weight gradient of the j-th feature, represents the average value of the fault matching confidence.

[0048] It should be noted that is used to control the weight update step size, and setting , which is a common default value in the tradition; is used to quantify the feature contribution degree, so as to accurately locate the features that need to be optimized. Among them, represents the average value of the feature data corresponding to the j-th feature, represents the middle value in the initial fault feature range corresponding to the j-th feature under any optimized fault type, that is, the ideal value. The larger the value of , the greater the gap between the feature data of the j-th feature in the specification and the ideal value, the greater the corresponding weight gradient, the greater the feature contribution degree of the j-th feature, and the greater the corresponding adjustment degree.

[0049] (3) Obtain the upper limit value and the lower limit value in the initial fault feature range corresponding to any one of the features, obtain the first difference between the average value of the feature data of any one of the features and the lower limit value, use the product of the preset learning rate and the first difference as the lower limit adjustment value, and use the difference between the lower limit value and the lower limit adjustment value as the new lower limit value; obtain the second difference between the average value of the feature data of any one of the features and the upper limit value, use the product of the preset learning rate and the second difference as the upper limit adjustment value, and use the sum of the upper limit value and the upper limit adjustment value as the new upper limit value; obtain the new fault feature range corresponding to any one of the features according to the new upper limit value and the new lower limit value.

[0050] In an embodiment, taking the j-th feature under any optimized fault type as an example, the update calculation formula for the lower limit value of the initial fault feature range corresponding to the j-th feature is:

[0051] Among them, represents the new lower limit value corresponding to the j-th feature, represents the lower limit value of the initial fault feature range corresponding to the j-th feature, represents the preset learning rate, represents the average value of the feature data corresponding to the j-th feature.

[0052] The update calculation formula for the upper limit value of the initial fault feature range corresponding to the j-th feature is as follows: Wherein, represents the new upper limit value corresponding to the j-th feature, represents the upper limit value of the initial fault feature range corresponding to the j-th feature, represents the preset learning rate, represents the mean value of the feature data corresponding to the j-th feature.

[0053] It should be noted that, is used to control the adjustment speed of the fault feature range. In the embodiments of the present invention, it is set as .

[0054] The initial feature parameter set is updated by using the new fault feature range and the new feature weight corresponding to each feature under each of the optimized fault types to obtain a new feature parameter set. Specifically: the initial fault feature range and the initial feature weight corresponding to each feature under the corresponding preset fault type in the initial feature parameter set are replaced with the new fault feature range and the new feature weight corresponding to each feature under each of the optimized fault types to obtain a new feature parameter set.

[0055] For example, if the number of times of the final fault type of insulation moisture absorption reaches the preset number threshold, it is found through the update of the initial feature parameter set that the initial feature weight corresponding to the average hydrogen concentration in the insulation moisture absorption fault type slightly increases, which indicates that the influence of the feature of the recent average hydrogen concentration on the insulation moisture absorption fault type becomes larger. Therefore, the contribution degree of the average hydrogen concentration feature to the insulation moisture absorption fault is updated and improved; at the same time, the initial fault feature range corresponding to the average hydrogen concentration becomes larger after the update, which indicates that similar faults may be more likely to occur in the future. Thus, the initial fault feature range is expanded through the update, so that it is easier to detect similar faults by using the updated initial feature parameter set (that is, the new feature parameter set).

[0056] After that, the new feature parameter set is used as the initial feature parameter set, and the final fault type of the target device in each cycle is continuously obtained. When an optimized fault type appears, the initial feature parameter set (that is, the new feature parameter set) is re-updated according to the update method of the initial feature parameter set until the final fault type of the target device in the previous cycle of the current cycle is obtained, and according to the final fault type of the target device in the previous cycle of the current cycle, the fault feature range and the feature weight corresponding to each feature under each preset fault type in the current cycle are obtained.

[0057] It should be noted that if no optimized fault type is obtained based on the final fault type of the target device in the previous cycle of the current cycle, it indicates that as of the previous cycle of the current cycle, the cumulative number of any preset fault type being the final fault type does not meet the preset number threshold. Therefore, there is no need to update the initial feature parameter set. Instead, directly use the initial fault feature range and initial feature weight corresponding to each feature under each preset fault type in the initial feature parameter set as the fault feature range and feature weight corresponding to each feature under each preset fault type in the current cycle. Conversely, if any optimized fault type is obtained, update the initial feature parameter set, and use the initial fault feature range and initial feature weight corresponding to each feature under each preset fault type in the updated initial feature parameter set as the fault feature range and feature weight corresponding to each feature under each preset fault type in the current cycle.

[0058] It should be noted that all the initial feature parameter sets after the first update of the initial feature parameter set refer to the new feature parameter sets obtained after the update.

[0059] Thus, the fault feature range and feature weight corresponding to each feature under each preset fault type in the current cycle can be obtained. Further, using the fault feature range and feature weight corresponding to each feature under each preset fault type in the current cycle, match each feature data in the device real-time state feature vector obtained for the target device in the current cycle, that is, detect whether each feature data in the device real-time state feature vector is within the fault feature range of the corresponding feature under any preset fault type. If it is, the match is successful, and the preset fault type with the successful match is recorded as the target fault type; otherwise, the match is not successful. If one target fault type is matched, use the matched preset fault type as the final fault type of the target device in the current cycle, and obtain the fault matching confidence of the matched preset fault type; if multiple target fault types are matched, determine the final fault type of the target device in the current cycle according to the fault confidence of each target fault model.

[0060] Step S103, for any target fault type, respectively obtain the feature compliance of each feature under any target fault type according to the relationship between each feature data in the device real-time state feature vector and the fault feature range corresponding to each feature under any target fault type. Combine the feature compliance and feature weight of each feature under any target fault type to obtain the fault matching confidence of any target fault type.

[0061] When analyzing the confidence level of each target fault type in the embodiments of the present invention, not only the feature weights corresponding to each feature under the target fault type are considered, but also the feature compliance of each feature data in the device real-time state feature vector of the target device in the current period needs to be considered, that is, the degree of compliance of a single feature data with the standard feature under the target fault type. Therefore, taking any target fault type as an example, according to the relationship between each feature data in the device real-time state feature vector and the fault feature range corresponding to each feature under any target fault type, the feature compliance of each feature under any target fault type is obtained respectively.

[0062] Among them, obtaining the feature compliance of each feature under any target fault type respectively according to the relationship between each feature data in the device real-time state feature vector and the fault feature range corresponding to each feature under any target fault type includes: For any feature under any target fault type, calculate the absolute value of the difference between the feature data corresponding to the feature in the device real-time state feature vector and the intermediate value within the fault feature range corresponding to the feature under any target fault type, and record it as the deviation degree; Calculate the absolute value of the difference between the upper limit value or the lower limit value of the fault feature range corresponding to the feature under any target fault type and the intermediate value, and record it as the reference feature value; calculate the ratio of the deviation degree to the reference feature value, and use the difference between the constant 1 and the ratio as the feature compliance of the feature.

[0063] In an embodiment, taking the jth feature under any target fault type as an example, the calculation formula for the feature compliance of the jth feature is:

[0064] Among them, represents the feature compliance of the jth feature, 1 represents a constant, represents the feature data corresponding to the jth feature in the device real-time state feature vector, represents the intermediate value within the fault feature range corresponding to the jth feature under any target fault type, represents the absolute value of the difference between the upper limit value (or the lower limit value) of the fault feature range corresponding to the jth feature under any target fault type and the intermediate value, that is, the reference feature value.

[0065] It should be noted that represents the distance between the feature data of the jth feature and the intermediate value within the fault feature range corresponding to it under any target fault type. The closer the distance is, the is smaller, and the obtained feature compliance The larger the value is, the higher the degree of conformity between the j-th feature and the standard feature of any target fault type, and the higher its contribution in the subsequent calculation of fault matching confidence.

[0066] Similarly, the feature conformity of each feature under any target fault type is obtained, and then the feature conformity and feature weight of each feature under any target fault type are combined to obtain the fault matching confidence of any target fault type: according to the feature weight of each feature under any target fault type, the feature conformity of all features under any target fault type is weighted and summed to obtain the fault matching confidence of any target fault type.

[0067] In one embodiment, the calculation formula for the fault matching confidence of any target fault type is:

[0068] in, represents the fault matching confidence of any target fault type, represents the feature weight of the jth feature under any target fault type, It represents the feature conformity of the jth feature under any target fault type.

[0069] Step S104 : obtaining the fault matching confidence of each target fault type, and determining the final fault type of the target device in the current cycle according to the fault matching confidence of each target fault type.

[0070] According to the above-mentioned method for obtaining the fault matching confidence of any target fault type, the fault matching confidence of each target fault type is obtained according to the real-time device status feature vector of the target device in the current cycle, and based on the fault matching confidence of each target fault type, the target fault type corresponding to the maximum fault matching confidence is used as the final fault type of the target device in the current cycle.

[0071] At this point, the final fault type of the target device in the current cycle and the corresponding fault matching confidence can be obtained, and then according to the updating method of the initial feature parameter set in step S102, that is, the number of times each preset fault type is the final fault type is continued to be accumulated and recorded, and at least one preset fault type corresponding to the cumulative number that meets the preset number threshold is obtained, which is recorded as the optimized fault type. The initial feature parameter set is updated according to the optimized fault type to obtain the fault feature range and feature weight corresponding to each feature under each preset fault type in the next cycle of the current cycle, and combined with the device real-time status feature vector of the target device in the next cycle of the current cycle, the final fault type of the target device in the next cycle of the current cycle is predicted.

[0072] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A substation fault prediction method based on multi-level device collaboration, characterized in that The method includes: During the operation of the substation, multi-source real-time monitoring data of the target device in the current cycle is collected by multi-level devices, and feature extraction is performed on the multi-source real-time monitoring data to obtain a device real-time status feature vector composed of at least two feature data; The fault feature range and feature weight corresponding to each feature in the current cycle under each preset fault type are respectively obtained. Using the fault feature range corresponding to each feature under each preset fault type, each feature data in the device real-time status feature vector is matched. If at least two preset fault types are matched, the matched preset fault types are recorded as target fault types; For any target fault type, according to the relationship between each feature data in the device real-time status feature vector and the fault feature range corresponding to each feature under the any target fault type, the feature compliance of each feature under the any target fault type is respectively obtained, and combining the feature compliance and feature weight of each feature under the any target fault type, the fault matching confidence of the any target fault type is obtained; The fault matching confidence of each target fault type is obtained, and according to the fault matching confidence of each target fault type, the final fault type of the target device in the current cycle is determined.

2. The substation fault prediction method based on multi-level device collaboration according to claim 1, wherein The performing feature extraction on the multi-source real-time monitoring data to obtain a device real-time status feature vector composed of at least two feature data includes: The multi-source real-time monitoring data includes an infrared thermal imaging image sequence, a vibration signal, an acoustic emission signal, and a hydrogen concentration sequence in an oil valve; The gradient value of each pixel point in each image in the infrared thermal imaging image sequence is respectively calculated to obtain the absolute value of the maximum gradient value, and the absolute value of the maximum gradient value is normalized to obtain the maximum temperature gradient value; The vibration signal is subjected to Fourier transform to obtain a spectrogram, the entropy value of the spectrogram is obtained, and the entropy value is normalized to obtain the vibration spectrum entropy; The energy value corresponding to the acoustic emission signal is obtained, and the energy value is normalized to obtain the acoustic emission energy; The hydrogen concentration sequence is subjected to moving average filtering to obtain a filtered sequence, the average value of the filtered sequence is calculated, and the average value is normalized to obtain the average hydrogen concentration; The maximum temperature gradient value, the vibration spectrum entropy, the acoustic emission energy, and the average hydrogen concentration are used as feature data to form a device real-time status feature vector.

3. The substation fault prediction method based on multi-level device collaboration according to claim 1, wherein The respectively obtaining the fault feature range and feature weight corresponding to each feature in the current cycle under each preset fault type includes: A digital twin model of the target device is constructed, and multiple fault simulations are performed on the digital twin model to obtain a set of device simulation status feature vectors under each preset fault type. According to the set of device simulation status feature vectors under each preset fault type, an initial feature parameter set composed of the initial fault feature range and initial feature weight corresponding to each feature under each preset fault type is obtained; According to the set of initial characteristic parameters and the device state characteristic vectors of the target device in each period, respectively obtain the final fault type of the target device in each period and the corresponding fault matching confidence degree, cumulatively record the number of times that each preset fault type is the final fault type, and obtain at least one preset fault type corresponding to the cumulative number meeting the preset number threshold, denoted as the optimized fault type; For any optimized fault type, obtain all the device state characteristic vectors and all the fault matching confidence degrees corresponding to the any optimized fault type, and update the initial fault characteristic range and the initial characteristic weight corresponding to each feature under the any optimized fault type according to all the device state characteristic vectors and all the fault matching confidence degrees corresponding to the any optimized fault type, so as to obtain the new fault characteristic range and the new characteristic weight corresponding to each feature under the any optimized fault type; Use the new fault characteristic range and the new characteristic weight corresponding to each feature under each optimized fault type to update the set of initial characteristic parameters, obtain a new set of characteristic parameters, use the new set of characteristic parameters as the set of initial characteristic parameters, continue to obtain the final fault type of the target device in each period, and repeat the update method of the set of initial characteristic parameters until the final fault type of the target device in the previous period of the current period is obtained, and obtain the fault characteristic range and the characteristic weight corresponding to each feature under each preset fault type in the current period according to the final fault type of the target device in the previous period of the current period.

4. The substation fault prediction method based on multi-level device collaboration according to claim 3, characterized in that The obtaining of the set of initial characteristic parameters composed of the initial fault characteristic range and the initial characteristic weight corresponding to each feature under each preset fault type according to the set of device simulation state characteristic vectors under each preset fault type includes: For any feature under any preset fault type, based on the feature simulation data corresponding to the any feature in the set of device simulation state feature vectors under the any preset fault type, obtain the average feature simulation data and the standard deviation of the feature simulation data, and set as the initial fault feature range corresponding to any feature under any preset fault type, where represents the average feature simulation data, represents the standard deviation of the feature simulation data; According to the difference in the characteristic simulation data of any feature under any preset fault type and under other preset fault types, obtain the discrimination ability index of the any feature, obtain the discrimination ability index of each feature under any preset fault type, normalize the discrimination ability indices of all features under any preset fault type to obtain the corresponding normalized values, and use the difference between the constant 1 and each normalized value as the initial characteristic weight corresponding to each feature under any preset fault type respectively; Obtain the initial fault characteristic range and the initial characteristic weight corresponding to each feature under each preset fault type, and form a set of initial characteristic parameters.

5. The substation fault prediction method based on multi-level device collaboration according to claim 4, characterized in that The obtaining of the discrimination ability index of any feature according to the difference in the characteristic simulation data of any feature under any preset fault type and under other preset fault types includes: Take any preset fault type other than the above-mentioned preset fault types as the reference type, sort all the feature simulation data of any feature under any preset fault type in ascending order to obtain the target sequence, and sort all the feature simulation data of any feature under the reference type in ascending order to obtain the reference sequence; calculate the absolute value of the difference between the two elements at each same position between the target sequence and the reference sequence, and obtain the mean value of the absolute value of the difference as the feature separation degree of any feature between any preset fault type and the reference type. According to the feature separation degree of any feature between any preset fault type and each reference type, obtain the minimum feature separation degree as the discrimination ability index of any feature.

6. The substation fault prediction method based on multi-level device collaboration according to claim 4, characterized in that, The method of updating the initial fault feature range and the initial feature weight corresponding to each feature under any optimized fault type according to all the device state feature vectors and all the fault matching confidence levels corresponding to any optimized fault type to obtain the new fault feature range and the new feature weight corresponding to each feature under any optimized fault type includes: According to all the device state feature vectors corresponding to any optimized fault type, obtain the mean value of the feature data of each feature, and calculate the mean value of the fault matching confidence levels according to all the fault matching confidence levels corresponding to any optimized fault type. For any feature under any optimized fault type, obtain the middle value in the initial fault feature range corresponding to any feature, calculate the absolute value of the difference between the mean value of the feature data of any feature and the middle value, and take the ratio of the absolute value of the difference to the middle value as the weight gradient of any feature. Take the product of the weight gradient, the mean value of the fault matching confidence levels, and the preset weight learning rate as the weight adjustment value, and take the sum of the initial feature weight corresponding to any feature and the weight adjustment value as the new feature weight corresponding to any feature.

7. The substation fault prediction method based on multi-level device collaboration according to claim 6, characterized in that, The method of updating the initial fault feature range and the initial feature weight corresponding to each feature under any optimized fault type according to all the device state feature vectors corresponding to any optimized fault type to obtain the new fault feature range and the new feature weight corresponding to each feature under any optimized fault type further includes: Obtain the upper limit value and the lower limit value in the initial fault feature range corresponding to any feature, obtain the first difference between the mean value of the feature data of any feature and the lower limit value, take the product of the preset learning rate and the first difference as the lower limit adjustment value, and take the difference between the lower limit value and the lower limit adjustment value as the new lower limit value; Obtain the second difference between the mean value of the feature data of any feature and the upper limit value, take the product of the preset learning rate and the second difference as the upper limit adjustment value, and take the sum of the upper limit value and the upper limit adjustment value as the new upper limit value; Obtain the new fault feature range corresponding to any feature according to the new upper limit value and the new lower limit value.

8. The substation fault prediction method based on multi-level device collaboration according to claim 3, characterized in that Updating the initial feature parameter set using the new fault feature ranges and new feature weights corresponding to each feature under each of the optimized fault types to obtain a new feature parameter set, including: Replacing the initial fault feature ranges and initial feature weights corresponding to each feature under the corresponding preset fault type in the initial feature parameter set with the new fault feature ranges and new feature weights corresponding to each feature under each of the optimized fault types to obtain a new feature parameter set.

9. The substation fault prediction method based on multi-level device collaboration according to claim 1, characterized in that According to the relationship between each feature data in the real-time state feature vector of the device and the fault feature range corresponding to each feature under any one of the target fault types, respectively obtaining the feature compliance of each feature under any one of the target fault types, including: For any one feature under any one of the target fault types, calculating the absolute value of the difference between the feature data corresponding to the any one feature in the real-time state feature vector of the device and the intermediate value within the fault feature range corresponding to the any one feature under any one of the target fault types, and denoting it as the deviation degree; Calculating the absolute value of the difference between the upper limit value or the lower limit value of the fault feature range corresponding to the any one feature under any one of the target fault types and the intermediate value, and denoting it as the reference feature value; calculating the ratio of the deviation degree to the reference feature value, and taking the difference between the constant 1 and the ratio as the feature compliance of the any one feature.

10. The substation fault prediction method based on multi-level device collaboration according to claim 1, wherein Combining the feature compliance and feature weights of each feature under any one of the target fault types to obtain the fault matching confidence of any one of the target fault types, including: According to the feature weights of each feature under any one of the target fault types, performing a weighted sum of the feature compliance of all features under any one of the target fault types to obtain the fault matching confidence of any one of the target fault types.

11. The substation fault prediction method based on multi-level device collaboration according to claim 1, wherein, According to the fault matching confidence of each target fault type, determining the final fault type of the target device in the current cycle, including: According to the fault matching confidence of each target fault type, taking the target fault type corresponding to the maximum fault matching confidence as the final fault type of the target device in the current cycle.

12. The substation fault prediction method based on multi-level device collaboration according to claim 1, wherein After matching each feature data in the real-time state feature vector of the device using the fault feature range corresponding to each feature under each preset fault type, further including: If a preset fault type is matched, taking the matched preset fault type as the final fault type of the target device in the current cycle, and obtaining the fault matching confidence of the matched preset fault type.

Images