Power transformation main equipment analysis method based on map enhancement

Abstract

The invention discloses a power transformation main equipment analysis method based on map enhancement, and relates to the technical field of power transformation main equipment on-line analysis, and the method comprises the following steps: S1, establishing main equipment data correlation; s2, performing symbiotic high-level feature representation of the time series data; s3, constructing an anomaly detection model; s4, time sequence data anomaly detection; and S5, online updating of the model. According to the method, knowledge graph relevance is utilized, a sparse learning model integrating a shared dictionary and group constraints is provided, the dictionary is continuously updated along with access of time series data to adapt to an adaptive learning technology of external environment changes and the like, robustness of data analysis is achieved, space-time relevance of collected data is fully mined, the shared dictionary is introduced, and the robustness of data analysis is improved. Similar parts in a training data set are coded in the sub-dictionary in a centralized mode, and the problem that a training sample set is insufficient is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of online analysis of substation main equipment, and in particular to a method for analyzing substation main equipment based on graph enhancement. Background Art

[0002] Using sensitive components in electronic devices to perceive external information can reach or exceed the functions of human sensory organs. However, in actual application scenarios, such as sensors in UHV substations, they will be affected by various coupled physical field factors and external environmental interference, and the collected data will show abnormal fluctuations. This data anomaly is currently a common problem faced by all walks of life and will directly affect the performance of subsequent data analysis models and decision algorithms. Therefore, how to effectively implement data analysis has important theoretical significance and practical value. The online monitoring data analysis of power equipment can be formalized as anomaly detection in time series data, where abnormal instances include steep changes, steps, null values, etc. There are many studies on anomaly detection at present. For example, some unbalanced classification problems are often solved by anomaly detection algorithms. The simplest method for anomaly detection is statistical methods, such as the Tukey method. Another simple, intuitive and usually effective anomaly detection method is to use some clustering algorithms (such as Gaussian mixture models and DBSCAN) to solve density estimation tasks. Then, any instance located in a low-density area can be considered an anomaly. Some supervised learning algorithms can also be used for anomaly detection. The two most popular ones are isolation forests and SVM. These algorithms are more suitable for singular value detection, but are also generally applicable to anomaly detection.

[0003] In actual application scenarios, especially in the industrial field, anomaly detection of sensor data not only pursues high accuracy and timeliness, but also takes into account the influence of various external interferences brought about by the changes in sensor data over time. However, there are two problems with the methods proposed by the existing technology. First, they are too simple, resulting in high anomaly omission rate and false alarm rate. Second, the model is too complex to be processed in real time. In addition, most technologies cannot achieve autonomous learning, resulting in the inability to adapt to changes in environmental factors in real time. Summary of the invention

[0004] The purpose of the present invention is to provide a substation main equipment analysis method based on spectrum enhancement to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: A method for analyzing substation main equipment based on graph enhancement, comprising the following steps:

[0006] S1. Establish data correlation of main equipment: establish graph association based on the same voltage level, manufacturer, and model information of substation main equipment, retrieve knowledge graph through SPARQL language, obtain operation status data of various sensors at corresponding time nodes, and establish mapping relationship between sensor fault types that have appeared in history and online monitoring data;

[0007] S2. Symbiotic high-level feature representation of time series data: Sample the time series top-level oil temperature data of the main substation equipment for a long period of time to obtain the high-dimensional feature representation of the short-term time series data, and further enhance the data representation capability to obtain the symbiotic high-level feature representation of the time series top-level oil temperature of the substation equipment;

[0008] S3. Anomaly detection model construction: Establish a time series data anomaly detection model based on sparse cognition, introduce hierarchical group dilution constraints of shared tags and dilution coefficients in model building, learn high-level semantic polynomials of long time series data, and improve the classification accuracy of each category of abnormal sequence data;

[0009] S4. Time series data anomaly detection: Use data high-level semantic polynomials to classify test time series data. is the test time series data, y j is the sampled symbiotic feature representation, and for each sampled symbiotic feature representation y in the sequence j Perform hierarchical group dilution coding and determine the category to which the sampling sequence belongs based on the reconstruction error;

[0010] S5. Online update of the model: An update mechanism is used to resolve interference from external factors and internal equipment factors, and to continuously adapt to changes in scenarios. A conservative update strategy is used, that is, the shared dictionary is updated only when the current collected data is judged to be normal, otherwise the output is abnormal and enters the next time node to continue to establish data correlation with the main device.

[0011] Preferably, in step S1, a graph association is established based on the same voltage level, manufacturer, and same model information of the substation main equipment, and the knowledge graph is retrieved through the SPARQL language to obtain the operating status data of various sensors at the corresponding time nodes, so that a mapping relationship is established between the sensor failure types that have occurred in the history and the online monitoring data, the SELECT function is used to obtain data that meets the monitoring conditions, the DESCRIBE function is used to obtain RDF data related to the monitoring data, and a detailed description of the online monitoring of the equipment associated with the substation main equipment is formed.

[0012] Preferably, in step S2, for a long period of time-series top oil temperature data of the main substation equipment, different step sizes are used for sampling to obtain a series of short-time time series data, and for each short-time time series data, a one-dimensional convolution operation is used to obtain a high-dimensional feature representation of the time series data.

[0013] Preferably, in step S2, the long-term time series data is further randomly sampled to obtain a short-term time series pair to enhance the data representation capability, and a symbiotic high-level feature representation is generated for this local short-term time series pair. At the same time, a symbiotic high-level feature representation of the top-level oil temperature of the time series of other equipment associated with the main substation equipment can also be obtained.

[0014] Preferably, in step S3, the feature set representation of each category is obtained based on the co-occurrence high-level feature representation of the time series data. Where C represents the number of categories, and Represents the feature representation set of the i-th type of data set, represents a co-occurrence feature representation generated by sampling in the i-th long-term series of the i-th data set, N i Indicates the total number of co-occurrence features generated by the long-term series.

[0015] Preferably, in step S3, a time series data anomaly detection model based on sparse cognition is established as follows:

[0016]

[0017] In the formula, represents the reconstruction error of the i-type data on the corresponding dictionary, D0 represents the shared dictionary to encode the similar parts between categories, Represents the hierarchical group dilution coding constraint, so that the non-zero coding coefficients are concentrated in a certain group and satisfy the sparsity. The above two items make the i-th category data encoded by the shared dictionary or the corresponding sub-dictionary. Make each sub-dictionary unrelated and ensure that the dilution coding coefficients under each dictionary are different. C represents the number of categories, D0, Represents the sub-dictionary and dilution coding of class i data.

[0018] Preferably, in step S4, the judgment formula for the category to which the sampling sequence belongs is:

[0019]

[0020] In the formula, Indicates that the current sampling sequence j is a shared marker, Indicates that the shared tags and the category tags corresponding to the sub-dictionary correspond to the topics of the long-term series, y jis the co-occurrence feature representation obtained by sampling, C represents the number of categories, and argmax|| || represents the function used to find the independent variable that maximizes the function value.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. The present invention utilizes the relevance of knowledge graphs, proposes a sparse learning model that integrates shared dictionaries and group constraints, and uses adaptive learning technology that continuously updates dictionaries to adapt to changes in the external environment as time series data is accessed, thereby achieving robustness in data analysis;

[0023] 2. The present invention fully exploits the spatiotemporal correlation of collected data and introduces a shared dictionary so that similar parts in the training data set are concentratedly encoded in this sub-dictionary. In addition, the introduction of the shared dictionary can also solve the problem of insufficient training sample sets, that is, a certain type of data can be represented by the shared part of the entire training set. The anomaly detection model introduces hierarchical group sparse coding constraints so that the sparse coefficients are concentrated in a sub-dictionary, and then the sub-dictionary is optimized by the sparse coefficients. The sub-dictionary includes a shared dictionary, so that the learned sub-dictionary can better characterize the characteristics of the local block. The anomaly detection model introduces shared tags corresponding to the shared dictionary and introduces high-level semantic polynomials for classification, instead of classification again through sparse coding and additional classifiers. On the basis of the anomaly detection model, a model update mechanism is proposed to adapt to the latest collected data and enhance the robustness of anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of the overall method flow provided by an embodiment of the present invention;

[0025] Figure 2 A flow chart of a sensor measurement data anomaly detection technology provided by an embodiment of the present invention;

[0026] Figure 3 A schematic diagram of abnormal data detection provided by an embodiment of the present invention;

[0027] Figure 4 A schematic diagram of abnormal detection of sensor measurement data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] See also Figures 1 to 4The present invention provides a technical solution: a method for analyzing substation main equipment based on graph enhancement, comprising the following steps:

[0030] S1. Establish data correlation of main equipment: establish graph association based on the same voltage level, manufacturer, and model information of substation main equipment, retrieve knowledge graph through SPARQL language, obtain operation status data of various sensors at corresponding time nodes, and establish mapping relationship between sensor fault types that have appeared in history and online monitoring data;

[0031] S2. Symbiotic high-level feature representation of time series data: Sampling the long-term time series top-level oil temperature data of the main substation equipment obtains the high-dimensional feature representation of the short-term time series data, and further enhances the data representation capability to obtain the symbiotic high-level feature representation of the time series top-level oil temperature of the substation equipment;

[0032] S3. Anomaly detection model construction: Establish a time series data anomaly detection model based on sparse cognition, introduce hierarchical group dilution constraints of shared tags and dilution coefficients in model building, learn high-level semantic polynomials of long time series data, and improve the classification accuracy of each category of sequence data anomalies;

[0033] S4. Time series data anomaly detection: Use data high-level semantic polynomials to classify test time series data. Represents the test sequence data, y j is the sampled symbiotic feature representation, and for each sampled symbiotic feature representation y in the sequence j Perform hierarchical group dilution coding and determine the category to which the sampling sequence belongs based on the reconstruction error;

[0034] S5. Online update of the model: An update mechanism is used to resolve interference from external factors and internal equipment factors to continuously adapt to scene changes. The update strategy adopts a conservative strategy, that is, the shared dictionary is updated only when the current collected data is judged to be normal, otherwise the output is abnormal and enters the next time node to continue to establish the data correlation of the main device.

[0035] In step S1, a graph association is established based on the same voltage level, manufacturer, and model information of the main substation equipment, and the knowledge graph is retrieved through the SPARQL language to obtain the operating status data of various sensors at the corresponding time node, so that the sensor fault types that have appeared in the history are mapped with the online monitoring data, and the SELECT function is used to obtain data that meets the monitoring conditions, and the DESCRIBE function is used to obtain RDF data related to the monitoring data, and a detailed description of online monitoring of the equipment associated with the main substation equipment is formed;

[0036] In step S2, for a long period of time series top oil temperature data of the main substation equipment, different step sizes are used for sampling to obtain a series of short-time time series data, and for each short-time time series data, a one-dimensional convolution operation is used to obtain a high-dimensional feature representation of the time series data;

[0037] In step S2, the long-term time series data is further randomly sampled to obtain a short-term time series pair, so as to enhance the data representation capability, and generate a symbiotic high-level feature representation for this local short-term time series pair. At the same time, the symbiotic high-level feature representation of the top-level oil temperature of other equipment associated with the main substation equipment can also be obtained;

[0038] Specifically, the co-occurrence feature representation can adopt the absolute value of concatenation or difference.

[0039] In step S3, the feature set representation of each category is obtained based on the co-occurrence high-level feature representation of the time series data Where C represents the number of categories, and Represents the feature representation set of the i-th type of data set, represents a co-occurrence feature representation generated by sampling in the i-th long-term series of the i-th data set, N i Indicates the total number of co-occurrence features generated by the long-term series;

[0040] In step S3, the time series data anomaly detection model based on sparse cognition is established as follows:

[0041]

[0042] In the formula, represents the reconstruction error of the i-type data on the corresponding dictionary, D0 represents the shared dictionary to encode the similar parts between categories, Represents the hierarchical group dilution coding constraint, so that the non-zero coding coefficients are concentrated in a certain group and satisfy the sparsity. The above two items make the i-th category data encoded by the shared dictionary or the corresponding sub-dictionary. Make each sub-dictionary unrelated and ensure that the dilution coding coefficients under each dictionary are different. C represents the number of categories, D0, The sub-dictionary and dilution coding representing the i-type data;

[0043] The optimization of the model can be divided into two parts: hierarchical group sparse coding and dictionary learning. When the dictionary is fixed, it is transformed into a sparse coding problem. When the sparse coefficient is fixed, it is transformed into a dictionary learning problem. The optimization of the dictionary can be achieved by using the gradient descent method.

[0044] In step S4, the judgment formula for the category to which the sampling sequence belongs is:

[0045]

[0046] In the formula, Indicates that the current sampling sequence j is a shared marker, Indicates that the shared tags and the category tags corresponding to the sub-dictionary correspond to the topics of the long-term series, y j is the co-occurrence feature representation obtained by sampling, C represents the number of categories, argmax|| || represents the function used to find the independent variable that makes the function value reach the maximum value, and finally the high-level semantic polynomial of the entire long-term series data can be expressed as:

[0047]

[0048] in, β is the regularization parameter, C is the number of categories, Y is the test sequence data, c j represents the current sampling sequence j mark, Indicates that the current sampling sequence j is a shared marker, Indicates c j and Perform an OR operation to obtain the final shared label c of the current category of the sampling sequence.

[0049] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0050] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing substation main equipment based on graph enhancement, characterized in that: The method comprises the following steps: S1. Establish data correlation of main equipment: establish graph association based on the same voltage level, manufacturer, and model information of substation main equipment, retrieve knowledge graph through SPARQL language, obtain operation status data of various sensors at corresponding time nodes, and establish mapping relationship between sensor fault types that have appeared in history and online monitoring data; S2. Symbiotic high-level feature representation of time series data: Sample the time series top-level oil temperature data of the main substation equipment for a long period of time to obtain the high-dimensional feature representation of the short-term time series data, and further enhance the data representation capability to obtain the symbiotic high-level feature representation of the time series top-level oil temperature of the substation equipment; S3. Anomaly detection model construction: Establish a time series data anomaly detection model based on sparse cognition, introduce hierarchical group dilution constraints of shared tags and dilution coefficients in model building, learn high-level semantic polynomials of long time series data, and improve the classification accuracy of each category of abnormal sequence data; S4. Time series data anomaly detection: Use data high-level semantic polynomials to classify test time series data. is the test time series data, y j is the sampled symbiotic feature representation, and for each sampled symbiotic feature representation y in the sequence j Perform hierarchical group dilution coding and determine the category to which the sampling sequence belongs based on the reconstruction error; S5. Online update of the model: An update mechanism is used to resolve interference from external factors and internal equipment factors, and to continuously adapt to changes in scenarios. A conservative update strategy is used, that is, the shared dictionary is updated only when the current collected data is judged to be normal, otherwise the output is abnormal and enters the next time node to continue to establish data correlation with the main device.

2. The method for analyzing substation main equipment based on graph enhancement according to claim 1 is characterized in that: In the step S1, a graph association is established based on the same voltage level, manufacturer, and same model information of the substation main equipment, and the knowledge graph is retrieved through the SPARQL language to obtain the operating status data of various sensors at the corresponding time node, so that a mapping relationship is established between the sensor failure types that have appeared in the history and the online monitoring data, and the SELECT function is used to obtain data that meets the monitoring conditions, and the DESCRIBE function is used to obtain RDF data related to the monitoring data, and a detailed description of the online monitoring of the equipment associated with the substation main equipment is formed.

3. The method for analyzing substation main equipment based on graph enhancement according to claim 1 is characterized in that: In step S2, for a long period of time series top oil temperature data of the main substation equipment, different step sizes are used for sampling to obtain a series of short-time time series data, and for each short-time time series data, a one-dimensional convolution operation is used to obtain a high-dimensional feature representation of the time series data.

4. The method for analyzing substation main equipment based on graph enhancement according to claim 3 is characterized in that: In the step S2, the long-term time series data is further randomly sampled to obtain a short-term time series pair to enhance the data representation capability, and a symbiotic high-level feature representation is generated for this local short-term time series pair. At the same time, a symbiotic high-level feature representation of the top-level oil temperature of the time series of other equipment associated with the main substation equipment can also be obtained.

5. The method for analyzing substation main equipment based on graph enhancement according to claim 1 is characterized in that: In step S3, the feature set representation of each category is obtained based on the co-occurrence high-level feature representation of the time series data. Where C represents the number of categories, and Represents the feature representation set of the i-th type of data set, represents a co-occurrence feature representation generated by sampling in the i-th long-term series of the i-th data set, N i Indicates the total number of co-occurrence features generated by the long-term series.

6. The method for analyzing substation main equipment based on graph enhancement according to claim 1 is characterized in that: In step S3, the time series data anomaly detection model based on sparse cognition is established as follows: In the formula, represents the reconstruction error of the i-type data on the corresponding dictionary, D0 represents the shared dictionary to encode the similar parts between categories, Represents the hierarchical group dilution coding constraint, so that the non-zero coding coefficients are concentrated in a certain group and satisfy the sparsity. The above two items make the i-th category data encoded by the shared dictionary or the corresponding sub-dictionary. Make each sub-dictionary unrelated and ensure that the dilution coding coefficients under each dictionary are different. C represents the number of categories. Represents the sub-dictionary and dilution coding of class i data.

7. The method for analyzing substation main equipment based on graph enhancement according to claim 1 is characterized in that: In step S4, the judgment formula for the category to which the sampling sequence belongs is: In the formula, Indicates that the current sampling sequence j is a shared marker, Indicates that the shared tags and the category tags corresponding to the sub-dictionary correspond to the topics of the long-term series, y j is the co-occurrence feature representation obtained by sampling, C represents the number of categories, and argmax|| || represents the function used to find the independent variable that maximizes the function value.

Images