Multi-dimensional data fusion and intelligent analysis system for smart park

Abstract

The invention relates to the technical field of data analysis, in particular to a multi-dimensional data fusion and intelligent analysis system for a smart park, which is characterized in that after various types of data are artificially divided into time sequence features, spatial features and attribute features, a controller respectively performs similarity calculation on the three features and then performs fusion; the time sequence features and the spatial features can cover most of data, the attribute features are used for capturing data which are not suitable for time and space distribution, and regular paths are introduced into the time sequence, so that unified path planning can be utilized to avoid data reference errors caused by path asynchronization and further avoid fusion failures; fusion is carried out by fully considering multiple features such as time, space and attributes of the data, information loss or wrong fusion caused by data isomerism is avoided, so that the accuracy of data fusion is improved, the capture capability of the data fusion method on complex data relationships is higher, and the data fusion efficiency is improved. And the fused data can more truly reflect the actual operation state of the park.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a smart park multi-dimensional data fusion and intelligent analysis system. Background Art

[0002] The construction of smart parks involves many fields and complex systems, generating massive amounts of multi-source heterogeneous data, such as real-time sensor data generated by IoT devices, image data from video surveillance systems, and structured data in park management business systems. In order to achieve efficient management, intelligent decision-making, and optimized operations in the park, it is necessary to build a system that can effectively integrate these multi-dimensional data and perform deep intelligent analysis. This technical solution aims to propose a set of innovative solutions to overcome the limitations of existing technologies in data fusion and analysis, and meet the high-precision, high-real-time, and intelligent requirements of smart parks for data processing.

[0003] There are significant differences in data formats, data types, semantic expressions, etc. of different data sources. For example, sensor data may be a continuous stream of values, which is highly real-time but has relatively simple semantics; while business data is usually in a structured tabular form, containing rich semantic information, but with a low update frequency. How to convert these heterogeneous data into a unified representation for effective fusion and analysis is the primary problem to be solved. Therefore, this application proposes a multi-dimensional data fusion and intelligent analysis system for a smart park that is capable of integrating and analyzing various types of data generated within the park. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a smart park multi-dimensional data fusion and intelligent analysis system that can integrate and analyze various types of data generated in the park.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A multi-dimensional data fusion and intelligent analysis system for a smart park includes a controller and a memory. The memory stores various types of data. After the various types of data are artificially divided into time series features, spatial features, and attribute features, the controller calculates the similarity of the three types of features and then fuses them, including: For the similarity calculation of time series features, dynamic time warping is used algorithm: in, They are the time series feature vectors of two data sources, and min means finding the minimum value is the regular path of the time series, Indicates that the path is regular Next, time series Corresponding to the time series The element value of the kth element in ; Representing time series The kth element value in ; the minimum cumulative distance obtained by dynamic time warping is used as the similarity of the time series ; For the similarity calculation of spatial features, a similarity measurement method based on graph structure is adopted: first, a graph model of spatial features is constructed and , by calculating the Laplacian matrix eigenvalues ​​and eigenvectors of the graph and comparing the similarity of the eigenvectors, we can get the spatial feature similarity ; For the similarity calculation of attribute features, we use the information entropy-based algorithm: =1- in, are the time attribute column feature vectors of the two data sources, and r is the dimension of the attribute feature. For example, Including the power of the equipment, the brand of the equipment, the installation date, Similarly; r is the dimension of the attribute feature, then r=3 (power, brand, installation date), and They are The kth element in ; for example, = , = ,So =200W, =180W, where k ranges from 1 to r, and is used to traverse each element in the attribute feature vector. Information entropy Indicates the similarity of features, with a value between 0 and 1. The closer to 1, the higher the similarity. The global similarity matrix M is obtained by combining the similarities of each modality. : M= = + + in, , , is the weight coefficient of each modality similarity, and + + =1; , , The actual value can be inferred through a computational model or can be set manually.

[0006] Calculate the credibility weight of data sources based on the global similarity matrix : = Where n is the total number of data sources, Indicates the sum of the global similarities between the i-th data source and all other data sources; It means to perform double summation of the global similarity between all data sources; Finally, deep fusion is performed to splice the multimodal features of each data source to obtain a fused feature vector = , the fused data F is: F= The controller performs real-time output based on the fused data F.

[0007] Preferably, there are three or more controllers; Each controller calculates the fused data F separately. If the output data F of each controller is the same value, the data F is used; if there are different data F, the data F with the highest repeatability is selected; if the output data F of each controller is different, a warning is issued and the average value is taken.

[0008] Preferably, a generative adversarial network model is constructed, wherein the adversarial network model includes a generator and a discriminator; The task of the generator is to learn the distribution of normal data and generate data samples similar to normal data; the discriminator is responsible for distinguishing between real data and data generated by the generator; During training, the generator tries to minimize the discriminator's accuracy, while the discriminator tries to maximize its accuracy. The fused data F is input into the trained anti-network model, and the discriminator outputs the probability that the data is abnormal. When the probability exceeds the preset value, the data is judged to be abnormal.

[0009] Preferably, the memory stores various types of data for preliminary data cleaning and adjustment of a predetermined format.

[0010] Preferably, the predetermined format includes a date format and a time format, the date format is: year-month-day; the time format is: hour-minute-second; after the adjustment is completed, the formats generated by the park equipment are uniformly adjusted to the same date format and time format.

[0011] Preferably, if the time counting methods used by the park equipment are inconsistent, the data is converted into the same counting scheme and the same numerical representation.

[0012] Preferably, if various types of data contain text information, the case is unified and all equipment names and area names in the park are converted into uniform uppercase or lowercase forms; special characters in the text information are cleaned or converted, and non-standard line breaks, tabs, etc. are uniformly processed into standard character formats to ensure the integrity and standardization of the text data.

[0013] Preferably, if there are missing values ​​in each type of data, the missing part is filled, and the filling method includes using the mean, median or mode to fill; if the missing value does not affect the use of the data, a placeholder is selected for filling.

[0014] Preferably, the predetermined format includes a spatial format, and the spatial format includes: converting different coordinate systems used by various types of data into a WGS84 coordinate system.

[0015] Preferably, the predetermined format includes an attribute format and a semantic format. The attribute format includes: uniformly converting different data expressions used by various types of data into floating point expressions; the semantic format includes: standardizing attribute words with the same meaning but different expressions.

[0016] The beneficial effects of the present invention are as follows: by artificially dividing various types of data into time series features, space features and attribute features, the time series features and space features can cover most of the data, while the attribute features are used to capture data that are not suitable for time and space distribution, and by introducing regular paths in the time series, unified path planning can be used to avoid data reference errors caused by path asynchrony and subsequent fusion failure; by fully considering the time, space and attributes of the data for fusion, when processing the equipment operation data and environmental monitoring data in the park, the algorithm can more accurately integrate the information on the changes of equipment status over time, the spatial distribution of equipment and the equipment's own attributes (such as model, power, working status, etc.) in different data sources, avoiding information loss or erroneous fusion caused by data heterogeneity, thereby improving the accuracy of data fusion, and its ability to capture complex data relationships is stronger, so that the fused data can more truly reflect the actual operation status of the park. DETAILED DESCRIPTION

[0017] In order to explain the technical content, achieved objectives and effects of the present invention in detail, the following is an explanation in conjunction with the implementation modes.

[0018] A multi-dimensional data fusion and intelligent analysis system for a smart park includes a controller and a memory. The memory stores various types of data. After the various types of data are artificially divided into time series features, spatial features, and attribute features, the controller calculates the similarity of the three types of features and then fuses them, including: For the similarity calculation of time series features, dynamic time warping is used algorithm: in, They are the time series feature vectors of two data sources, and min means finding the minimum value is the regular path of the time series, Indicates that the path is regular Next, time series Corresponding to the time series The element value of the kth element in ; Representing time series The kth element value in ; the minimum cumulative distance obtained by dynamic time warping is used as the similarity of the time series ; For the similarity calculation of spatial features, a similarity measurement method based on graph structure is adopted: first, a graph model of spatial features is constructed and , by calculating the Laplacian matrix eigenvalues ​​and eigenvectors of the graph and comparing the similarity of the eigenvectors, we can get the spatial feature similarity ; For the similarity calculation of attribute features, we use the information entropy-based algorithm: =1- in, are the time attribute column feature vectors of the two data sources, r is the dimension of the attribute feature, and They are The kth element in ; The global similarity matrix M is obtained by combining the similarities of each modality. : M= = + + in, , , is the weight coefficient of each modality similarity, and + + =1; Calculate the credibility weight of data sources based on the global similarity matrix : = Where n is the total number of data sources, Indicates the sum of the global similarities between the i-th data source and all other data sources; It means to perform double summation of the global similarity between all data sources; Finally, deep fusion is performed to splice the multimodal features of each data source to obtain a fused feature vector = , the fused data F is: F= The controller performs real-time output based on the fused data F.

[0019] From the above description, it can be seen that by artificially dividing various types of data into time series features, spatial features and attribute features, time series features and spatial features can cover most of the data, while attribute features are used to capture data that are not suitable for time and space distribution. In addition, by introducing regular paths in the time series, unified path planning can be used to avoid data reference errors caused by path asynchrony and subsequent fusion failures. By fully considering the time, space and attributes of the data for fusion, when processing the equipment operation data and environmental monitoring data in the park, the algorithm can more accurately integrate the information about the changes in equipment status over time, the spatial distribution of equipment, and the equipment's own attributes (such as model, power, working status, etc.) from different data sources, avoiding information loss or erroneous fusion caused by data heterogeneity, thereby improving the accuracy of data fusion. It has a stronger ability to capture complex data relationships, so that the fused data can more truly reflect the actual operation status of the park.

[0020] Furthermore, there are three or more controllers; Each controller calculates the fused data F separately. If the output data F of each controller is the same value, the data F is used; if there are different data F, the data F with the highest repeatability is selected; if the output data F of each controller is different, a warning is issued and the average value is taken.

[0021] From the above description, it can be seen that since the data types, scopes, and objects involved in the data fusion process vary greatly, it is easy to make mistakes during the calling process. Therefore, it is necessary to call the calculation independently to prevent errors. If the values ​​of the parallel calculations are different, it proves that there is a problem in the data calling process, and a warning will occur. If there is no error in the calculation, the average value is taken first and used, and then the problem of different outputs is found later.

[0022] Furthermore, a generative adversarial network model is constructed, which includes a generator and a discriminator; The task of the generator is to learn the distribution of normal data and generate data samples similar to normal data; the discriminator is responsible for distinguishing between real data and data generated by the generator; During training, the generator tries to minimize the discriminator's accuracy, while the discriminator tries to maximize its accuracy. The fused data F is input into the trained anti-network model, and the discriminator outputs the probability that the data is abnormal. When the probability exceeds the preset value, the data is judged to be abnormal.

[0023] From the above description, it can be seen that by detecting abnormal situations in the park, timely alarms can be issued.

[0024] Furthermore, the memory stores various types of data for preliminary data cleaning and adjustment of predetermined formats.

[0025] From the above description, it can be seen that by adjusting the data to a predetermined format, it is possible to facilitate reading and subsequent extraction and fusion.

[0026] Furthermore, the predetermined format includes a date format and a time format, the date format is: year-month-day; the time format is: hour-minute-second; after the adjustment is completed, the formats generated by the park equipment are uniformly adjusted to the same date format and time format.

[0027] Furthermore, if the time counting methods used by the equipment in the park are inconsistent, the data will be converted into the same counting scheme and the same numerical representation.

[0028] Furthermore, if various types of data contain text information, the case is unified and all equipment names and area names in the park are converted to unified uppercase or lowercase forms; special characters in the text information are cleaned or converted, and non-standard line breaks, tabs, etc. are uniformly processed into standard character formats to ensure the integrity and standardization of the text data.

[0029] Furthermore, if there are missing values ​​in each type of data, the missing part will be filled, and the filling method includes using the mean, median or mode to fill; if the missing value does not affect the use of the data, a placeholder is selected to fill it.

[0030] Furthermore, the predetermined format includes a spatial format, and the spatial format includes: converting different coordinate systems used by various types of data into a WGS84 coordinate system.

[0031] Furthermore, the predetermined format includes an attribute format and a semantic format. The attribute format includes: uniformly converting different data expressions used by various types of data into floating point expressions; the semantic format includes: standardizing attribute words with the same meaning but different expressions.

[0032] Example A multi-dimensional data fusion and intelligent analysis system for a smart park includes a controller and a memory. The memory stores various types of data, and the various types of data include data D from n different data sources, where D= For each data source, extract its multimodal features and artificially divide them into time series features = , spatial characteristics = and attribute characteristics = ( , , The k contained in is the dimension of each modal feature, which is independent and does not interfere with each other).

[0033] S1. Standardize each modal feature so that its mean is 0 and its variance is 1.

[0034] S2, the controller calculates the similarity of these three features and then fuses them, including: For the similarity calculation of time series features, dynamic time warping is used algorithm: in, They are the time series feature vectors of two data sources, and min means finding the minimum value is the regular path of the time series, which is a The mapping of Indicates that the path is regular Next, time series Corresponding to the time series The element value of the kth element in ; Representing time series The value of the kth element in ; for example, There are m elements, There are n elements, then can be represented as a sequence: = ,in, )express Zhongyu The index of the element where the kth element is aligned; the minimum cumulative distance obtained by dynamic time warping can find the regularization path that minimizes the sum of squared differences among all possible alignments. , get the similarity of two time series ; The purpose of this formula is to more accurately capture the similarities between data and provide a more reliable basis for subsequent data fusion and analysis; Assume further that there are two time series = and = , the warped path found by the dynamic time warping algorithm , = This is just a simple example for easy understanding. The actual calculation will be more complicated. It means Corresponding to The element value of the second element (i.e. 3) in is 2. This matching method can handle the offset and scaling of two time series on the time axis and measure their similarity more accurately.

[0035] For the similarity calculation of spatial features, a similarity measurement method based on graph structure is adopted: first, a graph model of spatial features is constructed and , by calculating the Laplacian matrix eigenvalues ​​and eigenvectors of the graph and comparing the similarity of the eigenvectors, we can get the spatial feature similarity ; The Plath matrix feature is a relatively common algorithm in the industry, so I will not explain it in detail here; For the similarity calculation of attribute features, we use the information entropy-based algorithm: =1- in, are the time attribute column feature vectors of the two data sources, r is the dimension of the attribute feature, and They are The kth element in ; for example, Including the power of the equipment, the brand of the equipment, the installation date, Similarly; r is the dimension of the attribute feature, then r=3 (power, brand, installation date), and They are The kth element in ; for example, = , = ,So =200W, =180W, where k ranges from 1 to r, and is used to traverse each element in the attribute feature vector. Information entropy Indicates the similarity of features, with a value between 0 and 1. The closer to 1, the higher the similarity. The whole formula is based on the principle of information entropy. By calculating the information entropy correlation value of each element in the two attribute feature vectors, summing and normalizing them, the similarity of the two data sources in terms of attribute characteristics is obtained. ; Information entropy is an indicator of information uncertainty. Here, logarithmic operations are used to quantify the uncertainty of the information carried by each element in the attribute feature. , That is, to perform logarithmic operation on the attribute value so that it can participate in the calculation of information entropy to determine its contribution in measuring the similarity of attribute features.

[0036] The global similarity matrix M is obtained by combining the similarities of each modality. : M= = + + in, , , is the weight coefficient of each modality similarity, and + + =1; Calculate the credibility weight of data sources based on the global similarity matrix : = Among them, n is the total number of data sources; n determines the scope of the data sources involved in the calculation. For example, if there are 5 different types of data sources in the park (such as environmental monitoring, equipment operation, personnel flow, energy consumption and security data, etc.), then n=5; which means that when calculating the credibility weight of each data source, the similarities between the 5 data sources should be comprehensively considered.

[0037] It means to sum the global similarity between the ith data source and all other data sources; the purpose of this step is to measure the overall similarity between the ith data source and the entire data source set. If this sum is large, it means that the data source has a high similarity with other data sources as a whole, which in turn implies that its credibility in data fusion may be high, because data sources with high similarity with other data sources are often more consistent with the overall distribution and laws of the data, and their data quality and reliability are relatively more guaranteed.

[0038] It represents the double summation of the global similarity between all data sources; it calculates the total similarity between all data sources in the entire data source set, and serves as a normalized denominator to ensure the calculated credibility weight It can be reasonably distributed within the range of 0 to 1, so that the sum of the credibility weights of all data sources is 1. In the subsequent data fusion process, the data of each data source can be weightedly fused (that is, the final deep fusion) according to the relative credibility of each data source, so as to obtain a more accurate and reliable fusion result.

[0039] S3: Perform deep fusion to combine the multimodal features of each data source to obtain a fused feature vector = , the fused data F is: F= The controller performs real-time output based on the fused data F.

[0040] There are more than three controllers; Each controller calculates the fused data F separately. If the output data F of each controller is the same value, the data F is used; if there are different data F, the data F with the highest repeatability is selected; if the output data F of each controller is different, a warning is issued and the average value is taken.

[0041] S4. Build a generative adversarial network model (GAN), which includes a generator and a discriminator; The task of the generator is to learn the distribution of normal data and generate data samples similar to normal data; the discriminator is responsible for distinguishing between real data and data generated by the generator; During training, the generator tries to minimize the discriminator's accuracy, while the discriminator tries to maximize its accuracy. The fused data F is input into the trained anti-network model, and the discriminator outputs the probability that the data is abnormal. When the probability exceeds the preset value, the data is judged to be abnormal.

[0042] The most difficult part of park data integration is actually the data. The data generated by each device is different, and each company's products have a specific recording method. Therefore, the data needs to be processed in a unified manner, and the storage device stores various types of data for preliminary data cleaning and adjustment of the predetermined format.

[0043] If various data types contain text information, the case should be unified, and all equipment names and area names in the park should be converted to unified uppercase or lowercase. For example, the brand and model of the equipment should have a unified character encoding format (such as GTX-4080 and gtx-4080, unified as GTX-4080). Special characters in text information should be cleaned or converted, and non-standard line breaks, tabs, etc. should be unified into standard character formats to ensure the integrity and standardization of text data.

[0044] If there are missing values ​​in each type of data, the missing part will be filled using the mean, median or mode. If the missing value does not affect the use of the data, a placeholder (such as "NA" or "None") will be selected to fill it.

[0045] The predefined formats include: The date format is: year-month-day; The time format is: hours-minutes-seconds; after the adjustment is completed, the formats generated by the park equipment are uniformly adjusted to the same date format and time format. If the time counting method used by the park equipment is inconsistent, the data will be converted to the same counting scheme and the same numerical representation.

[0046] The spatial format is: the different coordinate systems used by various types of data (such as geodetic coordinate system, plane rectangular coordinate system, etc.) are uniformly converted into the WGS84 coordinate system. This can be achieved through the coordinate conversion algorithm to ensure that all information involving spatial locations (such as equipment location, personnel location, building location, etc.) is expressed in the same coordinate framework, which is convenient for subsequent spatial analysis and calculation.

[0047] The attribute format is: convert the different data expressions used by various types of data into floating-point expressions; for example, if the power attribute of a device is expressed in integer form in one data source and in floating-point form in another data source, it is necessary to unify them into one data type, such as floating-point numbers, to enable accurate numerical calculation and analysis.

[0048] The semantic format is: standardize the attribute words with the same meaning but different expressions. For example, different data sources may use different words such as "running", "working", and "on" to describe the operating status of a device. These expressions need to be unified into a standard term, such as "operating status", and a clear value range is defined to ensure the semantic consistency of the data.

[0049] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention specification, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A smart park multi-dimensional data fusion and intelligent analysis system, characterized in that: The system includes a controller and a memory, wherein the memory stores various types of data. After artificially dividing the various types of data into time series features, spatial features, and attribute features, the controller calculates the similarity of the three types of features and then fuses them, including: For the similarity calculation of time series features, dynamic time warping is used algorithm: in, They are the time series feature vectors of two data sources, and min means finding the minimum value is the regular path of the time series, Indicates that the path is regular Next, time series Corresponding to the time series The element value of the kth element in ; Representing time series The kth element value in ; the minimum cumulative distance obtained by dynamic time warping is used as the similarity of the time series ; For the similarity calculation of spatial features, a similarity measurement method based on graph structure is adopted: first, a graph model of spatial features is constructed and , by calculating the Laplacian matrix eigenvalues ​​and eigenvectors of the graph and comparing the similarity of the eigenvectors, we can get the spatial feature similarity ; For the similarity calculation of attribute features, we use the information entropy-based algorithm: =1- in, are the time attribute column feature vectors of the two data sources, r is the dimension of the attribute feature, and They are The kth element in ; The global similarity matrix M is obtained by combining the similarities of each modality. : M= = + + in, , , is the weight coefficient of each modality similarity, and + + =1; Calculate the credibility weight of data sources based on the global similarity matrix : = Where n is the total number of data sources, Indicates the sum of the global similarities between the i-th data source and all other data sources; It means to perform double summation of the global similarities between all data sources; Finally, deep fusion is performed to splice the multimodal features of each data source to obtain a fused feature vector = , the fused data F is: F= The controller performs real-time output based on the fused data F.

2. The smart park multi-dimensional data fusion and intelligent analysis system according to claim 1 is characterized in that: There are more than three controllers; Each controller calculates the fused data F separately. If the output data F of each controller is the same value, the data F is used; if there are different data F, the data F with the highest repeatability is selected; if the output data F of each controller is different, a warning is issued and the average value is taken.

3. The smart park multi-dimensional data fusion and intelligent analysis system according to claim 1 is characterized in that: Build a generative adversarial network model, which includes a generator and a discriminator; The task of the generator is to learn the distribution of normal data and generate data samples similar to normal data; the discriminator is responsible for distinguishing between real data and data generated by the generator; During training, the generator tries to minimize the discriminator's accuracy, while the discriminator tries to maximize its accuracy. The fused data F is input into the trained anti-network model, and the discriminator outputs the probability that the data is abnormal. When the probability exceeds the preset value, the data is judged to be abnormal.

4. The smart park multi-dimensional data fusion and intelligent analysis system according to claim 1 is characterized in that: The memory stores various types of data for preliminary data cleaning and adjustment of the predetermined format.

5. The smart park multi-dimensional data fusion and intelligent analysis system according to claim 4 is characterized in that: The preset format includes date format and time format. The date format is: year-month-day; the time format is: hour-minute-second. After the adjustment is completed, the formats generated by the park equipment are uniformly adjusted to the same date format and time format.

6. The smart park multi-dimensional data fusion and intelligent analysis system according to claim 5 is characterized in that: If the time counting methods used by the park equipment are inconsistent, the data should be converted to the same counting scheme and the same numerical representation.

7. The smart park multi-dimensional data fusion and intelligent analysis system according to claim 4 is characterized in that: If various types of data contain text information, the case will be unified and all equipment names and area names in the park will be converted to unified uppercase or lowercase. Special characters in the text information will be cleaned or converted, and non-standard line breaks, tabs, etc. will be uniformly processed into standard character formats to ensure the integrity and standardization of the text data.

8. The smart park multi-dimensional data fusion and intelligent analysis system according to claim 4 is characterized in that: If there are missing values ​​in various types of data, the missing parts will be filled using the mean, median or mode. If the missing values ​​do not affect the use of the data, placeholders will be selected to fill them.

9. The smart park multi-dimensional data fusion and intelligent analysis system according to claim 4 is characterized in that: The predetermined format includes a spatial format, and the spatial format includes: converting different coordinate systems used by various types of data into a WGS84 coordinate system.

10. The smart park multi-dimensional data fusion and intelligent analysis system according to claim 4, characterized in that: The predetermined formats include attribute formats and semantic formats. The attribute formats include: uniformly converting different data expressions used by various types of data into floating-point expressions; the semantic formats include: standardizing attribute terms with the same meaning but different expressions.

Images