Abnormal power utilization identification method and device based on multi-modal neural network model

Abstract

The embodiment of the invention discloses an abnormal power utilization identification method and device based on a multi-mode neural network model. A specific embodiment of the method comprises the steps of performing virtual object construction according to a power supply topological structure and a power utilization portrait corresponding to a meta object in at least one meta object to obtain a virtual object set; for the virtual object, executing the following processing steps: collecting target power consumption data corresponding to the virtual object through a data relay corresponding to the virtual object; performing data preprocessing on the target power utilization data; dynamically determining a fusion channel corresponding to the preprocessed power consumption data according to the data state of the preprocessed power consumption data; and in response to the pre-processed power consumption data sent by the fusion channel and received by the server side, the server side performs abnormal power consumption identification on the pre-processed power consumption data through the multi-modal neural network model. According to the embodiment, effective identification of abnormal power consumption behaviors is realized, and the power consumption safety risk is reduced.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of computer technology, and in particular, to a method and device for abnormal power consumption recognition based on a multimodal neural network model. Background Art

[0002] With the continuous growth of global energy demand and the increasingly severe environmental problems, electric power energy has become one of the key energies for efficient energy utilization. Therefore, the social electricity consumption continues to rise. However, the accompanying problem is the increasing generation of abnormal power consumption behaviors (such as electricity theft behaviors). When exceeding the line load, abnormal power consumption behaviors also increase the power consumption safety risks.

[0003] The above information disclosed in this background art section is only used to enhance the understanding of the background of the inventive concept. Therefore, it may include information that does not form the prior art known to those of ordinary skill in the art. Summary of the Invention

[0004] The content part of the present disclosure is used to briefly introduce the inventive concepts, which will be described in detail in the following detailed implementation part. The content part of the present disclosure is not intended to identify the key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0005] Some embodiments of the present disclosure propose a method and device for abnormal power consumption recognition based on a multimodal neural network model to solve the technical problems mentioned in the above background art section.

[0006] In a first aspect, some embodiments of the present disclosure provide an abnormal power consumption recognition method based on a multimodal neural network model. The method includes: constructing virtual objects according to a power supply topology structure and power consumption portraits corresponding to meta-objects in at least one meta-object, where the meta-object is a power consumption object at the meta-granularity in the above power supply topology structure, and the virtual object is composed of at least two meta-objects having an indirect topological relationship and corresponding approximate power consumption behaviors; for each virtual object in the above virtual object set, perform the following processing steps: collect target power consumption data corresponding to the above virtual object through the data repeater corresponding to the above virtual object, where the above data repeater includes: a data concentrator and a radio frequency transceiver, and the data repeater is used to collect power consumption data of at least two meta-objects corresponding to the above virtual object in a wired manner, and the radio frequency transceiver is used to wirelessly send the data to the server side; perform data preprocessing on the above target power consumption data to obtain preprocessed power consumption data; dynamically determine a fusion channel corresponding to the above preprocessed power consumption data according to the data state of the above preprocessed power consumption data, where the fusion channel includes: a main channel and a secondary channel, and the signal-to-noise ratio corresponding to the main channel is greater than that of the secondary channel; in response to the server side receiving the above preprocessed power consumption data sent through the fusion channel, the server side performs abnormal power consumption recognition on the above preprocessed power consumption data through a multimodal neural network model to generate a power consumption recognition result for the above virtual object.

[0007] Second aspect, some embodiments of the present disclosure provide an abnormal power consumption recognition device based on a multimodal neural network model. The device includes: a virtual object construction unit configured to construct virtual objects based on a power supply topology structure and power consumption portraits corresponding to meta-objects in at least one meta-object, to obtain a set of virtual objects. Here, a meta-object is a power consumption object at the meta-granularity in the above-mentioned power supply topology structure, and a virtual object is composed of at least two meta-objects having an indirect topological relationship and corresponding approximate power consumption behaviors; an execution unit configured to, for each virtual object in the above-mentioned set of virtual objects, perform the following processing steps: collect target power consumption data corresponding to the above-mentioned virtual object through a data repeater corresponding to the above-mentioned virtual object. Here, the data repeater includes a data concentrator and a radio frequency transceiver. The data repeater is used to collect power consumption data of at least two meta-objects corresponding to the above-mentioned virtual object in a wired manner, and the radio frequency transceiver is used to wirelessly send the data to the server side; perform data preprocessing on the above-mentioned target power consumption data to obtain preprocessed power consumption data; dynamically determine a fusion channel corresponding to the above-mentioned preprocessed power consumption data according to the data state of the above-mentioned preprocessed power consumption data. Here, the fusion channel includes a main channel and a secondary channel, and the signal-to-noise ratio corresponding to the main channel is greater than that of the secondary channel; in response to the server side receiving the above-mentioned preprocessed power consumption data sent through the fusion channel, the server side performs abnormal power consumption recognition on the above-mentioned preprocessed power consumption data through a multimodal neural network model to generate a power consumption recognition result for the above-mentioned virtual object.

[0008] Third aspect, some embodiments of the present disclosure provide an electronic device, including: one or more processors; a storage device having stored thereon one or more programs, which when executed by the one or more processors, cause the one or more processors to implement the method described in any implementation manner of the first aspect above.

[0009] Fourth aspect, some embodiments of the present disclosure provide a computer-readable medium having stored thereon a computer program, where the program, when executed by a processor, implements the method described in any implementation manner of the first aspect above.

[0010] The above-mentioned various embodiments of the present disclosure have the following beneficial effects: Through the abnormal power consumption recognition method based on the multi-modal neural network model in some embodiments of the present disclosure, the effective recognition of abnormal power consumption behaviors is achieved, ensuring power consumption safety. Specifically, with the popularization of power consumption, the number of power-consuming devices connected to the power grid has increased sharply, and thus a large number of abnormal power consumption behaviors (such as electricity theft behaviors) have emerged. When the line load is exceeded, abnormal power consumption behaviors also increase the power consumption safety risk. Therefore, in some embodiments of the present disclosure, the abnormal power consumption recognition method based on the multi-modal neural network model first constructs virtual objects according to the power supply topology structure and the power consumption portraits corresponding to the meta-objects in at least one meta-object, obtaining a virtual object set, where the meta-object is the power-consuming object at the meta-granularity in the above-mentioned power supply topology structure, and the virtual object is composed of at least two meta-objects with an indirect topological relationship and corresponding approximate power consumption behaviors. In practice, conventional power consumption behavior analysis methods often adopt methods such as data analysis for a single power-consuming object or data analysis for a group of power-consuming objects with the same geographical location. For the former, due to the large number of power-consuming objects and the fact that power-consuming objects are often accompanied by approximate continuous power consumption behaviors, the amount of data to be analyzed is extremely large. For the latter, although a group of power-consuming objects with the same geographical location are aggregated and then data analysis is performed, the power consumption habits among a group of power-consuming objects with the same geographical location may be different, thus resulting in the inability to guarantee the recognition accuracy of abnormal power consumption. Therefore, by combining the power supply topology structure and the power consumption portrait, the present disclosure combines the meta-objects (power-consuming objects) with an indirect topological relationship and performing power consumption behaviors for analysis, which can reduce the amount of data processing while ensuring the recognition accuracy. Secondly, for each virtual object in the above-mentioned virtual object set, the following processing steps are executed: First step, through the data repeater corresponding to the above-mentioned virtual object, collect the target power consumption data corresponding to the above-mentioned virtual object, where the above-mentioned data repeater includes: a data concentrator and a radio frequency transceiver, and the data repeater is used to collect the power consumption data of at least two meta-objects corresponding to the above-mentioned virtual object in a wired manner, and the radio frequency transceiver is used to send the data to the server side wirelessly. In practice, since the power-consuming objects at the edge end have different power-consuming device layout methods, uniformly replacing them with wireless devices (such as radio meters) has a very high replacement cost. Therefore, for at least two meta-objects corresponding to the virtual object, collecting power consumption data in a wired manner and then sending it to the server side through the radio frequency transceiver can effectively reduce the device layout cost while ensuring the stability of power consumption data transmission. Second step, perform data preprocessing on the above-mentioned target power consumption data to obtain preprocessed power consumption data. In practice, if the target power consumption data is directly used, due to the real-time generation of power consumption data, there is still a problem of a large amount of data processing. Therefore, through data preprocessing, the subsequent amount of data processing is reduced, alleviating the data processing pressure on the server.In the third step, according to the data status of the preprocessed power consumption data, a fusion channel corresponding to the preprocessed power consumption data is dynamically determined, where the fusion channel includes: a main channel and a secondary channel, and the signal-to-noise ratio corresponding to the main channel is greater than that of the secondary channel. In practice, the selection of channels is involved in the wireless data transmission process. A poor channel will not only affect the transmission efficiency but also may cause data loss during the transmission process. And continuously occupying a high signal-to-noise ratio channel will result in low channel utilization rate, and at the same time, the data transmission pressure on the high signal-to-noise ratio channel is relatively large. Therefore, the present disclosure uses the method of channel fusion to fuse and use the low signal-to-noise ratio channel and the high signal-to-noise ratio channel, which not only improves the channel utilization rate but also can ensure the effective transmission of data and reduce the data transmission pressure on the high signal-to-noise ratio channel. In the fourth step, in response to the server receiving the preprocessed power consumption data sent through the fusion channel, the server uses a multi-modal neural network model to perform abnormal power consumption identification on the preprocessed power consumption data to generate a power consumption identification result for the virtual object. By adopting a multi-modal neural network model, the fusion analysis of heterogeneous power consumption data is carried out to improve the accuracy of abnormal power consumption identification. Through this method, the effective identification of abnormal power consumption behavior is realized, and the power consumption safety risk is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In combination with the accompanying drawings and with reference to the following specific embodiments, the above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent. Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the elements and elements are not necessarily drawn to scale.

[0012] Figure 1 is a flowchart of some embodiments of an abnormal power consumption identification method based on a multi-modal neural network model according to the present disclosure;

[0013] Figure 2 is a schematic diagram of the process of collecting power consumption data;

[0014] Figure 3 is a schematic diagram of the structure of a power supply topology;

[0015] Figure 4 is a schematic diagram of the process of receiving and transmitting data through a data repeater;

[0016] Figure 5 is a schematic diagram of the process of transmitting preprocessed power consumption data through a fusion channel;

[0017] Figure 6 is a schematic diagram of the model structure of a multi-modal neural network model;

[0018] Figure 7 is a schematic diagram of the structure of some embodiments of an abnormal power consumption identification device based on a multi-modal neural network model according to the present disclosure;

[0019] Figure 8 It is a schematic structural diagram of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed implementation manners

[0020] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not used to limit the protection scope of the present disclosure.

[0021] In addition, it should be noted that for the sake of convenience of description, only parts related to the relevant invention are shown in the drawings. Without conflict, the embodiments in the present disclosure and the features in the embodiments can be combined with each other.

[0022] It should be noted that the concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependent relationships.

[0023] It should be noted that the modifications of "one" and "multiple" mentioned in the present disclosure are illustrative rather than restrictive. Those skilled in the art should understand that unless otherwise clearly specified in the context, it should be understood as "one or more".

[0024] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only for illustrative purposes and are not used to limit the scope of these messages or information.

[0025] The present disclosure will be described in detail below with reference to the drawings and in combination with the embodiments.

[0026] Reference Figure 1 , a flow 100 of some embodiments of an abnormal power consumption recognition method based on a multi-modal neural network model according to the present disclosure is shown. The abnormal power consumption recognition method based on the multi-modal neural network model includes the following steps:

[0027] Step 101, constructing virtual objects according to the power supply topology structure and the power consumption portraits corresponding to the meta-objects in at least one meta-object to obtain a virtual object set.

[0028] In some embodiments, an execution subject (e.g., a computing device) of an abnormal power consumption recognition method based on a multi-modal neural network model may construct virtual objects according to a power supply topology structure and power consumption portraits corresponding to meta-objects in at least one meta-object, to obtain a set of virtual objects. The power supply topology structure refers to a topology structure composed of power supply lines, as well as power consumption devices and data acquisition devices connected by the power supply lines. A meta-object is a power consumption object with a meta-granularity in the above-mentioned power supply topology structure. Specifically, the meta-granularity refers to the object granularity of a single power consumption entity. For example, enterprise A may be a power consumption object with a meta-granularity. Resident A may also be a power consumption object with a meta-granularity. A virtual object is composed of at least two meta-objects that have an indirect topological relationship and corresponding approximate power consumption behaviors. The indirect topological relationship means that the meta-objects are not directly connected, but there is at least one intermediate point in between to associate the meta-objects. In practice, the meta-objects with approximate power consumption behaviors can be combined through similarity calculation in combination with the power consumption portraits to obtain virtual objects.

[0029] As an example, refer to Figure 2 the schematic diagram of the process of collecting power consumption data shown in Figure 2 which shows two ways of collecting power consumption data. Specifically, the first is constrained by geographical location, that is, multiple residents (power consumption objects with a meta-granularity) in Unit A building are regarded as a whole (virtual object), and through the data sending device set in Unit A building, the power consumption data is uniformly transmitted to the server end 201 for abnormal power consumption analysis. This method has a relatively low hardware layout cost, that is, only a single data sending device needs to be set in Unit A building to send the data to the server end 201. However, in this case, since there are differences in the power consumption habits of multiple residents in Unit A building, in this situation, when identifying abnormal power consumption, the accuracy of the identification result cannot be guaranteed. The second is to send the power consumption data of a single resident (power consumption object with a meta-granularity) to the server end 201 for data analysis through the data sending device set in the single resident. The accuracy of this method can be effectively guaranteed, but it is necessary to lay out a separate data sending device for each power consumption object with a meta-granularity, and the layout cost is extremely high. At the same time, the server end 201 needs to identify abnormal power consumption for power consumption objects with a finer granularity (meta-granularity), and the data analysis pressure is extremely high.

[0030] It should be noted that the above-mentioned computing device can be hardware or software. When the computing device is hardware, it can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or it can be implemented as a single server or a single terminal device. When the computing device is embodied as software, it can be installed in the hardware devices listed above. It can be implemented as multiple software or software modules for providing distributed services, for example, or it can be implemented as a single software or software module. No specific limitation is made here. In practice, the execution subject of the above-mentioned computing device can be an abnormal power consumption identification system, that is, the above-mentioned abnormal power consumption identification method based on the multimodal neural network model is applied to the above-mentioned abnormal power consumption identification system. Specifically, the abnormal power consumption identification system includes: an edge end and a server end. The edge end can be provided with a data repeater to collect and send power consumption data. The server end can be used to construct virtual objects and identify abnormal power consumption of power consumption data.

[0031] In some optional implementations of some embodiments, the execution subject constructs a virtual object according to the power supply topology and the power consumption portrait corresponding to the meta-object in at least one meta-object to obtain a virtual object set, including:

[0032] In the first step, for each meta-object in the at least one meta-object, the following profile feature construction steps are performed:

[0033] In the first sub-step, dynamic portrait attributes are extracted from the electricity consumption portrait corresponding to the above meta-object to obtain a dynamic electricity consumption portrait.

[0034] Among them, the dynamic electricity consumption portrait represents the portrait attributes that change dynamically over time. In practice, in order for the electricity consumption portrait to accurately characterize the meta-object, it is often necessary to combine electricity consumption data to construct a complex electricity consumption portrait. However, in the process of abnormal electricity consumption analysis, static portrait attributes (for example, name, gender, etc.) have little reference significance for abnormal electricity consumption analysis and identification, but will increase the subsequent data processing volume. Therefore, it is necessary to extract the attributes of the electricity consumption portrait corresponding to the meta-object to obtain a dynamic electricity consumption portrait. Specifically, the above-mentioned execution subject can eliminate the portrait attributes that do not change dynamically over time in the electricity consumption portrait, so as to realize the dynamic portrait attribute extraction of the electricity consumption portrait and obtain a dynamic electricity consumption portrait. Furthermore, in order to avoid the extra occupation of real-time portrait attribute extraction during the peak usage period of computing resources, pre-extraction can be performed during the low-peak usage period of computing resources.

[0035] The second sub-step is to construct the image features of the above-mentioned dynamic electricity consumption image to generate dynamic electricity consumption image features.

[0036] In practice, one-hot encoding or word embedding can be used to generate dynamic power consumption portrait features for the dynamic power consumption portrait. Specifically, the dynamic power consumption portrait features can be represented by a one-dimensional feature vector to facilitate subsequent similarity calculation.

[0037] In the second step, according to the dynamic power consumption portrait features corresponding to the meta-objects, object clustering is performed on the meta-objects in the at least one meta-object to obtain a sequence of meta-object groups.

[0038] In practice, the K-means clustering algorithm can be used to perform object clustering on the meta-objects in the at least one meta-object to obtain a sequence of meta-object groups. Wherein, a meta-object group is a clustering cluster.

[0039] In the third step, according to the sequence of meta-object groups, the following virtual object construction steps are performed:

[0040] In the first sub-step, the first meta-object group in the sequence of meta-object groups is determined as the target meta-object group.

[0041] As an example, refer to Figure 3 the structural schematic diagram of the power supply topology shown in Figure 3 which shows a power supply topology in a tree structure. In practice, other types of topologies such as star-shaped can also be used. Specifically, Figure 3 the corresponding sequence of meta-object groups may include: meta-object group A, meta-object group B, meta-object group C. Meta-object group A includes: meta-object A1, meta-object A2, meta-object A3, meta-object A4. Meta-object group B includes: meta-object B1, meta-object B2, meta-object B3, meta-object B4, meta-object B5, meta-object B6. Meta-object group C includes: meta-object C1, meta-object C2, meta-object C3, meta-object C4, meta-object C5. The above-mentioned execution entity can use meta-object group A as the target meta-object group.

[0042] In the second sub-step, with a unit increment as the step size and the target meta-object in the target meta-object group as the starting node, a reverse step-by-step traversal is performed on the above-mentioned power supply topology to determine the node list corresponding to each target meta-object in the target meta-object group.

[0043] Wherein, the unit increment refers to the distance between two adjacent nodes in the power supply topology. Taking Figure 3 as an example, the distance between the node corresponding to meta-object A1 and node T1 is the unit increment. The node list refers to the nodes with the connection relationship to the meta-object corresponding node with the target meta-object corresponding node as the root node.

[0044] As an example, further refer to Figure 3Schematic diagram of the power supply topology shown, where for the (target) meta-object A1 in the target meta-object group, the corresponding node list is [node T4, node T2, node T1]. For the (target) meta-object A2 in the target meta-object group, the corresponding node list is [node T4, node T2, node T1]. For the (target) meta-object A3 in the target meta-object group, the corresponding node list is [node T7, node T4, node T2, node T1]. For the (target) meta-object A4 in the target meta-object group, the corresponding node list is [node T8, node T2, node T1].

[0045] The third sub-step is to determine the node list intersections of each node list in the node list group corresponding to the target meta-object group.

[0046] As an example, the node list intersection can be [node T4, node T2, node T1] ∩ [node T4, node T2, node T1] ∩ [node T7, node T4, node T2, node T1] ∩ [node T8, node T2, node T1] = [node T2 and node T1].

[0047] The fourth sub-step is to, in response to the presence of a target node in the node list intersection and the node distance between the target node and the target meta-object with the farthest distance in the target meta-object group being less than the preset node distance, construct a virtual object according to the above target meta-object group.

[0048] Among them, the target node is the node in the node list that is closest to each target meta-object in the target meta-object group.

[0049] As an example, further taking the target meta-object group (meta-object group A) as an example, the target node can be node T2. Among them, the node distance (3) between the target node (node T2) and the target meta-object with the farthest distance in the target meta-object group (node A3) is less than the preset node distance (5). Therefore, the above execution entity can take the meta-object A1, meta-object A2, meta-object A3, and meta-object A4 in the target meta-object group as a whole to obtain a virtual object.

[0050] The fifth sub-step is to end the above virtual object construction step in response to the meta-object group sequence after removing the target meta-object group being empty.

[0051] The sixth sub-step is to, in response to the non-existence of a target node in the node list intersection or the node distance between the target node and the target meta-object with the farthest distance in the target meta-object group being greater than or equal to the preset node distance, perform group splitting on the target meta-object group to obtain at least one split meta-object group, and add the at least one split meta-object group to the meta-object group sequence after removing the target meta-object group.

[0052] As an example, when the target meta-object group is meta-object group C, for the (target) meta-object C1 in the target meta-object group, the corresponding node list is [node T4, node T2, node T1]. For the (target) meta-object C2 in the target meta-object group, the corresponding node list is [node T6, node T3, node T1]. For the (target) meta-object C3 in the target meta-object group, the corresponding node list is [node T10, node T7, node T4, node T2, node T1]. For the (target) meta-object C4 in the target meta-object group, the corresponding node list is [node T10, node T7, node T4, node T2, node T1]. For the (target) meta-object C5 in the target meta-object group, the corresponding node list is [node T11, node T9, node T6, node T3, node T1]. Therefore, the intersection of the node lists can be [node T1]. At this time, the distance (5) between the target meta-object C3 and node T1 is greater than or equal to the preset node distance. Therefore, it is necessary to split the (target) meta-object group C3. Specifically, the intersection node in the node list can be used as the split point, and the split results are [meta-object C1, meta-object C3, meta-object C4], [meta-object C2, meta-object C5] as at least one split meta-object group.

[0053] In the fourth step, in response to the fact that the meta-object group sequence after removing the target meta-object group is not empty, the meta-object group sequence after removing the target meta-object group is used as the meta-object group sequence, and the above virtual object construction steps are executed again.

[0054] The content of "in some optional implementation manners of some embodiments" above is an inventive point of the present disclosure. In practice, by combining the electricity consumption portrait and the power supply topology structure, the construction of virtual objects can be realized. Compared with the fine-grained method, it can effectively reduce the subsequent data processing volume. At the same time, compared with the method of dividing by geographical location, it can improve the accuracy of subsequent recognition. In addition, in order to effectively collect the target electricity consumption data corresponding to the virtual object, the present disclosure sets a preset node distance to constrain the communication distance between the target node and the meta-objects included in the virtual object, reducing the subsequent acquisition cost of the target electricity consumption data. At the same time, the data repeater is set at the target node to realize the collection and transmission of the target electricity consumption data. In this way, the target electricity consumption data is effectively collected, and at the same time, the deployment cost of fully deploying wireless devices can be effectively reduced, and the interference problem caused by densely deploying wireless devices can be effectively solved.

[0055] Step 102, for each virtual object in the virtual object set, execute the following processing steps:

[0056] Step 1021, collect the target electricity consumption data corresponding to the virtual object through the data repeater corresponding to the virtual object.

[0057] In some embodiments, the above-mentioned execution entity may collect target power consumption data corresponding to a virtual object through a data repeater corresponding to the virtual object. Among them, the above-mentioned data repeater includes: a data concentrator and a radio frequency transceiver. The data repeater is used to collect power consumption data of at least two meta-objects corresponding to the above-mentioned virtual object in a wired manner. The radio frequency transceiver is used to wirelessly send data to the server side. In practice, considering the stability of wired transmission, a certain proportion of power consumption users (meta-objects) have already laid corresponding limited lines, signal interference between densely arranged wireless transmission devices, and the equipment cost of replacing all wireless transmission devices, therefore, power consumption data is obtained in a wired manner between the data concentrator and at least two meta-objects included in the virtual object. Wireless data transmission is performed between the server side and the virtual object using a radio frequency transceiver.

[0058] As an example, refer to Figure 4 The schematic diagram of the process of receiving and sending data through a data repeater shown in the figure. Among them, the data repeater 401 includes a data concentrator 4011 and a radio frequency transceiver 4012. The virtual object may be composed of meta-object A1, meta-object A2, meta-object A3, and meta-object A4. The data concentrator 4011 may collect power consumption data corresponding to meta-object A1, meta-object A2, meta-object A3, and meta-object A4 included in the virtual object as target power consumption data in a wired connection manner. And through the radio frequency transceiver, the data is sent to the server side 201 in a wireless transmission manner.

[0059] In some optional implementation manners of some embodiments, the above-mentioned execution entity collects the target power consumption data corresponding to the above-mentioned virtual object through the data repeater corresponding to the above-mentioned virtual object, including:

[0060] The first step is to determine the communication links corresponding to the virtual object with the data repeater as the receiving end point, and obtain at least two communication links.

[0061] Among them, the data repeater corresponding to the above-mentioned virtual object is set on the target node corresponding to the above-mentioned virtual object. At least two communication links are links used for data transmission between the target node corresponding to the virtual object and at least two meta-objects corresponding to the virtual object. In practice, since when determining the virtual object, the links between the meta-objects included in the virtual object and the target node have been determined through reverse traversal, and the data repeater is set on the target node, therefore, the links between the meta-objects included in the virtual object and the target node in the reverse traversal stage can be used as communication links, and since the virtual object includes at least two meta-objects, at least two communication links can be obtained.

[0062] The second step is to determine the delay value corresponding to each communication link in the above-mentioned at least one communication link.

[0063] Among them, the time delay value refers to the time length for data to be transmitted from the meta-objects included in the virtual object to the data repeater set at the target node. Since at least two meta-objects included in the virtual object are not necessarily located in the same geographical location, there may be a certain time difference when obtaining the corresponding power consumption data, which will increase the data alignment cost in the subsequent data processing process. Therefore, in the acquisition stage, the synchronous acquisition of power consumption data is controlled by the time delay value to facilitate the direct alignment of data. Specifically, a PING command can be executed from the target node to determine the time delay value when sending data to the meta-objects of the virtual object through the communication link.

[0064] The third step is to asynchronously collect the target power consumption data corresponding to the virtual object through the at least one communication link according to the time delay value corresponding to the communication link.

[0065] As an example, the virtual object includes: meta-object A1, meta-object A2, meta-object A3, meta-object A4. The time delay value corresponding to the target node and meta-object A1 is △t1. The time delay value corresponding to the target node and meta-object A2 is △t2. The time delay value corresponding to the target node and meta-object A3 is △t3. The time delay value corresponding to the target node and meta-object A4 is △t4. Therefore, based on the meta-object with the smallest corresponding time delay value, and at intervals corresponding to the time delay values, the power consumption data corresponding to the meta-objects included in the virtual object is asynchronously collected through the corresponding communication links to obtain the target power consumption data.

[0066] Step 1022: Perform data preprocessing on the target power consumption data to obtain the preprocessed power consumption data.

[0067] In some embodiments, the execution subject performs data preprocessing on the target power consumption data to obtain the preprocessed power consumption data. For example, the target power consumption data can be preprocessed by means such as supplementing with the default value 0 to obtain the preprocessed power consumption data.

[0068] Optionally, the target power consumption data is an index value matrix represented in matrix form and varying with the time dimension. The matrix dimension of the target power consumption data is N×T×9. N is the number of meta-objects corresponding to the virtual object. T is the time scale length. "9" refers to 9 evaluation indicators corresponding to the power consumption data, namely: phase A current, phase B current, phase C current, phase A voltage, phase C voltage, phase D voltage, phase A power factor, phase B power factor, and phase C power factor. Considering that current, voltage, and power factor are continuous values varying with time, in order to reduce the data processing volume, when stored in matrix form, the current, voltage, and power factor have been discretized by means of fixed-frequency sampling. Therefore, T discrete values for current, voltage, and power factor can be obtained. During the acquisition process, the current, voltage, and power factor are stored using double-precision floating-point values.

[0069] In some alternative implementations of some embodiments, the above-mentioned execution entity performs data preprocessing on the above-mentioned target power consumption data to obtain preprocessed power consumption data, including:

[0070] First step, split the target power consumption data into a sequence of matrices to be processed.

[0071] Among them, the matrix dimension of the matrix to be processed is N×T×1. That is, each matrix to be processed corresponds to an evaluation index. For example, the sequence of matrices to be processed includes: matrix to be processed A1, matrix to be processed A2, matrix to be processed A3, matrix to be processed A4, matrix to be processed A5, matrix to be processed A6, matrix to be processed A7, matrix to be processed A8, matrix to be processed A9. Among them, the matrix to be processed A1 corresponds to the A-phase current, the matrix to be processed A2 corresponds to the B-phase current, the matrix to be processed A3 corresponds to the C-phase current, the matrix to be processed A4 corresponds to the A-phase voltage, the matrix to be processed A5 corresponds to the B-phase voltage, the matrix to be processed A6 corresponds to the C-phase voltage, the matrix to be processed A7 corresponds to the A-phase power factor, the matrix to be processed A8 corresponds to the B-phase power factor, and the matrix to be processed A9 corresponds to the C-phase power factor.

[0072] Second step, for each matrix to be processed in the above-mentioned sequence of matrices to be processed, perform the following fusion steps:

[0073] The first sub-step is to perform precision conversion on the above-mentioned matrix to be processed to obtain a matrix after precision conversion.

[0074] Among them, the precision of the matrix values corresponding to the matrix after precision conversion is less than the precision of the matrix values corresponding to the matrix to be processed. In practice, the matrix values in the matrix after precision conversion can be single-precision floating-point numbers. Therefore, the storage occupancy of the matrix values can be effectively compressed, and the subsequent data transmission efficiency can be improved.

[0075] The second sub-step is to fit a set of target curves according to the above-mentioned matrix after precision conversion.

[0076] Among them, the number of target curves in the set of target curves is N. In practice, the least squares method can be used for curve fitting.

[0077] The third sub-step is to determine a reference value according to the set of target curves.

[0078] In practice, since the meta-objects included in the virtual object have approximate power consumption behaviors, the curves corresponding to the current, voltage, and power factor of the meta-objects have approximate properties. Therefore, each value in the target curve can be converted into the form of reference value + increment. Specifically, the reference value can correspond to the midline of the set of target curves.

[0079] The fourth sub-step is to perform increment value conversion on the above-mentioned matrix after precision conversion with the above-mentioned reference value as the baseline to obtain a converted matrix.

[0080] In practice, the matrix values in the precision-converted matrix can be subtracted by the reference value to obtain the converted matrix. In this way, the single-precision floating-point number is converted into the form of reference value + increment value.

[0081] The fifth sub-step is to merge the reference value as the header with the above-mentioned converted matrix to obtain the updated matrix.

[0082] Among them, the matrix dimension of the above-mentioned updated matrix is 1×(N×T + 1). Since the reference value is added as the header, the matrix dimension is increased by 1. At the same time, converting the N×T matrix into a one-dimensional matrix of 1×(N×T + 1) can reduce the storage of the reference value at the header by N - 1 times.

[0083] The third step is to combine the obtained set of updated matrices to obtain the preprocessed power consumption data corresponding to the above virtual object.

[0084] Among them, the matrix dimension of the preprocessed power consumption data is 1×(N×T + 1)×9.

[0085] Step 1023: Dynamically determine the fusion channel corresponding to the preprocessed power consumption data according to the data status of the preprocessed power consumption data.

[0086] In some embodiments, the above-mentioned execution entity can dynamically determine the fusion channel corresponding to the preprocessed power consumption data according to the data status of the preprocessed power consumption data. Among them, the data status represents the zero occupancy ratio of the preprocessed power consumption data in each dimension. The fusion channels include: the main channel and the secondary channel. The signal-to-noise ratio of the main channel is greater than that of the secondary channel. Specifically, data with a low zero occupancy ratio can be transmitted through the main channel, and data with a high zero occupancy ratio can be transmitted through the secondary channel.

[0087] In some optional implementation manners of some embodiments, the above-mentioned execution entity dynamically determines the fusion channel corresponding to the above-mentioned preprocessed power consumption data according to the data status of the above-mentioned preprocessed power consumption data, including:

[0088] The first step is to obtain the available channel status information from the above-mentioned server side.

[0089] Among them, the available channel status information represents the available channels between the data repeater and the server side, and the signal-to-noise ratio corresponding to the available channels.

[0090] The second step is to determine the data zero occupancy ratio of the above-mentioned preprocessed power consumption data in each matrix dimension to obtain the data status.

[0091] In practice, the matrix dimension of the preprocessed power consumption data is 1×(N×T + 1)×9, that is, the zero occupancy ratios corresponding to 9 1×(N×T + 1) matrices can be determined from the dimensions of 9 evaluation indicators.

[0092] In the third step, at least one available channel group is determined according to the above data status and the above available channel status information.

[0093] The available channel group includes: available channel A, available channel B, and available channel C. Among them, the signal-to-noise ratio corresponding to available channel A is greater than the signal-to-noise ratio corresponding to available channel B, the signal-to-noise ratio corresponding to available channel B is greater than the signal-to-noise ratio of channel C, available channel A and available channel B form a primary channel, and available channel C forms a secondary channel. In practice, a channel with a high signal-to-noise ratio can be selected as available channel A, a signal with a signal-to-noise ratio lower than that corresponding to available channel A can be selected as available channel B, and a channel with a low signal-to-noise ratio can be selected as available channel C. In practice, in order to reduce the problem of data packet loss caused by channels, during the transmission process, a channel with a high signal-to-noise ratio is usually selected, which results in too much data traffic being carried by the channel with a high signal-to-noise ratio and is also not conducive to channel utilization from a global perspective. Therefore, through the method of channel combination, according to the data status, adaptive selection and transmission are carried out, which not only improves the channel utilization rate but also enables faster data transmission at an acceptable packet loss rate.

[0094] In some optional implementation manners of some embodiments, before the above-mentioned preprocessed power consumption data is received by the server through the fusion channel and the preprocessed power consumption data is subjected to abnormal power consumption identification through a multimodal neural network model to generate a power consumption identification result for the above-mentioned virtual object, the method further includes:

[0095] In the first step, the data corresponding to the low data zero occupancy ratio in the above-mentioned preprocessed power consumption data is transmitted through available channel A included in the above-mentioned primary channel.

[0096] In the second step, the data corresponding to the high data zero occupancy ratio in the above-mentioned preprocessed power consumption data is transmitted through available channel B included in the above-mentioned primary channel.

[0097] In the third step, the reference value in the above-mentioned preprocessed power consumption data is transmitted through available channel C included in the above-mentioned secondary channel.

[0098] As an example, see Figure 5Schematic diagram of the process of transmitting preprocessed electricity consumption data through a fusion channel. Among them, available channels may include: Channel 1, Channel 2, Channel 3, ……, Channel N-1, Channel N. Among them, the signal-to-noise ratio of Channel 1 is the highest, and the signal-to-noise ratio of Channel N is the lowest. In practice, due to the highest signal-to-noise ratio of Channel 1, Channel 1 may be preferentially selected for data transmission during other data transmission processes. At this time, when selecting Channel 1, the data transmission speed may be slower because Channel 1 bears more data traffic. Therefore, when constructing a fusion channel, Channel 2 and Channel 3 are selected as available channel A and available channel B respectively, and Channel N-1 is selected as available channel C. In this way, the data with a low zero occupancy ratio in the preprocessed electricity consumption data 501 is transmitted through available channel A (Channel 2) to avoid key data packet loss. At the same time, the data with a high zero occupancy ratio in the preprocessed electricity consumption data 501 is transmitted through available channel B (Channel 3). Because the zero occupancy ratio is relatively high, from a probability perspective, even if the signal-to-noise ratio is low, the tolerance for data loss is higher than that of the data with a low zero occupancy ratio when data is lost. Finally, the reference value in the preprocessed electricity consumption data is transmitted through available channel C (Channel N-1). Although the signal-to-noise ratio of available channel C is low and the probability of data packet loss is greater, the amount of reference value data is small, and even if it is resent repeatedly, it will not generate a large transmission cost. In this way, the preprocessed electricity consumption data 501 is transmitted to the server side 201. Through this method, channels are multiplexed to achieve efficient data transmission.

[0099] Step 1024, in response to the server side receiving the preprocessed electricity consumption data sent through the fusion channel, the server side uses a multi-modal neural network model to perform abnormal electricity consumption identification on the preprocessed electricity consumption data to generate an electricity consumption identification result for the virtual object.

[0100] In some embodiments, the above execution entity may, in response to the server side receiving the preprocessed electricity consumption data sent through the fusion channel, use a multi-modal neural network model to perform abnormal electricity consumption identification on the preprocessed electricity consumption data to generate an electricity consumption identification result for the virtual object.

[0101] Optionally, the multi-modal neural network model includes: 3 feature extraction blocks, a feature fusion block, and an electricity consumption result classifier. The 3 feature extraction blocks are used to extract features from different modal data. The feature extraction block includes: 3 feature extraction networks, and the parameters between the 3 feature extraction networks are shared.

[0102] In some optional implementation manners of some embodiments, the above execution entity, in response to the server side receiving the above preprocessed electricity consumption data sent through the fusion channel, the above server side uses a multi-modal neural network model to perform abnormal electricity consumption identification on the above preprocessed electricity consumption data to generate an electricity consumption identification result for the above virtual object, including:

[0103] First step: Through the above three feature extraction blocks, parallel feature extraction is performed on different modal data in the preprocessed electricity consumption data, and a set of data features to be fused is obtained.

[0104] Second step: Through the above feature fusion block, feature fusion is performed on the above set of data features to be fused, and fused features are obtained.

[0105] Third step: According to the above fused features and the above electricity consumption result classifier, an electricity consumption recognition result for the above virtual object is generated.

[0106] As an example, refer to Figure 6 the schematic diagram of the model structure of the multi-modal neural network model shown in the figure. Among them, the multi-modal neural network model includes: three feature extraction blocks (feature extraction block 601, feature extraction block 602, and feature extraction block 603), a feature fusion block 604, and an electricity consumption result classifier 605. The three feature extraction blocks are used to perform feature extraction on different modal data. Since the electricity consumption data is constructed from nine evaluation indicators, namely, phase A current, phase B current, phase C current, phase A voltage, phase C voltage, phase D voltage, phase A power factor, phase B power factor, and phase C power factor, and there are only differences in the phases among the three-phase currents (the same applies to the three-phase voltages and three-phase power factors), considering this point, the corresponding multi-modal neural network model selects three groups of feature extraction blocks. Each feature extraction block includes three feature extraction networks with shared parameters (feature extraction block 601 includes: feature extraction network 6011, feature extraction network 6012, and feature extraction network 6013; feature extraction block 602 includes: feature extraction network 6021, feature extraction network 6022, and feature extraction network 6023; feature extraction block 603 includes: feature extraction network 6031, feature extraction network 6032, and feature extraction network 6033). The network structures of the nine feature networks are the same. In practice, the feature extraction network can select TimesNet as the backbone network. Among them, TimeNet is composed of multiple time blocks (TimeBlock) connected by skip connections. Specifically, the time blocks (TimeBlock) at the relative positions in the three feature extraction networks included in the feature extraction block share parameters. The feature fusion block 604 splices the outputs of the three feature extraction blocks in a feature splicing manner. The electricity consumption result classifier 605 is composed of fully connected layers connected in series.

[0107] As another inventive point of the present disclosure, the above-mentioned multimodal neural network model realizes effective feature extraction for power consumption data and abnormal power consumption identification. By means of parameter sharing, the training complexity of the nine feature extraction networks in the model training stage can be effectively reduced. At the same time, the data corresponding to the nine evaluation indicators have obvious temporal characteristics. Using the temporal type of TimeNet as the backbone network can effectively perform feature extraction. In summary, through this multimodal neural network model, effective feature extraction for power consumption data and abnormal power consumption identification are realized.

[0108] The above-mentioned various embodiments of the present disclosure have the following beneficial effects: Through the abnormal power consumption recognition method based on the multi-modal neural network model in some embodiments of the present disclosure, the effective recognition of abnormal power consumption behaviors is realized, ensuring power consumption safety. Specifically, with the popularization of power consumption, the number of power-consuming devices connected to the power grid has increased sharply, and thus a large number of abnormal power consumption behaviors (such as electricity theft behaviors) have emerged. When the line load is exceeded, abnormal power consumption behaviors also increase the power consumption safety risk. Therefore, in some embodiments of the present disclosure, the abnormal power consumption recognition method based on the multi-modal neural network model first constructs virtual objects according to the power supply topology structure and the power consumption portraits corresponding to the meta-objects in at least one meta-object, obtaining a set of virtual objects, where the meta-object is the power-consuming object at the meta-granularity in the above-mentioned power supply topology structure, and the virtual object is composed of at least two meta-objects with an indirect topological relationship and corresponding approximate power consumption behaviors. In practice, conventional power consumption behavior analysis methods often adopt methods such as data analysis for a single power-consuming object or data analysis for a group of power-consuming objects with the same geographical location. For the former, due to the large number of power-consuming objects and the fact that power-consuming objects are often accompanied by approximate continuous power consumption behaviors, the amount of data to be analyzed is extremely large. For the latter, although a group of power-consuming objects with the same geographical location are aggregated and then data analysis is carried out, the power consumption habits among a group of power-consuming objects with the same geographical location may be different, thus resulting in the inability to guarantee the recognition accuracy of abnormal power consumption. Therefore, by combining the power supply topology structure and the power consumption portrait, the present disclosure combines meta-objects (power-consuming objects) with an indirect topological relationship and power consumption behaviors for analysis, which can reduce the amount of data processing while ensuring the recognition accuracy. Secondly, for each virtual object in the above-mentioned set of virtual objects, the following processing steps are executed: The first step is to collect the target power consumption data corresponding to the above-mentioned virtual object through the data repeater corresponding to the above-mentioned virtual object. Among them, the above-mentioned data repeater includes: a data concentrator and a radio frequency transceiver. The data repeater is used to collect the power consumption data of at least two meta-objects corresponding to the above-mentioned virtual object in a wired manner, and the radio frequency transceiver is used to send the data to the server side wirelessly. In practice, since the power-consuming objects at the edge end have different power-consuming device layout methods, replacing them uniformly with wireless devices (such as radio meters) has a very high replacement cost. Therefore, for at least two meta-objects corresponding to the virtual object, collecting power consumption data in a wired manner and then sending it to the server side through the radio frequency transceiver can effectively reduce the device layout cost and ensure the stability of power consumption data transmission. The second step is to perform data preprocessing on the above-mentioned target power consumption data to obtain preprocessed power consumption data. In practice, if the target power consumption data is directly used, due to the real-time generation of power consumption data, there is still a problem of a large amount of data processing. Therefore, through data preprocessing, the amount of subsequent data processing is reduced, alleviating the data processing pressure on the server.Step 3: Dynamically determine a fusion channel corresponding to the preprocessed power consumption data according to the data status of the preprocessed power consumption data. The fusion channel includes a primary channel and a secondary channel, and the signal-to-noise ratio of the primary channel is greater than that of the secondary channel. In practice, the selection of a channel is involved in the wireless data transmission process. A poor channel will not only affect the transmission efficiency but may also cause data loss during the transmission process. Continuously occupying a channel with a high signal-to-noise ratio will result in low channel utilization, and at the same time, the data transmission pressure on the high signal-to-noise ratio channel is relatively high. Therefore, the present disclosure uses a channel fusion method to fuse and use a low signal-to-noise ratio channel and a high signal-to-noise ratio channel, which not only improves the channel utilization rate but also ensures the effective transmission of data and reduces the data transmission pressure on the high signal-to-noise ratio channel. Step 4: In response to the server receiving the preprocessed power consumption data sent through the fusion channel, the server uses a multi-modal neural network model to identify abnormal power consumption of the preprocessed power consumption data to generate a power consumption identification result for the virtual object. By using a multi-modal neural network model, the fusion analysis of heterogeneous power consumption data is performed to improve the accuracy of abnormal power consumption identification. Through this method, the effective identification of abnormal power consumption behavior is realized, and the power consumption safety risk is reduced.

[0109] Further referring to Figure 7 , as an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of an abnormal power consumption identification device based on a multi-modal neural network model. These device embodiments correspond to Figure 1 the method embodiments shown, and the abnormal power consumption identification device based on the multi-modal neural network model can be specifically applied to various electronic devices.

[0110] As Figure 7As shown, the abnormal power consumption recognition device 700 based on a multi-modal neural network model in some embodiments includes: a virtual object construction unit 701 and an execution unit 702. Among them, the virtual object construction unit 701 is configured to construct virtual objects according to the power supply topology structure and the power consumption portraits corresponding to the meta-objects in at least one meta-object, and obtain a set of virtual objects. Here, the meta-object is the power consumption object at the meta-granularity in the above power supply topology structure, and the virtual object is composed of at least two meta-objects with an indirect topological relationship and corresponding approximate power consumption behaviors; the execution unit 702 is configured to perform the following processing steps for each virtual object in the above set of virtual objects: collect the target power consumption data corresponding to the virtual object through the data repeater corresponding to the virtual object. Here, the data repeater includes: a data concentrator and a radio frequency transceiver. The data repeater is used to collect the power consumption data of at least two meta-objects corresponding to the virtual object in a wired manner, and the radio frequency transceiver is used to wirelessly send the data to the server side; perform data preprocessing on the above target power consumption data to obtain preprocessed power consumption data; dynamically determine a fusion channel corresponding to the preprocessed power consumption data according to the data state of the preprocessed power consumption data. Here, the fusion channel includes: a main channel and a secondary channel, and the signal-to-noise ratio corresponding to the main channel is greater than that of the secondary channel; in response to the server side receiving the preprocessed power consumption data sent through the fusion channel, the server side uses a multi-modal neural network model to perform abnormal power consumption recognition on the preprocessed power consumption data to generate a power consumption recognition result for the virtual object.

[0111] It can be understood that the units described in the abnormal power consumption recognition device 700 based on the multi-modal neural network model correspond to the respective steps in the method described in the reference Figure 1 Therefore, the operations, features, and beneficial effects described above for the method also apply to the abnormal power consumption recognition device 700 based on the multi-modal neural network model and the units included therein, and will not be repeated here.

[0112] Next, refer to Figure 8 , which shows a schematic structural diagram of an electronic device (for example, a computing device) suitable for implementing some embodiments of the present disclosure. Figure 8 The electronic device shown is only an example and should not bring any restrictions to the functions and usage scopes of the embodiments of the present disclosure. As Figure 8As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus. Among them, the memory may include a non-volatile storage medium and an internal memory. The non-volatile storage medium can store an operating system and computer programs. The computer programs include program instructions that, when executed, can cause the processor to execute any one of the front-end page monitoring methods. The processor is used to provide computing and control capabilities to support the operation of the entire computer device. The internal memory provides an environment for the operation of the computer programs in the non-volatile storage medium. When the computer programs are executed by the processor, the processor can be caused to execute any one of the front-end page monitoring methods. The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 8 the structure shown in is only a block diagram of some structures related to the solution of the present disclosure, and does not constitute a limitation on the computer device to which the solution of the present disclosure is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0113] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0114] Among them, in one embodiment, the above-mentioned processor is used to run a computer program stored in the memory to implement the following steps: constructing virtual objects according to the power supply topology structure and the power consumption portraits corresponding to the meta-objects in at least one meta-object to obtain a set of virtual objects, where the meta-object is a power consumption object at the meta-granularity in the above-mentioned power supply topology structure, and the virtual object is composed of at least two meta-objects with an indirect topological relationship and corresponding approximate power consumption behaviors; for each virtual object in the above-mentioned set of virtual objects, perform the following processing steps: collecting target power consumption data corresponding to the above-mentioned virtual object through the data repeater corresponding to the above-mentioned virtual object, where the above-mentioned data repeater includes: a data concentrator and a radio frequency transceiver, and the data repeater is used to collect the power consumption data of at least two meta-objects corresponding to the above-mentioned virtual object in a wired manner, and the radio frequency transceiver is used to send the data to the server side wirelessly; performing data preprocessing on the above-mentioned target power consumption data to obtain preprocessed power consumption data; dynamically determining a fusion channel corresponding to the above-mentioned preprocessed power consumption data according to the data state of the above-mentioned preprocessed power consumption data, where the fusion channel includes: a main channel and a secondary channel, and the signal-to-noise ratio corresponding to the main channel is greater than that of the secondary channel; in response to the server side receiving the above-mentioned preprocessed power consumption data sent through the fusion channel, the server side uses a multi-modal neural network model to perform abnormal power consumption identification on the above-mentioned preprocessed power consumption data to generate a power consumption identification result for the above-mentioned virtual object.

[0115] The embodiments of the present disclosure also provide a computer-readable storage medium, on which a computer program is stored, and the computer program includes program instructions, and the method implemented when the program instructions are executed can refer to the various embodiments of the front-end page monitoring method of the present disclosure.

[0116] Among them, the above-mentioned computer-readable storage medium may be an internal storage unit of the above-mentioned computer device in the foregoing embodiment, such as the hard disk or memory of the above-mentioned computer device. The above-mentioned computer-readable storage medium may also be an external storage device of the above-mentioned computer device, such as a plug-in hard disk equipped on the above-mentioned computer device, a smart media card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc.

[0117] It should be noted that in this text, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, such that a process, method, article or system comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or system. Without further limitation, an element defined by the statement "comprising one..." does not exclude the existence of additional identical elements in the process, method, article or system comprising such element.

[0118] The above description is only some preferred embodiments of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with technical features having similar functions (but not limited to) disclosed in the embodiments of the present disclosure.

Claims

1. A method for identifying abnormal electricity consumption based on a multimodal neural network model, characterized in that: include: Constructing a virtual object according to the power supply topology and the power consumption portrait corresponding to the meta-object in at least one meta-object to obtain a set of virtual objects, wherein the meta-object is a meta-granularity power consumption object in the power supply topology, and the virtual object is composed of at least two meta-objects that have an indirect topological relationship and correspond to similar power consumption behaviors; For each virtual object in the virtual object set, the following processing steps are performed: The target power consumption data corresponding to the virtual object is collected through the data repeater corresponding to the virtual object, wherein the data repeater includes: a data concentrator and a radio frequency transceiver, the data repeater is used to collect the power consumption data of at least two meta-objects corresponding to the virtual object in a wired manner, and the radio frequency transceiver is used to send the data to the server in a wireless manner; Performing data preprocessing on the target power consumption data to obtain preprocessed power consumption data; According to the data state of the pre-processed power consumption data, dynamically determine a fusion channel corresponding to the pre-processed power consumption data, wherein the fusion channel includes: a main channel and a secondary channel, and the signal-to-noise ratio corresponding to the main channel is greater than the signal-to-noise ratio of the secondary channel; In response to the server side receiving the pre-processed power usage data sent through the fusion channel, the server side identifies abnormal power usage on the pre-processed power usage data through a multimodal neural network model to generate a power usage identification result for the virtual object.

2. The method according to claim 1, characterized in that The step of constructing a virtual object according to the power supply topology structure and the power consumption portrait corresponding to the meta-object in at least one meta-object to obtain a virtual object set includes: For each meta-object in the at least one meta-object, the following portrait feature construction steps are performed: Extracting dynamic portrait attributes of the power consumption portrait corresponding to the meta-object to obtain a dynamic power consumption portrait, wherein the dynamic power consumption portrait represents the portrait attributes that change dynamically over time; Constructing a profile feature for the dynamic electricity consumption profile to generate a dynamic electricity consumption profile feature; According to the dynamic power consumption portrait features corresponding to the meta-objects, clustering the meta-objects in the at least one meta-object to obtain a meta-object group sequence; According to the meta-object group sequence, the following virtual object construction steps are performed: Determine the first meta-object group in the meta-object group sequence as the target meta-object group; Taking a unit increment as a step length and taking a target meta-object in the target meta-object group as a starting node, performing a reverse stepping traversal on the power supply topology structure to determine a node list corresponding to each target meta-object in the target meta-object group; Determine the node list intersection of each node list in the node list group corresponding to the target meta-object group; In response to the existence of a target node in the intersection of the node lists, and the node distance between the target node and the target meta-object farthest away in the target meta-object group is less than a preset node distance, constructing a virtual object according to the target meta-object group, wherein the target node is the node in the node list that is closest to each target meta-object in the target meta-object group; In response to the meta-object group sequence after the target meta-object group is removed being empty, ending the virtual object construction step; In response to the target node not existing in the intersection of the node lists, or the node distance between the target node and the target meta-object with the farthest distance in the target meta-object group is greater than or equal to a preset node distance, the target meta-object group is group-segmented to obtain at least one segmented meta-object group, and the at least one segmented meta-object group is added to the meta-object group sequence after the target meta-object group is removed; In response to the fact that the meta-object group sequence of the target meta-object group to be removed is not empty, the meta-object group sequence of the target meta-object group to be removed is used as the meta-object group sequence, and the virtual object construction step is performed again.

3. The method according to claim 2, characterized in that The collecting target power consumption data corresponding to the virtual object through the data repeater corresponding to the virtual object includes: Determine a communication link corresponding to the virtual object with a data repeater as a receiving endpoint, and obtain at least two communication links, wherein the data repeater corresponding to the virtual object is set on a target node corresponding to the virtual object, and the at least two communication links are links for data transmission between the target node corresponding to the virtual object and at least two meta-objects corresponding to the virtual object; Determining a delay value corresponding to each communication link in the at least one communication link; According to the delay value corresponding to the communication link, the target power consumption data corresponding to the virtual object is asynchronously collected through the at least one communication link.

4. The method according to claim 3, characterized in that: The target electricity consumption data is a matrix of indicator values ​​that is represented in matrix form and changes with the time dimension. The matrix dimension of the target electricity consumption data is N×T×9, where N is the number of meta-objects corresponding to the virtual object, and T is the length of the time scale. as well as The performing data preprocessing on the target power consumption data to obtain preprocessed power consumption data includes: The target electricity consumption data is split into a sequence of matrices to be processed, wherein the matrix dimension of the matrix to be processed is N×T×1; For each matrix to be processed in the sequence of matrices to be processed, the following fusion steps are performed: Performing precision conversion on the matrix to be processed to obtain a precision-converted matrix, wherein the precision of the matrix value corresponding to the precision-converted matrix is ​​less than the precision of the matrix value corresponding to the matrix to be processed; According to the precision converted matrix, a target curve set is obtained by fitting, wherein the number of target curves in the target curve set is N; Determine the benchmark value based on the target curve set; Taking the reference value as a baseline, performing incremental value conversion on the precision converted matrix to obtain a converted matrix; Merging the reference value as a header with the transformed matrix to obtain an updated matrix, wherein the matrix dimension of the updated matrix is ​​1×(N×T+1); The obtained updated matrix sets are combined to obtain the pre-processed power consumption data corresponding to the virtual object, wherein the matrix dimension of the pre-processed power consumption data is 1×(N×T+1)×9.

5. The method according to claim 4, characterized in that The dynamically determining a fusion channel corresponding to the preprocessed power usage data according to the data state of the preprocessed power usage data comprises: Acquire available channel state information from the server, wherein the available channel state information represents an available channel between the data repeater and the server, and a signal-to-noise ratio corresponding to the available channel; Determine the data zero ratio of the pre-processed power consumption data in each matrix dimension to obtain a data state; At least an available channel group is determined according to the data status and the available channel status information, wherein the available channel group includes: available channel A, available channel B and available channel C, wherein the signal-to-noise ratio corresponding to available channel A is greater than the signal-to-noise ratio corresponding to available channel B, the signal-to-noise ratio corresponding to available channel B is greater than the signal-to-noise ratio corresponding to channel C, available channel A and available channel B constitute a main channel, and available channel C constitutes a secondary channel.

6. The method according to claim 5, characterized in that Before the server receives the pre-processed power usage data sent through the fusion channel and identifies abnormal power usage on the pre-processed power usage data through a multimodal neural network model to generate a power usage identification result for the virtual object, the method further includes: Transmitting data corresponding to a zero percentage of low data in the pre-processed power usage data through an available channel A included in the main channel; Transmitting data corresponding to a zero percentage of high data in the pre-processed power usage data through an available channel B included in the main channel; The reference value in the pre-processed power usage data is transmitted through the available channel C included in the secondary channel.

7. The method according to claim 6, characterized in that The multimodal neural network model includes: 3 feature extraction blocks, a feature fusion block and a power consumption result classifier, the 3 feature extraction blocks are used to extract features from different modal data, the feature extraction blocks include: 3 feature extraction networks, and parameters are shared between the 3 feature extraction networks; and In response to the server receiving the pre-processed power usage data sent through the fusion channel, the server performs abnormal power usage identification on the pre-processed power usage data through a multimodal neural network model to generate a power usage identification result for the virtual object, including: By means of the three feature extraction blocks, parallel feature extraction is performed on different modal data in the preprocessed power consumption data in parallel to obtain a feature set of data to be fused; By means of the feature fusion block, feature fusion is performed on the feature set of the data to be fused to obtain fused features; According to the fusion feature and the power consumption result classifier, a power consumption recognition result for the virtual object is generated.

8. An abnormal power consumption identification device based on a multimodal neural network model, characterized in that: include: A virtual object construction unit is configured to construct a virtual object according to the power supply topology structure and the power consumption portrait corresponding to the meta-object in at least one meta-object, so as to obtain a set of virtual objects, wherein the meta-object is a meta-granularity power consumption object in the power supply topology structure, and the virtual object is composed of at least two meta-objects that have an indirect topological relationship and correspond to similar power consumption behaviors; The execution unit is configured to perform the following processing steps for each virtual object in the virtual object set: collecting target power consumption data corresponding to the virtual object through a data repeater corresponding to the virtual object, wherein the data repeater includes: a data concentrator and a radio frequency transceiver, the data repeater is used to collect power consumption data of at least two meta-objects corresponding to the virtual object in a wired manner, and the radio frequency transceiver is used to send the data to a server side in a wireless manner; performing data preprocessing on the target power consumption data to obtain preprocessed power consumption data; dynamically determining a fusion channel corresponding to the preprocessed power consumption data according to a data state of the preprocessed power consumption data, wherein the fusion channel includes: a main channel and a secondary channel, and the signal-to-noise ratio corresponding to the main channel is greater than the signal-to-noise ratio of the secondary channel; in response to the server side receiving the preprocessed power consumption data sent through the fusion channel, the server side performs abnormal power consumption identification on the preprocessed power consumption data through a multimodal neural network model to generate a power consumption identification result for the virtual object.

9. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable medium, characterized in that A computer program is stored thereon, wherein when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images