State grid marketing business auditing method and system based on multi-source monitoring

Abstract

The application discloses a state grid marketing business auditing method and system based on multi-source monitoring, comprising S1: collecting multi-source monitoring data of automation and video monitoring, and performing key entity correlation preprocessing to obtain business behavior data; S2: obtaining an electricity mutation index from the business behavior data, obtaining an event sequence vector based on the business behavior data, constructing an event time sequence convolution network model based on the event sequence vector, and outputting running event correlation degree of the business behavior data based on the event time sequence convolution network model; S3: based on the electricity mutation index and the running event correlation degree, outputting an initial risk value through construction of a collaborative risk fusion network based on sparse excitation, and performing risk conduction and traceability analysis based on the initial risk value to obtain a business dynamic risk portrait of an auditing object; and S4: based on the business dynamic risk portrait of the auditing object, realizing marketing business auditing through intelligent auditing task distribution and accurate auditing strategy recommendation.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method and system for auditing the marketing business of the State Grid Corporation of China based on multi-source monitoring. Background Technology

[0002] Auditing State Grid's marketing operations is a crucial link in ensuring accurate electricity bill collection, maintaining fair power supply and consumption, and preventing operational risks. In the early stages of auditing, it primarily relied on manual spot checks and experience-based judgment. Auditors conducted verifications by reviewing paper work orders and comparing electricity bill ledgers, which was inefficient, had limited coverage, heavily depended on individual experience, and lacked standardized procedures.

[0003] With the advancement of information technology construction, we have entered a stage based on traditional information systems. Various marketing business systems and electricity consumption information collection systems have been gradually launched, realizing the electronic recording and storage of business data. Auditing work has begun to be able to perform some simple rule judgments by querying the database (such as threshold alarms for sudden increases or decreases in electricity consumption). This method has improved efficiency to a certain extent, but it is essentially still a reactive data retrieval after the fact. Moreover, each system operates independently, forming data silos and making it impossible to perform correlation analysis.

[0004] Currently, traditional auditing methods for power grid marketing operations rely on a single dimension for risk identification, often depending on threshold judgments of single indicators (such as sudden changes in electricity consumption). This fails to correlate abnormal electricity usage with abnormal events in the business process, such as records of violations at specific times. Furthermore, risk assessment only makes static judgments on individual audit targets, neglecting the transmission effect of risks in business networks such as customer relationships and financial transaction chains. This makes it impossible to identify potential group risks or risk sources. Moreover, the assignment of audit tasks is mostly based on simple rules or random checks, failing to achieve precise prioritization based on dynamic risks. In determining audit strategies, there is also a heavy reliance on the personal experience of auditors, lacking data-driven, targeted strategy recommendations. This leads to suboptimal allocation of audit resources and unclear audit priorities, affecting overall audit efficiency and accuracy. Therefore, this paper proposes a State Grid marketing operations auditing method based on multi-source monitoring. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution:

[0006] The State Grid's marketing business auditing method based on multi-source monitoring includes:

[0007] S1: Collect multi-source monitoring data from automated and video surveillance systems, perform key entity association preprocessing, and obtain business behavior data;

[0008] S2: Obtain the power fluctuation index from the business behavior data, obtain the event sequence vector based on the business behavior data, construct the event temporal convolutional network model based on the event sequence vector, and output the running event correlation degree of the business behavior data based on the event temporal convolutional network model;

[0009] S3: Based on the power fluctuation index and the correlation of operational events, an initial risk value is output by constructing a collaborative risk fusion network based on sparse incentives. Based on the initial risk value, risk transmission and source tracing analysis are performed to obtain a dynamic risk profile of the audit object.

[0010] The collaborative risk fusion network based on sparse excitation includes a temporal convolutional layer, an optimized fully connected layer with an adaptive feature focusing mechanism, and an output layer.

[0011] S4: Based on the dynamic risk profile of the audit object, marketing business audit is realized through intelligent assignment of audit tasks and precise recommendation of audit strategies.

[0012] The process of obtaining the power fluctuation index is as follows:

[0013] Analyze the electricity consumption sequence in the business behavior data and set a time window. ;

[0014] For each time window The rate of change of electricity within the calculated window is defined as the initial electricity fluctuation index when the rate of change exceeds a preset threshold. ;

[0015] Then, the duration factor of charge mutation is introduced. Based on mutation duration factor The initial power fluctuation index is corrected to obtain the final power fluctuation index. : , where a is the duration weighting coefficient.

[0016] Duration factor of electrical fluctuation The acquisition process is as follows:

[0017] Acquire instantaneous acceleration of battery power over time. The start and end points of the abrupt change state are determined based on the acceleration threshold. For the first time, it exceeded the positive threshold calculated based on historical data. When a mutation event begins, that moment is recorded as the starting point. When acceleration It fell back and stabilized at another negative threshold derived from historical data. The mutation state ends when the following time is reached; this time is recorded as the termination point. Then the mutation duration factor The calculation formula is: ;

[0018] in, The duration factor of the charge change. It is the moment when the mutation state ends. It is the moment when the mutation state begins.

[0019] The process of obtaining the event sequence vector is as follows:

[0020] Based on the business behavior data to be analyzed The time of occurrence As the absolute center reference point, defined by A continuous-time analysis window centered on [the target] with a total time span of 2L. ;

[0021] Based on time analysis window The system scans and categorizes business operation events occurring within the time analysis window. After the statistics are completed, the results are organized into a fixed data structure, arranged in ascending order according to the event type number, forming a vector containing K specific values, which is defined as the event sequence vector. .

[0022] The construction process of the event-series convolutional network model is as follows:

[0023] Based on event sequence vectors , event sequence vector The total time window is divided into For each time period, a K-dimensional counting vector is generated, forming... The sequence matrix is ​​used as the input form of the temporal convolutional network model;

[0024] Then, an event-time convolutional layer is constructed, using one-dimensional dilated convolution. For an input sequence X and a convolution kernel W, the convolution output at time t is: ;

[0025] in, This represents a convolution operation with an inflation factor of d. Where is the kernel size and d is the dilation factor. Extracted temporal features;

[0026] Then, an event-time convolutional layer is constructed using one-dimensional dilated convolution, and an optimized fully connected layer and output layer with an adaptive feature focusing mechanism are built to extract the temporal features from the event-time convolutional layer. After flattening, the data is fed into an optimized fully connected layer with an adaptive feature focusing mechanism for synthesis. The weighted event correlation is then output through the output layer. .

[0027] The construction process of the optimized fully connected layer with the introduction of a feature adaptive focusing mechanism is as follows:

[0028] Temporal characteristics Perform nonlinear transformation and introduce Activation function, output and Importance weight vectors with identical dimensions ;

[0029] Then, based on the importance weight vector α, the original flattened features are... Perform recalibration: , This indicates the element-wise multiplication operation. The optimized time series features;

[0030] For the optimized time series features By using a linear transformation and combining it with the Sigmoid activation function, it is mapped to a scalar value between 0 and 1: ,in, It optimizes the output features after ReLU activation of the fully connected layer. The weights of the fully connected layer, This is a bias term for the fully connected layer, used to optimize the construction of the fully connected layer.

[0031] The construction process of the collaborative risk fusion network based on sparse incentives is as follows:

[0032] By concatenating the power fluctuation index and the correlation of operational events, a collaborative sensing vector is obtained. Then, a sparse excitation transformation is applied to the collaborative sensing vector H, introducing a threshold-constrained Tanh function to obtain the sparsely excited business collaborative feature vector. ;

[0033] By slicing, the business collaboration feature vector is... The middle corresponds to the original Parts and corresponding to the original The partial separation yields the results of the electricity risk characteristics after competitive transformation. The result of the event risk characteristics after competitive transformation ,based on and Obtain the extreme values ​​of power risk characteristics Extreme values ​​of event risk characteristics ,based on and Define risk decision rules to obtain a collaborative risk fusion network, and output the final initial risk value. .

[0034] based on and The defined risk decision rule is expressed as:

[0035] ;

[0036] This represents the electricity fluctuation index of the i-th audit object. This represents the correlation degree of the runtime events of the i-th audit object. This represents the maximum value between the power fluctuation index and the correlation between operational events. This represents the initial risk value of the i-th audit object.

[0037] The process of obtaining a dynamic business risk profile is as follows:

[0038] A static analysis network is constructed based on an attribute graph network. The nodes of the static analysis network are defined, and each audit object is defined as a network node. And establish edges based on business relationships Weights are set based on the strength of the association. ;

[0039] Risk simulation using a linear threshold is performed on the static analysis network to obtain the final risk representation of the i-th node. , corresponding to the i-th network node Clearly define network nodes With neighboring network nodes Relationships, obtain the risk contribution of all neighboring network nodes: ;in, For network nodes Risk contribution Indicates the risk transmission attenuation factor. This represents the risk weight from node j to node i, ultimately resulting in a dynamic business risk profile of the audited entity. .

[0040] The State Grid Marketing Business Audit System, based on multi-source monitoring, includes:

[0041] Data acquisition module: Collects multi-source monitoring data from automated and video surveillance systems, performs key entity association preprocessing, and obtains business behavior data;

[0042] Association Acquisition Module: Obtains the power fluctuation index from business behavior data, obtains event sequence vectors based on business behavior data, constructs an event temporal convolutional network model based on the event sequence vectors, and outputs the operational event correlation degree of business behavior data based on the event temporal convolutional network model;

[0043] Risk Decision Module: Based on the power fluctuation index and the correlation of operational events, an initial risk value is output by constructing a collaborative risk fusion network based on sparse incentives. Based on the initial risk value, risk transmission and source tracing analysis are performed to obtain a dynamic risk profile of the audit object.

[0044] The collaborative risk fusion network based on sparse excitation includes a temporal convolutional layer, an optimized fully connected layer with an adaptive feature focusing mechanism, and an output layer.

[0045] Business Audit Module: Based on the dynamic risk profile of the audit object, marketing business audit is realized through intelligent assignment of audit tasks and precise recommendation of audit strategies.

[0046] The present invention has the following beneficial effects:

[0047] First, a dual risk monitoring system covering abnormal electricity consumption behavior and abnormal business processes is constructed by calculating the electricity fluctuation index and the correlation of operational events. Risk transmission is carried out based on attribute graph network. The risk propagation process is simulated by formula, the impact of edge weights on risk transmission is quantified, and the risk contribution is calculated to realize the visual analysis of risk tracing. At the business application level, intelligent allocation of audit resources and accurate recommendation of strategies are realized based on dynamic risk profiles.

[0048] Secondly, by setting risk thresholds to classify risk levels, audit resources are prioritized for high-risk areas. Based on the risk contribution analysis results, differentiated audit strategies are developed for equipment-related risks and business operation risks. Finally, by establishing a complete closed loop from data collection and risk analysis to decision support, the efficiency and accuracy of marketing business audits are improved.

[0049] Finally, by acquiring dynamic risk profiles of audit targets, not only are risk assessment results provided for individual audit targets, but also, through the accumulation and analysis of massive amounts of audit data, macro-level insights into regional marketing conditions, customer group characteristics, and business hotspots are formed. The multi-dimensional statistical data and visualization analysis results generated by the solution can clearly present the risk distribution patterns of different regions and different customer types, providing data support for the optimal allocation of marketing resources. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the steps of the State Grid marketing business audit method based on multi-source monitoring proposed in this invention.

[0051] Figure 2 This is a system block diagram of the State Grid Marketing Business Audit System based on multi-source monitoring proposed in this invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] like Figure 1 As shown, the State Grid marketing business auditing method based on multi-source monitoring proposed in this invention includes:

[0055] S1: Collect multi-source monitoring data from automated and video surveillance systems, perform key entity association preprocessing, and obtain business behavior data;

[0056] First, multi-source monitoring data acquisition is performed. For automated data, focusing on the power marketing scenario, real-time power consumption and equipment operation status indicators are collected at the minute-level time granularity through the DL / T645 communication protocol interface. At the same time, video streams are collected at a frame rate of 25 frames / second for video monitoring equipment deployed in business halls, power consumption sites, and metering device installation points, generating a video frame sequence V(t), where t is the timestamp.

[0057] After the data collection is completed, key entity association preprocessing is carried out. Key entities include electricity customers and electrical equipment.

[0058] For multi-source monitoring data, the data ownership is determined by the customer number C and the device number D identification fields. For video surveillance data, the YOLO object detection algorithm is used to identify people and equipment in the picture, and then the identity of business personnel is associated through facial recognition. After associating the device with the device number D through the device appearance features, the entity association relationship is established.

[0059] For customer C (customer number C), retrieve a series of smart meter data, video frames of their business transactions at the service center, and business system operation records, and correlate them, specifically:

[0060] For automated data, which is inherently structured, the system directly reads the two identifier fields, Customer C and Device D, which are inherent in the data record, to clearly know which customer the data belongs to and which device generated it.

[0061] For video surveillance data, which is unstructured image, the system first uses the YOLO object detection algorithm to analyze each frame of video footage, identify the people and devices contained in the footage, and then uses facial recognition technology to compare the facial features of the identified people with the business personnel database and associate them with customer C.

[0062] Then, for the identified devices (such as electricity meters), image recognition technology is used to read their appearance features (such as nameplates) and associate them with the specific device number D. The system uses customer C and device D as the core association key. When it is necessary to analyze a specific customer C, the system will perform the following operations:

[0063] Retrieve all automated data (electricity E) and device operating status identifiers (S) generated by devices (such as smart meters) under the customer's name;

[0064] Search the video data for all video frames that are associated with this customer C or device D after entity identification.

[0065] The system retrieves all relevant business operation records for this customer from the business system. Using timestamp t as a reference, the system aligns the retrieved data and bundles together automation data, video frames, and business records that occurred within the same time window and involved the same group of entities, customer C and device D, ultimately forming business behavior data. This data integrates the temporal information of the behavior of various entities from multiple sources.

[0066] S2: Obtain the power fluctuation index from the business behavior data, obtain the event sequence vector based on the business behavior data, construct the event temporal convolutional network model based on the event sequence vector, and output the running event correlation degree of the business behavior data based on the event temporal convolutional network model;

[0067] The process of obtaining the power fluctuation index is as follows:

[0068] First, we analyze the electricity consumption sequence E(C,t) in the business behavior data, where C represents the electricity customer and t is the timestamp. To capture sudden changes in electricity consumption, a time window is set. ;

[0069] Specifically, time window Based on the business scenario, a 1-hour time window is selected for residential customers and a 15-minute time window is selected for large industrial customers to improve monitoring accuracy.

[0070] For each time window, the rate of change of electricity within the window is calculated. When the rate of change exceeds a preset threshold (1), it is defined as a sudden change in electricity. At this time, the initial electricity sudden change index is... The calculation formula is:

[0071] ;

[0072] in, This refers to the amount of electricity consumed by customer C at time t. It is customer C in time Battery life at that time;

[0073] For example, a residential customer's electricity consumption at time t is 10 kWh. ( =1 hour) The electricity consumption was 2kWh. Substituting it into the formula, we can get EI=4, which indicates that the customer's electricity consumption has changed significantly.

[0074] Since the duration of power fluctuations can also reflect the severity of the risk, a power fluctuation duration factor is introduced. , used to correct the initial charge mutation index;

[0075] Furthermore, mutation duration factor The acquisition process is as follows:

[0076] First, we calculate the instantaneous acceleration of the energy consumption time series. Since the energy consumption data is a sequence that changes over time, we need to calculate the instantaneous rate of change of energy consumption at each moment, i.e., the speed at which the energy consumption changes. However, to more sensitively capture the start and end points of change, we also calculate the change in the rate of change itself, i.e., the instantaneous acceleration. The formula for calculating is:

[0077] ;

[0078] in, This represents the instantaneous acceleration of the change in charge at time t. Indicates the rate of change. The instantaneous rate of change of charge at time t is expressed by the following formula: , It is the charge value at time t. It is the instantaneous rate of change of charge at the previous time t−1;

[0079] The start and end points of abrupt changes are determined based on acceleration thresholds. For the first time, it exceeded a positive threshold derived from historical data. At that time, a mutation event is considered to have begun, and that moment is recorded as the starting point. Subsequently, even though the battery level remained high, the growth had stagnated (acceleration dropped to near zero) and even began to decelerate. It fell back and stabilized at another negative threshold derived from historical data. When the mutation state ends, this moment is recorded as the termination point. Then the mutation duration factor The calculation formula is:

[0080] ;

[0081] in, This is the duration factor of the electrical fluctuation. The end time of the abrupt change state is determined by the above acceleration criterion. It is the moment when the abrupt change begins, determined by the above acceleration criterion;

[0082] Then based on the mutation duration factor The initial power fluctuation index is corrected to obtain the final power fluctuation index. :

[0083]

[0084] Where 'a' is the duration weighting coefficient;

[0085] The process of obtaining the event sequence vector is as follows:

[0086] First, define a time frame, using the business behavior data you want to analyze at present. The time of occurrence Using the absolute center reference point, then tracing forward a length along the time axis, and simultaneously extending backward a length, both of which are L, a value of L is defined. A continuous-time analysis window centered on [the target] with a total time span of 2L. ;

[0087] Specifically, the setting of this window ensures that the scope of analysis can simultaneously cover the preparatory state before the business activity occurs and the subsequent impact after it occurs;

[0088] Then based on the time analysis window The system scans and categorizes business operation events occurring within the time analysis window. After the statistics are completed, the results are organized into a fixed data structure, arranged in ascending order according to the event type number, forming a vector containing K specific values. This vector is defined as the input event sequence vector for business behavior B. ;

[0089] The process of constructing an event-series convolutional network model is as follows:

[0090] Receive the above K-dimensional event sequence vector To capture temporal information more precisely, the event sequence vector is... The total time window is divided into For each time period, a K-dimensional counting vector is generated, forming... The sequence matrix is ​​used as the input form of the temporal convolutional network model;

[0091] Then, an event-time convolutional layer is constructed, using one-dimensional dilated convolution. For an input sequence X and a convolution kernel W, the convolution output at time t is:

[0092] ;

[0093] in, This represents a convolution operation with an inflation factor of d. Where is the kernel size, and d is the dilation factor, which increases exponentially with increasing network depth. Extracted temporal features;

[0094] Specifically, by outputting the convolutional output at time t through this formula, high-level convolutional neurons can observe the input sequence with an exponentially expanded receptive field, thereby capturing long-term dependencies throughout the entire time window with fewer network layers, a significant advantage over ordinary convolutional or recurrent networks.

[0095] Then, we construct and optimize the fully connected layer and the output layer to extract the temporal features from the event-series convolutional layer. After flattening, the features are fed into the optimized fully connected layer for synthesis. Traditional temporal convolutional networks typically flatten the features and feed them directly into the fully connected layer. However, the optimized fully connected layer of the event-temporal convolutional network model introduces a specific multi-level, cross-scale feature interaction and adaptive weighting feature adaptive focusing mechanism between the temporal convolutional network layer and the final output layer. This enables the model to synthesize temporal features more intelligently, rather than simply using linear weighting.

[0096] The construction process of the optimized fully connected layer with the introduction of a feature adaptive focusing mechanism is as follows:

[0097] A small weight generation network is constructed to analyze these features. This small weight generation network generates weights from the feature vectors. Perform nonlinear transformation and introduce The activation function aims to evaluate the relative importance of each feature element in the vector. The network ultimately outputs a value that is related to the value of each feature element in the vector. Importance weight vectors with identical dimensions ;

[0098] Specifically, this small weight generation network maps original features to attention weights by inputting features, performing nonlinear transformations, and then normalizing them using Softmax. The resulting weight vector α is used to recalibrate the original features, allowing subsequent network layers to focus on key features.

[0099] After obtaining the importance weight vector α, the original flattened features are... To perform recalibration, the formula is:

[0100] , This represents the element-wise multiplication operation;

[0101] Specifically, this step focuses on the concrete implementation of the function, amplifying the key features that contribute more to the current task (determining the correlation between business behavior and events), while suppressing irrelevant or redundant features. The result is a refined and optimized time-series feature. ;

[0102] The optimized time series features The features are fed into a nonlinear combination process to capture the complex interactions between them. Through a linear transformation combined with the Sigmoid activation function, the high-dimensional features are mapped to a scalar value between 0 and 1. The process can be represented as follows:

[0103] ,in, These are the output features after ReLU activation of the fully connected layer. The weights of the fully connected layer, This is the bias term for the fully connected layer, completing the optimization of the fully connected layer construction;

[0104] Then based on the output of the fully connected layer Enter the output layer to output the execution of event correlation. , is represented as: ;

[0105] Here, σ is the Sigmoid function, which ensures the correlation of the output weights running events. It is a probability value that can be interpreted as the strength of the association;

[0106] Specifically, this involves transforming raw time-series data into a precise, interpretable, and dynamic correlation of runtime events. This provides core feature inputs for building high-quality dynamic risk profiles of businesses, and is key to realizing the transformation of State Grid's marketing business audit from manual investigation and experience-driven to automatic early warning and data-driven intelligent transformation.

[0107] S3: Based on the power fluctuation index and the correlation of operating events, a collaborative risk fusion network based on sparse incentives is constructed, and risk decision rules are set to output the initial risk value. Based on the initial risk value, risk transmission and source tracing analysis are performed to obtain the business dynamic risk profile of the audit object.

[0108] The construction process of the collaborative risk fusion network based on sparse incentives is as follows:

[0109] First, the power surge index and the correlation degree of the operation event are concatenated to obtain the collaborative sensing vector. ;

[0110] Then, a sparse activation transformation is applied to the co-perception vector H. Specifically, an activation function that produces sparsity, namely the threshold-constrained Tanh function, is used. The goal is to make most neurons output zero or close to zero, with only a few features most critical to the current risk situation being strongly activated. This simulates the efficient information processing mode in the brain where most neurons are at rest and a few key neurons are firing. The formula is expressed as:

[0111] ;

[0112] in, This represents the Tanh function with a threshold constraint. For sparse weights, For sparse bias, This is the business collaboration feature vector after sparse excitation. Each element value is between -1 and 1. Values ​​with absolute values ​​close to 1 represent significant features that are strongly activated or suppressed, while values ​​with absolute values ​​close to 0 are considered insignificant features that have been sparsified.

[0113] From business collaboration feature vectors We extract quantifiable evidence to determine whether electricity risk characteristics or event risk characteristics dominated the competition in this fusion.

[0114] By slicing, the business collaboration feature vector is... The middle corresponds to the original Parts and corresponding to the original Separate the parts, let the vector part The dimension is , The dimension is The subvector consisting of the first d elements This is the result of the electricity risk characteristics after competitive transformation. ;

[0115] Business collaboration feature vector From the middle The subvector consisting of elements up to the end This is the result of the event risk characteristics undergoing competitive transformation. ;

[0116] Calculate the results of the two competing transformations after the above separation respectively. and The maximum absolute value of all elements represents the strongest degree to which this type of risk characteristic is activated or suppressed after competition. The calculation formula is as follows:

[0117] , ;

[0118] in, This represents the extreme value of the power risk characteristic. The closer the value is to 1, the more strongly one or more dimensions of the power risk are activated (positive value close to 1) or suppressed (negative value close to -1), indicating that this type of risk signal is very significant in the current situation.

[0119] This represents the extreme value of the event risk characteristics. The closer the value is to 1, the more strongly one or more dimensions of the event risk are activated (positive value close to 1) or suppressed (negative value close to -1), indicating that this type of risk signal is very significant in the current context.

[0120] get and Then, risk decision-making rules are defined based on relative magnitudes to obtain a collaborative risk fusion network, which outputs the final initial risk value. The rule is as follows:

[0121] ;

[0122] This represents the electricity fluctuation index of the i-th audit object. This represents the correlation degree of the runtime events of the i-th audit object. This represents the initial risk value of the i-th audit object;

[0123] Specifically, when This indicates that after feature competition, the electricity consumption risk is more significant than the event risk. Therefore, the system determines that the current risk is mainly driven by abnormal electricity consumption behavior, hence the initial risk value... Directly using the power mutation index ;

[0124] when When the risk level is high, it indicates that the event risk is more significant. The system determines that the risk mainly originates from abnormal events in the business process; therefore, the initial risk value is... Directly using event correlation ;

[0125] when When the value is equal to the significance of the two types of risks, the system adopts a conservative strategy, selecting the larger of the two original risk values ​​as the initial risk value. ,Right now , ;

[0126] Then based on the initial risk value The process of conducting risk transmission and source tracing analysis is as follows:

[0127] A static analysis network is constructed based on attribute graph networks, defining network nodes, with each audit object defined as a network node. And establish edges based on business relationships (such as fund transfers). Weights are set based on the strength of the association (such as the normalization of transfer amounts). ;

[0128] Specifically, the process of defining network nodes is to create a mapping of real-world business entities in the digital world, with each node becoming a carrier and transmission point of risk energy;

[0129] Then, the risk transmission is simulated on the network, and business data is analyzed to establish connections between nodes;

[0130] For example, by analyzing fund flow records, network nodes can be discovered. Corresponding clients and network nodes If there are frequent fund transfers between customers, then they are considered to have a business relationship, and an edge is established between the two nodes, denoted as . ;

[0131] The existence of an edge only indicates a connection, but the strength of the connection varies. Therefore, each edge must be assigned a weight. To quantify this intensity, for example, the weight can be set based on the total amount of transfers between two customers within a certain period, and then normalized so that the weight value falls between 0 and 1. The larger the value, the closer the business ties between the two nodes are, and the higher the potential likelihood and intensity of risk transmission.

[0132] After the static network is constructed, the dynamic propagation process of risk is simulated, and the ultimate source of risk is analyzed. A linear threshold is used for risk simulation, and the calculation formula is as follows:

[0133] ;

[0134] in, This represents the final risk representation of the i-th audit object. This represents the initial risk value of the i-th audit object. This represents the initial risk value of the j-th audit object. Indicates the risk transmission attenuation factor. Partially represents the sum of risks propagated from all neighboring network nodes to node i. This represents the risk weight from node j to node i;

[0135] The network node corresponding to the i-th audit object Clearly identify all network nodes How much of the ultimate risk originates from one of its neighboring network nodes? That is, to determine network nodes With neighboring network nodes Based on the relationship, the risk contribution of all neighboring network nodes is obtained, and the calculation formula is as follows: ;

[0136] in, For nodes The risk contribution level represents the comprehensive risk level of the audited object itself and its local network environment, and is the direct basis for risk ranking and early warning.

[0137] Finally, by combining the final risk representation with the risk contribution of source tracing and diagnosis, a dynamic business risk profile of the audited entity is obtained. It consists of a single tuple, represented as .

[0138] S4: Based on the dynamic risk profile of the audit object, marketing business audit is realized through intelligent assignment of audit tasks and precise recommendation of audit strategies;

[0139] Obtain dynamic risk profiles of audit targets. After that, the audit tasks are first intelligently assigned based on the final risk assessment. The audit targets are classified into risk levels, and risk thresholds are set. The final risk will be represented Above the risk threshold Those classified as high-risk, or below or equal to the risk threshold The risk level is classified as low risk. Different risk levels correspond to different audit task priorities. High-risk tasks have the highest priority and will be assigned to auditors first.

[0140] At the same time, the risk contribution of neighboring network nodes is also considered. If the risk revealed is mainly from related parties in a certain business process, then the focus of the audit should be clearly defined when assigning tasks. For example, if a neighbor with a high risk contribution is a certain business processing node, the task will focus on auditing the business processing process.

[0141] Next, precise recommendations for audit strategies are made, targeting different ultimate risk representations. and risk contribution Risk sources, matching corresponding audit strategies, and final risk representation Above the risk threshold And the risk contribution of neighboring network nodes The audit targets that indicate risks originate from electrical equipment, and a strategy of on-site inspection of the parameters and operating status of the electrical equipment is recommended.

[0142] For the final risk representation Below the risk threshold And the risk contribution of neighboring network nodes For audit targets that indicate risks originating from business operations, strategies such as reviewing business processing records and retrieving relevant video surveillance are recommended to ensure that the most effective strategies are used for each audit task, thereby achieving high efficiency and accuracy in marketing business audits.

[0143] Example 2

[0144] like Figure 2 As shown, the State Grid Marketing Business Audit System based on multi-source monitoring includes:

[0145] Data acquisition module: Collects multi-source monitoring data from automated and video surveillance systems, performs key entity association preprocessing, and obtains business behavior data;

[0146] Association Acquisition Module: Obtains the power fluctuation index from business behavior data, obtains event sequence vectors based on business behavior data, constructs an event temporal convolutional network model based on the event sequence vectors, and outputs the operational event correlation degree of business behavior data based on the event temporal convolutional network model;

[0147] Risk Decision Module: Based on the power fluctuation index and the correlation of operational events, an initial risk value is output by constructing a collaborative risk fusion network based on sparse incentives. Based on the initial risk value, risk transmission and source tracing analysis are performed to obtain a dynamic risk profile of the audit object.

[0148] The collaborative risk fusion network based on sparse excitation includes a temporal convolutional layer, an optimized fully connected layer with an adaptive feature focusing mechanism, and an output layer.

[0149] Business Audit Module: Based on the dynamic risk profile of the audit object, marketing business audit is realized through intelligent assignment of audit tasks and precise recommendation of audit strategies.

[0150] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0151] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0152] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A State Grid marketing business audit method based on multi-source monitoring, characterized in that, include: S1: Collect multi-source monitoring data from automated and video surveillance systems, perform key entity association preprocessing, and obtain business behavior data; S2: Obtain the power fluctuation index from the business behavior data, obtain the event sequence vector based on the business behavior data, construct the event temporal convolutional network model based on the event sequence vector, and output the running event correlation degree of the business behavior data based on the event temporal convolutional network model; S3: Based on the power fluctuation index and the correlation of operational events, an initial risk value is output by constructing a collaborative risk fusion network based on sparse incentives. Based on the initial risk value, risk transmission and source tracing analysis are performed to obtain a dynamic risk profile of the audit object. The construction process of the collaborative risk fusion network based on sparse incentives is as follows: By concatenating the power fluctuation index and the correlation of operational events, a collaborative sensing vector is obtained. Then, a sparse excitation transformation is applied to the collaborative sensing vector H, introducing a threshold-constrained Tanh function to obtain the sparsely excited business collaborative feature vector. ; By slicing, the business collaboration feature vector is... The middle corresponds to the original Parts and corresponding to the original The partial separation yields the results of the electricity risk characteristics after competitive transformation. The result of the event risk characteristics after competitive transformation ,based on and Obtain the extreme values ​​of power risk characteristics Extreme values ​​of event risk characteristics ,based on and Define risk decision rules to obtain a collaborative risk fusion network, and output the final initial risk value. ; The collaborative risk fusion network based on sparse excitation includes a temporal convolutional layer, an optimized fully connected layer with an adaptive feature focusing mechanism, and an output layer. S4: Based on the dynamic risk profile of the audit object, marketing business audit is realized through intelligent assignment of audit tasks and precise recommendation of audit strategies.

2. The State Grid marketing business auditing method based on multi-source monitoring according to claim 1, characterized in that, The process of obtaining the power fluctuation index is as follows: Analyze the electricity consumption sequence in the business behavior data and set a time window. ; For each time window The rate of change of electricity within the calculated window is defined as the initial electricity fluctuation index when the rate of change exceeds a preset threshold. ; Then, the duration factor of energy fluctuation is introduced. Based on mutation duration factor The initial charge fluctuation index is corrected to obtain the final charge fluctuation index. : , where a is the duration weighting coefficient.

3. The State Grid marketing business auditing method based on multi-source monitoring according to claim 2, characterized in that, Duration factor of electrical fluctuation The acquisition process is as follows: Acquire instantaneous acceleration of battery power over time. The start and end points of the abrupt change state are determined based on the acceleration threshold. For the first time, it exceeded the positive threshold calculated based on historical data. When a mutation event begins, that moment is recorded as the starting point. When acceleration It fell back and stabilized at another negative threshold derived from historical data. The mutation state ends when the following time is reached; this time is recorded as the termination point. Then the mutation duration factor The calculation formula is: ; in, The duration factor of the charge change. It is the moment when the mutation state ends. It is the moment when the mutation state begins.

4. The State Grid marketing business auditing method based on multi-source monitoring according to claim 3, characterized in that, The process of obtaining the event sequence vector is as follows: Based on the business behavior data to be analyzed The time of occurrence As the absolute center reference point, defined by A continuous-time analysis window centered on [the target] with a total time span of 2L. ; Based on time analysis window The system scans and categorizes business operation events occurring within the time analysis window. After the statistics are completed, the results are organized into a fixed data structure, arranged in ascending order according to the event type number, forming a vector containing K specific values, which is defined as the event sequence vector. .

5. The State Grid marketing business auditing method based on multi-source monitoring according to claim 1, characterized in that, The construction process of the event-series convolutional network model is as follows: Based on event sequence vectors , event sequence vector The total time window is divided into For each time period, a K-dimensional counting vector is generated, forming... The sequence matrix is ​​used as the input form of the temporal convolutional network model; Then, an event-time convolutional layer is constructed, using one-dimensional dilated convolution. For an input sequence X and a convolution kernel W, the convolution output at time t is: ; in, This represents a convolution operation with an expansion factor of d. Where is the kernel size and d is the dilation factor. Extracted temporal features; Then, an event-time convolutional layer is constructed using one-dimensional dilated convolution, and an optimized fully connected layer and output layer with an adaptive feature focusing mechanism are built to extract the temporal features from the event-time convolutional layer. After flattening, the data is fed into an optimized fully connected layer with an adaptive feature focusing mechanism for synthesis. The weighted event correlation is then output through the output layer. .

6. The State Grid marketing business auditing method based on multi-source monitoring according to claim 5, characterized in that it introduces... The construction process of the optimized fully connected layer using the feature adaptive focusing mechanism is as follows: Temporal characteristics Perform nonlinear transformation and introduce Activation function, output and Importance weight vectors with identical dimensions ; Then, based on the importance weight vector α, the original flattened features are... Perform recalibration: , This indicates the element-wise multiplication operation. The optimized time series features; For the optimized time series features By mapping the weighted running event correlation degree to a scalar value between 0 and 1 through a linear transformation and combined with the Sigmoid activation function, the output is determined. Complete the construction of the optimized fully connected layer.

7. The State Grid marketing business auditing method based on multi-source monitoring according to claim 1, characterized in that, based on and The defined risk decision rule is expressed as: ； This represents the electricity fluctuation index of the i-th audit object. This represents the correlation degree of the runtime events of the i-th audit object. This represents the maximum value between the power fluctuation index and the correlation between operational events. This represents the initial risk value of the i-th audit object.

8. The State Grid marketing business auditing method based on multi-source monitoring according to claim 7, characterized in that, The process of obtaining a dynamic business risk profile is as follows: A static analysis network is constructed based on an attribute graph network. The nodes of the static analysis network are defined, and each audit object is defined as a network node. And establish edges based on business relationships Weights are set based on the strength of the association. ; Risk simulation using a linear threshold is performed on the static analysis network to obtain the final risk representation of the i-th node. , corresponding to the i-th network node Clearly define network nodes With neighboring network nodes Relationships, obtain the risk contribution of all neighboring network nodes: ;in, For nodes Risk contribution Indicates the risk transmission attenuation factor. This represents the risk weight from node j to node i, ultimately resulting in a dynamic business risk profile of the audited entity. .

9. A State Grid marketing business audit system based on multi-source monitoring, implemented according to any one of claims 1-8, characterized in that, include: Data acquisition module: Collects multi-source monitoring data from automated and video surveillance systems, performs key entity association preprocessing, and obtains business behavior data; Association Acquisition Module: Obtains the power fluctuation index from business behavior data, obtains event sequence vectors based on business behavior data, constructs an event temporal convolutional network model based on the event sequence vectors, and outputs the operational event correlation degree of business behavior data based on the event temporal convolutional network model; Risk Decision Module: Based on the power fluctuation index and the correlation of operational events, an initial risk value is output by constructing a collaborative risk fusion network based on sparse incentives. Based on the initial risk value, risk transmission and source tracing analysis are performed to obtain a dynamic risk profile of the audit object. The collaborative risk fusion network based on sparse excitation includes a temporal convolutional layer, an optimized fully connected layer with an adaptive feature focusing mechanism, and an output layer. Business Audit Module: Based on the dynamic risk profile of the audit object, marketing business audit is realized through intelligent assignment of audit tasks and precise recommendation of audit strategies.

Images