A user intention recognition method, device and medium for power service

Abstract

This specification discloses a method, device, and medium for user intent recognition in electricity services, relating to the field of artificial intelligence technology, to address the problem of existing electricity service push notifications in business halls being detached from user intent. The method includes: acquiring real-time 3D visitor point cloud data of the current business hall using millimeter-wave radar deployed there; processing the 3D visitor point cloud data to extract the 3D coordinates of key skeletal joints corresponding to each visitor, and constructing a human feature map structure based on the spatial topological connection relationship of the key skeletal joints, extracting its spatiotemporal features to identify visitor user behavior based on these features; mapping user behavior based on the visitor's location within the current business hall to obtain a visitor's user intent profile, facilitating the matching of the user intent profile with the electricity service profile to obtain the electricity service matching the profile.

Description

Technical Field

[0001] This specification relates to the field of power service technology, and in particular to a method, device and medium for identifying user intent in power services. Background Technology

[0002] As the service level of the power industry continues to improve, traditional business halls are gradually transforming towards intelligent and personalized services. Proactively identifying customer needs and providing matching services has become key to enhancing customer experience and business efficiency.

[0003] Currently, traditional electricity service delivery relies primarily on the proactive observation and experience of branch staff. When staff notice customers lingering at self-service terminals, loitering near information boards, or exhibiting anxiety in the waiting area, they proactively approach them to inquire and offer assistance. However, the service quality of this method is highly dependent on the staff's subjective experience and attention span. During peak hours, staff cannot attend to all users, easily leading to delayed service responses. Furthermore, close observation and questioning may cause privacy intrusion and psychological stress for some customers. To improve automation, some branches use camera-based visual analysis technology to automatically identify customer queuing status and emotional characteristics. However, this method is heavily influenced by the environment; its accuracy is unstable in complex branch environments, and it lacks understanding of user intent, making it difficult to accurately deliver electricity services.

[0004] Therefore, how to obtain accurate and real-time user intent regarding power services in complex environments has become an urgent problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this specification provides one or more embodiments of a user intent recognition method, device, and medium for power services.

[0006] One or more embodiments of this specification employ the following technical solutions: This specification provides one or more embodiments of a user intent recognition method for electricity services, the method comprising: The three-dimensional visitor point cloud data of the current business hall is acquired in real time by using millimeter-wave radar deployed in the current business hall; The three-dimensional visitor point cloud data is processed to extract the three-dimensional coordinates of the key human skeletal joints corresponding to each visitor, and a human feature map structure is constructed based on the spatial topological connection relationship of the key skeletal joints; wherein the human feature map structure uses the key skeletal joints as nodes and the spatial topological connection relationship between each key skeletal joint as edges. Extract the spatiotemporal features of the human body feature map structure to identify visitor user behavior based on the spatiotemporal features; Based on the visitor's location in the current business hall, the user behavior is mapped to obtain the visitor's user intent profile, so as to match the user intent profile with the power service profile and then obtain the power service that matches the power service profile.

[0007] Optionally, in one or more embodiments of this specification, the three-dimensional visitor point cloud data of the current business hall is acquired in real time by a millimeter-wave radar deployed in the current business hall, specifically including: Control the millimeter-wave radar deployed in the current business hall to scan the current business hall and obtain the raw point cloud data within the current business hall; Based on the spatial pose calibration information of the millimeter-wave radar, the original point cloud data is transformed from the radar local coordinate system to the preset global coordinate system of the business hall to generate three-dimensional dynamic point cloud data under a unified coordinate system; wherein, the three-dimensional dynamic point cloud data is used to represent the physical spatial distribution of each object in the current business hall. Static background filtering is applied to the three-dimensional dynamic point cloud data to obtain the filtered point cloud data of the current business hall; Based on the preset constraint data, the filtered point cloud data is screened to obtain three-dimensional visitor point cloud data; wherein, the preset constraint data includes at least: human body constraint data and motion feature constraint data; the three-dimensional visitor point cloud data corresponds to the visitor user of the current business hall.

[0008] Optionally, in one or more embodiments of this specification, the three-dimensional visitor point cloud data is processed to extract the three-dimensional coordinates of key human skeletal joints corresponding to each visitor, and a human feature map structure is constructed based on the spatial topological connection relationship of the key skeletal joints, specifically including: Based on the data density of the 3D visitor point cloud data, spatial clustering processing is performed on the 3D visitor point cloud data to obtain an initial individual point cloud data cluster corresponding to a single visitor; The initial human point cloud cluster is subjected to centroid normalization and scale standardization to obtain an individual point cloud data cluster for a single visitor; The individual point cloud data clusters are input into a pre-trained lightweight joint regression network to output the three-dimensional coordinates of each key human skeletal joint in the individual point cloud data clusters. Based on the spatial topological connection relationship of the key skeletal joints, the three-dimensional coordinates are structurally connected to generate a human feature map structure; wherein, the human feature map structure is composed of the coordinate sequence of the key skeletal joints and their connecting edge vectors.

[0009] Optionally, in one or more embodiments of this specification, the spatiotemporal features of the human body feature map structure are extracted to identify visitor user behavior based on the spatiotemporal features, specifically including: The human feature map structure of consecutive frames is sorted based on time series to construct a spatiotemporal graph sequence; wherein the spatiotemporal graph sequence uses the key skeletal joints as vertices, skeletal connections as edges, and the joint motion state as vertex attributes; Extract the spatial attitude features and temporal motion features of the spatiotemporal graph sequence; The spatial attitude features and temporal motion features are fused in a multimodal manner to generate a spatiotemporal fusion feature vector; The spatiotemporal fusion feature vector is input into a pre-trained dual-branch behavior recognition model to obtain the visitor's user behavior category and corresponding confidence level; wherein, the dual-branch behavior recognition model includes: a security event detection branch for identifying abnormal behavior events, and a service intent recognition branch for identifying business interaction intent; If the confidence level of the user behavior category is greater than a preset threshold, then the user behavior category is determined to be the visitor's user behavior.

[0010] Optionally, in one or more embodiments of this specification, the user behavior is mapped based on the visitor's location in the current business hall to obtain the visitor's user intent profile, specifically including: Based on the human feature map structure of the continuous frames and the area division data of the current business hall, the functional area of ​​the business hall where the visitor is located and the length of time he stays in the functional area of ​​the business hall are determined. Based on the user behavior and the dwell time, a user behavior vector of the visitor is generated; Based on the functional areas of the business hall, a corresponding intent mapping rule base is determined, and the intent mapping rule base is queried based on the user behavior vector to construct the visitor's user intent tag; wherein, the intent mapping rule base is dynamically updated based on the historical service records of the functional areas of the business hall and user feedback data. The user intent tags and the corresponding confidence scores of the user behavior categories are encapsulated into structured data to obtain the visitor's user intent profile.

[0011] Optionally, in one or more embodiments of this specification, based on the human feature map structure of consecutive frames and the current business hall area division data, determining the functional area of ​​the business hall where the visitor is located and the duration of stay in the functional area of ​​the business hall specifically includes: The three-dimensional coordinates of the central joint of the human torso in the human feature map structure are obtained. The three-dimensional coordinates of the central joint of the human torso in consecutive frames are matched with the area division data of the current business hall to determine the functional area of ​​the business hall where the visitor is located in each frame. Count the number of consecutive frames the visitor stays in each of the functional areas of the business hall; The duration of a visitor's stay in the functional area of ​​the business hall is determined based on the number of consecutive stay frames and the sampling frequency of the millimeter-wave radar.

[0012] Optionally, in one or more embodiments of this specification, matching the user intent profile with the power service profile to obtain the power service matching the power service profile specifically includes: Calculate the semantic matching degree between the user intent tags of the user intent profile and the service intent tags in each of the power service profiles; Based on the semantic matching degree, candidate power services that match the user intent profile are determined; Based on the confidence level corresponding to the user behavior category, the candidate power services are sorted in descending order to obtain power services that match the power service profile.

[0013] Optionally, in one or more embodiments of this specification, after matching the user intent profile with the power service profile to obtain the power service matching the power service profile, the method further includes: Determine the operating status of each business terminal device in the functional area of ​​the business hall to identify the candidate business terminal devices that are in a ready state; Based on the hand joint data of the human feature map structure, the interaction state of the visitor relative to the candidate business terminal device is identified; wherein, the interaction state includes: operation state and non-operation state; If the interaction state is an operation state, the power service will be pushed to the current operation interface of the candidate business terminal device in the form of a non-blocking interface element push. If the interaction state is a non-operation state, the distance between the visitor and each public information display device in the business hall functional area is determined based on the three-dimensional coordinates of the central joint of the visitor's human torso. Based on the distance, the nearest public information display device is determined, and the power service is pushed to the public information display device and the visitor's user terminal.

[0014] This specification provides one or more embodiments of a user intent recognition device for electricity services, the device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the methods described above.

[0015] This specification provides one or more embodiments of a non-volatile computer storage medium storing computer-executable instructions, the computer-executable instructions being configured to execute any of the methods described above.

[0016] The above-described at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: By deploying millimeter-wave radar, seamless visitor identification within the service hall is achieved, effectively overcoming the response lag and environmental interference issues inherent in traditional methods relying on manual observation or visual recognition. The human features of 3D visitor point cloud data are converted into a human feature map structure, and then highly discriminative spatiotemporal behavioral features are extracted, enabling accurate identification of user behavior. Based on the visitor's location context within the current service hall, their behavior is mapped to obtain a user intent profile, achieving a combination of behavior and location. This solves the problem of the same behavior having different meanings in different areas, improving the accuracy of user intent profiling. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart illustrating a user intent recognition method for electricity services provided in an embodiment of this specification; Figure 2 This specification provides a schematic diagram of the structure of a user intent recognition device for power services, as illustrated in an embodiment of the present specification. Figure 3 This is a schematic diagram of the structure of a non-volatile storage medium provided in the embodiments of this specification. Detailed Implementation

[0018] This specification provides a method, device, and medium for identifying user intent in power services.

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0020] like Figure 1 The illustrated embodiment of this specification provides a structural diagram of a user intent recognition method for electricity services. Figure 1 As can be seen, in one or more embodiments of this specification, a user intent recognition method for electricity services specifically includes the following steps: S101: The three-dimensional visitor point cloud data of the current business hall is acquired in real time by using millimeter-wave radar deployed in the current business hall.

[0021] To achieve non-invasive and accurate perception of visitor behavior within the service hall, in this embodiment of the specification, a set of millimeter-wave radars, pre-calibrated in position and orientation, are pre-deployed in key areas of the service hall, such as the entrance, waiting area, self-service area, and service counter area, either above or to the side. These millimeter-wave radars work collaboratively in frequency-modulated continuous wave mode, continuously emitting electromagnetic waves into the monitoring space and receiving echo signals reflected by targets such as human bodies. Therefore, in this embodiment of the specification, by deploying millimeter-wave radars in the service hall, real-time 3D visitor point cloud data of the service hall is acquired, thereby avoiding the problem of unstable accuracy due to environmental influences in traditional visual analysis technologies such as camera-based recognition.

[0022] Specifically, in one or more embodiments of this specification, the three-dimensional visitor point cloud data of the current business hall is acquired in real time by a millimeter-wave radar deployed in the current business hall, specifically including: The millimeter-wave radar deployed in the current business hall is controlled to scan the hall, thereby acquiring the raw point cloud data within it. In other words, by controlling the millimeter-wave radar deployed in the current business hall to scan the area emitting frequency-modulated continuous wave radar signals, the echo signals reflected back from all objects within the hall can be received simultaneously. The echo signals are processed using the current standardized radar signal processing procedure to obtain the raw point cloud data of the current business hall. This procedure is an existing process, and in a certain scenario, it specifically includes: mixing, filtering, and analog-to-digital conversion of the echo signals to obtain the digital intermediate frequency (IF) signal. The IF signal is then subjected to range-dimensional and Doppler-dimensional fast Fourier transforms sequentially. The target range is analyzed based on the range-dimensional fast Fourier transform, and the target radial velocity is analyzed based on the Doppler-dimensional fast Fourier transform. Meanwhile, to obtain the target's azimuth in space, the target's spatial angle can be calculated by combining the phase information of the receiving channel. For each detected target point, the obtained distance, radial velocity, horizontal azimuth angle, and elevation angle can be correlated and bound to obtain the four-dimensional radar data corresponding to the digital intermediate frequency signal, which serves as the raw point cloud data.

[0023] Because multiple radar sensors are typically deployed in the service hall to cover blind spots and expand the field of view, the raw point cloud data generated by each radar resides in its own local sensor coordinate system. To achieve a unified view across the entire hall, it is necessary to access the spatial pose calibration information of the millimeter-wave radar obtained through measurement and calibration during the installation phase. This spatial pose calibration information describes the three-dimensional position and orientation of each radar in the global world coordinate system of the service hall. Based on the spatial pose calibration information, a rigid body transformation matrix is ​​used to uniformly transform all points in the raw point cloud data of each radar from their respective local coordinate systems to the same global world coordinate system. Under this global world coordinate system, coordinate transformation and data fusion are performed on the raw point cloud data to eliminate redundant points in overlapping areas and fill in the blind spots of individual radars, generating three-dimensional dynamic point cloud data in a unified coordinate system. This three-dimensional dynamic point cloud data represents the physical spatial target distribution of the service hall at a given moment; that is, the distribution of all detected targets within the entire physical space of the service hall at a given time. Each three-dimensional dynamic point cloud data includes world coordinates and instantaneous radial velocity.

[0024] Because the aforementioned 3D dynamic point cloud data contains a large amount of data from static background objects such as walls, pillars, fixed counters, tables, and chairs, this specification's embodiments perform static background filtering on the acquired 3D dynamic point cloud data to obtain the initial 3D visitor point cloud data for the current business hall. In one feasible approach, by analyzing multiple frames of data to statistically determine the probability of points appearing at each spatial location, 3D dynamic point cloud data with a stable occurrence exceeding a preset high frequency is identified as static background. By filtering this static background, the initial 3D visitor point cloud data for the current business hall is obtained. In another feasible approach, a pre-entered static 3D model of the business hall is compared and filtered against the 3D dynamic point cloud data to remove static backgrounds from the 3D dynamic point cloud data, thus obtaining the initial 3D visitor point cloud data for the current business hall.

[0025] Although some interfering data is filtered out based on a static background, the data may still contain non-human moving targets, such as moving cleaning robots, swaying plants, or papers blown by the wind. Therefore, in this embodiment, the initial 3D visitor point cloud data is filtered based on preset constraint data, which includes at least human constraint data and motion feature constraint data, to obtain 3D visitor point cloud data. At this time, the 3D visitor point cloud data corresponds to the visitor user in the current business hall. It should be noted that the preset human constraint data is determined based on prior knowledge of standard human body size and shape, such as: the range of the constraint point cloud cluster in the height direction should be within the range of human height; the width and thickness should conform to the proportion of the human torso and limbs; the overall shape of the point cloud cluster can be matched with the human body model through features such as contour moments, etc. The motion feature constraint data includes: analyzing whether its overall motion trajectory conforms to the laws of human gait, etc. The point cloud cluster in the above example can be understood as being obtained from the initial 3D visitor point cloud data based on spatial clustering.

[0026] S102: Process the three-dimensional visitor point cloud data to extract the three-dimensional coordinates of the key human skeletal joints corresponding to each visitor, and construct a human feature map structure based on the spatial topological connection relationship of the key skeletal joints; wherein the human feature map structure uses the key skeletal joints as nodes and the spatial topological connection relationship between the key skeletal joints as edges.

[0027] After obtaining the 3D visitor point cloud data, in order to convert the acquired 3D visitor point cloud data into a standard digital model with spatial topological relationships, thereby laying a data foundation for subsequent visitor behavior analysis, the embodiments of this specification preprocess the 3D visitor point cloud data to extract the 3D coordinates of the key human skeletal joints corresponding to each visitor. Then, based on the spatial topological connection relationship of the key skeletal joints, the 3D coordinates are connected with the key skeletal joints as nodes and the spatial topological connection relationship between each key skeletal joint as edges to construct a human feature map structure. It should be noted that, in one feasible way, the extraction of the 3D coordinates of the key human skeletal joints corresponding to each visitor can be achieved based on the following process: obtaining one or more pre-created standard 3D human skeleton templates generated based on standard anthropometry data and containing various common poses. This template contains preset key skeletal joints and spatial topological connection relationships. After preprocessing the input 3D visitor point cloud data such as denoising, it is matched with the 3D human skeleton template for nearest point matching, for example, by iteratively optimizing rigid body transformation, so that the 3D visitor point cloud data and the 3D human skeleton template are aligned in 3D space, thereby minimizing the distance error between the corresponding points of the two. After alignment, based on the three-dimensional coordinates of each key skeletal joint predefined in the three-dimensional human skeleton template, the optimized rigid body transformation matrix is ​​mapped to the coordinate system of the three-dimensional visitor point cloud data to obtain the three-dimensional coordinates of the key skeletal joints of each visitor.

[0028] Specifically, in another feasible embodiment of this specification, the above-mentioned processing of the three-dimensional visitor point cloud data to extract the three-dimensional coordinates of the key skeletal joints of the human body corresponding to each visitor, and the construction of a human body feature map structure based on the spatial topological connection relationship of the key skeletal joints, specifically includes the following process: To perform independent analysis on each visitor, spatial clustering is performed on the 3D visitor point cloud data based on its data density to obtain an initial individual point cloud data cluster corresponding to each visitor. This spatial clustering method can be density-based, such as clustering based on the proximity distance between points and a minimum point count threshold within a region to obtain the initial individual point cloud data cluster corresponding to each visitor.

[0029] The initial human point cloud cluster is subjected to centroid normalization and scale normalization to obtain individual point cloud data clusters. Centroid normalization involves calculating the 3D arithmetic mean coordinates of all points in the cluster to obtain its geometric center. Then, the centroid coordinates are subtracted from the coordinates of all points within the cluster, ensuring that the centroid of the processed point cloud cluster coincides with the origin of the 3D coordinate system, thus eliminating spatial randomness. Scale normalization involves calculating the vertical height range of the point cloud cluster, or the overall bounding box size of the point cloud, linearly scaling the overall size of the cluster to a preset standard human body size range. After centroid and scale normalization, individual point cloud data clusters that are normalized in both position and size are obtained.

[0030] Individual point cloud data clusters are input into a pre-trained lightweight joint regression network to output the 3D coordinates of key skeletal joints in each individual point cloud data cluster. This pre-trained lightweight joint regression network is trained on a deep learning network using a dataset containing a large number of 3D human point clouds and their corresponding precisely labeled joints. The architecture of this deep learning network can be a PointNet network or other point cloud processing networks. After the individual point cloud data clusters are fed into this pre-trained lightweight joint regression network, the network extracts local and global features of the point cloud layer by layer through its multilayer perceptron and other structures, and directly regresses and predicts the 3D coordinates of each key skeletal joint in the network's output layer. It is understood that, to describe the spatial position of major limb connections, key skeletal joints include at least: the top of the head, neck, both shoulders, both elbows, both hands, the center of the torso, both hips, both knees, and both feet.

[0031] Based on the spatial topological connections of key skeletal joints, 3D coordinates are structurally connected to generate a human feature map structure. For example, the connection rules are defined as follows: the top of the head connects to the neck, the neck connects to both shoulders and the center of the torso, both shoulders connect to the elbows on the same side, the elbows connect to the hands on the same side, the center of the torso connects to both hips, both hips connect to the knees on the same side, and the knees connect to the feet on the same side. Through these connections, isolated joints are linked together to form a complete digital human skeleton, obtaining the human feature map structure. This human feature map structure consists of a sequence of coordinates of key skeletal joints (i.e., a 3D coordinate array of joints arranged in a predetermined order) and its connecting edge vectors (i.e., direction vectors defining the connections between joint index pairs).

[0032] S103: Extract the spatiotemporal features of the human body feature map structure to identify visitor user behavior based on the spatiotemporal features.

[0033] As the background technology indicates, current traditional electricity service delivery relies primarily on the proactive observation and experience of branch staff. When staff notice customers lingering at self-service terminals, loitering near information boards, or exhibiting anxiety in the waiting area, they proactively approach them to inquire and offer assistance. However, the service quality of this method is highly dependent on the staff's subjective experience and attention span. During peak hours, staff cannot attend to all users, easily leading to delayed service responses. Furthermore, this observation and inquiry process may cause some customers to feel privacy intrusion and psychological stress. To improve automation, some branches use camera-based visual analysis technology to automatically identify customer queuing status and emotional characteristics. However, this method is heavily influenced by the environment; its accuracy is unstable in complex branch environments, and it lacks understanding of user intent, making it difficult to accurately deliver electricity services. Therefore, the challenge lies in accurately understanding visitors' true behavioral intentions in the complex and dynamic environment of an electricity branch without relying on cameras, facial recognition, or proactive user interaction, thereby supporting personalized and proactive electricity service delivery. In the embodiments of this specification, the spatiotemporal features of the human body feature map structure are extracted, thereby identifying the visitor's user behavior based on the spatiotemporal features, so as to recommend corresponding power services based on the user behavior, so that the recommended power services meet the user's actual intentions.

[0034] Specifically, in one or more embodiments of this specification, the spatiotemporal features of the human body feature map structure are extracted to identify visitor user behavior based on the spatiotemporal features, specifically including: The structure of human feature maps in consecutive frames is sorted based on time series to construct a spatiotemporal graph sequence. This sequence uses key skeletal joints as vertices, skeletal connections as edges, and the motion state of the joints as vertex attributes. Essentially, the spatial graphs of consecutive frames are concatenated along the time dimension, and temporal edges are established between the same joints in adjacent frames to obtain the spatiotemporal graph sequence. Spatial pose features and temporal motion features are extracted from the spatiotemporal graph sequence. Spatial pose features include: torso tilt angle, arm spread, head pitch angle, and center of gravity projection position. Temporal motion features include: trajectory curvature of specified joints, instantaneous velocity change amplitude, periodic intensity of limb movements, and overall motion entropy. Each feature can be obtained based on its corresponding feature calculation method, such as fitting the motion trajectory of key joints in multiple consecutive frames and calculating its average curvature to obtain the trajectory curvature of a specified joint. After obtaining the spatial pose features and temporal motion features, they are fused using multimodal methods, such as direct concatenation or joint fusion based on a decision layer, to generate a spatiotemporal fused feature vector.

[0035] The spatiotemporal fusion feature vector is input into a pre-trained dual-branch behavior recognition model to obtain the visitor's user behavior category and corresponding confidence score. The pre-trained dual-branch behavior recognition model is a neural network architecture with a shared underlying feature extractor and two independent output heads. The two independent output heads correspond to a security event detection branch for identifying abnormal behavior events and a service intent recognition branch for identifying business interaction intentions, respectively. If the confidence score corresponding to a user behavior category is greater than a preset threshold, the user behavior category is determined to be a visitor's user behavior. Specifically, after inputting the spatiotemporal fusion feature vector into the pre-trained dual-branch behavior recognition model, feature vectors are extracted through a backbone network consisting of multiple stacked fully connected layers and activation function layers. After outputting the feature vector from the shared backbone network, the data flow branches into two independent and structurally similar but target-differentiated branches: a network security event detection branch and a service intent recognition branch. Each branch consists of one to two fully connected layers, and the confidence score of the corresponding abnormal event category or the corresponding business intent category is obtained through the output layer.

[0036] S104: Based on the visitor's location in the current business hall, the user behavior is mapped to obtain the visitor's user intent profile, so as to match the user intent profile with the power service profile, and then obtain the power service that matches the power service profile.

[0037] Traditional video surveillance solutions are highly susceptible to environmental influences, while technologies such as Wi-Fi / Bluetooth positioning cannot capture posture information. Therefore, existing technologies typically remain at the level of simple behavioral statistics, such as passenger flow counting, or general action recognition, such as fall detection, making it difficult to accurately connect with the specific business needs of the power industry for precise power services. Traditional push notification methods, on the other hand, suffer from problems such as information harassment, inappropriate timing of push notifications, and interference with the user's current tasks.

[0038] Specifically, in one or more embodiments of this specification, the user behavior is mapped based on the visitor's location in the current business hall to obtain the visitor's user intent profile, specifically including: Based on the structure of continuous frame human body feature maps and the current area division data of the service hall, the functional area of ​​the service hall where the visitor is located and the duration of their stay in that area can be determined. The current area division data divides the physical space of the service hall into multiple functional areas with clearly defined pre-defined business semantics, such as a number dispensing and waiting area, a low-voltage service counter area, a high-voltage and new energy service consultation area, and a self-service terminal area. Each area has its corresponding three-dimensional spatial boundary definition. Therefore, based on the location corresponding to the continuous frame human body feature map structure, the functional area of ​​the service hall where the visitor is located can be determined. By tracking the sequence of areas the visitor is in over a period of time, the continuous duration of the visitor's stay in a specific functional area of ​​the service hall can be accurately calculated.

[0039] To facilitate subsequent processing, user behavior and dwell time are encoded separately. These encoded data are then concatenated to generate a visitor's user behavior vector. Based on the functional areas of the service hall, a corresponding intent mapping rule base is determined. The visitor's user intent tags are then constructed by querying this intent mapping rule base based on the user behavior vector. It's important to note that the intent mapping rule base is dynamically updated based on historical service records of the service hall's functional areas and user feedback data. For example, in one scenario: historical interaction data occurring in this area is continuously collected, including the type of service ultimately processed, customer satisfaction ratings, service personnel notes, and other user feedback data, and correlated with the captured user behavior vectors. Data mining or machine learning algorithms are used to obtain the true business intent corresponding to different user behavior vectors in this specific area. For example, the rule base might contain a rule like: in the new energy exhibition area, if the behavior vector shows stopping to watch and dwelling time > 120 seconds, there is an 85% probability of having an initial interest in photovoltaic business. Alternatively, the currently generated user behavior vector can be used as query input to perform matching searches in the corresponding rule base, such as cosine similarity or direct matching of predefined rule conditions. After a successful match, based on the output of the rule base, construct one or more user intent tags that best describe the visitor's current potential needs, such as intent to inquire about peak-valley electricity changes, possible need for assistance operating a self-service machine, or strong interest in electric vehicle charging stations.

[0040] After obtaining user intent tags, the confidence scores corresponding to the user intent tags and user behavior categories are encapsulated into structured data to obtain a visitor's user intent profile. This user intent profile is output in a machine-readable standardized format, enabling accurate matching and delivery of electricity services in the electricity business scenario. It provides clear customer needs and helps improve the accuracy of electricity service delivery.

[0041] Furthermore, in one or more embodiments of this specification, based on the human feature map structure of the continuous frames and the current area division data of the business hall, the functional area of ​​the business hall where the visitor is located and the duration of stay in the functional area of ​​the business hall are determined, specifically including: To ensure the stability of determining the functional area of ​​the service hall where the visitor is located, in this feasible embodiment, the three-dimensional coordinates of the central joint of the human torso in the human feature map structure are obtained. These three-dimensional coordinates are used as a stable reference point. The three-dimensional coordinates of the central joint of the human torso in consecutive frames are then matched with the current service hall area division data to determine the functional area of ​​the service hall where the visitor is located in each frame. Specifically, for the t-th frame, the three-dimensional coordinates of the central joint of the visitor's human torso in that frame are read. This coordinate point is compared with the three-dimensional boundaries of all predefined functional areas to determine whether the coordinate point is contained within the three-dimensional boundaries. If it is determined that the visitor is in that specific functional area of ​​the service hall at that frame time, the functional areas of the service hall corresponding to the visitor are sorted according to the time sequence to determine the frame interval in which the visitor is continuously located in the same functional area, obtaining the number of consecutive stay frames of the visitor in each functional area. Using the sampling frequency of the millimeter-wave radar as the time reference, the length of time the visitor stays in the functional area of ​​the service hall is determined based on the ratio of the number of consecutive stay frames to the sampling frequency.

[0042] In this process, the center node of the torso was selected as the reference point to reduce positioning jitter caused by limb swings and ensure the stability of position determination. The sampling frequency of the millimeter-wave radar was used as the time base for conversion, allowing the duration calculation to be directly traced back to the clock of the physical device, avoiding timing errors caused by software processing delays and queue waiting. Furthermore, defining a dwell time by counting the number of consecutive dwell frames effectively filtered out invalid brief crossings, visitors quickly moving through an area, or instantaneous backtracking during brief probe visits at the boundary, ensuring the reliability of the behavior analysis.

[0043] After obtaining the user intent profile based on the above steps, in order to push the electricity service corresponding to the user intent profile to the corresponding visitor, this embodiment of the specification matches the user intent profile with the electricity service profile to obtain the electricity service that matches the electricity service profile. Then, based on the corresponding business hall linkage device of the visitor, the electricity service is pushed and displayed to the visitor. The business hall linkage device refers to available devices in the current business hall, such as interactive screens, intelligent robots, teller auxiliary screens, self-service terminals, etc. After sending the electricity service to the business hall linkage device corresponding to the visitor, the content of the electricity service can be rendered and displayed according to the screen size, resolution, and interactive capabilities of the device.

[0044] Specifically, in one or more embodiments of this specification, matching the user intent profile with the power service profile to obtain the power service that matches the power service profile includes: Calculate the semantic matching degree between the user intent tags in the user intent profile and the service intent tags in each power service profile. For example, convert the user intent tags and the service intent tags of each power service into semantic vectors, and obtain the semantic matching degree by calculating the cosine similarity or the reciprocal of the Euclidean distance between the user intent vector and each service intent vector.

[0045] After calculating the semantic matching degree of all relevant services, candidate power services that match the user's intent profile are determined based on the semantic matching degree and a preset semantic matching degree threshold. The candidate power services are then sorted in descending order according to the corresponding confidence level of the user behavior category, with the top-ranked power services being selected as those that match the power service profile.

[0046] Furthermore, in one or more embodiments of this specification, after matching the user intent profile with the power service profile to obtain the power service matching the power service profile, the method further includes: A status scan is performed on the available service terminal equipment in the service hall's functional areas to obtain the operating status of each terminal equipment. The operating status includes at least: equipment power status, network connection status, current load, and fault indicators. Based on the above operating status, equipment that is powered on, online, idle, and fault-free is identified as candidate service terminal equipment in a ready state.

[0047] To determine the real-time relationship between visitors and these devices, after identifying potential service terminals, the interaction state of the visitor relative to the terminal is identified based on hand joint data from a human feature map structure. This interaction state includes both active and non-active states. The minimum Euclidean distance between the hand joint data and the potential service terminal, along with the trajectory corresponding to consecutive frames of hand joint data, determines the visitor's interaction state. If the interaction state is active, the power service is pushed to the current interface of the potential service terminal in a non-interrupting interface element format. This non-interrupting push format includes a semi-transparent sidebar, a bottom notification bar, or a floating QR code, allowing the visitor to view or respond to the power service without interrupting the current business process. If the interaction state is non-active, the distance between the visitor and each public information display device in the service hall's functional area is determined based on the three-dimensional coordinates of the visitor's torso center joint. The nearest public information display device is then identified based on this distance, and the power service is pushed to both the public information display device and the visitor's user terminal.

[0048] This process pushes services to a visitor's current interface only when the visitor is actually using the terminal, effectively avoiding mis-push notifications and information overload caused by traditional methods based on approximate location. Furthermore, it uses non-blocking interface elements for push notifications, ensuring that users are not forcibly interrupted during their transactions. This balances proactive service with smooth operation, significantly improving the user experience. When a visitor is not using any terminal, the system automatically switches to the nearest public display screen or an authorized user terminal for fallback push notifications, achieving full-scenario coverage and service continuity.

[0049] like Figure 2 As shown in the diagram, this specification provides a schematic diagram of a user intent recognition device for power services. In one or more embodiments of this specification, a user intent recognition device for power services includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the methods described above.

[0050] like Figure 3 As shown in the diagram, this specification provides a schematic diagram of the structure of a non-volatile storage medium. Figure 3 As can be seen, in one or more embodiments of this specification, a non-volatile storage medium stores computer-executable instructions 301, which are capable of executing any of the methods described above.

[0051] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, and therefore described more simply; relevant parts can be referred to the descriptions of the method embodiments.

[0052] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0053] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for recognizing user intent in electricity services, characterized in that, The method includes: The three-dimensional visitor point cloud data of the current business hall is acquired in real time by using millimeter-wave radar deployed in the current business hall; The three-dimensional visitor point cloud data is processed to extract the three-dimensional coordinates of the key human skeletal joints corresponding to each visitor, and a human feature map structure is constructed based on the spatial topological connection relationship of the key skeletal joints; wherein the human feature map structure uses the key skeletal joints as nodes and the spatial topological connection relationship between each key skeletal joint as edges. Extract the spatiotemporal features of the human body feature map structure to identify visitor user behavior based on the spatiotemporal features; Based on the visitor's location in the current business hall, the user behavior is mapped to obtain the visitor's user intent profile, so as to match the user intent profile with the power service profile and then obtain the power service that matches the power service profile.

2. The user intent recognition method for electricity services according to claim 1, characterized in that, By deploying millimeter-wave radar in the current business hall, real-time 3D visitor point cloud data of the current business hall is acquired, specifically including: Control the millimeter-wave radar deployed in the current business hall to scan the current business hall and obtain the raw point cloud data within the current business hall; Based on the spatial pose calibration information of the millimeter-wave radar, the original point cloud data is transformed from the radar local coordinate system to the preset global coordinate system of the business hall to generate three-dimensional dynamic point cloud data under a unified coordinate system; wherein, the three-dimensional dynamic point cloud data is used to represent the physical spatial distribution of each object in the current business hall. Static background filtering is applied to the three-dimensional dynamic point cloud data to obtain the filtered point cloud data of the current business hall; Based on the preset constraint data, the filtered point cloud data is screened to obtain three-dimensional visitor point cloud data; wherein, the preset constraint data includes at least: human body constraint data and motion feature constraint data; the three-dimensional visitor point cloud data corresponds to the visitor user of the current business hall.

3. The user intent recognition method for electricity services according to claim 1, characterized in that, The 3D visitor point cloud data is processed to extract the 3D coordinates of key skeletal joints corresponding to each visitor, and a human feature map structure is constructed based on the spatial topological connectivity of the key skeletal joints, specifically including: Based on the data density of the 3D visitor point cloud data, spatial clustering processing is performed on the 3D visitor point cloud data to obtain an initial individual point cloud data cluster corresponding to a single visitor; The initial human point cloud cluster is subjected to centroid normalization and scale standardization to obtain an individual point cloud data cluster for a single visitor; The individual point cloud data clusters are input into a pre-trained lightweight joint regression network to output the three-dimensional coordinates of each key human skeletal joint in the individual point cloud data clusters. Based on the spatial topological connection relationship of the key skeletal joints, the three-dimensional coordinates are structurally connected to generate a human feature map structure; wherein, the human feature map structure is composed of the coordinate sequence of the key skeletal joints and their connecting edge vectors.

4. The user intent recognition method for electricity services according to claim 3, characterized in that, Extracting the spatiotemporal features of the human body feature map structure to identify visitor user behavior based on the spatiotemporal features specifically includes: The human feature map structure of consecutive frames is sorted based on time series to construct a spatiotemporal graph sequence; wherein the spatiotemporal graph sequence uses the key skeletal joints as vertices, skeletal connections as edges, and the joint motion state as vertex attributes; Extract the spatial attitude features and temporal motion features of the spatiotemporal graph sequence; The spatial attitude features and temporal motion features are fused in a multimodal manner to generate a spatiotemporal fusion feature vector; The spatiotemporal fusion feature vector is input into a pre-trained dual-branch behavior recognition model to obtain the visitor's user behavior category and corresponding confidence level; wherein, the dual-branch behavior recognition model includes: a security event detection branch for identifying abnormal behavior events, and a service intent recognition branch for identifying business interaction intent; If the confidence level of the user behavior category is greater than a preset threshold, then the user behavior category is determined to be the visitor's user behavior.

5. The user intent recognition method for electricity services according to claim 4, characterized in that, Based on the visitor's location in the current business hall, the user behavior is mapped to obtain the visitor's user intent profile, specifically including: Based on the human feature map structure of the continuous frames and the area division data of the current business hall, the functional area of ​​the business hall where the visitor is located and the length of time he stays in the functional area of ​​the business hall are determined. Based on the user behavior and the dwell time, a user behavior vector of the visitor is generated; Based on the functional areas of the business hall, a corresponding intent mapping rule base is determined, and the intent mapping rule base is queried based on the user behavior vector to construct the visitor's user intent tag; wherein, the intent mapping rule base is dynamically updated based on the historical service records of the functional areas of the business hall and user feedback data. The user intent tags and the corresponding confidence scores of the user behavior categories are encapsulated into structured data to obtain the visitor's user intent profile.

6. The user intent recognition method for electricity services according to claim 5, characterized in that, Based on the human feature map structure of consecutive frames and the current business hall area division data, the functional area of ​​the business hall where the visitor is located and the duration of stay in the functional area of ​​the business hall are determined, specifically including: The three-dimensional coordinates of the central joint of the human torso in the human feature map structure are obtained. The three-dimensional coordinates of the central joint of the human torso in consecutive frames are matched with the area division data of the current business hall to determine the functional area of ​​the business hall where the visitor is located in each frame. Count the number of consecutive frames the visitor stays in each of the functional areas of the business hall; The duration of a visitor's stay in the functional area of ​​the business hall is determined based on the number of consecutive stay frames and the sampling frequency of the millimeter-wave radar.

7. A user intent recognition method for electricity services according to claim 6, characterized in that, Matching the user intent profile with the power service profile to obtain the power service that matches the power service profile, specifically including: Calculate the semantic matching degree between the user intent tags of the user intent profile and the service intent tags in each of the power service profiles; Based on the semantic matching degree, candidate power services that match the user intent profile are determined; Based on the confidence level corresponding to the user behavior category, the candidate power services are sorted in descending order to obtain power services that match the power service profile.

8. A user intent recognition method for electricity services according to claim 6, characterized in that, After matching the user intent profile with the power service profile to obtain the power service that matches the power service profile, the method further includes: Determine the operating status of each business terminal device in the functional area of ​​the business hall to identify the candidate business terminal devices that are in a ready state; Based on the hand joint data of the human feature map structure, the interaction state of the visitor relative to the candidate business terminal device is identified; wherein, the interaction state includes: operation state and non-operation state; If the interaction state is an operation state, the power service will be pushed to the current operation interface of the candidate business terminal device in the form of a non-blocking interface element push. If the interaction state is a non-operation state, the distance between the visitor and each public information display device in the business hall functional area is determined based on the three-dimensional coordinates of the central joint of the visitor's human torso. Based on the distance, the nearest public information display device is determined, and the power service is pushed to the public information display device and the visitor's user terminal.

9. A user intent recognition device for electricity services, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in any one of claims 1-8.

10. A non-volatile storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are capable of performing the method described in any one of claims 1-8.

Images