An advertisement putting point position intelligent screening method based on AI and big data analysis

Abstract

The application relates to the technical field of commercial information processing, and discloses an advertisement putting point position intelligent screening method based on AI and big data analysis, which comprises the following steps: collecting original sensing data streams and hashing and desensitizing identity information by each point position edge computing node; extracting group density and residence time length features by a local dimension reduction algorithm, constructing a current semantic feature vector, and calculating the deviation of the current semantic feature vector from a historical state; when the deviation exceeds a threshold value, increment update messages are asynchronously sent to the cloud; the cloud calculates real-time commercial value weights by using a space-time graph convolution network model and issues instructions; through the logic cooperation of edge side sensing and cloud asynchronous driving, the application reduces channel congestion caused by node concurrency, eliminates the time delay generated by centralized decision-making, and ensures that the issued putting instructions are accurately matched with the customer group residence window.

Description

Technical Field

[0001] This invention belongs to the field of business information processing technology, and in particular relates to an intelligent method for selecting advertising placement locations based on AI and big data analysis. Background Technology

[0002] With the continuous expansion of distributed delivery networks, acquiring location resources through a cloud-centralized processing architecture has become an industry consensus. This architecture relies on data collection terminals to acquire raw passenger flow data and uploads the entire sensor data stream, including video signals, to the cloud center. The server then centrally processes feature extraction and matching calculations and issues control commands. However, the effectiveness of this architecture implicitly depends on the backbone network having near-infinite throughput and instantaneous response capabilities. When offline nodes are distributed on a large scale, the massive heterogeneous data generated by high-frequency dynamic passenger flow can easily reach the physical limits of network transmission during concurrent uploads, causing channel congestion and queue backlogs. This forced setting of data flow converging from the bottom layer to the center causes a time lag in the scheduling decisions output by the cloud. By the time the delivery command is routed to the terminal, the physical passenger flow state that triggered the decision has changed. Simply expanding computing resources or increasing communication bandwidth cannot fundamentally eliminate the spatiotemporal inconsistencies caused by large-scale data transfer.

[0003] Besides the physical link constraints at the hardware level, existing analysis methods also have limitations in balancing real-time performance and communication overhead at the data processing and synchronization logic level. For example, Chinese invention patent application CN109993589A discloses a machine vision-based method for analyzing the passenger flow data of advertising machines. This method relies on high-definition cameras to extract real-time visual features from continuous video streams. The logic of this technical path implicitly anchors to the local perception of individual points. When large-scale point matrix linkage is implemented, there is a lack of quantitative screening and filtering mechanisms for the evolution of features. This results in redundant feature data interacting frequently in the global network. It is impossible to get rid of the dependence on high-bandwidth communication conditions. In resource-constrained or passenger flow-changing scenarios, it is difficult to ensure that cloud scheduling decisions and physical space windows are accurately synchronized.

[0004] Therefore, how to reconstruct the boundary of responsibilities between edge nodes and the cloud, and achieve real-time synchronization and accurate screening of deployment point status with low communication overhead, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides a method for intelligent selection of advertising placement locations based on AI and big data analysis, comprising the following steps:

[0006] Step S101: The edge computing nodes at each advertising placement point acquire the original sensor data stream within a preset spatial range and use a preset one-way irreversible hash algorithm to desensitize the entity identification information in the original sensor data stream.

[0007] In step S102, the edge computing node uses a pre-set local dimensionality reduction algorithm to extract the population density feature value and dwell time feature value of the spatial region within the current time window, and maps them to generate the current semantic feature vector representing the real-time passenger flow attributes of the location.

[0008] Step S103: The edge computing node calculates the Euclidean distance between the current semantic feature vector and the pre-stored historical feature vector to determine the deviation value of the state evolution of the representation point.

[0009] Step S104: When the deviation value exceeds the preset asynchronous trigger threshold, the edge computing node encapsulates the current semantic feature vector and the point identifier into an incremental update message and asynchronously uploads it to the cloud, overwriting the historical mirror status of the corresponding point stored in the cloud.

[0010] In step S105, the cloud inputs the received incremental update message into the preset spatiotemporal graph convolutional network model, determines the real-time commercial value weight of each advertising placement point through spatiotemporal correlation convolution operation, and issues advertising placement instructions to the target advertising placement point according to the real-time commercial value weight.

[0011] Preferably, step S102 includes the following sub-steps: Step S1021, the edge computing node uses a preset sliding time window to extract the desensitized original sensor data stream and counts the temporal aggregated features within the sliding time window; Step S1022, the edge computing node inputs the temporal aggregated features into a preset dimensionality reduction mapping matrix, and transforms the high-dimensional features into a low-dimensional manifold space through linear mapping to generate the current semantic feature vector, the dimension of the current semantic feature vector being lower than the dimension of the original sensor data stream.

[0012] Preferably, after step S104, the method further includes: receiving incremental update messages in the cloud and extracting the current semantic feature vector from the incremental update messages to update the attribute parameters of the corresponding nodes in the global point topology map, thereby offsetting the decision lag caused by the delay in the backhaul of the original sensor data stream under the centralized processing mode.

[0013] Preferably, step S105 includes: constructing a spatiotemporal graph topology of a distributed advertising network in the cloud, wherein each advertising placement point corresponds to a node in the spatiotemporal graph topology; the spatiotemporal graph convolutional network model extracts the spatial correlation features and temporal evolution rules between nodes by performing synchronous convolution operations on the time axis and the spatial axis, and generates a probability distribution result representing the global commercial value distribution.

[0014] Preferably, the edge computing node is further provided with a high-frequency filtering module. Step S101 further includes: the high-frequency filtering module preprocesses the original sensing data stream, removes transient noise signals with a duration of less than 50ms, and reduces the invalid computational load of the local dimensionality reduction algorithm.

[0015] Preferably, the cloud monitors the load parameters of each edge computing node and the real-time congestion rate of the backbone network, and dynamically adjusts the asynchronous trigger threshold corresponding to each edge computing node, so that the edge-side state perception sensitivity and the global communication bandwidth utilization rate are maintained in a balanced range of 0.8 to 1.2.

[0016] Preferably, issuing advertising delivery instructions to target advertising locations includes: matching the target customer group tags and real-time commercial value weights of the target advertising locations in the cloud, and issuing advertising delivery instructions within the dwell time window corresponding to the target customer group tags.

[0017] Preferably, the encapsulation into an incremental update message in step S104 includes: the edge computing node performing a one-way irreversible hash operation on the location identifier, and encapsulating the hashed identifier with the current semantic feature vector with a timestamp.

[0018] Preferably, after step S105, the method further includes: the edge computing node receiving the advertising delivery instruction, parsing the content control parameters in the advertising delivery instruction, and calling the locally stored digital material resources to display targeted advertisements.

[0019] Compared to existing technologies, the intelligent advertising placement method based on AI and big data analysis of this invention has the following advantages:

[0020] 1. In the intelligent selection of advertising placement locations, the interaction between the local dimensionality reduction operator of the edge node and the asynchronous triggering mechanism based on distance deviation transforms the originally high-frequency and heavy-load raw sensor data stream into lightweight discrete state update pulses in situ at the edge. This breaks the rigid dependence of the centralized decision-making process on the uplink bandwidth of the backbone network, effectively alleviates the channel congestion caused by the concurrent communication of large-scale distributed nodes, and enables the system to maintain a stable state update frequency under the existing communication bandwidth limitations.

[0021] 2. The local semantic feature extraction executed on the edge side and the global topology synchronization logic driven asynchronously in the cloud work together to enable the topology node attributes maintained in the cloud to reflect changes in passenger flow characteristics in the physical space in real time. This eliminates the decision lag caused by the delay in the return of massive amounts of raw data in the traditional centralized processing mode, and ensures that the issued control commands are accurately matched with the dwell time window of the offline customer groups. This solves the mismatch between the allocation of commercial resources and the physical displacement of passenger flow from the perspective of logical timing.

[0022] 3. The high-frequency filtering module set up on the edge side and the dimensionality reduction execution logic pre-set in the cloud jointly establish a stability defense system. The edge side uses a time sliding window mechanism to filter feature vector distortions caused by non-target passenger flow. When the cloud detects communication interruption or computing power degradation on the edge side, it automatically extracts historical static benchmark data to participate in the calculation, ensuring that the distributed system can still output continuous and deterministic filtering results under complex network environment and fluctuating computing power constraints. Attached Figure Description

[0023] Figure 1 This is a flowchart of the intelligent filtering process for advertising locations using cloud-edge asynchronous collaboration in this invention;

[0024] Figure 2 This is a diagram of the spatiotemporal graph convolutional network architecture that supports real-time business value assessment according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0026] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0027] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0028] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0029] A method for intelligently selecting advertising placement locations based on AI and big data analysis includes the following steps:

[0030] Step S101: The edge computing nodes at each advertising placement point acquire the original sensor data stream within a preset spatial range and use a preset one-way irreversible hash algorithm to desensitize the entity identification information in the original sensor data stream.

[0031] In step S102, the edge computing node uses a pre-set local dimensionality reduction algorithm to extract the population density feature value and dwell time feature value of the spatial region within the current time window, and maps them to generate the current semantic feature vector representing the real-time passenger flow attributes of the location.

[0032] Step S103: The edge computing node calculates the Euclidean distance between the current semantic feature vector and the pre-stored historical feature vector to determine the deviation value of the state evolution of the representation point.

[0033] Step S104: When the deviation value exceeds the preset asynchronous trigger threshold, the edge computing node encapsulates the current semantic feature vector and the point identifier into an incremental update message and asynchronously uploads it to the cloud, overwriting the historical mirror status of the corresponding point stored in the cloud.

[0034] In step S105, the cloud inputs the received incremental update message into the preset spatiotemporal graph convolutional network model, determines the real-time commercial value weight of each advertising placement point through spatiotemporal correlation convolution operation, and issues advertising placement instructions to the target advertising placement point according to the real-time commercial value weight.

[0035] Preferably, step S102 includes the following sub-steps: Step S1021, the edge computing node uses a preset sliding time window to extract the desensitized original sensor data stream and counts the temporal aggregated features within the sliding time window; Step S1022, the edge computing node inputs the temporal aggregated features into a preset dimensionality reduction mapping matrix, and transforms the high-dimensional features into a low-dimensional manifold space through linear mapping to generate the current semantic feature vector, the dimension of the current semantic feature vector being lower than the dimension of the original sensor data stream.

[0036] Preferably, after step S104, the method further includes: receiving incremental update messages in the cloud and extracting the current semantic feature vector from the incremental update messages to update the attribute parameters of the corresponding nodes in the global point topology map, thereby offsetting the decision lag caused by the delay in the backhaul of the original sensor data stream under the centralized processing mode.

[0037] Preferably, step S105 includes: constructing a spatiotemporal graph topology of a distributed advertising network in the cloud, wherein each advertising placement point corresponds to a node in the spatiotemporal graph topology; the spatiotemporal graph convolutional network model extracts the spatial correlation features and temporal evolution rules between nodes by performing synchronous convolution operations on the time axis and the spatial axis, and generates a probability distribution result representing the global commercial value distribution.

[0038] Preferably, the edge computing node is further provided with a high-frequency filtering module. Step S101 further includes: the high-frequency filtering module preprocesses the original sensing data stream, removes transient noise signals with a duration of less than 50ms, and reduces the invalid computational load of the local dimensionality reduction algorithm.

[0039] Preferably, the cloud monitors the load parameters of each edge computing node and the real-time congestion rate of the backbone network, and dynamically adjusts the asynchronous trigger threshold corresponding to each edge computing node, so that the edge-side state perception sensitivity and the global communication bandwidth utilization rate are maintained in a balanced range of 0.8 to 1.2.

[0040] Preferably, issuing advertising delivery instructions to target advertising locations includes: matching the target customer group tags and real-time commercial value weights of the target advertising locations in the cloud, and issuing advertising delivery instructions within the dwell time window corresponding to the target customer group tags.

[0041] Preferably, the encapsulation into an incremental update message in step S104 includes: the edge computing node performing a one-way irreversible hash operation on the location identifier, and encapsulating the hashed identifier with the current semantic feature vector with a timestamp.

[0042] Preferably, after step S105, the method further includes: the edge computing node receiving the advertising delivery instruction, parsing the content control parameters in the advertising delivery instruction, and calling the locally stored digital material resources to display targeted advertisements.

[0043] Example 1: When high-concurrency dynamic passenger flow occurs in the core business district of a large transportation hub, the distributed advertising placement points output massive amounts of raw sensor data streams. If the centralized architecture is relied upon to continuously transmit the full amount of data, the uplink of the backbone network will be limited by physical transmission limits, resulting in data queue accumulation. The cloud will experience scheduling delays due to waiting for the underlying data to be transmitted back, causing the target customer group to physically shift when the advertising placement instructions output by the cloud are routed to the target advertising placement points, leading to computing power dissipation and mismatch of business information. Edge computing nodes deployed at each advertising placement point acquire the raw sensor data stream within a preset spatial range. They use a pre-set one-way irreversible hash algorithm to desensitize the entity identification information in the raw sensor data stream. The edge computing nodes use a pre-set local dimensionality reduction algorithm to extract the group density feature value and dwell time feature value of the spatial region within the current time window, and map them to generate a current semantic feature vector representing the real-time passenger flow attributes of the point. The current semantic feature vector replaces the raw spatial data.

[0044] Edge computing nodes calculate the Euclidean distance between themselves and pre-stored historical feature vectors to determine the deviation value representing the evolution of the location's state. When the deviation value exceeds a preset asynchronous trigger threshold, the edge computing node encapsulates the current semantic feature vector and the location identifier into an incremental update message and asynchronously uploads it to the cloud. The continuous raw sensor data stream is transformed into discrete pulses at the edge, covering the historical mirror state of the corresponding location stored in the cloud. The cloud inputs the received incremental update message into a preset spatiotemporal graph convolutional network model, determines the real-time commercial value weight of each advertising placement location through spatiotemporal correlation convolution operations, and issues advertising placement instructions to the target advertising placement locations based on the real-time commercial value weight. The cloud processing layer avoids the load of cleaning the underlying raw data, and the global communication network changes from continuous concurrent reporting of full data to state mapping driven by asynchronous incremental messages. Each issued advertising placement instruction is aligned with the actual dwell time window of the offline target customer group in both time and space dimensions.

[0045] Example 2: Constructing a distributed advertising placement screening test environment, selecting 500 edge computing nodes to simulate a spatial point matrix. The test uses a publicly available passenger flow trajectory dataset, superimposing Gaussian white noise with a signal-to-noise ratio of 15dB onto the passenger flow trajectory dataset, while randomly removing some sensing frames to simulate signal drift and personnel occlusion disturbances in the physical environment. The asynchronous trigger threshold setting needs to balance the synchronization accuracy of the cloud state and the uplink bandwidth consumption of the edge. When the asynchronous trigger threshold approaches the lower limit, the system responds to the underlying disturbance and triggers invalid message transmission; when the asynchronous trigger threshold approaches the upper limit, a time discontinuity occurs between the cloud state and the offline physical state. Based on the historical passenger flow mutation rate statistical distribution model, the asynchronous trigger threshold is set to 0.35. In the first control group, the system uploads all original sensor data. According to the data stream, the uplink bandwidth was measured at 12.4Gbps, and the end-to-end scheduling latency was measured at 1453ms. A large number of advertising delivery instructions missed the target customer dwell time window. In the test group, the edge computing nodes acquired the original sensor data stream superimposed with Gaussian white noise. The entity identification information in the original sensor data stream was desensitized using a pre-set one-way irreversible hash algorithm. The edge computing nodes used a pre-set local dimensionality reduction algorithm to extract the group density feature value and dwell time feature value of the spatial region within the current time window, and mapped to generate the current semantic feature vector representing the real-time passenger flow attributes of the location. The local dimensionality reduction algorithm filtered out the Gaussian white noise. The edge computing nodes calculated the Euclidean distance between the current semantic feature vector and the pre-stored historical feature vector to determine the deviation value of the status evolution of the location.

[0046] In the second control group, the asynchronous trigger threshold was set to 0.05. The edge computing node sent messages with an uplink bandwidth of 8.1 Gbps and an end-to-end scheduling latency of 678 ms. In the third control group, the asynchronous trigger threshold was set to 0.85, the uplink bandwidth was 0.4 Gbps, resulting in the loss of 73.2% of peak passenger flow, and an advertising placement command matching accuracy of 41.2%. In the fourth control group, the step of calculating the deviation was removed, and the edge computing node was set to periodically upload the current semantic feature vector with an upload period of 5000 ms. The uplink bandwidth was 1.8 Gbps, resulting in a scheduling blind spot lasting longer than 4900 ms. In the experimental group, the asynchronous trigger threshold was set to 0.35. When the deviation exceeded the asynchronous trigger threshold, the edge computing node matched the current semantic feature vector with the location identifier. The data is encapsulated as incremental update messages and asynchronously uploaded to the cloud, overwriting the historical mirror status of the corresponding locations stored in the cloud. The uplink bandwidth is measured at 0.6Gbps. The cloud inputs the received incremental update messages into a pre-set spatiotemporal graph convolutional network model. The real-time commercial value weight of each advertising placement location is determined through spatiotemporal correlation convolution operations. Based on the real-time commercial value weight, advertising placement instructions are issued to the target advertising placement locations. The end-to-end scheduling latency is measured at 113ms, and the advertising placement instruction matching accuracy is measured at 94.6%. The above test data confirms the synergistic effect of local dimensionality reduction feature extraction and calculation deviation triggering asynchronous upload, reducing the system's uplink bandwidth load. This solution replaces high-frequency continuous data streams with incremental messages, controls end-to-end scheduling latency, and ensures that the issued advertising placement instructions coincide with the actual dwell time window of the offline target customer group in the spatiotemporal dimension.

[0047] Example 3: When the system faces the task of generating real-time commercial value weights based on edge sensing data, the edge computing nodes deployed at each advertising placement point are set with a sliding time window of 500ms to capture the original sensor data stream that has been de-identified by a one-way irreversible hash algorithm. In this step, the edge computing nodes load a lightweight human target detection and multi-target tracking algorithm on the underlying processor. This algorithm scans the environmental contour image frame by frame to remove identity information, locks the pixel position of the bottom center of each customer flow entity, and the system calls the pre-calibrated camera intrinsic and extrinsic parameter matrices to accurately map the aforementioned two-dimensional pixel coordinates of the image to the physical plane two-dimensional coordinate points from the top view, thereby constructing a continuous and... The target motion trajectory of the time-series label is calculated. The edge computing node counts the number of target coordinate points in the spatial region within the sliding time window and calculates the population density feature value. The edge computing node also counts the average duration of target coordinate points with spatial displacement less than 0.5m and calculates the dwell time feature value. The edge computing node concatenates the population density feature value and the dwell time feature value into a two-dimensional input vector and inputs this two-dimensional input vector into a preset multilayer perceptron. The multilayer perceptron includes an input layer, two hidden layers with 64 computing nodes each, and an output layer. The multilayer perceptron applies linear weighting and nonlinear activation functions and outputs a one-dimensional array containing 128 floating-point values. This one-dimensional array constitutes the current semantic feature vector.

[0048] Based on the principle of orthogonal transformation for data dimensionality reduction, the core physical variance of the feature space principal component direction mapping system is used as an alternative to the aforementioned nonlinear multilayer perceptron dimensionality reduction method. Edge computing nodes extract population density and dwell time feature values ​​from historical observation windows to construct a sample matrix. They calculate the covariance matrix of the sample matrix and extract the principal component feature vectors corresponding to the largest eigenvalues. Based on this, a one-dimensional dimensionality reduction mapping matrix is ​​constructed. The edge computing nodes perform matrix multiplication and linear mapping operations with the vector formed by the current temporal aggregated features and the dimensionality reduction mapping matrix, directly outputting the current semantic feature vector in scalar form, reducing the ineffective computational load on the edge side. The cloud receives incremental update messages asynchronously uploaded by the edge computing nodes, using each advertising placement location as a graph. The nodes construct a spatial topology graph and determine the edge relationships based on the actual geographical distance between each advertising placement location. When the geographical distance between two advertising placement locations is less than 50m, the cloud assigns the adjacency matrix element of the spatial topology graph to the reciprocal of that geographical distance. Based on the law of gravitational potential energy distance decay, the topology parameter divergence phenomenon caused by nodes approaching each other infinitely is eliminated. The cloud sets the adjacency matrix element value range calculation formula to 1 / (d+ϵ), where the variable d represents the actual geographical distance between two nodes, and the constant ϵ represents the spatial smoothing factor used to suppress gradient explosion, which is fixed at 1.5 meters. When the geographical distance between two advertising placement locations is greater than or equal to 50m, the cloud assigns the adjacency matrix element to 0.

[0049] The pre-built spatiotemporal graph convolutional network model in the cloud contains three cascaded spatiotemporal convolutional blocks. Each spatiotemporal convolutional block sequentially connects a spatial graph convolutional layer and a temporal convolutional layer. The spatial graph convolutional layer multiplies the current semantic feature vector corresponding to each graph node with the adjacency matrix to extract the aggregated features of the spatial neighborhood of the point. The temporal convolutional layer has a one-dimensional convolutional kernel with a stride of 1. The temporal convolutional layer calculates a one-dimensional sliding convolution on the aggregated features of the spatial neighborhood along the timestamp sequence of each incremental update message arriving at the cloud, and outputs a spatiotemporal fusion feature matrix. Considering that the incremental update messages returned by the asynchronous triggering mechanism are non-uniformly distributed on the time axis, before performing the above one-dimensional sliding convolution operation, the temporal convolutional layer performs linear interpolation and forward zero-padded operations on the timestamp sequence of the message with a preset standard clock period as the equidistant stride through the built-in time series alignment logic. Specifically, when the time interval between two adjacent actual messages is greater than the standard clock cycle, the system uses the current semantic feature vectors of the two adjacent messages to calculate a linear transition state to fill the missing time node. This transforms the originally physically asynchronous and non-uniformly distributed discrete feature pulses into a standardized sequence input that is equidistant and continuous in the time dimension, thereby satisfying the homogeneity requirement of the standard one-dimensional convolutional kernel for the input data matrix. The end of the spatiotemporal graph convolutional network model is connected to a fully connected layer. The fully connected layer receives the spatiotemporal fusion feature matrix output by the third spatiotemporal convolutional block. The Sigmoid activation function is applied to map the spatiotemporal fusion feature matrix into a floating-point scalar with a value between 0.0 and 1.0. The cloud determines the floating-point scalar as the real-time commercial value weight of each advertising placement point and issues advertising placement instructions to the target advertising placement point according to the real-time commercial value weight.

[0050] Example 4: When the system faces the situation of its first offline deployment in a large commercial complex physical space that has never been deployed before, the edge computing nodes deployed at each advertising placement point initiate a 7-day on-site pre-baseline calibration procedure. The edge computing nodes acquire the original sensor data streams that are not superimposed with commercial advertising activities during this period, use a pre-set one-way irreversible hash algorithm to desensitize entity identity information, use a pre-set local dimensionality reduction algorithm to extract the group density feature value and dwell time feature value of the spatial area within a specific time window, and map them to generate an initial semantic feature vector representing the basic passenger flow attributes of the point. The edge computing nodes calculate the arithmetic mean of the initial semantic feature vectors for the same time period within 7 consecutive natural days, output the arithmetic mean feature vector, and write the arithmetic mean feature vector into the local non-volatile storage medium. The arithmetic mean feature vector is determined as a pre-stored historical feature vector for subsequent Euclidean distance calculation, and a basic data table recording the basic passenger flow parameters of this specific physical space is generated.

[0051] The cloud synchronously initiates offline calibration and data filling procedures for the pre-built spatiotemporal graph convolutional network model. The cloud acquires sample datasets containing initial semantic feature vectors uploaded by each edge computing node, and constructs a supervised learning task by combining historical advertising conversion rate tags. The cloud constructs an initial spatial topology graph with each advertising placement location as a graph node. The cloud uses the sample dataset to calculate the backpropagation gradient of the alternating spatial graph convolutional layer and temporal convolutional layer in the spatiotemporal graph convolutional network model. The network weight parameters are continuously updated according to the preset loss function until the loss function converges to within the predetermined error threshold. The cloud solidifies the calibrated spatiotemporal graph convolutional network model and connects it to the production scheduling network to generate a real-time commercial value weight mapping benchmark adapted to the specific spatial topology features.

[0052] Example 5: When the system faces the task of asynchronously calibrating the parameters of each advertising placement point within a specific physical space, the edge computing nodes associated with each advertising placement point acquire raw sensor data streams over 24 consecutive natural hours. The edge computing nodes calculate the deviation value within each minute window and generate a deviation value sample sequence reflecting the fluctuations in passenger flow within that physical space. Based on this, the edge computing nodes calculate the sample standard deviation σ of the deviation value sample sequence and select a sensitivity factor k within the range of 2.0 to 3.0. The edge computing nodes then apply the formula... The asynchronous trigger threshold for that location is calculated and established, thereby generating a judgment benchmark that is adapted to the deployment environment of a specific location.

[0053] The frequency of physical domain event triggering and the channel capacity of the communication network follow a dynamic balance. The cloud executes a closed-loop adaptive adjustment procedure for the asynchronous triggering threshold, setting a statistical time window of 10 minutes. The total number of incremental update messages actually reported by edge computing nodes within the statistical time window is calculated, and the ratio of this total to the total amount within the sliding time window is used to determine the dimensionless sensing sensitivity S. Simultaneously, the cloud collects the actual throughput data traffic of the corresponding edge computing nodes, and calculates the ratio of this actual throughput data traffic to the theoretical bandwidth limit pre-allocated to the node's backbone network, determining the dimensionless bandwidth utilization rate B. The cloud calculates the evaluation coefficient R, with the equation R = S / B. If the evaluation coefficient R is greater than 1.2 for two consecutive statistical time windows, the cloud sends a parameter adjustment instruction to the target edge computing node, increasing the internal preset sensitivity factor k in increments of 0.05 until the evaluation coefficient R falls back to the range of 0.8 to 1.2. Conversely, if the evaluation coefficient R is lower than 0.8, the sensitivity factor k is decreased by the same increment. This fixed increment or decrement of 0.05 is controlled through closed-loop control in offline mode. Simulation tests confirmed that if the step size exceeds 0.1, the system state equation is prone to overshoot during feedback adjustment, causing high-frequency oscillations in the sensitivity parameter at the interval boundary. If the step size is below 0.01, the control loop response is sluggish and cannot promptly suppress channel congestion caused by large-scale passenger flow. 0.05 is the optimal control empirical value that balances convergence speed and system stability under the constraint of the specific width of the balance interval, ensuring that the system can complete parameter optimization within a finite number of iterations. Based on the state-space Markov time-series evolution mechanism, the basic passenger flow characteristics of the physical scene undergo systematic drift over long periods of time. The edge computing node is built-in and executes the spatial baseline reconstruction procedure, extracting the arithmetic mean feature vector collected during the low-load period in the early morning of each day according to the natural day cycle. Using an exponential moving average algorithm with a smoothing coefficient of 0.1, the extracted arithmetic mean feature vector is weighted and fused with the locally pre-stored historical feature vector, refreshing and overwriting the historical mirror data in the local non-volatile storage medium, and suppressing the long-term aging of the measurement benchmark.

[0054] When the cloud platform needs to calibrate the spatiotemporal feature extraction accuracy of a pre-defined spatiotemporal graph convolutional network model, it obtains incremental update messages containing the current semantic feature vector from each edge computing node. The cloud sets a temporal convolutional kernel with a stride of 1 within the spatiotemporal convolutional block of the spatiotemporal graph convolutional network model to capture the dynamic evolution of passenger flow between adjacent message sequences. The cloud aggregates the group density and dwell time feature values ​​within the spatial topology graph neighborhood through spatial graph convolutional layers, and uses a fully connected layer combined with a sigmoid activation function to calculate the processing result for this spatial topology graph neighborhood. Therefore, the cloud determines this processing result as a floating-point scalar between 0.0 and 1.0 and establishes this floating-point scalar as... The real-time commercial value weight of the corresponding location; to support the filtering logic of issuing advertising delivery instructions to the target advertising delivery location, after determining the real-time commercial value weight, the cloud calls the preset filtering gateway module. The filtering gateway module obtains the business feature matrix of the candidate ad set, and calculates the Hadamard product of the real-time commercial value weight and the business feature matrix to produce the matching degree score of each candidate ad. The filtering gateway module selects the top N candidate ads according to the descending order of the matching degree score, and encapsulates them into an advertising delivery instruction containing multimedia material index information. The cloud then distributes the advertising delivery instruction to the corresponding edge computing node to complete the intelligent filtering process of advertising content for the physical space location.

[0055] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of protection of this application, and all of these forms are within the protection scope of this application.

Claims

1. A method for intelligent selection of advertising placement locations based on AI and big data analysis, characterized in that, Includes the following steps: Step S101: The edge computing nodes at each advertising placement point acquire the original sensor data stream within a preset spatial range and use a preset one-way irreversible hash algorithm to desensitize the entity identification information in the original sensor data stream. In step S102, the edge computing node uses a pre-set local dimensionality reduction algorithm to extract the population density feature value and dwell time feature value of the spatial region within the current time window, and maps them to generate the current semantic feature vector representing the real-time passenger flow attributes of the location. Step S103: The edge computing node calculates the Euclidean distance between the current semantic feature vector and the pre-stored historical feature vector to determine the deviation value of the state evolution of the representation point. Step S104: When the deviation value exceeds the preset asynchronous trigger threshold, the edge computing node encapsulates the current semantic feature vector and the point identifier into an incremental update message and asynchronously uploads it to the cloud, overwriting the historical mirror status of the corresponding point stored in the cloud. In step S105, the cloud inputs the received incremental update message into the preset spatiotemporal graph convolutional network model, determines the real-time commercial value weight of each advertising placement point through spatiotemporal correlation convolution operation, and issues advertising placement instructions to the target advertising placement point according to the real-time commercial value weight.

2. The intelligent selection method for advertising placement locations based on AI and big data analysis according to claim 1, characterized in that, Step S102 includes the following sub-steps: Step S1021, the edge computing node uses a preset sliding time window to extract the desensitized original sensor data stream and counts the temporal aggregated features within the sliding time window; Step S1022, the edge computing node inputs the temporal aggregated features into a preset dimensionality reduction mapping matrix, and transforms the high-dimensional features into a low-dimensional manifold space through linear mapping to generate the current semantic feature vector, the dimension of which is lower than the dimension of the original sensor data stream.

3. The intelligent selection method for advertising placement locations based on AI and big data analysis according to claim 1, characterized in that, Step S104 and thereafter includes: receiving incremental update messages in the cloud and extracting the current semantic feature vector in the incremental update messages to update the attribute parameters of the corresponding nodes in the global point topology map, so as to offset the decision lag caused by the delay in the backhaul of the original sensor data stream under the centralized processing mode.

4. The intelligent selection method for advertising placement locations based on AI and big data analysis according to claim 1, characterized in that, Step S105 includes: constructing a spatiotemporal graph topology of a distributed advertising delivery network in the cloud, wherein each advertising delivery point corresponds to a node in the spatiotemporal graph topology; the spatiotemporal graph convolutional network model extracts the spatial correlation features and temporal evolution rules between nodes by performing synchronous convolution operations on the time axis and the spatial axis, and generates a probability distribution result representing the global commercial value distribution.

5. The intelligent selection method for advertising placement locations based on AI and big data analysis according to claim 1, characterized in that, The edge computing node is also equipped with a high-frequency filtering module. Step S101 further includes: the high-frequency filtering module preprocesses the original sensing data stream, removes transient noise signals with a duration of less than 50ms, and reduces the invalid computational load of the local dimensionality reduction algorithm.

6. The intelligent selection method for advertising placement locations based on AI and big data analysis according to claim 1, characterized in that, The cloud monitors the load parameters of each edge computing node and the real-time congestion rate of the backbone network, and dynamically adjusts the asynchronous trigger threshold corresponding to each edge computing node to keep the edge-side state perception sensitivity and global communication bandwidth utilization within a balanced range of 0.8 to 1.

2.

7. The intelligent selection method for advertising placement locations based on AI and big data analysis according to claim 1, characterized in that, Sending advertising delivery instructions to target advertising locations includes: matching the target customer tags and real-time commercial value weights of the target advertising locations in the cloud, and sending advertising delivery instructions within the dwell time window corresponding to the target customer tags.

8. The intelligent selection method for advertising placement locations based on AI and big data analysis according to claim 1, characterized in that, The step S104 of encapsulating the incremental update message includes: the edge computing node performs a one-way irreversible hash operation on the point identifier, and encapsulates the hashed identifier with the current semantic feature vector with a timestamp.

9. The intelligent selection method for advertising placement locations based on AI and big data analysis according to claim 1, characterized in that, Step S105 and beyond also includes: the edge computing node receiving the advertising delivery instruction, parsing the content control parameters in the advertising delivery instruction, and calling the locally stored digital material resources to display targeted advertisements.

Images