Advertisement player content intelligent scheduling system based on cloud platform

Abstract

The invention relates to the technical field of advertisement scheduling, in particular to an intelligent advertisement player content scheduling system based on a cloud platform, and the system comprises a feature vector mapping module which is used for constructing an advertisement content feature vector; the global perception acquisition module is used for outputting audience characteristics and equipment operation environment data; the multi-factor weight calculation module is used for executing weighting operation by using the frequency attenuation factor and the environment adaptability correction coefficient to generate a real-time scheduling priority index; and the dynamic queue rearrangement module is used for executing to-be-played content queue sorting. In the invention, the task urgency item weight is corrected by using the environment illumination intensity data to eliminate the strong light interference, the frequency attenuation factor is introduced to carry out fatigue compensation on the audience feature matching degree, and the queue sorting logic is adjusted in combination with the network link quality factor, so that the transformation of the putting strategy from static carousel to environment adaptive dynamic scheduling is realized; the problems of playing lagging and audience sight escape in complex illumination and network fluctuation scenes are solved.

Description

A cloud-based intelligent content scheduling system for advertising machines Technical Field

[0001] This invention relates to the field of advertising scheduling technology, and in particular to an intelligent scheduling system for advertising machine content based on a cloud platform. Background Technology

[0002] The field of advertising scheduling technology involves a system architecture that uses digital means to automatically manage the playback order, duration, and frequency of content on commercial display terminals. Its core lies in using the interaction between server clusters and terminal devices to realize the storage, distribution, and playback control of advertising materials. It is widely used in information dissemination and commercial promotion in public scenarios such as shopping malls, buildings, and transportation hubs.

[0003] Traditional advertising machine content intelligent scheduling systems typically adopt a mechanical rotation mode based on a preset schedule. This ignores the nonlinear interference of sudden changes in ambient light intensity on screen visibility, lacks dynamic quantification of the diminishing marginal utility of audience characteristics and content matching as playback frequency increases, and cannot establish a real-time mapping mechanism between content bitrate and transmission bandwidth when network link quality experiences high-frequency fluctuations. This results in low effective reach of advertising content in complex business scenarios and ineffective redundant consumption of terminal computing resources. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a cloud-based intelligent scheduling system for advertising machine content.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a cloud-based intelligent scheduling system for advertising machine content includes: a feature vector mapping module, used to parse material metadata and perform spatial mapping, construct advertising content feature vectors, and establish a queue of content to be played based on time window constraints; a global perception acquisition module, used to drive cameras and light sensors to capture audience features and ambient light intensity data, and collect network bandwidth data to generate device operating environment data; a multi-factor weight calculation module, used to calculate the cosine similarity between advertising content feature vectors and audience features to obtain audience feature matching degree, generate a frequency decay factor based on the complement of the playback frequency ratio, generate a network link quality factor and an environmental adaptability correction coefficient based on device operating environment data mapping, perform weighted summation on the product of audience feature matching degree and frequency decay factor and the task urgency term based on the remaining time ratio, and use the network link quality factor and environmental adaptability correction coefficient to correct the summation result to generate a real-time scheduling priority index; and a dynamic queue reordering module, used to sort the queue of content to be played in descending order according to the real-time scheduling priority index, and output a target content loading instruction according to the sorting result.

[0006] As a further aspect of the present invention, the feature vector mapping module specifically includes: a metadata parsing submodule, used to obtain the tag data, duration attribute, and resolution parameters of the advertising material, perform word segmentation and stop word removal on the tag data, and generate a metadata descriptor sequence; a spatial mapping submodule, used to call a pre-trained semantic embedding model to perform vectorization transformation on the metadata descriptor sequence, mapping the discrete tag data to a high-dimensional continuous vector space, and generating an advertising content feature vector; and a queue initialization submodule, used to obtain the current system timestamp and the preset advertising delivery schedule, filter out advertising materials within the effective delivery time window, and associate the candidate material index and its corresponding advertising content feature vector according to the filtering results to establish a queue of content to be played.

[0007] As a further aspect of the present invention, the all-domain perception acquisition module specifically includes: a visual feature extraction submodule, used to activate the camera to capture video stream frame images, extract facial key point data in the frame images using a face recognition algorithm, analyze the age distribution, gender ratio, and gaze duration of the audience, and generate audience features; and an environmental data generation submodule, used to read the analog voltage signal output by the light sensor and convert it into an ambient light intensity value, while monitoring the uplink and downlink transmission rates and packet loss rates of the network interface, and combining and encapsulating the ambient light intensity value, transmission rate, and packet loss rate to generate device operating environment data.

[0008] As a further aspect of the present invention, the multi-factor weight calculation module specifically includes: a similarity calculation submodule, used to obtain the advertising content feature vector and audience features, calculate the cosine value of the angle between the two vectors in multi-dimensional space, and normalize the calculation result as the audience feature matching degree; a parameter factor generation submodule, used to obtain the number of times the advertising material has been played and the preset maximum playback limit, calculate the ratio between the two and take the complement to obtain the frequency decay factor, and at the same time, according to the preset mapping rule table, map the network indicators in the device operating environment data to the network link quality factor, and map the illumination indicators to the environmental adaptability correction coefficient; and an index weighted correction submodule, used to obtain the remaining effective duration and total duration of the advertising task, calculate the task urgency items, and perform multi-dimensional weighted correction operation by combining the audience feature matching degree, frequency decay factor, network link quality factor and environmental adaptability correction coefficient to generate a real-time scheduling priority index.

[0009] As a further aspect of the present invention, the dynamic queue rearrangement module specifically includes: a sequence descending sorting submodule, used to traverse each material node in the queue of content to be played, associate the real-time scheduling priority index with the corresponding material node, and use a quick sorting algorithm to rearrange all material nodes in the queue in descending order of the index; and an instruction encapsulation and output submodule, used to obtain the storage path and checksum of the first material node in the rearranged queue, and encapsulate the material loading path, decoding parameters and playback timing requirements in combination with the current rendering engine state to generate the target content loading instruction.

[0010] As a further aspect of the present invention, the process of generating the advertising content feature vector specifically includes: obtaining a sequence of metadata descriptors; calculating the weight value of each descriptor in the sequence using the TF-IDF algorithm to construct a weighted bag-of-words model; converting the metadata descriptor sequence into a word embedding matrix using the Word2Vec algorithm; performing a weighted average pooling operation on the word embedding matrix in combination with the weight values ​​to extract the semantic core features of the material; and performing L2 regularization on the pooled feature vector to eliminate the influence of vector magnitude differences on similarity calculation, thereby generating the advertising content feature vector.

[0011] As a further aspect of the present invention, the audience feature extraction process specifically includes: acquiring real-time monitoring frame images captured by a camera; performing target detection on the frame images using a multi-task convolutional neural network to locate the bounding box of the face region; extracting facial geometric features composed of 68 feature points within the bounding box of the face region; inputting these features into an attribute classifier to predict the age and gender labels of the audience; and calculating the gaze attention based on the gaze angle deviation; and encoding the age label, gender label, and gaze attention according to a preset dimensional splicing rule to generate audience features.

[0012] As a further aspect of the present invention, the process of generating the frequency decay factor specifically includes: obtaining the cumulative number of times the current advertising material has been played within the current campaign period. and the preset maximum number of allowed playbacks Based on the formula: ; Calculate the frequency attenuation factor; where, Represents the frequency decay factor. The steepness coefficient represents the rate of decay. This represents the threshold bias that triggers decay. is the base of the natural logarithm.

[0013] As a further aspect of the present invention, the calculation process of the real-time scheduling priority index specifically includes: obtaining audience feature matching degree, frequency decay factor, task remaining time ratio, network link quality factor, and environmental adaptability correction coefficient; based on the formula: Perform fusion computation to obtain the real-time scheduling priority index; where, Represents the real-time scheduling priority index. Represents the degree of matching with audience characteristics. Represents the frequency decay factor. The percentage of remaining time for the representative task. Represents the interest matching weight coefficient. The weighting coefficient represents the urgency of the task. Represents the network link quality factor. This represents the environmental adaptability correction factor.

[0014] As a further aspect of the present invention, the process of generating the target content loading instruction specifically includes: obtaining the cloud storage URL address and local cache hash value of the sorted target material, and detecting whether the corresponding material file exists in the local cache; if the local cache exists and the hash value matches, then a direct read instruction is generated; if the local cache does not exist or the hash value does not match, then the cloud URL address, the segmented download strategy, and the breakpoint resume flag are encapsulated to generate a network loading instruction, and the network loading instruction is encapsulated as a target content loading instruction.

[0015] Compared with existing technologies, the advantages and positive effects of this invention are as follows: In this invention, by constructing a priority dynamic calculation model based on multi-dimensional perception data, using ambient light intensity data to correct the weight of task urgency items to eliminate strong light interference, introducing a frequency attenuation factor to compensate for audience feature matching degree fatigue, and combining network link quality factors to adjust the sorting logic of the queue of content to be played in real time, the advertising content delivery strategy has been transformed from static carousel to environmentally adaptive dynamic scheduling. This solves the problems of playback stuttering and audience gaze escape in complex lighting and network fluctuation scenarios, and significantly improves the dissemination efficiency and resource utilization of commercial display terminals. Attached Figure Description

[0016] Figure 1 is a block diagram of the intelligent scheduling system for advertising machine content based on the cloud platform of the present invention; Figure 2 is a flowchart of the operation of the feature vector mapping module of the present invention; Figure 3 is a flowchart of the operation of the global perception and acquisition module of the present invention; Figure 4 is a flowchart of the operation of the multi-factor weight calculation module of the present invention; Figure 5 is a flowchart of the operation of the dynamic queue rearrangement module of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0018] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0019] Please refer to Figures 1 and 2. The present invention provides a technical solution: an intelligent scheduling system for advertising machine content based on a cloud platform includes: a feature vector mapping module, used to parse material metadata and perform spatial mapping, construct advertising content feature vectors, and establish a queue of content to be played based on time window constraints.

[0020] Specifically, the feature vector mapping module includes: a metadata parsing submodule, used to obtain the tag data, duration attribute, and resolution parameters of the advertising materials, perform word segmentation and stop word removal on the tag data, and generate a metadata descriptor sequence; a spatial mapping submodule, used to call a pre-trained semantic embedding model to perform vectorization transformation on the metadata descriptor sequence, mapping the discrete tag data to a high-dimensional continuous vector space, and generating advertising content feature vectors; and a queue initialization submodule, used to obtain the current system timestamp and the preset advertising delivery schedule, filter out advertising materials within the effective delivery time window, and associate the candidate material index and its corresponding advertising content feature vector based on the filtering results to establish a queue of content to be played.

[0021] The process of generating the ad content feature vector specifically includes: obtaining a sequence of metadata descriptors; using the TF-IDF algorithm to calculate the weight value of each descriptor in the sequence and constructing a weighted bag-of-words model; using the Word2Vec algorithm to convert the metadata descriptor sequence into a word embedding matrix; combining the weight values ​​to perform a weighted average pooling operation on the word embedding matrix to extract the semantic core features of the material; and performing L2 regularization on the pooled feature vector to eliminate the influence of vector magnitude differences on similarity calculation, thereby generating the ad content feature vector.

[0022] The feature vector mapping module, as the core data preprocessing component of the intelligent scheduling system, plays a crucial role in transforming unstructured advertising materials into high-dimensional semantic vectors that computers can understand. The metadata parsing submodule receives raw material metadata data packages from the cloud-based content management system through a standardized data interface. These metadata data packages are typically encapsulated in JSON format and include the ad title, descriptive tags, duration attributes, resolution parameters, and file encoding format. The metadata parsing submodule first performs a cleaning operation on the input tag data fields, using pre-defined regular expressions to remove Hypertext Markup Language tags, special symbols, and invisible characters. Subsequently, this submodule calls a Chinese word segmentation engine to segment the text content. To improve the purity of feature extraction, the submodule loads a stop word library containing 1800 common stop words, iterates through the segmentation results, and removes matches, ultimately generating a clean sequence of metadata descriptors. For example, for the input text "This is a shock-absorbing running shoe designed specifically for marathons," the processed output descriptor sequence would be "marathon," "design," "shock absorption," and "running shoes."

[0023] The aforementioned Chinese word segmentation engine refers to a statistical or rule-based natural language processing tool used to segment continuous Chinese character sequences into independent lexical units according to semantic logic. Common algorithms include prefix dictionary-based matching algorithms or statistical algorithms based on hidden Markov models.

[0024] The spatial mapping submodule is responsible for mapping the discrete symbol sequences to a continuous vector space. This submodule integrates a pre-trained Word2Vec deep learning model. This model uses a Skip-gram architecture to predict context from the center word. The model architecture includes an input layer, hidden layers, and an output layer: the input layer size corresponds to the vocabulary dimension, set to 50,000 high-frequency words in this embodiment; the hidden layer is 128-dimensional, serving as the dimension space for the feature vectors, and does not use an activation function to maintain linear mapping properties; the output layer uses a hierarchical Softmax structure to optimize computational efficiency. During training, cross-entropy is used as the loss function, and the weight matrix is ​​updated using stochastic gradient descent. The learning rate is set to 0.025, and the window size is set to 5. In the inference phase, the spatial mapping submodule first calculates the word frequency and inverse document frequency (IVF) of each word in the descriptor sequence in the corpus. The product of the word frequency and IVF is used as a weight coefficient to perform weighted average pooling on the 128-dimensional word vectors corresponding to each descriptor output by the Word2Vec model. Specifically, each word vector is multiplied by its corresponding weight, and the results are accumulated along each dimension. Finally, the sum of the weights is divided to synthesize an initial feature vector representing the global semantics of the ad creative. To eliminate the influence of the vector magnitude on subsequent similarity calculations, the submodule further performs L2 regularization on the initial feature vector. This involves calculating the Euclidean norm of the vector and dividing each element of the vector by the norm, ensuring that the final generated ad content feature vector has a magnitude of 1 and is distributed on a unit hypersphere.

[0025] The Word2Vec deep learning model mentioned above refers to a shallow neural network model used to generate word embeddings. It can map words to real vectors, making semantically similar words closer together in the vector space, thereby capturing the semantic relationships between words.

[0026] The TF-IDF algorithm mentioned above refers to a statistical method for evaluating the importance of a word to a document in a document set or corpus, where word frequency represents the frequency of a word's occurrence in a document, and inverse document frequency represents a measure of the word's general importance.

[0027] The queue initialization submodule is primarily responsible for time-domain filtering and data association. This submodule synchronizes high-precision timestamps from the network time protocol server in real time and reads the ad delivery schedule pre-stored in the Redis cache database. This schedule includes the start and end dates of each ad creative and the effective delivery time slots each day. The submodule iterates through all creative indices with generated feature vectors, comparing whether the current system timestamp falls within the effective delivery time window. If the current time is 10:30:00 AM on January 25, 2026, and the delivery time slot of a certain creative covers this time point, it is determined as a candidate creative. The submodule establishes a key-value pair mapping between the unique identifier of the selected candidate creative and its corresponding 128-dimensional ad content feature vector, constructs a queue of content to be played, and serializes and stores this queue structure in memory for subsequent modules to call.

[0028] The aforementioned Redis caching database refers to a memory-based key-value storage system with extremely high read and write speeds. It is often used to cache frequently accessed data to reduce the access pressure on the backend database and supports various data structures such as strings, hashes, and lists.

[0029] Please refer to Figures 1 and 3. The global perception acquisition module is used to drive the camera and light sensor to capture audience characteristics and ambient light intensity data, and to collect network bandwidth data to generate device operating environment data.

[0030] Specifically, the global perception acquisition module includes: a visual feature extraction submodule, which is used to activate the camera to capture video stream frame images, extract facial key point data in the frame images using a face recognition algorithm, analyze the age distribution, gender ratio, and gaze duration of the audience, and generate audience features; and an environmental data generation submodule, which is used to read the analog voltage signal output by the light sensor and convert it into an ambient light intensity value, while monitoring the uplink and downlink transmission rates and packet loss rate of the network interface, and combining and encapsulating the ambient light intensity value, transmission rate, and packet loss rate to generate device operating environment data.

[0031] The audience feature extraction process specifically includes: acquiring real-time monitoring frame images captured by the camera; performing target detection on the frame images using a multi-task convolutional neural network to locate the bounding box of the face region; extracting facial geometric features composed of 68 feature points within the bounding box of the face region, inputting them into an attribute classifier to predict the age and gender labels of the audience, and calculating the gaze attention based on the gaze angle deviation; and encoding the age label, gender label, and gaze attention according to a preset dimensional splicing rule to generate audience features.

[0032] The global perception acquisition module constructs a digital mapping of the physical environment and the state of the audience through a multi-dimensional sensor array. The visual feature extraction submodule drives a high-resolution industrial-grade camera to continuously capture video streams of the monitored area via a Universal Serial Bus 3.0 interface. This submodule has built-in frame extraction logic, extracting keyframes at a frequency of 5 frames per second and sending them into the processing pipeline to balance computational load and real-time performance. The core of the submodule integrates a multi-task convolutional neural network for face detection and feature point localization. This network consists of three cascaded networks: P-Net for quickly generating candidate windows; R-Net for filtering false detection windows through bounding box regression; and O-Net for finally outputting accurate face bounding boxes and key points. Based on the face region output by O-Net, the visual feature extraction submodule further runs a deep attribute analysis network based on the ResNet-50 architecture. The input layer of this network receives a face image normalized to 224 by 224 pixels, and extracts deep texture features through 5 convolutional stages (including convolutional layers, batch normalization layers, linear rectified functions, and max pooling layers). In the fully connected layer, the network branches output prediction results in three dimensions: age group classification, gender classification, and head pose angle. Gaze intensity is calculated based on head pose angle; when the absolute values ​​of both yaw and pitch angles are less than 15 degrees, it is considered emmetropia, and the gaze duration counter is incremented; otherwise, it is considered non-gaze. Finally, the submodule concatenates the predicted age and gender labels with the normalized gaze intensity values ​​to form an audience feature data package.

[0033] The aforementioned Multi-Task Convolutional Neural Network (MTCNN) refers to a face detection algorithm based on deep learning. It simultaneously performs face detection, bounding box regression, and facial landmark localization tasks through three cascaded convolutional networks, achieving high detection accuracy and speed.

[0034] The ResNet-50 architecture mentioned above refers to a deep residual network model that solves the gradient vanishing and degradation problems in deep neural networks by introducing residual connections. It contains 50 convolutional layers and fully connected layers and is commonly used for image classification and feature extraction tasks.

[0035] The environmental data generation submodule is responsible for collecting and quantifying physical environmental parameters. This submodule connects to a high-sensitivity light sensor and reads ambient light intensity data via an I2C bus. The sensor outputs raw 16-bit binary values, which the submodule converts to lux-based light intensity values. Simultaneously, this submodule polls the device's communication module or wireless network interface using a Simple Network Management Protocol (SMMP) to obtain the current downlink transmission rate and network packet loss rate. To eliminate data jitter, the submodule employs a moving average filtering algorithm, taking the average of the most recent 10 samples as the current network status indicator. Finally, the ambient light intensity value, average transmission rate, and packet loss rate are encapsulated into JSON-formatted device operating environment data and published via the internal message bus.

[0036] Table 1 Parameter Table of the Global Perception and Acquisition Module As shown in Table 1, the system is configured with high-resolution visual acquisition parameters and clear judgment thresholds to ensure the accuracy of input data.

[0037] The aforementioned moving average filtering algorithm is a digital signal processing technique that smooths data by calculating the arithmetic mean of data within a fixed-length window in a data sequence, effectively suppressing random noise interference.

[0038] Please refer to Figures 1 and 4. The multi-factor weight calculation module is used to calculate the cosine similarity between the ad content feature vector and the audience features to obtain the audience feature matching degree. It generates a frequency decay factor based on the complement of the playback frequency ratio, generates a network link quality factor and an environmental adaptability correction coefficient based on the device operating environment data mapping, performs a weighted summation on the product of the audience feature matching degree and the frequency decay factor and the task urgency term based on the remaining time ratio, and uses the network link quality factor and the environmental adaptability correction coefficient to correct the summation result to generate a real-time scheduling priority index.

[0039] Specifically, the multi-factor weight calculation module includes: a similarity calculation submodule, used to obtain the feature vector of the advertising content and the audience features, calculate the cosine value of the angle between the two vectors in the multi-dimensional space, and normalize the calculation result as the audience feature matching degree; a parameter factor generation submodule, used to obtain the number of times the advertising material has been played and the preset maximum playback limit, calculate the ratio between the two and take the complement to obtain the frequency decay factor, and at the same time, according to the preset mapping rule table, map the network indicators in the device operating environment data to the network link quality factor, and map the illumination indicators to the environmental adaptability correction coefficient; and an index weighted correction submodule, used to obtain the remaining effective duration and total duration of the advertising task, calculate the task urgency items, and perform multi-dimensional weighted correction calculation by combining the audience feature matching degree, frequency decay factor, network link quality factor and environmental adaptability correction coefficient to generate a real-time scheduling priority index.

[0040] The generation process of the frequency decay factor specifically includes: obtaining the cumulative number of times the current ad creative has been played within the current campaign period. and the preset maximum number of allowed playbacks Based on the formula: ; Calculate the frequency attenuation factor; where, Represents the frequency decay factor. The steepness coefficient represents the rate of decay. This represents the threshold bias that triggers decay. is the base of the natural logarithm.

[0041] The calculation process for the real-time scheduling priority index specifically includes: obtaining audience feature matching degree, frequency decay factor, task remaining time percentage, network link quality factor, and environmental adaptability correction coefficient; based on the formula: Perform fusion computation to obtain the real-time scheduling priority index; where, Represents the real-time scheduling priority index. Represents the degree of matching with audience characteristics. Represents the frequency decay factor. The percentage of remaining time for the representative task. Represents the interest matching weight coefficient. The weighting coefficient represents the urgency of the task. Represents the network link quality factor. This represents the environmental adaptability correction factor.

[0042] The multi-factor weight calculation module transforms perceived multi-dimensional data into executable scheduling instructions through complex numerical calculation logic. The similarity calculation submodule receives a 128-dimensional advertisement content feature vector (denoted as ) from the feature vector mapping module. ) and the audience feature vector from the global perception acquisition module (denoted as ) Here, the age and gender tags in the audience characteristics need to be pre-mapped into a vector space of the same dimension as the ad vector through one-hot encoding or embedding layers. The submodule calculates the cosine similarity between the two, specifically by calculating... and The dot product is divided by the product of the magnitudes of the two components. Since L2 regularization has been performed in the previous module and the magnitudes are all 1, the cosine similarity is simplified to a dot product operation. The calculation result ranges from -1 to +1, and the submodule normalizes it to the interval of 0 to 1 to generate the audience feature matching degree. For example, if the dot product result is 0.75, then The value is directly taken as 0.75.

[0043] The cosine similarity mentioned above refers to evaluating the similarity between two vectors by calculating the cosine value of the angle between them. The closer the cosine value is to 1, the more consistent the directions of the two vectors are, that is, the higher the similarity.

[0044] The parameter factor generation submodule handles frequency control and environmental correction logic. This is specifically for the frequency decay factor. The generation process begins with this submodule querying the local database to obtain the cumulative number of plays for the current content within the current distribution period. and the preset maximum number of allowed playbacks By introducing a variant formula of the sigmoid function to perform nonlinear mapping, the calculation logic is as follows: first, calculate the cumulative number of plays. With maximum allowed number of plays The ratio; then subtract the preset trigger decay threshold bias from this ratio. The result is then multiplied by the steepness factor that controls the decay rate. Then the natural constants were calculated. The product result is raised to the power of the product; finally, 1 is divided by (1 plus the result of the power operation) to obtain the frequency attenuation factor. In this calculation logic, the steepness coefficient... The threshold bias is set to 10. It was set to 0.8. Verification was performed by substituting actual values: assuming the current cumulative number of plays... The maximum number of times allowed to play is 50. If the ratio is 100, then the ratio is 0.5. Calculate the intermediate term. The natural constant raised to the power of -3 is approximately 0.0498. Final calculation... The results indicate that when the number of plays reaches half of the limit, the frequency decay factor remains at a high level (approximately 0.95) and does not significantly suppress priority.

[0045] The aforementioned sigmoid function is a mathematical function with an "S"-shaped curve that can map any real number to between 0 and 1. It is often used as an activation function in neural networks or in logistic regression classifiers.

[0046] Network link quality factor Environmental adaptability correction factor The mapping is generated based on a lookup table. The submodule has a built-in preset mapping rule table. When the network packet loss rate is less than 0.5% and the downlink speed is greater than 20 megabits per second, The mapping is 1.0; if the packet loss rate is between 0.5% and 2%, The linearity decreases to 0.6. Similarly, if the ambient light intensity is within the suitable viewing range (300 to 800 lux). Mapped to 1.0; if direct sunlight (greater than 2000 lux) causes screen glare, The value was adjusted to 0.8 to lower the delivery priority.

[0047] The index-weighted correction submodule performs the final fusion calculation. This submodule obtains the remaining valid duration of the advertising task. With total duration The percentage of remaining time for the calculation task Then, a real-time scheduling priority index is generated according to the formula. .

[0048] This real-time scheduling priority index The generation logic is as follows: ;in, Represents the interest matching weight coefficient. The weighting coefficient represents the urgency of the task. Represents the degree of matching with audience characteristics. Represents the frequency decay factor. The percentage of remaining time for the representative task. Represents the network link quality factor. This represents the environmental adaptability correction factor. In this embodiment, Set to 0.7. The value is set to 0.3, and the sum of the two is 1. The calculation process is illustrated with a textual example: Assume the audience characteristic matching degree of the current advertising creative is... The calculated frequency attenuation factor is 0.9. The percentage of remaining task time is 0.952 (based on the aforementioned calculation). The value is 0.5, indicating a good network connection. ), suitable environment ( Substituting the above values ​​into the formula: First, calculate the interest items: Secondly, calculate the urgent items: The final total index is calculated as follows: The calculated result of 1.19976 is used as the real-time scheduling priority index for this material.

[0049] Please refer to Figures 1 and 5. The dynamic queue reordering module is used to sort the queue of content to be played in descending order according to the real-time scheduling priority index, and output the target content loading instruction according to the sorting result.

[0050] Specifically, the dynamic queue rearrangement module includes: a sequence descending sorting submodule, which iterates through each media node in the queue of content to be played, associates the real-time scheduling priority index with the corresponding media node, and uses a quicksort algorithm to rearrange all media nodes in the queue in descending order of the index; and an instruction encapsulation and output submodule, which obtains the storage path and checksum of the first media node in the rearranged queue, and encapsulates the media loading path, decoding parameters, and playback timing requirements in combination with the current rendering engine state to generate the target content loading instruction.

[0051] The process of generating the target content loading instruction specifically includes: obtaining the cloud storage URL address and local cache hash value of the sorted target material, and checking whether the corresponding material file exists in the local cache; if the local cache exists and the hash value matches, a direct read instruction is generated; if the local cache does not exist or the hash value does not match, the cloud URL address, the segmented download strategy, and the breakpoint resume flag are encapsulated to generate a network loading instruction, and the network loading instruction is encapsulated as the target content loading instruction.

[0052] The dynamic queue reordering module is responsible for translating abstract priority indices into concrete playback execution sequences. The sequence descending sorting submodule takes over the queue of content to be played in memory, which contains dozens to hundreds of ad creative nodes. This submodule traverses the queue, calculating the real-time scheduling priority index output by the multi-factor weighting module. Dynamically bind to the corresponding media node object. To ensure real-time scheduling, this submodule uses an optimized quicksort algorithm. In the sorting logic, the algorithm selects the media node in the middle of the queue. Using a baseline, the queue is divided into two sub-intervals: high priority and low priority. These sub-intervals are then recursively sorted. For nodes with the same numerical value, the submodule introduces a secondary sorting key, which sorts them according to the order in which the materials were added to the database, using a first-in, first-out (FIFO) approach to ensure the determinism of the sorting result. After sorting, the material node at the head of the queue is the optimal playback target at the current moment.

[0053] The quicksort algorithm described above is an efficient sorting algorithm that uses a divide-and-conquer strategy. It divides the array into two subarrays by selecting a pivot element, such that all elements in the left subarray are less than the pivot and all elements in the right subarray are greater than the pivot. Then, it recursively sorts the subarrays.

[0054] The instruction encapsulation and output submodule is responsible for converting the sorting results into executable instructions for the player. This submodule first extracts the cloud storage Uniform Resource Locator (URL) address of the first and second media nodes after sorting, along with a pre-calculated local file hash value. The submodule accesses the local storage file system to check for a cached file with a matching filename and hash value. If the local cache exists and the verification passes, the submodule generates a direct read instruction in JSON format, containing the playback action, local source path, and volume parameters. If the local cache is missing or the checksum does not match, the submodule generates a network loading instruction. When generating the network loading instruction, the submodule encapsulates a chunked download strategy based on the network bandwidth data obtained from the preceding module. For example, if the current bandwidth is sufficient (greater than 50 megabits per second), the instruction sets a larger HIP protocol range for requesting chunk sizes (e.g., 5 megabytes per chunk) to reduce the number of requests; if bandwidth is limited, smaller chunk sizes are set (e.g., 512 kilobytes per chunk), and a resume function is enabled. The final generated network loading instruction is sent to the underlying media rendering engine via inter-process communication, driving seamless switching of the advertising machine's screen content.

[0055] The hash value mentioned above (such as MD5) refers to a widely used cryptographic hash function that can produce a 128-bit (16-byte) hash value to ensure the integrity and consistency of information transmission. It is often used for file verification.

[0056] Inter-process communication (IPC) refers to techniques or methods for transmitting data or signals between at least two processes or threads, used for data exchange between programs running on the same computer or different computers connected to a network.

[0057] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A cloud-based intelligent content scheduling system for advertising machines, characterized in that, The system includes: a feature vector mapping module, used to parse material metadata and perform spatial mapping, construct advertising content feature vectors, and establish a queue of content to be played based on time window constraints; a global perception acquisition module, used to drive cameras and light sensors to capture audience features and ambient light intensity data, and collect network bandwidth data to generate device operating environment data; a multi-factor weight calculation module, used to calculate the cosine similarity between the advertising content feature vector and the audience features to obtain the audience feature matching degree, generate a frequency decay factor based on the complement of the playback frequency ratio, generate a network link quality factor and an environmental adaptability correction coefficient based on the device operating environment data mapping, perform a weighted summation on the product of the audience feature matching degree and the frequency decay factor and the task urgency term based on the remaining time percentage, and use the network link quality factor and the environmental adaptability correction coefficient to correct the summation result to generate a real-time scheduling priority index; and a dynamic queue reordering module, used to sort the queue of content to be played in descending order according to the real-time scheduling priority index, and output a target content loading instruction according to the sorting result.

2. The intelligent content scheduling system for advertising machines based on a cloud platform according to claim 1, characterized in that, The feature vector mapping module specifically includes: a metadata parsing submodule, used to obtain the tag data, duration attribute, and resolution parameters of the advertising material, perform word segmentation and stop word removal on the tag data, and generate a metadata descriptor sequence; a spatial mapping submodule, used to call a pre-trained semantic embedding model to perform vectorization transformation on the metadata descriptor sequence, mapping the discrete tag data to a high-dimensional continuous vector space, and generating the advertising content feature vector; and a queue initialization submodule, used to obtain the current system timestamp and the preset advertising delivery schedule, filter out advertising materials within the effective delivery time window, and associate the candidate material index and its corresponding advertising content feature vector according to the filtering results to establish the queue of content to be played.

3. The intelligent content scheduling system for advertising machines based on a cloud platform according to claim 1, characterized in that, The comprehensive perception acquisition module specifically includes: a visual feature extraction submodule, used to activate the camera to capture video stream frame images, extract facial key point data in the frame images using a face recognition algorithm, analyze the age distribution, gender ratio, and gaze duration of the audience, and generate audience features; and an environmental data generation submodule, used to read the analog voltage signal output by the light sensor and convert it into an ambient light intensity value, while monitoring the uplink and downlink transmission rates and packet loss rate of the network interface, and combining and encapsulating the ambient light intensity value, transmission rate, and packet loss rate to generate the device operating environment data.

4. The intelligent content scheduling system for advertising machines based on a cloud platform according to claim 1, characterized in that, The multi-factor weight calculation module specifically includes: a similarity calculation submodule, used to obtain the advertising content feature vector and audience features, calculate the cosine value of the angle between the two vectors in multi-dimensional space, and normalize the calculation result as the audience feature matching degree; a parameter factor generation submodule, used to obtain the number of times the advertising material has been played and the preset maximum playback limit, calculate the ratio between the two and take the complement to obtain the frequency decay factor, and at the same time, according to the preset mapping rule table, map the network indicators in the device operating environment data to the network link quality factor, and map the illumination indicators to the environmental adaptability correction coefficient; and an index weighted correction submodule, used to obtain the remaining effective duration and total duration of the advertising task, calculate the task urgency items, and perform multi-dimensional weighted correction operation in combination with the audience feature matching degree, the frequency decay factor, the network link quality factor and the environmental adaptability correction coefficient to generate the real-time scheduling priority index.

5. The intelligent content scheduling system for advertising machines based on a cloud platform according to claim 1, characterized in that, The dynamic queue rearrangement module specifically includes: a sequence descending sorting submodule, used to traverse each material node in the queue of content to be played, associate the real-time scheduling priority index with the corresponding material node, and use a quick sorting algorithm to rearrange all material nodes in the queue in descending order of the index; and an instruction encapsulation and output submodule, used to obtain the storage path and checksum of the first material node in the rearranged queue, and encapsulate the material loading path, decoding parameters and playback timing requirements in combination with the current rendering engine state to generate the target content loading instruction.

6. The intelligent content scheduling system for advertising machines based on a cloud platform according to claim 2, characterized in that, The process of generating the advertising content feature vector specifically includes: obtaining the metadata descriptor sequence, calculating the weight value of each descriptor in the sequence using the TF-IDF algorithm, and constructing a weighted bag-of-words model; converting the metadata descriptor sequence into a word embedding matrix using the Word2Vec algorithm, performing a weighted average pooling operation on the word embedding matrix in combination with the weight values, and extracting the semantic core features of the material; performing L2 regularization on the pooled feature vector to eliminate the influence of vector magnitude differences on similarity calculation, and generating the advertising content feature vector.

7. The intelligent content scheduling system for advertising machines based on a cloud platform according to claim 3, characterized in that, The audience feature extraction process specifically includes: acquiring real-time monitoring frame images captured by the camera; performing target detection on the frame images using a multi-task convolutional neural network to locate the bounding box of the face region; extracting facial geometric features composed of 68 feature points within the bounding box of the face region, inputting them into an attribute classifier to predict the age and gender labels of the audience, and calculating the gaze attention based on the gaze angle deviation; encoding the age label, gender label, and gaze attention according to a preset dimensional splicing rule to generate audience features.

8. The intelligent content scheduling system for advertising machines based on a cloud platform according to claim 4, characterized in that, The generation process of the frequency decay factor specifically includes: obtaining the cumulative number of times the current advertising creative has been played within the current campaign period. and the preset maximum number of allowed playbacks Based on the formula: ; Calculate the frequency attenuation factor; wherein, Represents the frequency attenuation factor, The steepness coefficient represents the rate of decay. This represents the threshold bias that triggers decay. is the base of the natural logarithm.

9. The intelligent content scheduling system for advertising machines based on a cloud platform according to claim 4, characterized in that, The calculation process of the real-time scheduling priority index specifically includes: obtaining the audience feature matching degree, the frequency decay factor, the proportion of remaining task time, the network link quality factor, and the environmental adaptability correction coefficient; based on the formula: Perform fusion calculation to obtain the real-time scheduling priority index; wherein, This represents the real-time scheduling priority index. This represents the degree of matching of the stated audience characteristics. Represents the frequency attenuation factor, The percentage of remaining time for the representative task. Represents the interest matching weight coefficient. The weighting coefficient represents the urgency of the task. This represents the network link quality factor. This represents the environmental adaptability correction factor.

10. The intelligent content scheduling system for advertising machines based on a cloud platform according to claim 5, characterized in that, The process of generating the target content loading instruction specifically includes: obtaining the cloud storage URL address and local cache hash value of the sorted target material, and detecting whether the corresponding material file exists in the local cache; if the local cache exists and the hash value matches, a direct read instruction is generated; if the local cache does not exist or the hash value does not match, the cloud URL address, the segmented download strategy, and the breakpoint resume flag are encapsulated to generate a network loading instruction, and the network loading instruction is encapsulated as the target content loading instruction.

Images