Internet marketing delivery control method and system based on terminal device state perception

Abstract

This application discloses a method and system for controlling internet marketing campaign delivery based on terminal device status awareness, relating to the field of internet information delivery. The method includes: real-time collection of the target terminal device's operating status and identification of the current user interaction task, obtaining candidate ad metadata and user interest models; quantifying device load by calculating a comprehensive resource pressure index, and conducting a comprehensive utility evaluation of candidate ads by combining ad interest matching degree, expected business revenue, and resource consumption; generating an initial delivery probability based on the evaluation results, and dynamically adjusting the delivery probability using a decision gating state driven by user interaction tasks, thereby obtaining the final delivery probability. This achieves precise ad display and resource optimization control on the device, improving delivery effectiveness while ensuring terminal operating efficiency.

Description

Technical Field

[0001] This application relates to the field of internet information delivery technology, and in particular to an internet marketing delivery control method and system based on terminal device status perception. Background Technology

[0002] With the widespread adoption of mobile internet, commercial advertising and marketing information delivery have become core components of many mainstream applications. Traditional internet advertising systems primarily rely on user profiling and recommendation algorithms, continuously collecting and analyzing users' browsing history and purchasing behavior data to infer user preferences and thus deliver personalized content.

[0003] However, this advertising model, which relies solely on user behavior data, has significant limitations. It typically ignores the real-time physical resource status of the terminal device and the user's current interaction scenario. As a result, in real-world applications, even when the user is on a call, navigating, or experiencing critical and resource-constrained moments such as extremely low device battery, the system may still frequently push advertisements. This not only severely disrupts the user's experience of handling primary tasks but may even pose serious security risks in certain scenarios (such as in-vehicle systems).

[0004] Although some advertising strategies on the market attempt to incorporate environmental awareness features, most of them use static thresholds or simple linear rules for rough control (such as setting the frequency of ads to be reduced when the battery level is below 30%), which cannot truly balance user experience and device safety with the commercial conversion efficiency of advertising. Summary of the Invention

[0005] This application provides a method, system, storage medium, computer program product, and electronic device for controlling internet marketing campaigns based on terminal device status awareness, in order to at least solve the problem in the prior art that internet marketing campaigns are difficult to balance campaign effectiveness and terminal user experience.

[0006] In a first aspect, embodiments of this application provide an internet marketing delivery control method based on terminal device state awareness. The method includes: real-time collection of current operating status data of a target terminal device and identification of the current user interaction task of the target terminal device; acquisition of advertising metadata of candidate advertisements to be delivered and a user interest model constructed for the target terminal device, wherein the advertising metadata includes the estimated resource consumption requirements and expected business revenue of the candidate advertisements; calculation of a comprehensive resource pressure index to quantify the overall operating load level of the target terminal device based on the current operating status data; determination of the interest matching degree of the candidate advertisements based on the user interest model, and evaluation and calculation of the comprehensive utility evaluation value of the candidate advertisements by combining the comprehensive resource pressure index, the interest matching degree, and the expected business revenue and estimated resource consumption requirements in the advertising metadata; calculation of the initial delivery probability of the candidate advertisements based on the comprehensive utility evaluation value, and determination of the decision gating state of the target terminal device based on the current user interaction task; gating adjustment of the initial delivery probability using the decision gating state to obtain the final delivery probability of the candidate advertisements, and control of the display of the candidate advertisements on the target terminal device based on the final delivery probability.

[0007] Secondly, embodiments of this application provide an internet marketing delivery control system based on terminal device status awareness. The system includes: a terminal status awareness unit, used to collect real-time data on the current operating status of a target terminal device and identify the current user interaction task of the target terminal device; an advertising information acquisition unit, used to acquire advertising metadata of candidate advertisements to be delivered and a user interest model constructed for the target terminal device, wherein the advertising metadata includes the estimated resource consumption requirements and expected business revenue of the candidate advertisements; a resource pressure calculation unit, used to calculate a comprehensive resource pressure index based on the current operating status data to quantify the overall operating load level of the target terminal device; and an advertising effectiveness evaluation unit, used to evaluate the advertising effectiveness based on the current operating status data. A user interest model determines the interest matching degree of the candidate advertisements, and, in conjunction with the comprehensive resource pressure index, the interest matching degree, and the expected business revenue and estimated resource consumption requirements in the advertisement metadata, evaluates and calculates the comprehensive utility assessment value of the candidate advertisements; a delivery decision generation unit is used to calculate the initial delivery probability for the candidate advertisements based on the comprehensive utility assessment value, and determine the decision gating state of the target terminal device based on the current user interaction task; a gating adjustment unit is used to adjust the initial delivery probability using the decision gating state to obtain the final delivery probability for the candidate advertisements, and controls the display of the candidate advertisements on the target terminal device based on the final delivery probability.

[0008] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the Internet marketing delivery control method based on terminal device state awareness according to any embodiment of this application.

[0009] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, characterized in that, when the program is executed by a processor, it implements the steps of the Internet marketing delivery control method based on terminal device state awareness according to any embodiment of this application.

[0010] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the internet marketing delivery control method based on terminal device state awareness according to any embodiment of this application.

[0011] The internet marketing campaign control method and system based on terminal device status awareness provided in this application can produce at least the following technical effects: (1) The real-time operating status of the terminal device is collaboratively modeled with the current user interaction task, and the underlying operating load level is uniformly quantified by calculating the comprehensive resource pressure index. Based on this processing method, the system no longer responds to the single state of the terminal in isolation, but can adjust the timing of advertising display in a more practical way from two dimensions: the overall carrying capacity of the device and the sensitivity of the current task. As a result, while ensuring the continuity of the main interaction tasks and the stability of terminal operation, the impact of unnecessary displays on the user's current operation rhythm is effectively reduced, thereby improving the scenario adaptability of the delivery control and the consistency of the terminal-side experience.

[0012] (2) The system jointly evaluates user interest matching degree, expected business revenue of advertising, estimated resource consumption of advertising, and the comprehensive resource pressure index to form a comprehensive utility evaluation mechanism for candidate advertisements. Furthermore, it achieves flexible hierarchical control by calculating the initial placement probability and superimposed decision gating status. Since this mechanism considers both "placement value" and "resource cost" in utility measurement, the selection results of candidate advertisements can better match the balance between actual placement revenue and terminal carrying capacity. Furthermore, by using task gating to block or allow the initial placement probability, advertising display is no longer a simple fixed trigger, but a closed-loop decision that dynamically changes with the scenario state. Based on this, the system can maintain the effectiveness of marketing reach while reducing inefficient and energy-intensive advertising exposure, achieving better coordination between business objectives and resource consumption in placement behavior.

[0013] This technical solution constructs a dynamic ad delivery control mechanism based on multi-dimensional perception of terminal status, centered on comprehensive utility evaluation, and constrained by task gating adjustment, achieving adaptive generation of the final ad delivery probability. Thus, internet marketing ad delivery has successfully transformed from static, coarse-grained, single-dimensional judgment into a refined decision-making process deeply coupled with the underlying terminal status, user interaction scenarios, and business objectives. This maximizes the overall business efficiency of ad delivery while ensuring terminal-side operational security and user-friendliness. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart is shown as an example of an internet marketing delivery control method based on terminal device state awareness according to an embodiment of this application; Figure 2 This document illustrates an example of an operation flowchart for identifying the current user interaction task of a target terminal device in a method according to an embodiment of this application. Figure 3 A flowchart illustrating an example of performing a parameter adaptive tuning step based on reinforcement learning and privacy computation in a method according to an embodiment of this application is shown. Figure 4 A system operation principle diagram of an example of an internet marketing placement control method based on terminal device status awareness according to an embodiment of this application is shown. Figure 5 This paper illustrates the comprehensive performance simulation results of different delivery control strategies under multi-dimensional evaluation indicators in the embodiments of this application. Figure 6 A time-series trajectory simulation diagram of the control response capability under different delivery control strategies in a dynamic scenario is shown. Figure 7 This diagram illustrates the performance trade-off between different delivery control strategies in terms of total exposure and effective clicks, based on a simulation experiment. Figure 8 A structural block diagram of an example of an internet marketing delivery control system based on terminal device status awareness, according to an embodiment of this application, is shown. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] It should be noted that in traditional recommendation models that rely on user behavior data, although precise user tags can theoretically improve ad reach, related research indicates that a large number of personalized recommendations that are detached from the user's current context often increase the perceived intrusion. Simply relying on behavioral data without respecting the user's current situation can easily trigger privacy concerns and avoidance behaviors.

[0018] To address this deficiency, current technologies have proposed intelligent push strategies that integrate features from multiple devices or incorporate external environmental data (such as geographic location and basic communication environment). However, these solutions largely focus on more accurately updating users' real-time interest models to match advertising content, essentially lacking a deep understanding and quantitative assessment of the terminal device's own computing load and energy consumption (such as CPU load and heat generation). This leads to situations where, even with a high degree of user interest matching, the system may still push ads frequently when the device is under high load or resources are extremely strained, severely impacting device performance and the overall user experience.

[0019] Furthermore, in specialized application scenarios with extremely high safety requirements, such as intelligent connected vehicles, some current technologies attempt to push advertisements based on objective vehicle status data (such as mileage and location). Theoretically, this approach can improve the relevance of commercial information; however, in practice, due to the lack of a dynamic safety blocking mechanism based on core task priorities, pop-up ads often appear at critical moments requiring high concentration, such as when the vehicle starts, reverses, or navigates complex road conditions. This interference with core interactive tasks not only obscures key information but also directly increases driving risks.

[0020] In particular, even though some current technologies have introduced basic environmental awareness control into ad scheduling, their underlying logic is extremely crude, typically relying solely on isolated static thresholds (such as mechanically reducing display frequency when battery levels fall below a certain percentage). This control method completely ignores the complex nonlinear coupling effects between multi-dimensional hardware challenges; furthermore, when executing interception, it fails to consider the estimated resource consumption characteristics of the candidate ads themselves during the loading and rendering phases (i.e., the energy consumption attributes of the ad content). In addition, existing evaluation schemes for ad playback effectiveness mostly rely on lagging scores based on sensor-collected user feedback, making dynamic intervention before ad delivery impossible.

[0021] In summary, although current technologies have made attempts to integrate precise ad recommendations with some environmental features, they generally lack a deep understanding of multi-dimensional terminal operating status and a comprehensive pressure quantification mechanism. Because a multi-dimensional coupling and trade-off model has not been established between device resource status, user interaction task security, and the resource needs of the ads themselves, current ad delivery control systems cannot truly achieve a dynamic balance between commercial benefits, user experience, and device security in complex real-time operating environments.

[0022] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0023] The technical solutions in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of users' personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0024] Figure 1 A flowchart illustrating an example of an internet marketing delivery control method based on terminal device state awareness according to an embodiment of this application is shown.

[0025] Regarding the execution subject of the method in the embodiments of this application, it can be any controller, processor, or processing platform with computing or processing capabilities, such as the processing controller in the terminal-side marketing delivery control platform. By running computer programs, instructions, or functional modules stored in the storage medium, it executes the Internet marketing delivery control method based on terminal device status awareness provided in the embodiments of this application.

[0026] In some examples, it can be integrated into electronic devices or terminals in the form of software, hardware, or a combination of both. The terminals or electronic devices can be various types of smart devices, such as mobile phones, tablets, in-vehicle systems, smart wearable devices, and other terminal devices with data processing capabilities.

[0027] like Figure 1 As shown, in step S110, the current operating status data of the target terminal device is collected in real time, and the current user interaction task of the target terminal device is identified.

[0028] In practical implementation, terminal devices (such as smartphones, tablets, smart car systems, or wearable devices) face complex and ever-changing physical environments and varying hardware and software loads during daily operation. The system can obtain real-time data on the terminal's current physical resource usage (e.g., but not limited to, battery consumption, processor load, memory usage, or network connection quality) by calling the operating system's underlying standard application programming interfaces (APIs) or hardware monitoring daemons, either through periodic polling or event triggering. Simultaneously, the system can identify the type of interactive task the user is currently focusing on by monitoring the operating system's event bus, active foreground processes, or user interface actions.

[0029] By establishing a state awareness mechanism on the device side, the limitations of traditional advertising systems that rely solely on cloud-based user profiles while ignoring the actual carrying capacity of the device side are effectively overcome. This enables the internet marketing control system to have the environmental interaction capabilities to sense the physical carrying capacity limits of the device and the user's current operational immersion.

[0030] In step S120, the advertising metadata of the candidate advertisement to be delivered and the user interest model constructed for the target terminal device are obtained. The advertising metadata includes the estimated resource consumption requirements and expected business revenue of the candidate advertisement.

[0031] In some implementations, the user interest model can be pre-trained and constructed locally on the device or in collaboration with the cloud, based on the user's historical ad interaction logs (such as click-through rate, browsing dwell time, ignore or close actions, etc.), to accurately represent the user's personalized preferences; no restrictions are imposed here. The ad metadata is a set of attribute tags issued along with the candidate ads. Among them, "estimated resource consumption requirements" is a key reference parameter, used to quantify the system costs, such as computing power, memory space, or network traffic, that the ad is expected to occupy throughout its loading, parsing, rendering, or video playback cycle; "expected business benefits" represents the commercial value that the ad can bring after successful exposure or conversion (such as estimated click-through rate or revenue per thousand impressions, etc.). By extracting the above multi-dimensional features, the system can not only evaluate the "interest fit" and "commercial profit potential" of ad content like traditional recommendation algorithms, but also incorporate the resource consumption attributes of the ad itself into the core evaluation system.

[0032] In step S130, based on the current operating status data, a comprehensive resource pressure index is calculated to quantify the overall operating load level of the target terminal equipment.

[0033] In practical application scenarios, since the operational status data collected by the terminal is often multi-dimensional and has heterogeneous units of measurement (such as the physical degree of heating temperature, the percentage of memory usage, the network bandwidth rate, etc.), the system needs to use a preset numerical normalization transformation and joint feature fusion logic to map the heterogeneous underlying physical status data into a unified risk scale.

[0034] Here, complex multi-dimensional status indicators are aggregated into an intuitive comprehensive resource pressure index. This index can be a continuous numerical indicator used to macroscopically quantify the overall "operational fatigue or critical load-bearing capacity" of equipment, greatly reducing the computational complexity of subsequent decision-making modules. When this index is high, it objectively indicates that the terminal is currently in a poor physical state such as high load, low power, or weak network. Based on this, the system can provide a quantitative early warning benchmark for high-energy-consuming advertising push requests, ensuring the operational stability and hardware security of terminal equipment from the underlying mechanism.

[0035] In step S140, the interest matching degree of the candidate advertisement is determined based on the user interest model, and the comprehensive utility evaluation value of the candidate advertisement is calculated by combining the comprehensive resource pressure index, the interest matching degree, and the expected business benefits and estimated resource consumption requirements in the advertisement meta information.

[0036] The system first uses a user interest model to correlate with the features of candidate ads to calculate the estimated interest match between the user and the ad content. Then, the system constructs a multi-objective optimization utility evaluation logic, jointly weighing parameters representing positive campaign value (i.e., interest match degree and expected business revenue) with parameters representing negative campaign costs (i.e., the coupling term of estimated resource consumption demand and comprehensive resource pressure index). This achieves a fundamental reconstruction of the traditional simple bidding or click-through rate ranking mechanism.

[0037] When the overall resource pressure index rises, the system adaptively amplifies the negative penalty weighting of estimated resource consumption demands during the evaluation process, significantly reducing the scores of ads that, while generating high commercial revenue, are extremely power- and data-intensive. Conversely, if the terminal device is in good condition and resources are sufficient, the system reverts to the conventional scoring logic based on interest and revenue. This allows for a complete dynamic trade-off and adaptive adaptation between commercial demands and the capabilities of the terminal hardware in the ad effectiveness evaluation.

[0038] In step S150, the initial delivery probability of the candidate advertisement is calculated based on the comprehensive utility evaluation value, and the decision gating state of the target terminal device is determined based on the current user interaction task.

[0039] In practice, the system employs a specific probability mapping mechanism to transform the calculated comprehensive utility evaluation value into an initial delivery probability. This probability initially determines the likelihood of candidate ads being displayed. Simultaneously, the system does not blindly execute delivery based on this probability. Instead, it initiates a parallel scenario security verification branch. This branch determines a decision gating state based on the identified attribute characteristics of the current user interaction task. This gating state is specifically used to characterize the anti-interference capability or do-not-disturb priority level of the current user's interaction scenario.

[0040] By separating continuous probability calculations from discrete scene gating, a dual-defense control mechanism combining soft probability allocation and hard gating is constructed. Through this mechanism, even if a candidate ad has an extremely high overall utility score (i.e., a high initial probability of being displayed), once the user's current interaction task is determined to be a highly sensitive task that cannot be disturbed, the decision gating state can trigger a forced blocking mechanism to ensure the safety bottom line of the control logic.

[0041] In step S160, the initial delivery probability is gating and adjusted using the decision gating state to obtain the final delivery probability for the candidate advertisement, and the display of the candidate advertisement on the target terminal device is controlled based on the final delivery probability.

[0042] In some implementations, the system uses the decision gating state as a safety adjustment factor, cascading it onto the initial delivery probability. If the gating state indicates that the current task is highly sensitive and requires absolute non-interference, the system significantly reduces the initial delivery probability or even resets it to zero to trigger a blocking mechanism. If the gating state indicates that the current scene is generally safe and display is allowed, the system retains or smoothly outputs the initial delivery probability, thereby deriving the final delivery probability. Finally, based on the magnitude of this final delivery probability and in conjunction with specific scheduling rules, the system pushes candidate ads into the underlying rendering engine for actual exposure and display control.

[0043] By adjusting the gating operation, the pain point of being abruptly interrupted by marketing information when users are focused on important operations (such as driving navigation, emergency calls, or immersive interactive scenarios) is effectively avoided. This ensures the continuity and security of the user experience to the greatest extent, making the entire internet marketing campaign not only precise and intelligent, but also highly adaptable to different scenarios and controllable in business.

[0044] Figure 2 A flowchart illustrating an example of identifying the current user interaction task of a target terminal device in a method according to an embodiment of this application is shown.

[0045] like Figure 2As shown, in step S210, the system kernel monitoring interface of the target terminal device is called to obtain system resource monitoring indicators in real time. The system resource monitoring indicators include at least the current remaining battery power, the current processor utilization rate, and the device temperature.

[0046] It should be noted that, since application-layer monitoring interfaces often suffer from data latency or are restricted by system policies, this embodiment preferably reads the raw register data of the hardware sensors directly and with high priority by calling the underlying driver files of the operating system kernel space (such as the sysfs virtual file system interface in Linux / Android systems). Among these, the remaining battery power reflects the device's basic operating range, the processor utilization rate reflects the current level of computing power congestion, and the device's temperature is a state indicator of the overall power consumption and heat dissipation pressure of the underlying hardware. By directly connecting to the system kernel monitoring interface, the system can obtain the highest fidelity, zero-latency profile of the underlying hardware with extremely low performance loss.

[0047] In step S220, the network status detection interface is called to obtain the currently available network downlink bandwidth.

[0048] In mobile internet scenarios, the network environment in which the terminal is located is highly time-varying and uncertain. The system dynamically obtains the currently available downlink bandwidth by calling the underlying network status detection interface (such as actively triggering a lightweight TCP / UDP probe, or passively listening to the throughput monitoring callback of the current network adapter).

[0049] Furthermore, the loading process of internet marketing advertisements (especially video streams or high-precision interactive ads) is a typical bandwidth-intensive operation. Real-time monitoring of network downlink bandwidth as an independent dimension can effectively prevent forced preloading of ads in weak network conditions or extremely congested channels, thereby preventing disruptive network resource contention between advertising services and users' primary tasks (such as web browsing and online meetings).

[0050] In step S230, the current remaining battery power, current processor utilization, device temperature, and current available network downlink bandwidth are time-aligned and fused to form the current operating status data.

[0051] It should be noted that the sampling frequencies of the hardware sensors at each monitoring node exhibit significant heterogeneity (for example, processor utilization may refresh every 100 milliseconds, while the temperature sensor may only report every 2 seconds). Directly concatenating these discrete data would lead to logical misalignment in the subsequent decision model inputs. Therefore, the system can employ a sliding time window and a zero-order hold algorithm to align the time sequence of multi-source asynchronous data.

[0052] Specifically, based on the system's set decision clock cycle Based on this, the system extracts a specific sampling time. The corresponding valid observations of each indicator are concatenated into a multidimensional state feature vector. This serves as the current running status data. The fusion process is as follows:

[0053] Equation (1) In the formula, , , as well as These represent slices at this time sequence. The system displays the current remaining battery power, current processor utilization, device temperature, and current available network downlink bandwidth after resampling and alignment. This timing alignment and vectorization process eliminates time-domain artifacts caused by inconsistent sampling frequencies of the underlying hardware interfaces, ensuring that the system can measure the overall physical load of the device at the same absolute timestamp.

[0054] In step S240, the operating system active window events of the target terminal device are monitored, and the concurrent occupancy status of the system-level audio channel and location service channel is jointly detected to extract the running characteristics of the currently executing application.

[0055] Specifically, simply determining which application is in the foreground window often fails to accurately reflect the user's true engagement level. For example, when a user is making a voice call or using car navigation in the background, although the main interface may be on the desktop or locked screen, the interaction is extremely important. Therefore, in addition to monitoring the operating system's Activity / Window focus switching events, the system forcibly penetrates and detects the hardware-level concurrent usage flags of the underlying AudioFlinger (audio projection channel) and LocationService (location service channel), comprehensively generating a runtime feature vector that includes multi-dimensional attributes such as foreground dwell time, microphone activation, and continuous GPS call status. This joint detection mechanism constructs a three-dimensional perception network to prevent evasion, completely solving the blind spots of traditional single front-end judgment logic in complex multi-task concurrent scenarios (such as "back-end navigation + front-end WeChat").

[0056] In step S250, the running features are classified according to the preset task mapping model to identify the target task type of the current user interaction task; wherein the target task type is one of the following: security constraint task, focus exclusive task, and background silent task.

[0057] Specifically, to transform complex application runtime characteristics into standardized control constraints, the system incorporates a task mapping model built on expert rules or lightweight machine learning algorithms. This model receives the extracted runtime feature vectors. And calculate the matching confidence of the current interaction task belonging to each type of target task. To facilitate engineering implementation, the classification confidence level can be evaluated using the following linear weighted mapping function:

[0058] Equation (2) In the formula, Indicates the first Confidence score for each task type ( These correspond to security-constrained, focus-exclusive, and background silent tasks, respectively. The total dimension of the running features, For the first The component values ​​of the feature (e.g., Boolean channel occupancy flags or normalized active time). For this feature, for the first Prior relevance weights for different task types This is the bias term. The system ultimately selects the confidence score. The highest-scoring option is selected as the target task type for output. For example, when an audio or location channel is detected to be continuously monopolized, the confidence level of a safety-constrained task (such as driving navigation or emergency calls) will increase and the task will be selected for output by the system.

[0059] Through the aforementioned multidimensional feature mapping, the system successfully reduced the abstract and complex user application performance into three main target task types with clear boundaries that can directly guide subsequent gating blocking or frequency suppression operations, thus realizing the transformation from behavior monitoring to intent characterization.

[0060] This application's embodiments construct a three-dimensional state perception matrix on the terminal side, interwoven with both bottom-up (from physical resources to network bandwidth) and top-down (from interface focus to channel concurrency) approaches. On one hand, relying on a rigorous time sequence alignment mechanism, it eliminates the time difference between monitoring indicators of multi-source heterogeneous hardware, extracting high-fidelity, synchronous snapshots of the device's underlying infrastructure. On the other hand, through system-level channel joint detection and a multi-dimensional feature mapping model, it accurately penetrates the user's true interaction intent in complex multi-tasking environments. Therefore, the system can not only keenly capture any potential device resource overload risks but also accurately define the rigid boundaries of the user's current operation to avoid disturbances, thus providing a basis for subsequent high-precision ad blocking, yielding, and flexible ad delivery decisions.

[0061] Regarding the implementation details of calculating the comprehensive resource pressure index in step S130, in some examples of embodiments of this application, firstly, the current remaining battery power, current processor utilization, and current available network downlink bandwidth in the current operating status data are used as physical status indicators. Each is obtained through the corresponding nonlinear risk mapping function. Normalization mapping is performed to obtain the corresponding stress risk factor with a value range of [0, 1]. Stress risk factors include power stress factor, processor stress factor, and network stress factor.

[0062] In actual equipment operation, different underlying physical resources (such as percentage of power, megabits per second of bandwidth, etc.) differ greatly in physical dimensions and measurement scales. This implementation uses a customized nonlinear risk mapping mechanism to uniformly compress and map these heterogeneous physical indicators into a dimensionless probability space of [0, 1]. When a physical resource is in a sufficient and safe range, its output risk factor gradually approaches 0; while when the resource approaches depletion or a high-risk boundary, the output risk factor will experience a nonlinear and sharp jump. Thus, this operation completely eliminates the dimensional barriers between multidimensional hardware indicators.

[0063] Then, based on the target task type of the current user interaction task, an attention stress factor is dynamically configured to characterize the user's attention load on the terminal. Its value range is [0, 1]; among which, when the target task type is identified as a focus-exclusive task, an attention pressure factor is assigned. Extreme parameters that exceed the preset focus threshold.

[0064] Specifically, the actual load on a terminal comes not only from the computing power consumption of the underlying silicon-based hardware, but also from the "cognitive resources" invested by the user at the software application level. For "focus-exclusive tasks" (e.g., when a user is playing a full-screen, high-frame-rate competitive game or watching a highly immersive video), the user's tolerance threshold for external distractions such as pop-ups is extremely low. In this case, the system will abandon the conventional smooth scoring and directly assign a score. Assign a penalty constant that approximates the limit (e.g., take...). For regular background silent tasks (such as listening to music with the screen off), a very low threshold is applied. This value, in turn, intuitively transforms the user's subjective interactive experience needs into an objective mathematical penalty factor, endowing the advertising decision-making system with a powerful scene adaptive perception capability, thereby ensuring an undisturbed experience in highly immersive scenarios from the source mechanism.

[0065] Next, the device heating temperature is extracted from the current operating status data, and a dynamic temperature penalty index is set based on the device heating temperature. Dynamic temperature penalty index The initial baseline value is 1, and the dynamic temperature penalty index is applied when the device's heating temperature exceeds the preset safe temperature threshold. The value increases exponentially and non-linearly.

[0066] In some implementations, the device's heating temperature is the ultimate macroscopic physical characteristic of the terminal's overall high-load operation, and can be calculated using the following temperature control penalty function: Equation (3) In the formula, The temperature at which the equipment heats up to be extracted is determined. The preset safe temperature threshold defined for the system (e.g., set to 40℃). To control the positive constant coefficient of temperature sensitivity, This is a natural constant. When the equipment temperature is within a safe and comfortable range, Maintaining a baseline constant of 1 ensures no additional amplification effect in subsequent calculations; however, once the temperature exceeds a preset safe temperature threshold, The exponential surge mechanism will be rapidly activated as the temperature difference increases.

[0067] By establishing the aforementioned independent exponential temperature control feedback barrier, the algorithm can respond before the operating system's underlying throttling mechanism, severely blocking resource-consuming ad loading requests at this time, thereby significantly extending the hardware's lifespan and effectively preventing system crashes.

[0068] Then, the comprehensive resource pressure index is calculated by invoking a joint probabilistic risk model that incorporates a dynamic temperature penalty index: Equation (4) In the formula, This represents the comprehensive resource pressure index. The total number of dimensions of the physical state indices used in the calculation; Indicates the first Item physical state index, This indicates the use of the corresponding nonlinear risk mapping function. The calculated pressure risk factor; For the first The normalized weight parameters corresponding to the physical state index and satisfy the following conditions: By using the multiplication term decay mechanism in the joint probability risk model and the exponential operation of the dynamic temperature penalty index, when the target terminal device experiences abnormally high temperature or is currently executing a focus-exclusive task, the risk weight of the local resource bottleneck is forcibly amplified, causing the comprehensive resource pressure index to exponentially approach the maximum value of 1.

[0069] In equation (4), the “failure model of series system” in the field of reliability engineering is referenced. Indicates the first The probability that the underlying physical resources are within a safety margin is expressed in exponential order. Used to nonlinearly determine the relative importance of the resource in the system evaluation; multiplication symbol By coupling all the underlying security probabilities together, it can be represented that if any one underlying resource fails (i.e., a certain base value goes to zero), the entire system faces collapse. At this point, the exponent... As a macroscopic scaler, it directly acts on the multiplicative terms. Once a high temperature is triggered, the exponentiation operation will rapidly compress the overall underlying security probability.

[0070] Finally, the overall security probability of the underlying hardware is compared with the cognitive security probability of the upper layer. By performing a second multiplication and fusion, and subtracting the total security probability of the system from 1, the final comprehensive resource pressure output can be obtained. Therefore, by using the multiplicative attenuation mechanism in the joint probabilistic risk model and the exponential operation of the dynamic temperature penalty index, when the target terminal device experiences abnormally high temperatures or is currently executing a focus-exclusive task, the risk weight of local resource bottlenecks is forcibly amplified, causing the comprehensive resource pressure index to exponentially approach its maximum value of 1. Compared to the traditional linear weighted algorithm, the algorithm in this embodiment adopts a nonlinear characteristic, which enables it to have higher boundary protection sensitivity when dealing with extreme load conditions.

[0071] This application's embodiments construct a nonlinear resource pressure measurement system characterized by "dual constraints of hardware and software, and local-global linkage." This mechanism utilizes a joint probabilistic risk model to deeply decouple and recouple heterogeneous underlying physical hardware losses (power consumption, processor load, network bandwidth) from surface-level macroscopic physical states (heat surges) and user interaction scenarios (focused, exclusive tasks) using mathematical decoupling. This enables the advertising delivery system to instantly capture resource bottlenecks in any single dimension within the terminal with extremely high sensitivity. Furthermore, it solves the device lag and experience fragmentation problems caused by the lack of global resource coordination on the terminal side in traditional advertising delivery engines. Thus, the entire internet marketing outreach process achieves a precise and rigorous dynamic balance between maximizing commercial value and ensuring device operational security.

[0072] In some examples of embodiments of this application, considering the significant differences in the negative impact mechanisms on system stability and user experience when different hardware resources within the terminal face exhaustion, the system employs a nonlinear risk mapping function tailored to the exhaustion characteristics of different physical resources. The calculations are performed using a differentiated mapping model, with the power stress factor, processor stress factor, and network stress factor respectively represented as follows: , as well as .

[0073] Specifically, firstly, the current remaining battery power is used as an input variable. The power pressure factor is obtained by calling the calculation formula based on reverse S-shaped attenuation. To prevent the battery level from falling below a preset safe battery warning threshold. At that time, risk manifestations that trigger an exponential surge are: Equation (5) In actual use of mobile devices, the user's psychological panic about battery consumption and the system's power consumption warnings often do not occur linearly. For example, a user may not notice the battery level dropping from 80% to 70%, but a drop from 20% to 10% can trigger significant anxiety. Therefore, this embodiment uses a variant of the Sigmoid function to construct a battery stress factor. To further clarify the surge characteristics of this inverse S-shaped curve, the formula can be used to calculate the battery level... Risk sensitivity gradient (absolute value of derivative) :

[0074] Equation (6) In the formula, This represents the rate of risk escalation caused by a small decrease in battery power. From this derivative formula, it can be seen that the risk escalates if and only if the current battery power... Approaching the safe power warning threshold At this time Approaching 0.5), risk sensitivity gradient Reached its peak.

[0075] Therefore, by employing a mapping mechanism based on inverse S-shaped attenuation, the advertising system can maintain extremely high activity (stress factor approaching 0) when the terminal has sufficient battery power, but will remain active once the battery level drops below a certain threshold. When the risk factor reaches a critical point (such as 20%), it can approach 1 as quickly as possible, like triggering an alarm, thereby suppressing the rendering of power-consuming advertisements and extending the device's maximum battery life.

[0076] Then, the current processor utilization rate is used as an input variable. Call with load-strengthening index The processor stress factor is obtained using the power function formula. To penalize the risk of performing nonlinear amplification in high processor utilization ranges: Equation (7) It should be noted that processor scheduling exhibits high congestion characteristics: when CPU utilization increases from 10% to 30%, the system remains smooth; however, increasing it from 70% to 90% can easily lead to thread blocking, UI frame drops, and even severe overheating and frequency throttling. Therefore, a load enhancement exponent greater than 1 is introduced into the formula. (For example, taking 3 or 4). When inputting variables When the exponent is small, its growth is extremely slow; while when Gradually approaching the maximum theoretical utilization constant At this time, the characteristics of the power function will cause the output result to rise sharply. Thus, it describes the vulnerability of the processor in the high load range, so that the system shows a very high tolerance for advertising computing power overhead when the device is lightly loaded, while implementing extremely severe risk penalties when heavily loaded, effectively preventing the negative experience of complex advertising animation rendering overwhelming the system's smoothness.

[0077] Therefore, the currently available network downlink bandwidth is used as an input variable. The network stress factor is obtained by calling the inverse linear calculation formula. To construct a direct mapping of bandwidth scarcity: Equation (8) In the formula, This represents the current remaining battery power. The safe power warning threshold, To control the constant parameter that ensures the slope of the decrease in electricity risk is greater than 0, It is a natural constant; Current processor utilization This is the processor's maximum theoretical utilization constant. A load enhancement index greater than 1; This represents the currently available network downlink bandwidth. This is used as a reference for the maximum bandwidth. All input variables are acquired in real time by the underlying sensors, and all output results are strictly constrained within the safe closed interval [0,1].

[0078] It should be noted that the consumption of network resources has a relatively direct and inversely proportional impact on the ad loading experience. When the device's downlink bandwidth... Approaching the reference maximum bandwidth value Network stress factor (e.g., when the physical bandwidth limit of the current network environment is reached) A value approaching 0 indicates unobstructed network connectivity; however, as available bandwidth increases... As the threshold decreases, this risk factor exhibits a strictly linear upward trend. This lightweight inverse linear mapping can accurately reflect the current network channel congestion level with extremely low computational overhead, avoiding the blocking of users' core business data packets caused by forcibly fetching large-volume video advertisements in weak network environments.

[0079] Through the embodiments of this application, differentiated nonlinear risk mapping models are deployed to address the depletion characteristics of different physical resources, deeply integrating the underlying principles of hardware engineering: an inverse S-shaped decay function is used to accurately characterize the sharp performance decline trend during low battery periods; a high-power function is used to sensitively capture the nonlinear congestion risk of the processor under high load; and a linear function is used to calibrate the scarcity of network bandwidth. With this personalized heterogeneous mapping mechanism, when facing multi-dimensional and complex hardware resource constraints, the terminal system can not only quickly trigger adaptive protection responses to resource bottlenecks approaching their limits, but also maintain service processing capacity and load balance when resources are abundant. This establishes an adaptive operation guarantee system on the terminal side that combines service flexibility with hardware resource security constraints.

[0080] Regarding the implementation details of evaluating and calculating the comprehensive utility evaluation value of candidate advertisements in step S140, in some examples of embodiments of this application, firstly, the degree of interest matching determined based on the user interest model is... This is represented as a user experience gain term.

[0081] In practical deployments, the system can pre-output dense vectors representing user preferences and advertising features using collaborative filtering or deep neural networks. By calculating the cosine similarity or inner product between the two, the degree of interest matching mapped within the interval [0, 1] can be obtained. This gain term, from the audience's perspective, quantitatively defines the accuracy of the personalized adaptation of candidate ad content to the current user. Using interest matching as an independent positive gain feature effectively filters out redundant ads with low relevance, ensuring a good information acquisition experience for users on their devices at the initial stage of the business process and reducing invalid exposure.

[0082] Then, the expected business revenue Compared to the system-defined normalized reference maximum value for returns The ratio of is represented as the business value gain term.

[0083] It should be noted that, given the significant differences in absolute values ​​between the expected revenue (e.g., eCPM, or expected revenue per thousand impressions) of different ad units in the candidate ad library, directly introducing such an unbounded scalar would lead to an imbalance in the weights of the subsequent evaluation model. Therefore, this embodiment introduces a global or time-window-based normalized reference maximum value for revenue. Through division operation By mapping the commercial value characteristics with economic dimensions to a standardized mathematical space of [0, 1], the commercial monetization indicators of advertising can be fairly weighted and compared with the aforementioned user experience indicators on the same numerical scale.

[0084] Next, the resource consumption demand will be estimated. Compared to the system's set normalized reference maximum resource consumption The ratio of the comprehensive resource pressure index Perform product coupling to construct a dynamic resource penalty term.

[0085] In a preferred embodiment, the resource consumption demand is estimated. The following results were obtained by weighted fusion calculation of multi-dimensional hardware cost prediction indicators carried in the advertising metadata: Equation (9) In the formula, , and These represent the expected processor computing power overhead, memory usage, and network communication traffic generated by the candidate ads during the loading and rendering phases, respectively. , and Convert weight parameters for preset resource types.

[0086] Based on this, the normalized energy consumption scalar of the advertisement itself is... The comprehensive resource pressure index reflects the real-time environment of the terminal. Perform nonlinear product coupling operations. When the terminal hardware is in good condition ( When the resource requirements of the advertising material (such as high-definition 3D interactive advertising) are extremely high, the product of this penalty term is still suppressed to near zero, thus fully releasing computing power; while when the terminal is in a high-load critical state, the internal penalty base of high-energy-consuming advertising is reduced. Linear amplification allows for the control of resource-sensitive advertising at its source.

[0087] Then, a multi-objective optimization function, including user experience gain, business value gain, and dynamic resource penalty, is invoked to calculate the comprehensive utility evaluation value of the candidate advertisement: Equation (10) In the formula, This is the overall utility evaluation value for candidate advertisements; The degree of interest matching is represented by a value in the range [0,1]. For expected business revenue, The maximum normalized reference value for revenue set for the system; To predict resource consumption requirements, a quantitative value representing the total computing power and traffic resources expected to be consumed by candidate ads during the loading and rendering phases is defined. The normalized reference maximum value for resource consumption set by the system; A comprehensive resource pressure index; , and All are non-negative adjustment weight coefficients, and satisfy the following conditions: .

[0088] In equation (10), the multi-objective optimization function ensures the biased selection of high-quality advertisements with "high matching and high returns" through the positive addition of the first two terms; and introduces dynamic constraints through the negative subtraction of the third term. Ultimately, these three independent sub-terms with different physical and business attributes adjust their weight coefficients ( Under the cooperative equilibrium of ), it collapses into a bounded scalar of comprehensive utility. In this way, through a dynamic resource penalty, the negative suppression weight of the estimated resource consumption demand of candidate ads is amplified simultaneously when the overall resource pressure index increases.

[0089] This application presents a three-dimensional collaborative evaluation mechanism that integrates user experience preferences, business revenue targets, and hardware capacity. This optimization mechanism effectively avoids the short-sighted evaluation defects of traditional recommendation systems that focus solely on click-through rates or single monetization revenues. It incorporates the terminal's real-time comprehensive resource pressure index as an adaptive damping term into the utility solution model. Through the product coupling design of static advertising characteristics and dynamic terminal states in the multi-objective optimization function, the system can sensitively and quantitatively reduce the computational power allocation weight of high-energy-consuming advertisements when the terminal's operating load intensifies; and automatically remove constraints to maximize monetization efficiency when resources are abundant. Thus, internet marketing distribution control achieves a leap from "one-sided static revenue orientation" to "global optimization under dynamic state constraints," ensuring the objective realization of platform business profitability while constructing an intelligent intervention barrier for terminal devices that combines business flexibility and overload protection.

[0090] Regarding the implementation details of determining the decision gating state of the target terminal device in step S150, in some examples of embodiments of this application, firstly, the comprehensive utility evaluation value is... As a nonlinear smoothing variable, it is input to a variable with adjustable bias. In the activation mapping function, the utility evaluation result is smoothly constrained to an initial deployment probability with a value range of [0, 1]. : Equation (11) In the formula, This is an adjustable bias term used to control the overall advertising frequency of the system. This is a smoothing coefficient greater than 0 used to adjust the sensitivity of the probability mapping curve.

[0091] It should be noted that the comprehensive utility assessment value Essentially, it is a real scalar obtained through multidimensional weighted calculation. Its numerical distribution is not directly equivalent to a probability in a statistical sense, and cannot be directly used as a trigger command for the system. Therefore, this embodiment introduces a Logistic nonlinear activation function with customized parameters. In this function model, the adjustable bias term... It acts as the activation centerline of the probability mapping: when At that time, the calculated initial deployment probability is strictly equal to 0.5; smoothing coefficient This determines the slope of the function curve near the bias point (i.e., its sensitivity to fluctuations in utility fraction).

[0092] Through this nonlinear smoothing mapping, the system not only successfully forces the evaluation scalar with no fixed boundary to converge to the standard probability space [0, 1], but also exposes characteristic parameters with macro-control capabilities to the operations and maintenance side. Operations and maintenance personnel or the cloud-based strategy center only need to dynamically fine-tune the bias terms. This allows for the overall raising or lowering of the system's benchmark advertising exposure level on the terminal without altering the underlying complex computing power model, achieving agile decoupling and rapid control of business deployment scale and terminal-side macro load.

[0093] Then, based on the target task type of the current user interaction task, a security defense verification of the terminal interface is performed: in response to the verification result that the target task type is determined to be a security-constrained task, a forced blocking mechanism to prevent interference is triggered, and the characteristic value of the decision gating state is changed. Forced to be set to 0; in response to the verification result that the target task type is not determined to be a security-constrained task, the characteristic value of the decision gating state is set. Set to release status flag 1.

[0094] It should be noted that during daily operation, terminal devices inevitably encounter certain task scenarios with extremely high security constraints on interaction continuity and interface exclusivity (e.g., full-screen in-vehicle navigation, emergency SOS calls, real-time video conferencing, or fingerprint payment verification). At this level, the system performs real-time context verification by polling the operating system's process focus and sensor concurrency status. Once the current interaction task matches a security constraint task identifier library issued locally or in the cloud, the system immediately activates the highest-priority defense mechanism, assigning a gating feature value through strong logic blocking. This discretized gating mechanism establishes a safety baseline that is independent of the commercial revenue evaluation model, eliminating the risk of terminal basic function failure or serious user experience incidents caused by commercial deployment at the underlying control engineering level.

[0095] Accordingly, regarding the implementation details of determining the final delivery probability of candidate advertisements in step S160, in some examples of embodiments of this application, the feature values ​​are converted through scalar multiplication. Compared with the initial delivery probability Cascaded coupling is performed to output the final delivery probability of candidate ads while ensuring absolute non-interference characteristics when the target terminal device performs security-constrained tasks. : Equation (12) In the formula, Indicates the final probability of delivery. Indicates the initial probability of deployment. The characteristic value of the decision-gated state.

[0096] In equation (12), by employing a product coupling method, the system control architecture is equivalent to an asynchronous AND gate controller, which converts the continuous analog quantity representing the business evaluation scale ( ) and discrete digital quantities representing the underlying security state of the system ( This has been deeply integrated. Through this cascading control, it is ensured that the terminal advertising engine strictly follows the probability distribution of multi-dimensional utility in normal scenarios; while in the millisecond period triggered by any security constraint scenario, all advertising requests are instantly multiplied by 0 and silently intercepted, thereby achieving seamless connection and strict constraints between business distribution and terminal security.

[0097] This application's embodiments construct a hierarchical delivery decision-making system that intertwines flexible probability mapping with hard gating blocking. Utilizing a parameter-adjustable nonlinear activation mapping model, complex cross-domain multi-objective utility scalars are smoothly converged into a baseline delivery probability with globally controllable attributes. Simultaneously, highly sensitive interaction scenarios at the terminal level are rigorously abstracted into discrete binary gating features. Through the scalar cascade coupling operation of these two aspects, the terminal advertising control engine achieves strict decoupling and dynamic re-integration between commercial monetization logic and system scenario security. This not only endows cloud-based strategy delivery with high flexibility in high-dimensional business macro-control but also establishes an inviolable user experience red line at the terminal-side low-level execution stage, thereby deploying an intelligent decision-making center that combines business flexibility with extremely high engineering determinism.

[0098] Regarding the implementation details of controlling the display of candidate advertisements on the target terminal device based on the final delivery probability in step S160, in some examples of the embodiments of this application, after the system obtains the final delivery probability of each candidate advertisement, it does not adopt a simple real-time concurrent display, but introduces an asynchronous scheduling mechanism based on queue management and time window yielding.

[0099] First, obtain a queue of candidate ads to be delivered, which is composed of multiple candidate ads. Sort the candidates in descending order according to the final delivery probability of each candidate ad, and select the candidate ads whose final delivery probability is greater than the preset display safety threshold to construct the target ad set.

[0100] It should be noted that in actual commercial advertising scenarios, ad retrieval engines typically issue a batch sequence containing multiple candidate materials. To optimize processing efficiency on the client side, this embodiment performs a strict priority division on this sequence, reordering the queue according to the final delivery probability to ensure that ads with the best overall performance in terms of utility and safety are given priority to enter the rendering-ready state. Simultaneously, a truncation operation is performed using a preset display safety threshold, directly discarding those tail ads with insufficient probability from memory. This sorting and truncation mechanism effectively narrows the candidate range of target ads, ensuring not only the priority delivery of high-quality ads but also freeing up memory space occupied by redundant ads at the underlying cache management level, avoiding the ineffective occupation of system resources by inefficient ads.

[0101] Then, extract the historical timestamp of the last successful ad display on the target terminal device, and combine it with the current comprehensive resource pressure index to initiate a resource adaptive dynamic fallback scheduling mechanism to calculate the minimum allowed delivery interval. : Equation (13) In the formula, This is the minimum allowed delivery interval. Pre-configure the stress-free baseline delivery interval for the system (e.g., set to a standard 30 seconds). The interval yield factor used to control yield sensitivity (e.g., set to 1.5 or 2.0). This represents the current comprehensive resource pressure index.

[0102] In equation (13), This represents the basic advertising cooling period of the equipment under ideal no-load conditions, while This constitutes a linear expansion multiplier. When terminal resources are abundant ( When the expansion multiplier approaches 1, the system maintains its basic cooling period; however, when the terminal resource pressure increases dramatically ( When this happens, the inflation multiplier will increase linearly to [a certain value]. This forces a proportionally longer minimum delivery interval.

[0103] Through the aforementioned dynamic fallback scheduling mechanism, the current resource pressure level of the terminal is directly mapped to a cooling penalty in the time dimension. This allows the system to adaptively reduce the frequency of ad fetching and rendering when the device load is high, providing sufficient buffer window for the release and heat dissipation of underlying hardware resources. Specifically, by using the dynamic fallback scheduling mechanism, when the overall resource pressure index of the target terminal device increases, the cooling gap between consecutive ad displays is forcibly extended based on a linearly expanding time window, thereby dispersing the peak consumption of computing power and network resources in the time dimension.

[0104] Then, the difference between the current system time and the historical timestamp is calculated, and if the difference is greater than or equal to the minimum delivery interval, the system will proceed accordingly. At that time, the top-ranked target ad is extracted from the target ad set and rendered and displayed on the target terminal device.

[0105] In some implementations, the elapsed time difference is calculated using the following formula: Equation (14) In the formula, This represents the calculated difference in elapsed time. This indicates the current system time that has been read. This represents the historical timestamp of the last successful ad display on the target terminal device.

[0106] At the execution level, the system will calculate Dynamically generated from the previous level Perform real-time comparison. If This indicates that the system is still in a dynamic cooling-off period, and display requests will be temporarily suspended or blocked; if and only if Only when this time barrier is breached will the system accurately extract the top-ranked ad from the target ad set and call the underlying rendering engine (such as WebView or native components) to execute the actual exposure and display. This completely eliminates the risk of resource exhaustion caused by concurrent ad requests, ensuring that each high-concurrency ad rendering action is smoothly distributed across safe time points, greatly improving the smoothness of the terminal interface and the responsiveness of the main process.

[0107] This application's embodiments construct a time-oriented adaptive resource limiting and scheduling distribution mechanism, transforming the comprehensive resource pressure index into a dynamic yield penalty on the time axis. This allows the cooldown window for ad display to exhibit elastic scaling characteristics based on fluctuations in the terminal's physical load. Combined with priority queue threshold truncation and sorting strategies, the system can force peak shaving and valley filling in the time domain by lengthening the display interval when terminal computing power is strained, effectively avoiding resource congestion and system lag caused by continuous ad rendering. When resource conditions permit, it can quickly restore the normal distribution rhythm. Thus, ad delivery control forms a complete defensive closed loop in both spatial (resource constraints) and temporal (interval yield) dimensions, ensuring business reach efficiency while endowing the terminal's underlying hardware system with extremely high shock resistance and operational robustness.

[0108] Figure 3 The diagram illustrates an example of an operation flowchart for performing a parameter adaptive tuning step based on reinforcement learning and privacy computation in the method according to an embodiment of this application. Here, in order to free the system from dependence on manual static rules and endow the model with the ability to self-evolve in complex hardware environments, the system introduces an edge-cloud collaborative federated reinforcement learning architecture to achieve closed-loop evolution of global control parameters.

[0109] like Figure 3 As shown, in step S310, after the target terminal device completes the display control of the candidate advertisement, the user's interaction feedback event on the candidate advertisement is monitored in real time within a preset time window, and the dynamic increment of the device heat temperature and the current processor utilization rate during the loading and rendering cycle of the candidate advertisement is extracted in a synchronous manner to quantify the underlying resource fluctuation penalty value.

[0110] The end of the ad display is not the end of the control chain, but rather the starting point for model iteration and feedback. To accurately characterize the instantaneous hardware impact of ad rendering, the system records various physical state indicators before and after ad loading and rendering in the background. In a preferred embodiment, the underlying resource fluctuation penalty value... It can be calculated using the state difference function:

[0111] Equation (15) In the formula, and These represent the peak device temperatures before ad loading and during the rendering cycle, respectively. and These represent the corresponding peak current processor utilization rates; and This is a preset resource fluctuation penalty normalization coefficient. In this way, the originally implicit abstract engineering phenomena such as "device overheating" and "instantaneous lag" can be rigorously transformed into discrete mathematical scalars that can be used for algorithm optimization, thereby providing an accurate hardware loss assessment benchmark for reinforcement learning models.

[0112] In step S320, a multi-objective reward function combining commercial benefits and hardware experience is constructed, which maps interactive feedback events representing positive user intentions to positive benefit incentives and maps underlying resource fluctuation penalty values ​​to negative experience suppression.

[0113] Here, the evolution direction of the reinforcement learning agent can be defined by the reward function, and the system can construct the comprehensive reward value for a single demonstration decision using the following formula. : Equation (16) In the formula, A binary discrete value representing a positive incentive (e.g., 1 for a user click or stays on the page, and 0 for closing or ignoring the page). This is the underlying resource fluctuation penalty value output by the higher-level quantification; and These are the incentive weights and suppression weights used to balance business exploration with a conservative hardware strategy.

[0114] By anchoring "improving business conversion rate" and "suppressing abnormal hardware fluctuations" within the same optimization space through the aforementioned multi-objective reward function, the agent must balance business efficiency and system operation security during trial and error exploration, avoiding falling into a local optimum that simply pursues click-through rate while ignoring the risk of device crashes.

[0115] In step S330, based on a multi-objective reward function that includes positive reward incentives and negative experience suppression, the locally deployed reinforcement learning agent is driven to iterate the execution policy in order to calculate the adjustment weight coefficients for the comprehensive utility evaluation function. , , and adjustable bias terms The local gradient is updated.

[0116] Specifically, a lightweight policy gradient algorithm model is deployed on the terminal side. The agent aims to maximize the long-term expected reward. To optimize the objective, the backpropagation mechanism is used to optimize the core parameter set (denoted as ) of the aforementioned comprehensive utility evaluation and probability mapping stages. The local update gradient is obtained by performing differentiation. Therefore, by performing gradient calculations locally on the device side, the system can deeply adapt to the aging state of a specific device and the personalized interaction habits of a single user, achieving customized parameter tuning for each device, while avoiding off-end transmission of raw sensitive behavioral data from a physical isolation perspective.

[0117] In step S340, the local differential privacy algorithm is invoked to inject random perturbation noise that satisfies the preset privacy budget into the local update gradient, thereby generating noisy gradient data.

[0118] To defend against malicious attackers potentially using gradient inversion attack to reverse-engineer a user's private interaction history, the system introduces a differential privacy (DP) mechanism. Specifically, it updates the extracted local gradients... Apply Gaussian noise interference:

[0119] Equation (17) In the formula, This is the final generated noisy gradient data; To conform to a mean of 0 and a variance of The Gaussian noise matrix, where the noise standard deviation is... Strictly controlled by the system's default privacy budget. ; It is the identity matrix. Therefore, this noise-adding mechanism provides theoretical-level privacy guarantees within a rigorous mathematical framework. -DP guarantee) ensures that even if the transmission channel is monitored, the outside world cannot parse the specific sample characteristics of a single terminal, thus severing the link between data availability and privacy leakage.

[0120] In step S350, the noisy gradient data is asynchronously uploaded to the cloud marketing strategy server through the federated learning secure aggregation channel, so that the cloud marketing strategy server can perform the fusion update and distribution of global control parameters without collecting the original interaction data and device physical characteristics, thereby realizing the adaptive parameter tuning of end-cloud collaboration.

[0121] The cloud-based policy server acts as a central aggregation node, receiving noisy gradient data asynchronously uploaded from numerous terminals. ( (This refers to the terminal device serial number). Using FedAvg or Secure Multi-Party Computation (SMPC) protocols, the global average gradient is calculated in the cloud, and then the global control parameter model is iteratively updated. Subsequently, the updated global parameter weights are distributed to each edge terminal, overwriting their local initial model. This architecture effectively alleviates the data silo effect and achieves a high-order collaborative evolution mechanism of "data localization and model remoteization." With this mechanism, the system can integrate exploration experience from massive heterogeneous terminal devices, enabling the advertising delivery decision engine to iteratively optimize strategies across devices and scenarios, thereby approaching the globally optimal delivery solution. Simultaneously, this architecture avoids the compliance and information leakage risks associated with centralized storage of user data, providing a foundation for the system to provide efficient decision support while protecting user privacy.

[0122] This application presents a high-order adaptive optimization system that integrates federated reinforcement learning and local differential privacy. This system jointly models the underlying physical resource fluctuation characteristics with the surface-level business interaction feedback to drive control parameters to autonomously optimize within the complex boundary between "maximizing revenue" and "minimizing hardware loss." By combining edge computing, gradient noise addition, and cloud aggregation in a federated flow mechanism, it abandons the outdated approach of traditional advertising engines that heavily rely on centralized cloud data mining, establishing an intelligent closed-loop mechanism of "downward computing power, experience sharing, and absolute privacy isolation." This not only enables the system's delivery decision-making strategy to undergo unsupervised long-term self-correction as the external environment changes and device status deteriorates, but also constructs the highest standard of user privacy protection barriers in both legal and engineering dimensions, achieving a leap from intelligent marketing technology to a secure and trustworthy computing paradigm.

[0123] Figure 4 The diagram illustrates an example of the system operation principle of an internet marketing placement control method based on terminal device state awareness according to an embodiment of this application. The entire dynamic placement decision-making chain is systematically divided into four highly decoupled yet closely linked macro-processing stages: input data, core processing, control and gating, and output and feedback.

[0124] like Figure 4 As shown, firstly, in the "input data" stage, the system collects status and tasks in real time through the underlying sensors and system interface, and simultaneously pulls the user's historical behavior logs and the advertising metadata of the candidate ads to be delivered, thereby gathering features in both the business end and the device end, providing multi-dimensional source data support for subsequent scenario-based perception and personalized decision-making.

[0125] In the "core processing" phase, the system constructs a deep state and utility modeling mechanism. On the one hand, it performs resource pressure calculations based on real-time hardware status to quantify the risk of load exhaustion at the terminal level; on the other hand, it combines user logs to perform user interest model matching. These two dynamic and static features, along with the advertising metadata (expected revenue and estimated consumption), are jointly fed into a multi-objective utility function for joint game theory.

[0126] Subsequently, the system enters the "control and gating" phase, where the utility scalar is smoothly mapped to a baseline probability through the initial probability evaluation module. Then, it seamlessly connects to the decision gating and yielding module, where a strict security constraint judgment logic is executed: if the current scenario is a security-sensitive scenario, a hardware-level blocking mechanism is directly triggered; if it is not a sensitive scenario, it is allowed and a yielding cooldown time is scheduled based on the terminal pressure status. Thus, within the algorithm framework, absolute isolation between the elasticity of business distribution and the bottom line of non-interference in operation is achieved.

[0127] Finally, in the "output and feedback" phase, the system executes the ad presentation action on the terminal interface based on the sorting and yield scheduling results. Simultaneously, the dynamic surges in underlying physical resources triggered during the ad display cycle, as well as the actual user interaction results, are collected again as closed-loop reward signals and flow to the update module to execute iterative end-to-cloud collaborative strategies. Through this continuous feedback and evolution mechanism, this internet marketing delivery control system can continuously adapt to the complex and ever-changing physical environment in actual engineering deployments, ultimately building a high-level intelligent defense and distribution network on the terminal side that balances business value realization with the underlying security of device operation.

[0128] To fully verify the effectiveness of the internet marketing placement control method based on terminal device status awareness proposed in this application, especially the robustness of the comprehensive resource pressure index modeling and dynamic gating mechanism in complex environments, this application embodiment constructs a hybrid simulation experimental environment based on the combination of "real data playback and physical state simulation".

[0129] In terms of data construction, to ensure the objectivity and authenticity of user interaction behavior, the experiment used a publicly available mobile advertising dataset as the basic reference for user interests and click behavior. Based on this, a device state enhancement generator was introduced, using Markov chains to simulate and generate a continuous stream of terminal physical state data. This state data stream specifically includes battery discharge curves to simulate the nonlinear voltage drop process of lithium-ion batteries under different loads, processor thermal scheduling characteristics to simulate temperature rise and frequency reduction phenomena under high loads, and task interruption streams randomly injected with high-priority events such as calls, navigation, or driving.

[0130] In the aforementioned hybrid simulation environment, three delivery control strategies were implemented in parallel for performance comparison. The first was the traditional greedy revenue baseline strategy (Baseline A), which ranks ads solely based on historical click-through rate estimates and ad bids, prioritizing short-term revenue without considering device status. The second was a fixed-rule environment-adaptive baseline strategy (Baseline B), which uses preset static thresholds for control, such as randomly discarding 50% of ad requests when battery power is below 20% or network downlink bandwidth is below 200KB / s. The third was the state-aware delivery strategy proposed in this application, which calculates a multi-dimensional Comprehensive Resource Pressure Index (SPC), combines it with the resource-aware utility evaluation value of candidate ads for comprehensive scoring, and finally executes delivery through dynamic probability mapping and decision gating mechanisms.

[0131] To quantitatively evaluate the overall performance of different strategies, the experiment defined three core evaluation metrics. The first is the effective click-through rate (eCTR), which is the ratio of total clicks to the total number of ad impressions, used to measure the commercial efficiency of the campaign. The second is the energy-to-revenue ratio (EER), defined as the amount of battery power consumed to generate one unit of ad revenue; a higher EER indicates that the algorithm is more battery-efficient. The third is the user interference index, calculated based on a task conflict model. When an ad pops up during a user's high-load or high-focus task (such as voice calls or immersive games), it is penalized as a high interference penalty, thus verifying the scenario adaptability of this solution in ensuring the continuity of core user operations.

[0132] Figure 5 This diagram illustrates the comprehensive performance simulation of different delivery control strategies under multi-dimensional evaluation indicators in the embodiments of this application. The simulation results use a five-dimensional radar chart to visually display the normalized performance scores of the three comparison methods on five key evaluation dimensions: click-through rate, conversion revenue, device smoothness, power saving, and user retention prediction. The larger the coverage area of ​​the radar chart, the better the comprehensive effectiveness and edge-cloud balance capability of the control strategy.

[0133] like Figure 5As shown, in terms of baseline comparison, the traditional greedy revenue-driven baseline strategy (represented by the gray dashed line of baseline A) scores highest on the conversion revenue axis, but scores extremely low on the power saving and device smoothness axes. The overall graph presents a sharp polygon heavily skewed towards the revenue side, indicating that this traditional strategy is solely driven by maximizing short-term commercial value, severely overdrawing the underlying physical resources of the terminal and foreshadowing a very high risk of user churn. On the other hand, the environment-adaptive baseline strategy based on fixed rules (represented by the blue solid line of baseline B), while showing some improvement in various hardware resource indicators and a relatively convergent and balanced graph, suffers from a significant numerical dip in click-through rate and conversion revenue due to its crude static threshold truncation mechanism and lack of refined perception and flexible scheduling of users' true interests. This results in a substantial decline in commercial monetization efficiency.

[0134] In comparison, the state-aware deployment control strategy proposed in this invention ( Figure 5 (The solid red line with shading indicates that this approach achieves maximum overall area coverage, demonstrating excellent overall utility. Although its total conversion revenue is slightly lower than the traditional greedy strategy (which remains above 92%), this solution scores near-perfect marks in device smoothness and power saving, and its effective click-through rate surpasses that of invalid and laggy clicks. Specifically, the overall utility evaluation model of this application introduces a dynamic resource penalty term. Through this non-linear penalty mechanism, the system can accurately and adaptively eliminate high-energy-consuming and low-matching-value advertising requests when terminal resources are under high pressure. This ensures the bottom line of core monetization revenue while improving the smoothness of terminal device operation by more than 40%, perfectly balancing the platform's short-term commercial monetization goals with long-term user interaction and retention experience.

[0135] Figure 6 This diagram illustrates the temporal trajectory simulation effect of different ad delivery control strategies on the control response capability in dynamic scenarios. The simulation experiment captured a continuous 60-minute real user interaction session, fully covering three typical terminal operating conditions: "normal use, starting a high-load large-scale game, and connecting to charging." The dual-axis trajectory in the figure clearly reflects the evolution of the terminal's Comprehensive Resource Pressure Index (SPC, represented by the left vertical axis and the blue shaded area in the figure) over time, as well as the corresponding triggered ad delivery actions (represented by the right vertical axis and the scatter plot in the figure).

[0136] like Figure 6As shown, during the normal usage phase from 0 to 20 minutes, the SPC remained stable at a low level, and all control strategies distributed ads at a normal frequency. However, during the high-load period from the 20th to the 40th minute (i.e., the simulated large-scale game scenario), with processor computing power depletion and heat accumulation, the SPC index rapidly and non-linearly expanded to a dangerous level above 0.8. Under this extreme condition, the traditional greedy reward baseline strategy (Baseline A, represented by the gray dots and red crosses in the figure) lacked underlying state awareness and continued to send dense ad requests regardless of hardware load (see the dense red "×" marked areas in the figure). This high-risk delivery not only severely depleted the computing power of the main task but also easily caused severe device lag or even thermal crashes.

[0137] In contrast, the state-aware delivery strategy proposed in this application (represented by the green diamonds and pentagrams in the figure) demonstrates extremely sensitive and superior adaptive resilience. During the aforementioned high-load period, as the SPC index surged, the comprehensive utility evaluation model of this solution simultaneously amplified the dynamic resource penalty for all candidate ads, resulting in a significant reduction in the comprehensive evaluation value of most regular ads. This, in turn, automatically sparsified the actual ad delivery density of the system (successfully suppressing approximately 85% of ordinary interference requests). It is noteworthy that during peak SPC periods, only a very small number of key ads with extremely high commercial conversion value and extremely low resource consumption (marked by the green pentagrams in the figure) were able to overcome the stringent pressure threshold and gain display opportunities.

[0138] Furthermore, when the device connects to charging at the 45-minute mark and the task load decreases, causing a sharp drop in SPC, the frequency of ad delivery in this solution quickly rebounds, accurately capturing the user's subsequent leisure time. This time-series trajectory fully demonstrates that the dynamic control mechanism of this application effectively avoids the disastrous user experience caused by traditional solutions forcibly rendering ads when devices are congested, achieving a refined closed-loop control of intelligent non-interference in high-load scenarios and flexible exposure promotion when resources are abundant.

[0139] Figure 7 This diagram illustrates the performance trade-offs between different ad delivery control strategies in terms of total impressions and effective clicks, based on simulation results. The horizontal axis of this coordinate system represents the total impressions achieved by the system in delivering ads, while the vertical axis represents the actual number of effective clicks triggered by users. The slope of the curves in this coordinate system visually reflects the effective click-through rate (CTR) of the ads.

[0140] like Figure 7As shown, the overall slope of the state-aware delivery strategy proposed in this application is significantly higher than that of the traditional greedy revenue baseline strategy (Baseline A), indicating that the single-exposure conversion efficiency of the proposed solution is greatly improved (overall CTR is improved by approximately 22%). Meanwhile, simulation data shows that compared to the baseline strategy, the total ad exposure of the proposed solution objectively decreases by approximately 18.5%. This decrease in the horizontal axis value is not due to abnormal traffic loss, but rather because the system accurately triggers decision-gating features when facing extreme scenarios such as users performing security-constrained tasks or terminal hardware resources nearing exhaustion. Forced blocking mechanism and based on comprehensive resource pressure index The dynamic backoff scheduling was implemented. Although the total exposure was reduced due to active constraints, the total number of effective clicks on the vertical axis remained at a high level (only a slight decrease of about 3%, with no statistically significant difference). This result strongly demonstrates that the solution successfully filtered out a large number of invalid and redundant exposures that are prone to causing device lag and have extremely low commercial conversion rates.

[0141] Furthermore, for specific marketing scenarios (such as advertisers focusing on indiscriminate brand display to a broad audience while downplaying click metrics) where budget consumption might be delayed due to prioritizing device state protection, the control architecture of this invention exhibits high engineering configurability. The system strategy side requires no changes to the underlying algorithm code; only the weighting coefficients of the dynamic resource penalty in the comprehensive utility evaluation function need to be adjusted by fine-tuning. Alternatively, it can directly intervene in the adjustable bias term in the activation mapping function used to control the overall advertising frequency of the system. This allows for flexible relaxation of the terminal's do-not-disturb and hardware protection thresholds. This parameterized, flexible intervention mechanism empowers the system to achieve rapid decoupling and adaptive trade-offs between diverse business needs and the physical boundaries of the terminal.

[0142] The simulation results demonstrate that the control method based on terminal device state awareness constructed in this application does not involve simple, brute-force global interception, but rather achieves a cross-temporal arbitrage scheduling of computing power and attention resources. This mechanism successfully and precisely shifts and redistributes marketing budgets and terminal hardware computing power from inefficient, redundant exposure periods characterized by poor user experience and near-limited device operating load to high-conversion windows where device performance is good and user attention is highly focused. This temporal peak shaving and valley filling, along with flexible reshaping, fully validates the scientific validity of the aforementioned comprehensive resource pressure index model and multi-objective dynamic utility optimization function, perfectly achieving the comprehensive control objective of balancing user experience (avoiding intrusion) with maximizing commercial monetization efficiency within the constraints of physical hardware boundaries.

[0143] To further analyze the relative contributions and technical synergies of the key functional modules of the system, this application's embodiments supplemented the module ablation verification of the control link. Simulation comparison tests revealed that after stripping away the comprehensive resource pressure index model (i.e., degenerating into a traditional greedy scoring based solely on user interest and business revenue), although the absolute total number of clicks showed a pseudo-increase due to the disorderly surge in exposure, the end-side resource friendliness showed a significant decrease, and user experience satisfaction under high load deteriorated sharply. This, in turn, confirms the decisive role of the underlying physical resource perception mechanism in maintaining the terminal's lifespan. Furthermore, after removing decision gating and yield strategies, the system still frequently triggered forced pop-ups in high-risk constraint scenarios such as simulated full-screen navigation or emergency calls, causing a surge in negative user blocking evaluations, establishing the irreplaceable role of gating features in safeguarding system operational security. Simultaneously, when the parameter adaptive tuning channel based on reinforcement learning and privacy computation was closed, the control engine gradually lost its ability to dynamically track user interest drift and device component aging, resulting in an irreversible decline in the system's long-tail revenue. Based on the above ablation experiment results, the necessity of the high coupling between the multi-dimensional state perception, discrete gating constraints, and edge-cloud co-evolution modules in the framework proposed in this application is fully demonstrated.

[0144] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0145] Figure 8 A structural block diagram of an example of an internet marketing delivery control system based on terminal device status awareness, according to an embodiment of this application, is shown.

[0146] like Figure 8 As shown, the Internet marketing delivery control system 800 based on terminal device status perception includes a terminal status perception unit 810, an advertising information acquisition unit 820, a resource pressure calculation unit 830, an advertising effectiveness evaluation unit 840, a delivery decision generation unit 850, and a gating adjustment unit 860.

[0147] The terminal status sensing unit 810 is used to collect the current operating status data of the target terminal device in real time and identify the current user interaction task of the target terminal device.

[0148] The advertising information acquisition unit 820 is used to acquire the advertising metadata of the candidate advertisement to be delivered and the user interest model constructed for the target terminal device. The advertising metadata includes the estimated resource consumption requirements and expected business revenue of the candidate advertisement.

[0149] The resource pressure calculation unit 830 is used to calculate a comprehensive resource pressure index to quantify the overall operating load level of the target terminal equipment based on the current operating status data.

[0150] The advertising utility evaluation unit 840 is used to determine the interest matching degree of the candidate advertisement based on the user interest model, and to evaluate and calculate the comprehensive utility evaluation value of the candidate advertisement by combining the comprehensive resource pressure index, the interest matching degree, and the expected business revenue and estimated resource consumption requirements in the advertisement meta information.

[0151] The delivery decision generation unit 850 is used to calculate the initial delivery probability for the candidate advertisement based on the comprehensive utility evaluation value, and to determine the decision gating state of the target terminal device based on the current user interaction task.

[0152] The gating adjustment unit 860 is used to adjust the initial delivery probability using the decision gating state to obtain the final delivery probability for the candidate advertisement, and to control the display of the candidate advertisement on the target terminal device based on the final delivery probability.

[0153] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. The execution instructions can be read and executed by electronic devices (including but not limited to computers, servers, or network devices) to perform the steps of any of the above-described internet marketing delivery control methods based on terminal device state awareness.

[0154] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the steps of any of the above-described internet marketing delivery control methods based on terminal device state awareness.

[0155] In some embodiments, this application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of an internet marketing delivery control method based on terminal device state awareness.

[0156] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0157] The electronic devices in this application can exist in various forms, including but not limited to: mobile communication devices, ultra-mobile personal computer devices, portable entertainment devices, or other airborne electronic devices with data interaction functions.

[0158] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0159] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for controlling internet marketing campaign delivery based on terminal device status awareness, applied to target terminal devices, characterized in that: The method includes: Real-time acquisition of the current operating status data of the target terminal device, and identification of the current user interaction task of the target terminal device; Obtain the advertising metadata of the candidate ads to be delivered and the user interest model constructed for the target terminal device. The advertising metadata includes the estimated resource consumption requirements and expected business revenue of the candidate ads. Based on the current operating status data, a comprehensive resource pressure index is calculated to quantify the overall operating load level of the target terminal device. Based on the user interest model, the interest matching degree of the candidate advertisement is determined, and combined with the comprehensive resource pressure index, the interest matching degree, and the expected business revenue and estimated resource consumption requirements in the advertisement metadata, the comprehensive utility evaluation value of the candidate advertisement is evaluated and calculated. The initial delivery probability for the candidate advertisement is calculated based on the comprehensive utility evaluation value, and the decision gating state of the target terminal device is determined based on the current user interaction task. The initial delivery probability is gating and adjusted using the decision gating state to obtain the final delivery probability for the candidate advertisement, and the display of the candidate advertisement on the target terminal device is controlled based on the final delivery probability.

2. The method according to claim 1, characterized in that, The real-time acquisition of the current operating status data of the target terminal device and the identification of the current user interaction task of the target terminal device include: Call the system kernel monitoring interface of the target terminal device to obtain system resource monitoring indicators in real time. The system resource monitoring indicators include at least the current remaining battery power, the current processor utilization rate, and the device temperature. Call the network status detection interface to obtain the currently available network downlink bandwidth; The current remaining battery power, the current processor utilization, the device temperature, and the current available network downlink bandwidth are time-aligned and fused together to form the current operating status data. Listen to the operating system active window events of the target terminal device, and jointly detect the concurrent occupancy status of the system-level audio channel and location service channel to extract the running characteristics of the currently executing application; The intent classification of the running features is performed based on a preset task mapping model to identify the target task type of the current user interaction task; wherein the target task type is one of security-constrained task, focus-exclusive task, and background silent task.

3. The method according to claim 2, characterized in that, The calculation of the comprehensive resource pressure index, based on the current operating status data, to quantify the overall operating load level of the target terminal device includes: The current remaining battery power, current processor utilization, and current available network downlink bandwidth in the current operating status data are used as physical status indicators. Each is obtained through the corresponding nonlinear risk mapping function. Normalization mapping is performed to obtain the corresponding stress risk factor with a value range of [0, 1]. The stress risk factors include power stress factor, processor stress factor, and network stress factor. Based on the target task type of the current user interaction task, dynamically configure an attention pressure factor to characterize the user's attention load on the terminal. Its value range is [0, 1]; wherein, when the target task type is identified as the focus-exclusive task, the attention pressure factor is assigned. Extreme parameters that exceed the preset focus threshold; Extract the device's heating temperature from the current operating status data, and set a dynamic temperature penalty index based on the device's heating temperature. The dynamic temperature penalty index The initial baseline value is 1, and when the device's heating temperature exceeds the preset safe temperature threshold, the dynamic temperature penalty index... The value increases exponentially and non-linearly. The comprehensive resource pressure index is calculated by invoking a joint probabilistic risk model that incorporates the dynamic temperature penalty index: ， In the formula, This represents the comprehensive resource pressure index. The total number of dimensions of the physical state indices used in the calculation; Indicates the first Item physical state index, For the first The normalized weight parameters corresponding to the physical state index and satisfy the following conditions: Through the multiplication term decay mechanism in the joint probability risk model and the exponential operation of the dynamic temperature penalty index, when the target terminal device experiences abnormally high temperature or is currently executing the focus-exclusive task, the risk weight of the local resource bottleneck is forcibly amplified, so that the comprehensive resource pressure index exponentially approaches the maximum value 1.

4. The method according to claim 3, characterized in that, The nonlinear risk mapping function A differentiated mapping model is used for calculation; wherein, the power pressure factor, the processor pressure factor, and the network pressure factor are respectively represented as follows: , as well as Specifically, it includes: Use the current remaining battery power as an input variable. The energy pressure factor is obtained by calling the calculation formula based on reverse S-shaped attenuation. To prevent the battery level from falling below a preset safe battery warning threshold. At that time, risk manifestations that trigger an exponential surge are: ， Use the current processor utilization rate as an input variable. Call with load-strengthening index The processor stress factor is obtained using the power function formula. To penalize the risk of performing nonlinear amplification in high processor utilization ranges: ， Using the currently available network downlink bandwidth as an input variable The network stress factor is obtained by calling the inverse linear calculation formula. To construct a direct mapping of bandwidth scarcity: ， In the formula, This represents the current remaining battery power. The safe power threshold is set. To control the constant parameter that ensures the slope of the decrease in electricity risk is greater than 0, It is a natural constant; Current processor utilization This is the processor's maximum theoretical utilization constant. A load enhancement index greater than 1; This represents the currently available network downlink bandwidth. This is the maximum bandwidth value for reference.

5. The method according to claim 4, characterized in that, The process of determining the interest matching degree of the candidate advertisement based on the user interest model, and evaluating and calculating the comprehensive utility assessment value of the candidate advertisement by combining the comprehensive resource pressure index, the interest matching degree, and the expected business revenue and estimated resource consumption requirements in the advertisement metadata, includes: The degree of interest matching determined based on the user interest model This is represented as a user experience gain term; Expected business revenue Compared to the system-defined normalized reference maximum value for returns The ratio of these values ​​represents the business value gain. Projected resource consumption demand Compared to the system's set normalized reference maximum resource consumption The ratio of the comprehensive resource pressure index Perform product coupling to construct a dynamic resource penalty term; The comprehensive utility evaluation value of the candidate advertisement is calculated by invoking a multi-objective optimization function that includes the user experience gain term, the business value gain term, and the dynamic resource penalty term: ， In the formula, This is the overall utility evaluation value for the candidate advertisements; The degree of interest matching is represented, and its value range is [0, 1]. For expected business revenue, The maximum normalized reference value for revenue set for the system; To estimate resource consumption requirements, a quantitative value representing the total computing power and traffic resources expected to be consumed by the candidate advertisement during the loading and rendering phases is provided. The normalized reference maximum value for resource consumption set by the system; A comprehensive resource pressure index; , and All are non-negative adjustment weight coefficients, and satisfy the following conditions: ; By means of the dynamic resource penalty item, when the comprehensive resource pressure index increases, the negative suppression weight of the estimated resource consumption demand of the candidate advertisement is simultaneously amplified.

6. The method according to claim 5, characterized in that, The step of calculating the initial delivery probability for the candidate advertisement based on the comprehensive utility evaluation value, and determining the decision gating state of the target terminal device based on the current user interaction task, includes: Comprehensive utility assessment value As a nonlinear smoothing variable, it is input to a variable with adjustable bias. In the activation mapping function, the utility evaluation result is smoothly constrained to the initial deployment probability with a value range of [0, 1]. : ， In the formula, This is an adjustable bias term used to control the overall advertising frequency of the system. A smoothing coefficient greater than 0 is used to adjust the sensitivity of the probability mapping curve; For the target task type of the current user interaction task, perform a security defense verification on the terminal interface: in response to the verification result that the target task type is determined to be a security-constrained task, trigger a forced blocking mechanism to prevent interference, and pass the feature value of the decision gating state. Forced to be set to 0; in response to the verification result that the target task type is not determined to be the security constraint type task, the feature value of the decision gating state is set to 0. Set the clearance status flag to 1; The step of adjusting the initial delivery probability using the decision gating state to obtain the final delivery probability for the candidate advertisement includes: Eigenvalues ​​are obtained through scalar product operations. Compared with the initial delivery probability Cascaded coupling is performed to output the final delivery probability for the candidate advertisements while ensuring that the target terminal device has absolute non-interference characteristics when performing the security-constrained task. : ， In the formula, Indicates the final probability of delivery. Indicates the initial probability of deployment. is the feature value of the decision-gated state.

7. The method according to claim 6, characterized in that, The step of controlling the display of the candidate advertisement on the target terminal device based on the final delivery probability includes: Obtain a queue of candidate ads to be delivered, which is composed of multiple candidate ads. Sort the candidates in descending order according to the final delivery probability of each candidate ad. Select the candidate ads whose final delivery probability is greater than the preset display safety threshold to construct a target ad set. Extract the historical timestamp of the last successful ad display on the target terminal device, and combine it with the current comprehensive resource pressure index to initiate a resource adaptive dynamic fallback scheduling mechanism to calculate the minimum allowed delivery interval. : ， In the formula, This is the minimum allowed delivery interval. The pre-configured stress-free baseline delivery time interval for the system, The interval yield coefficient is used to control the yield sensitivity. Calculate the elapsed time difference between the current system time and the historical timestamp, and when the elapsed time difference is detected to be greater than or equal to the minimum delivery interval... At that time, the top-ranked target advertisement is extracted from the target advertisement set and rendered and displayed on the target terminal device.

8. The method according to claim 6 or 7, characterized in that, The method further includes a parameter adaptive tuning step based on reinforcement learning and privacy computation, specifically including: After the target terminal device completes the display control of the candidate advertisement, the user's interaction feedback event on the candidate advertisement is monitored in real time within a preset time window, and the dynamic surge increment of the device's heat temperature and the current processor utilization rate during the loading and rendering cycle of the candidate advertisement is extracted in time to quantify it as the underlying resource fluctuation penalty value. Construct a multi-objective reward function that combines commercial benefits and hardware experience, mapping the interactive feedback events that represent positive user intentions to positive benefit incentives, and mapping the underlying resource fluctuation penalty value to negative experience suppression; Based on the multi-objective reward function that includes the positive reward incentive and the negative experience suppression, the locally deployed reinforcement learning agent is driven to perform policy iterations to calculate the adjustment weight coefficients for the comprehensive utility evaluation function. , , and adjustable bias terms Local gradient update; The local differential privacy algorithm is invoked to inject random perturbation noise that meets the preset privacy budget into the local update gradient, thereby generating noisy gradient data; Through a secure aggregation channel of federated learning, the noisy gradient data is asynchronously uploaded to the cloud marketing strategy server, so that the cloud marketing strategy server can perform the fusion update and distribution of global control parameters without collecting the original interaction data and device physical characteristics, thereby realizing the adaptive parameter tuning of end-to-cloud collaboration.

9. An internet marketing delivery control system based on terminal device status perception, characterized in that, The system includes: The terminal status sensing unit is used to collect the current operating status data of the target terminal device in real time and identify the current user interaction task of the target terminal device; An advertising information acquisition unit is used to acquire advertising metadata of candidate advertisements to be delivered and a user interest model constructed for the target terminal device. The advertising metadata includes the estimated resource consumption requirements and expected business revenue of the candidate advertisements. The resource pressure calculation unit is used to calculate a comprehensive resource pressure index based on the current operating status data, which is used to quantify the overall operating load level of the target terminal equipment. An advertising utility evaluation unit is used to determine the interest matching degree of the candidate advertisement based on the user interest model, and to evaluate and calculate the comprehensive utility evaluation value of the candidate advertisement by combining the comprehensive resource pressure index, the interest matching degree, and the expected business revenue and the estimated resource consumption demand in the advertisement metadata. The delivery decision generation unit is used to calculate the initial delivery probability for the candidate advertisement based on the comprehensive utility evaluation value, and to determine the decision gating state of the target terminal device based on the current user interaction task. The gating adjustment unit is used to adjust the initial delivery probability using the decision gating state to obtain the final delivery probability for the candidate advertisement, and to control the display of the candidate advertisement on the target terminal device based on the final delivery probability.

Images