Intelligent scene perception method and system based on time sequence scene classification model

By using an intelligent scene perception method based on a time-series scene classification model, changes in user scenarios are dynamically identified and predicted, enabling seamless switching and resource optimization of smart devices, thereby improving user experience and device performance.

CN121455347APending Publication Date: 2026-02-03SHENZHEN XINLONGPENG TECH CO LTD
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
CN202610013250.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-07
Publication Date
2026-02-03

AI Technical Summary

Technical Problem

Existing smart devices and systems have shortcomings in function switching, resource allocation, and interaction design, resulting in cumbersome operation, waste of resources, and reduced user experience, and an inability to dynamically respond to user needs.

Method used

An intelligent scene perception method based on a time-series scene classification model is adopted. By acquiring multi-dimensional data, a two-layer hybrid model is constructed to identify and predict scenes, and resource allocation and interaction strategies are dynamically adjusted.

Benefits of technology

It achieves seamless scene switching, improves resource utilization and interaction efficiency, extends battery life, and solves the problem of cumbersome operation in traditional systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121455347A_ABST
    Figure CN121455347A_ABST
Patent Text Reader

Abstract

The invention relates to an intelligent scene perception method and system based on a time sequence scene classification model. The method comprises the following steps: obtaining multi-dimensional data from a user behavior, an application framework layer, a sensor center and a system state; slicing the multi-dimensional data according to a time window, and fusing and converting the multi-dimensional data into feature vectors; performing scene recognition processing on the feature vector by adopting a double-layer hybrid model to output a clear scene, and performing strategy matching on the output clear scene based on a preset scene-strategy mapping table; by means of the method, through machine learning and behavior analysis, automatic recognition of user use scenes and intelligent adjustment of system parameters are achieved, the system can not only recognize the current scene, but also predict possible scene changes, resource preparation is conducted in advance, and the user experience is improved. The tedious problem that a traditional system needs to be manually adjusted and set by a user is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent scene perception technology, and more specifically, to an intelligent scene perception method and system based on a temporal scene classification model. Background Technology

[0002] With the rapid development of smart devices and software systems, users' demands for personalized, efficient, and seamless experiences are growing daily. However, existing technologies generally face three core challenges that hinder further improvements in system performance and user experience: 1. Insufficient feature stacking and scenario adaptation: Modern systems typically integrate multimodal functions (such as navigation, entertainment, and office work), but lack context awareness. Users must manually adjust modes when switching between different tasks, leading to cumbersome operations. For example, while driving, one needs to switch from navigation mode to music playback, or while reading, one needs to close background applications that might interfere with reading. This static functional design not only reduces efficiency but also increases cognitive load, potentially causing safety hazards, especially in emergency or complex scenarios.

[0003] 2. Resource allocation and load imbalance: System resource management often employs a "one-size-fits-all" strategy, failing to dynamically respond to actual needs. In light-load scenarios (such as standby or basic browsing), hardware (such as CPU and GPU) still runs at full speed, resulting in energy waste; while in high-load scenarios (such as gaming or multitasking), insufficient resource allocation leads to lag, latency, or even crashes. This mismatch stems from the limitations of traditional algorithms, which fail to intelligently schedule resources based on real-time task priorities and hardware status.

[0004] 3. Staticization of interaction logic: UI layouts and interaction designs are often fixed, making it difficult to focus on the core task at hand. For example, when watching a movie, control buttons may be scattered around the edges of the screen, requiring additional operations to access them; when reading, advertisements or recommended content may distract the viewer. This singular interaction pattern ignores the immediacy of user intent, reducing immersion and operational smoothness.

[0005] In summary, existing solutions (such as rule-based resource scheduling or preset UI templates) can no longer meet the needs of dynamic environments. There is an urgent need to introduce intelligent scene perception methods and systems based on adaptive algorithms and context awareness technology to solve the problems of device functional rigidity and unreasonable resource allocation in different usage scenarios, and to realize "scenario-driven" intelligent resource allocation. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide an intelligent scene perception method based on a temporal scene classification model, and to provide an intelligent scene perception system based on a temporal scene classification model, in order to address the above-mentioned deficiencies of the prior art.

[0007] The technical solution adopted by this invention to solve its technical problem is: An intelligent scene perception method based on a temporal scene classification model is constructed, wherein the method includes the following steps: Acquire multi-dimensional data from user behavior, application framework layer, sensor hub, and system status; Multi-dimensional data is sliced ​​according to time windows and then fused into feature vectors. A two-layer hybrid model is used to process the feature vectors for scene recognition and output a clear scene. The first layer is a high-speed filter for recognizing clear scenes, and the second layer is a temporal scene classification model for recognizing complex scenes. The policy is matched to the output specific scenario based on the preset scenario-policy mapping table.

[0008] The intelligent scene perception method based on a temporal scene classification model described in this invention, wherein, User behavior data includes one or more of the following: clicks, swipes, and dwell time; Application framework layer data includes one or more of the following: currently active application, usage frequency, foreground application package name, and window stack; Sensor central data includes: IMU determining whether the device is stationary / moving / rotating, and light sensor determining one or more of the following: ambient light or darkness; System status data includes one or more of the following: network connection type, Bluetooth peripherals, and CPU / GPU load timings.

[0009] The intelligent scene perception method based on a temporal scene classification model described in this invention further includes using the temporal scene classification model to predict future scenes and preload resources based on the predicted future scenes, specifically including: Let the input sequence be ,in It is a multidimensional feature vector; LSTM at each time step The fused temporal feature vectors Compared to the previous hidden state Enter them together; Forget Gate Control Vector Determines the cell state at the previous moment The proportion of information to be retained in each part of the data, and the calculation of the forget gate control vector, use the following formula: ; Calculate input gate and candidate cell status The input gate determines how much new information will be stored, while the candidate cell state is the potential new memory content generated based on the current input. The calculation uses the following formula: ; ; The current cell state is generated by forgetting a portion of the old state and adding new candidate information. This achieves stable transmission of long-term dependent information; the cell state update calculation uses the following formula: ; Calculate the output gate And based on the cell state, generate the hidden state that will be output to the outside world. The output gate controls how much of the current memory needs to be output to the next layer or the next time step. The output gate control vector and the hidden state are calculated using the following formula: ; ; in, is the weight matrix, which corresponds to the linear transformation matrices of the forget gate, input gate, candidate state, and output gate, respectively; This is the bias vector.

[0010] The intelligent scene perception method based on a temporal scene classification model described in this invention further includes: The probability distribution of the current scene is calculated using the following formula, achieved through a fully connected layer and a Softmax function: ; in, A well-defined multidimensional probability vector, where the dimension k of the vector equals the total number of scenes the system can recognize, and the j-th element in the vector... The value represents the specific probability that the system will switch to the j-th scene at the next moment.

[0011] The intelligent scene perception method based on a temporal scene classification model described in this invention, wherein when a certain future scene... Predicted probability Exceeding the preset threshold When this happens, it will trigger asynchronous preloading of the resources required for that scenario.

[0012] The intelligent scene perception method based on a temporal scene classification model described in this invention includes a video playback scene, and the strategies corresponding to the video playback scene include: Display strategy: Lock the refresh rate to match the frame rate and enable HDR color mode; Performance strategy: The scheduler favors the "big cores" in the big.LITTLE architecture, the GPU frequency is increased to the mid-range, and the media decoder uses a dedicated CPU core; Interaction strategy: Automatically hide the navigation bar and place the playback control and brightness / volume slider in the screen's touch hotspot; Power consumption strategy: Suspend unnecessary background synchronization services and reduce memory frequency in non-display areas.

[0013] The intelligent scene perception method based on a temporal scene classification model described in this invention, wherein the step of slicing multi-dimensional data by time window and fusing it into feature vectors includes: Based on multidimensional data, extract corresponding multidimensional features, and combine the multidimensional features into a unified vector by slicing according to time windows; The multidimensional features include: Time characteristics: one or more of the following: time period, duration, and time pattern; Behavioral characteristics: one or more of the following: operation sequence, frequency distribution, and transition probability; System characteristics: one or more of the following: resource usage patterns and process combinations; Contextual features: location, environmental state, or one or more of these; The unified vector includes one or more of the following: application category, network throughput, average brightness, motion status indicator, and touch event frequency.

[0014] The intelligent scene perception method based on a temporal scene classification model described in this invention, wherein the step of combining multi-dimensional features into a unified vector by slicing by time window includes: Data preprocessing: Data cleaning: handling missing and outlier values ​​to ensure data quality; Standardization / normalization: unify the data scale to avoid the impact of differences in the units of measurement between features on the model; Time window slice: Determine the window size and step size; The data is divided into multiple time windows using the sliding window method, with each window containing multidimensional data of a fixed length. Feature extraction: Calculate the statistical characteristics of each time window: mean, standard deviation, maximum value, and minimum value; Feature fusion: Dimensionality reduction: Use PCA or t-SNE methods to fuse high-dimensional features into low-dimensional vectors; Feature selection: Selecting features and combining them.

[0015] The intelligent scene perception method based on a temporal scene classification model described in this invention further includes: Record user responses to system adjustments, analyze scene recognition accuracy and system performance, adjust adaptation rules and parameters, and periodically retrain the classification model using new data.

[0016] An intelligent scene perception system based on a temporal scene classification model is used to implement the intelligent scene perception method based on the temporal scene classification model described above. The system includes a data acquisition unit, a feature engineering unit, a scene recognition unit, and a strategy matching unit. The data acquisition unit acquires multi-dimensional data from user behavior, application framework layer, sensor hub, and system status. The feature engineering unit slices multi-dimensional data according to time windows and fuses them into feature vectors. The scene recognition unit uses a two-layer hybrid model to process the feature vectors for scene recognition and output a clear scene. The first layer is a high-speed filter for recognizing clear scenes, and the second layer is a temporal scene classification model for recognizing complex scenes. The strategy matching unit matches strategies to the output specific scenarios based on a preset scenario-strategy mapping table.

[0017] The beneficial effects of this invention are as follows: By applying the methods and approaches of this application, through machine learning and behavioral analysis, the system can automatically identify user scenarios and intelligently adjust system parameters. This system can not only identify the current scenario, but also predict possible scenario changes and prepare resources in advance, thus solving the cumbersome problem of traditional systems requiring users to manually adjust settings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort: Figure 1 This is a flowchart of an intelligent scene perception method based on a temporal scene classification model, according to a preferred embodiment of the present invention. Figure 2 This is a block diagram illustrating the principle of an intelligent scene perception system based on a temporal scene classification model, according to a preferred embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, a clear and complete description will be provided below in conjunction with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0020] The preferred embodiment of the present invention is an intelligent scene perception method based on a temporal scene classification model, such as... Figure 1 As shown, it includes the following steps: S11: Acquire multi-dimensional data from user behavior, application framework layer, sensor hub, and system status; User behavior data includes one or more of the following: clicks, swipes, and dwell time; Application framework layer data includes one or more of the following: currently active application, usage frequency, foreground application package name, and window stack; Sensor central data includes: IMU determining whether the device is stationary / moving / rotating, and light sensor determining one or more of the following: ambient light or darkness; System status data includes one or more of the following: network connection type, Bluetooth peripherals, and CPU / GPU load timings.

[0021] S12: Slice multi-dimensional data according to time windows and fuse them into feature vectors; Based on multidimensional data, extract corresponding multidimensional features, and combine the multidimensional features into a unified vector by slicing according to time windows; Multidimensional features include: Time characteristics: one or more of the following: time period, duration, and time pattern; Behavioral characteristics: one or more of the following: operation sequence, frequency distribution, and transition probability; System characteristics: one or more of the following: resource usage patterns and process combinations; Contextual features: location, environmental state, or one or more of these; The unified vector includes one or more of the following: application category, network throughput, average brightness, motion status flags, and touch event frequency.

[0022] The intelligent scene perception method based on a temporal scene classification model of the present invention, wherein combining multi-dimensional features into a unified vector by slicing by time window includes: Data preprocessing: Data cleaning: handling missing and outlier values ​​to ensure data quality; Standardization / normalization: unify the data scale to avoid the impact of differences in the units of measurement between features on the model; Time window slice: Determine the window size and step size; The data is divided into multiple time windows using the sliding window method, with each window containing multidimensional data of a fixed length. Feature extraction: Calculate the statistical characteristics of each time window: mean, standard deviation, maximum value, and minimum value; Feature fusion: Dimensionality reduction: Use PCA or t-SNE methods to fuse high-dimensional features into low-dimensional vectors; Feature selection: Selecting features and combining them.

[0023] S13: A two-layer hybrid model is used to process the feature vectors for scene recognition and output a clear scene. The first layer is a high-speed filter for recognizing clear scenes, and the second layer is a temporal scene classification model for recognizing complex scenes. The time-series scene classification model is also used to predict future scenes and preload resources based on the predicted future scenes, specifically including: Let the input sequence be ,in It is a multidimensional feature vector; LSTM at each time step The fused temporal feature vectors Compared to the previous hidden state Enter them together; Forget Gate Control Vector Determines the cell state at the previous moment The proportion of information to be retained in each part of the data, and the calculation of the forget gate control vector, use the following formula: ; Calculate input gate and candidate cell status The input gate determines how much new information will be stored, while the candidate cell state is the potential new memory content generated based on the current input. The calculation uses the following formula: ; ; The current cell state is generated by forgetting a portion of the old state and adding new candidate information. This achieves stable transmission of long-term dependent information; the cell state update calculation uses the following formula: ; Calculate the output gate And based on cell state Generate the hidden state that will eventually be output. The output gate controls how much of the current memory needs to be output to the next layer or the next time step. The output gate control vector and the hidden state are calculated using the following formula: ; ; Wherein, the time identifier t is a subscript, representing the sequence number of the current time step; As the input feature vector, multidimensional feature data of the model are input at time step t, such as application status, sensor fusion features, etc. The hidden state vector is the output of the LSTM's short-term memory at time step t, which serves as the representation of the external world at this moment and is passed to the next time step. As the cell state vector, LSTM serves as the long-term memory carrier at time step t, responsible for preserving the core information after screening across time steps in the sequence; Gating and State Vectors: This is the forgetting gate vector, with elements ranging from (0, 1), which determines the previous cell state. The retention ratio of each component; The input gate vector, with elements ranging from (0, 1), determines the candidate states. The proportion of each component incorporated into the new cell state; The candidate cell state vector is a new information content to be evaluated, generated from the current input and historical states. This is the output gate vector, with elements ranging from (0, 1), which controls the cell state. Each component affects the final output The proportion of contribution; Model parameters (trainable): , where is the weight matrix, and is the linear transformation matrix corresponding to the forget gate, input gate, candidate state, and output gate, respectively; These are the bias vectors, corresponding to the bias parameters in the four calculation processes mentioned above. Numerical operations and functions: Vector concatenation combines vectors a and b end-to-end in a dimension, forming a new higher-dimensional vector. For example, if... , ,but .

[0024] The Sigmoid activation function is defined as follows: This maps the input to the (0,1) interval, producing a gating effect.

[0025] The hyperbolic tangent activation function is defined as follows: This maps the input to the (-1,1) interval for data normalization.

[0026] The Hadamard product represents the element-wise multiplication of two vectors of the same dimension.

[0027] Then, through a fully connected layer and a Softmax function, the probability distribution of the current scene is calculated using the following formula: ; in, A well-defined multidimensional probability vector, where the dimension k of the vector equals the total number of scenes the system can recognize, and the j-th element in the vector... The value represents the specific probability that the system will switch to the j-th scene at the next moment.

[0028] In a certain future scenario Predicted probability Exceeding the preset threshold When this happens, it will trigger asynchronous preloading of the resources required for that scenario; The LSTM's formula set is responsible for digesting the historical sequence and producing a summary representing "past and present" at each time step t. .

[0029] The prediction formula then utilizes As input, a forward computation is performed with the goal of predicting the future, and the output represents a probability distribution representing how the future might unfold. .

[0030] This process is cyclical: at the next time step t+1, new observation data is input into the LSTM, updating the hidden state. And based on this, new predictions are made, and so on, continuously rolling forward.

[0031] S14: Match strategies to the output specific scenarios based on the preset scenario-policy mapping table; Taking video playback as an example, the strategies for video playback include: Display strategies include locking the refresh rate to match the frame rate and enabling HDR color mode. Performance strategies include: the scheduler favors the "big cores" in the big.LITTLE architecture, the GPU frequency is increased to the mid-range, and the media decoder uses a dedicated CPU core; Interaction strategy: Automatically hide the navigation bar and place playback controls and brightness / volume sliders in the screen's touch-sensitive areas; Power consumption strategies include suspending unnecessary background synchronization services and reducing memory frequency in non-display areas.

[0032] A better approach also includes: recording user responses to system adjustments, analyzing scene recognition accuracy and system performance, adjusting adaptation rules and parameters, and periodically retraining the classification model with new data.

[0033] By applying the methods and approaches of this application, and through machine learning and behavioral analysis, the system achieves automatic identification of user scenarios and intelligent adjustment of system parameters. The system can not only identify the current scenario, but also predict possible scenario changes and prepare resources in advance, thus solving the cumbersome problem of traditional systems requiring users to manually adjust settings.

[0034] Based on actual testing, it has the following advantages: Seamless scene switching: Scene recognition and system reconfiguration are completed imperceptibly within 500ms; Improved resource utilization: Through precise scheduling, average CPU utilization is reduced by 20% in typical mixed usage scenarios, while ensuring performance. Improved interaction efficiency: The first-screen reach rate of frequently used functions in corresponding scenarios has been increased to over 95%; Extended battery life: In typical office scenarios, overall system power consumption is reduced by 15-25%.

[0035] An intelligent scene perception system based on a temporal scene classification model is applied to implement the intelligent scene perception method based on the temporal scene classification model described above. Figure 2 As shown, the system includes a data acquisition unit 200, a feature engineering unit 201, a scene recognition unit 202, and a strategy matching unit 203; The data acquisition unit 200 acquires multi-dimensional data from user behavior, application framework layer, sensor hub, and system status. Feature engineering unit 201 slices multi-dimensional data by time window and fuses them into feature vectors; The scene recognition unit 202 uses a two-layer hybrid model to process the feature vectors for scene recognition and output a clear scene. The first layer is a high-speed filter for recognizing clear scenes, and the second layer is a temporal scene classification model for recognizing complex scenes. The strategy matching unit 203 matches the output specific scenario with a strategy based on a preset scenario-strategy mapping table.

[0036] The system applying this application achieves automatic identification of user scenarios and intelligent adjustment of system parameters through machine learning and behavioral analysis. The system can not only identify the current scenario, but also predict possible scenario changes and prepare resources in advance, solving the cumbersome problem of traditional systems requiring users to manually adjust settings.

[0037] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An intelligent scene perception method based on a temporal scene classification model, characterized in that, The method includes the following steps: Acquire multi-dimensional data from user behavior, application framework layer, sensor hub, and system status; Multi-dimensional data is sliced ​​according to time windows and then fused into feature vectors. A two-layer hybrid model is used to process the feature vectors for scene recognition and output a clear scene. The first layer is a high-speed filter for recognizing clear scenes, and the second layer is a temporal scene classification model for recognizing complex scenes. The policy is matched to the output specific scenario based on the preset scenario-policy mapping table.

2. The intelligent scene perception method based on a temporal scene classification model according to claim 1, characterized in that, User behavior data includes one or more of the following: clicks, swipes, and dwell time; Application framework layer data includes one or more of the following: currently active application, usage frequency, foreground application package name, and window stack; Sensor central data includes: IMU determining whether the device is stationary / moving / rotating, and light sensor determining one or more of the following: ambient light or darkness; System status data includes one or more of the following: network connection type, Bluetooth peripherals, and CPU / GPU load timings.

3. The intelligent scene perception method based on a temporal scene classification model according to claim 2, characterized in that, The time-series scene classification model is also used to predict future scenes and preload resources based on the predicted future scenes, specifically including: Let the input sequence be ,in It is a multidimensional feature vector; LSTM at each time step The fused temporal feature vectors Compared to the previous hidden state Enter them together; Forget Gate Control Vector Determines the cell state at the previous moment The proportion of information to be retained in each part of the data, and the calculation of the forget gate control vector, use the following formula: ; Calculate input gate and candidate cell status The input gate determines how much new information will be stored, while the candidate cell state is the potential new memory content generated based on the current input. The calculation uses the following formula: ; ; The current cell state is generated by forgetting a portion of the old state and adding new candidate information. This achieves stable transmission of long-term dependent information; the cell state update calculation uses the following formula: ; Calculate the output gate And based on the cell state, generate the hidden state that will be output to the outside world. The output gate controls how much of the current memory needs to be output to the next layer or the next time step. The output gate control vector and the hidden state are calculated using the following formula: ; ; in, is the weight matrix, which corresponds to the linear transformation matrices of the forget gate, input gate, candidate state, and output gate, respectively; This is the bias vector.

4. The intelligent scene perception method based on a temporal scene classification model according to claim 3, characterized in that, The method also includes: using a fully connected layer and a Softmax function, the probability distribution of the current scene is calculated using the following formula: ; in, A well-defined multidimensional probability vector, where the dimension k of the vector equals the total number of scenes the system can recognize, and the j-th element in the vector... The value represents the specific probability that the system will switch to the j-th scene at the next moment.

5. The intelligent scene perception method based on a temporal scene classification model according to claim 3, characterized in that, In a certain future scenario Predicted probability Exceeding the preset threshold When this happens, it will trigger asynchronous preloading of the resources required for that scenario.

6. The intelligent scene perception method based on a temporal scene classification model according to claim 1, characterized in that, The scenario includes a video playback scenario, and the strategies corresponding to the video playback scenario include: Display strategy: Lock the refresh rate to match the frame rate and enable HDR color mode; Performance strategy: The scheduler favors the "big cores" in the big.LITTLE architecture, the GPU frequency is increased to the mid-range, and the media decoder uses a dedicated CPU core; Interaction strategy: Automatically hide the navigation bar and place the playback control and brightness / volume slider in the screen's touch hotspot; Power consumption strategy: Suspend unnecessary background synchronization services and reduce memory frequency in non-display areas.

7. The intelligent scene perception method based on a temporal scene classification model according to claim 1, characterized in that, The step of slicing multi-dimensional data by time window and fusing it into feature vectors includes: Based on multidimensional data, extract corresponding multidimensional features, and combine the multidimensional features into a unified vector by slicing according to time windows; The multidimensional features include: Time characteristics: one or more of the following: time period, duration, and time pattern; Behavioral characteristics: one or more of the following: operation sequence, frequency distribution, and transition probability; System characteristics: one or more of the following: resource usage patterns and process combinations; Contextual features: location, environmental state, or one or more of these; The unified vector includes one or more of the following: application category, network throughput, average brightness, motion status indicator, and touch event frequency.

8. The intelligent scene perception method based on a temporal scene classification model according to claim 7, characterized in that, The method of combining multidimensional features into a unified vector by slicing by time window includes: Data preprocessing: Data cleaning: handling missing and outlier values ​​to ensure data quality; Standardization / normalization: unify the data scale to avoid the impact of differences in the units of measurement between features on the model; Time window slice: Determine the window size and step size; The data is divided into multiple time windows using the sliding window method, with each window containing multidimensional data of a fixed length. Feature extraction: Calculate the statistical characteristics of each time window: mean, standard deviation, maximum value, and minimum value; Feature fusion: Dimensionality reduction: Use PCA or t-SNE methods to fuse high-dimensional features into low-dimensional vectors; Feature selection: Selecting features and combining them.

9. The intelligent scene perception method based on a temporal scene classification model according to claim 1, characterized in that, The method further includes: Record user responses to system adjustments, analyze scene recognition accuracy and system performance, adjust adaptation rules and parameters, and periodically retrain the classification model using new data.

10. An intelligent scene perception system based on a temporal scene classification model, applied to implement the intelligent scene perception method based on a temporal scene classification model as described in any one of claims 1-9, characterized in that, The system includes a data acquisition unit, a feature engineering unit, a scene recognition unit, and a strategy matching unit; The data acquisition unit acquires multi-dimensional data from user behavior, application framework layer, sensor hub, and system status. The feature engineering unit slices multi-dimensional data according to time windows and fuses them into feature vectors. The scene recognition unit uses a two-layer hybrid model to process the feature vectors for scene recognition and output a clear scene. The first layer is a high-speed filter for recognizing clear scenes, and the second layer is a temporal scene classification model for recognizing complex scenes. The strategy matching unit matches strategies to the output specific scenarios based on a preset scenario-strategy mapping table.

Citation Information

Patent Citations

  • Identification method and device for application scenario of mobile terminal

    CN108965609A

  • HVAC optimization control method based on scene identification and demand prediction

    CN118729501A

  • Network security analysis method and system based on privatized generative artificial intelligence

    CN120050109A

  • Machine vision-based cruise type ring spinning broken yarn detection system

    CN120314204A

  • DCS system maintenance system based on networking intelligent monitoring and method thereof

    CN120449034A