AI-driven LED intelligent dimming method and system fusing user behavior and ambient light

Abstract

The application discloses a kind of user behavior and environment light fusion AI-driven LED intelligent dimming method and system, comprising: first, the environmental light data in target application scene is acquired and its characteristic representation is extracted, to characterize illumination intensity state and change trend.Meanwhile, the behavior data unit in target user behavior data is executed behavior representation reconstruction processing, to enhance its core contribution to dimming decision.Making subsequently, based on the behavior data unit after reconstruction and environmental light characteristic representation, corresponding dimming control parameter is generated by fusion decision.Finally, the output brightness of LED light source is adaptively adjusted according to the parameter, to realize a kind of intelligent dimming control of depth fusion user behavior intention and environmental light condition, improve the comfort, individualization and energy-saving effect of lighting experience.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, and more specifically, to an AI-driven intelligent LED dimming method and system that integrates user behavior with ambient light. Background Technology

[0002] Existing LED intelligent dimming technologies primarily rely on ambient light sensors or preset scene modes for control, lacking a deep understanding of users' real-time behavioral intentions. Ambient light control only responds to changes in external lighting and cannot adapt to the user's current activity; while dimming based on simple behavior triggers or fixed patterns struggles to handle complex, continuous, and dynamically changing sequences of user behavior. This results in rigid and insufficiently personalized adjustments in existing systems, failing to achieve precise and dynamic integration of lighting with the user's actual needs and environmental conditions, thus impacting comfort and energy efficiency. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-driven intelligent LED dimming method and system that integrates user behavior with ambient light.

[0004] In a first aspect, embodiments of the present invention provide an AI-driven intelligent LED dimming method that integrates user behavior with ambient light, comprising:

[0005] Acquire ambient light data in the target application scenario, wherein the ambient light data includes at least one ambient light acquisition parameter;

[0006] The ambient light data is subjected to ambient light feature extraction processing to obtain an ambient light feature representation corresponding to the ambient light data; the ambient light feature representation is used to characterize the light intensity state and light change trend in the target application scene.

[0007] Perform behavioral representation reconstruction processing on the behavioral data units in the target user behavior data to obtain the reconstructed behavioral data units;

[0008] Based on the reconstructed behavioral data unit and the ambient light feature representation, corresponding dimming control parameters are generated;

[0009] The output brightness of the LED light source is adaptively adjusted according to the dimming control parameters.

[0010] In one possible implementation, the step of performing behavior representation reconstruction processing on behavior data units in the target user behavior data to obtain reconstructed behavior data units includes:

[0011] The behavioral data units in the target user behavior data are encoded to obtain at least two behavioral feature representations corresponding to the behavioral data units; the at least two behavioral feature representations are used to reflect feature data of different categories; the behavioral data units are obtained by dividing the target user behavior data into behavioral sequences;

[0012] Perform a feature merging operation on at least two of the behavioral feature representations to obtain a merged feature representation corresponding to the behavioral data unit. Perform feature rearrangement and aggregation on the merged feature representation based on the correlation matrix to obtain a correlation-aggregated feature representation of the merged feature representation used to enhance the dimming decision contribution.

[0013] The behavioral association feature representation corresponding to the target user behavior data and the generated feature representation corresponding to the behavior data unit are subjected to hierarchical weighted synthesis processing to obtain the comprehensive behavioral representation corresponding to the behavior data unit; the generated feature representation is determined by the relevance aggregation feature representation and the merged feature representation; the behavioral association feature representation is used to characterize the association strength between the behavior data units;

[0014] A linear mapping calculation is performed on the comprehensive behavioral representation to obtain the behavioral adjustment matrix corresponding to the behavioral data unit;

[0015] The behavior adjustment matrix is ​​standardized to obtain the standardized adjustment matrix corresponding to the behavior adjustment matrix, and the behavior data unit is reconstructed based on the standardized adjustment matrix to obtain the reconstructed behavior data unit.

[0016] In one possible implementation, at least two of the behavioral feature representations include behavioral identification feature representations;

[0017] The encoding process of behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units includes:

[0018] Load the target user behavior data into the feature encoder in the behavior representation reconstruction model;

[0019] In the feature encoder, at least one behavior identifier unit of the behavior data unit in the target user behavior data is obtained; at least one behavior identifier unit is used to characterize the behavior expression attribute of the behavior data unit;

[0020] Obtain the behavior identifier unit matrix corresponding to at least one of the behavior identifier units, and perform local association calculation on the at least one behavior identifier unit matrix to obtain the local behavior association matrix corresponding to at least one behavior identifier unit matrix.

[0021] At least one of the local behavior association matrices is aggregated and dimensionality reduced to obtain the behavior identifier feature representation corresponding to the behavior data unit.

[0022] In one possible implementation, at least two of the behavioral feature representations include behavioral appearance feature representations;

[0023] The encoding process of behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units includes:

[0024] Load the target user behavior data into the feature encoder in the behavior representation reconstruction model;

[0025] In the feature encoder, at least one behavior visual mapping data of the behavior data unit in the target user behavior data is obtained; at least one of the behavior visual mapping data is used to reflect the behavior data unit with different behavior presentation modes;

[0026] Feature representation generation processing is performed on at least one of the behavioral visual mapping data to obtain a performance pattern matrix corresponding to each of the at least one behavioral visual mapping data.

[0027] At least one of the performance pattern matrices is aggregated and dimensionality reduced to obtain the behavioral appearance feature representation corresponding to the behavioral data unit.

[0028] In one possible implementation, at least two of the behavioral feature representations include behavioral type feature representations;

[0029] The encoding process of behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units includes:

[0030] Load the target user behavior data into the feature encoder in the behavior representation reconstruction model;

[0031] In the feature encoder, the target user behavior data is subjected to behavior sequence deconstruction processing to obtain at least one behavior semantic segment in the target user behavior data; at least one behavior semantic segment is used to constitute the target user behavior data, and at least one behavior semantic segment is a behavior semantic segment of different behavior types in the target user behavior data;

[0032] Obtain the behavior type matrix corresponding to the target behavior semantic segment from the behavior type matrices corresponding to at least one of the behavior semantic segments respectively, and use the behavior type matrix corresponding to the target behavior semantic segment as the behavior type feature representation of the behavior data unit in the target user behavior data; the target behavior semantic segment is the behavior semantic segment corresponding to the behavior data unit in at least one of the behavior semantic segments.

[0033] In one possible implementation, at least two of the behavioral feature representations include behavioral granularity feature representations;

[0034] The encoding process of behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units includes:

[0035] Load the target user behavior data into the feature encoder in the behavior representation reconstruction model;

[0036] In the feature encoder, behavioral representation mapping is performed on the behavioral data units in the target user behavior data to obtain the behavioral sub-unit vector corresponding to the behavioral data unit;

[0037] Determine the behavior time-series index of the behavior data unit in the target user behavior data, perform time-series weight mapping on the behavior time-series index of the behavior data unit, and obtain the time-series index vector corresponding to the behavior data unit.

[0038] Obtain the behavior boundary identifier vector corresponding to the behavior data unit, and perform joint feature encoding on the behavior sub-unit vector, the temporal index vector, and the behavior boundary identifier vector corresponding to the behavior data unit to obtain the behavior granularity feature representation corresponding to the behavior data unit.

[0039] In one possible implementation, the step of performing a feature merging operation on at least two of the behavioral feature representations to obtain a merged feature representation corresponding to the behavioral data unit, and performing feature rearrangement and aggregation based on the correlation matrix on the merged feature representation to obtain a correlation-aggregated feature representation of the merged feature representation used to enhance the dimming decision contribution, includes:

[0040] Load at least two of the behavioral feature representations into the feature joint modeling unit in the behavioral representation reconstruction model;

[0041] In the joint feature modeling unit, a feature merging operation is performed on at least two of the behavioral feature representations to obtain the merged feature representation corresponding to the behavioral data unit;

[0042] A linear mapping calculation is performed on the merged feature representation to obtain at least one correlation calculation matrix corresponding to the merged feature representation; at least one correlation calculation matrix includes a correlation reference vector, a feature contribution vector, and a correlation matching vector.

[0043] Obtain the dimension alignment vector corresponding to the association reference vector, and use the product of the association matching vector and the dimension alignment vector as the association degree evaluation vector; the association degree evaluation vector is used to characterize the correlation strength between at least two behavioral feature representations.

[0044] The correlation evaluation vector is standardized to obtain the standardized vector corresponding to the correlation evaluation vector. The product of the standardized vector and the feature contribution vector is used as the weighted aggregation vector representing the correlation of at least two behavioral features.

[0045] A linear mapping calculation is performed on the associated aggregated vector to obtain the correlation aggregated feature representation of the merged feature representation used to enhance the contribution of dimming decision.

[0046] In one possible implementation, the hierarchical weighted synthesis process of the behavior association feature representation corresponding to the target user behavior data and the generated feature representation corresponding to the behavior data unit to obtain the comprehensive behavior representation corresponding to the behavior data unit includes:

[0047] The behavioral association feature representation corresponding to the target user behavior data and the generated feature representation corresponding to the behavior data unit are loaded into the hierarchical weighted synthesis processing network in the behavior representation reconstruction model; the hierarchical weighted synthesis processing network includes a feature mapping unit, at least two hierarchical weighted synthesis processing units, and a feature synthesis processing unit; the at least two hierarchical weighted synthesis processing units include a target hierarchical weighted synthesis processing unit, and the target hierarchical weighted synthesis processing unit is any one of the at least one hierarchical weighted synthesis processing units;

[0048] In the feature mapping unit, the generated feature representation corresponding to the behavior data unit is subjected to feature mapping processing to obtain the mapped feature representation corresponding to the behavior data unit;

[0049] If the target hierarchical weighted synthesis processing unit is the first hierarchical weighted synthesis processing unit among at least two hierarchical weighted synthesis processing units, then the target hierarchical weighted synthesis processing unit performs behavior association analysis on the behavior association feature representation corresponding to the target user behavior data and the mapping feature representation corresponding to the behavior data unit to obtain the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit.

[0050] If the target hierarchical weighted synthesis processing unit is not the first hierarchical weighted synthesis processing unit among at least two hierarchical weighted synthesis processing units, then the target hierarchical weighted synthesis processing unit performs behavior association analysis on the behavior association feature representation corresponding to the target user behavior data and the hierarchical feature representation corresponding to the preceding hierarchical weighted synthesis processing unit to obtain the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit; the preceding hierarchical weighted synthesis processing unit is the hierarchical weighted synthesis processing unit preceding the target hierarchical weighted synthesis processing unit.

[0051] The hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit in at least two of the hierarchical weighted synthesis processing units is used as the hierarchical weighted synthesis processing vector corresponding to the behavior data unit; the target hierarchical weighted synthesis processing unit is the last hierarchical weighted synthesis processing unit in at least two of the hierarchical weighted synthesis processing units.

[0052] In the feature synthesis processing unit, the hierarchical weighted synthesis processing vector corresponding to the behavior data unit and the mapping feature representation corresponding to the behavior data unit are subjected to weight-based result integration processing to obtain the comprehensive behavior representation corresponding to the behavior data unit.

[0053] In one possible implementation, the number of behavioral data units is at least two, and the at least two behavioral data units include target behavioral data units; the behavioral association feature representation corresponding to the target user behavioral data includes behavioral influence weight coefficients of the target behavioral data units for the at least two behavioral data units respectively; the mapping feature representation corresponding to the behavioral data units includes the mapping feature representation corresponding to the target behavioral data units;

[0054] The step of performing behavior association analysis on the behavior association feature representation corresponding to the target user behavior data and the mapping feature representation corresponding to the behavior data unit through the target hierarchical weighted synthesis processing unit to obtain the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit includes:

[0055] The target hierarchical weighted synthesis processing unit obtains the behavior influence weight matrix of the target behavior data unit for at least two of the behavior data units respectively;

[0056] In the target hierarchical weighted synthesis processing unit, feature merging operations are performed on the mapping feature representation corresponding to the target behavior data unit and at least two behavior influence weight matrices respectively to obtain the merged behavior influence weight matrices corresponding to at least two behavior data units respectively.

[0057] Based on at least two of the said behavior influence weight coefficients, a weighted cumulative calculation is performed on at least two of the said combined behavior influence weight matrices to generate a behavior influence relationship representation corresponding to the target behavior data unit;

[0058] The behavior influence relationship representation corresponding to the target behavior data unit is subjected to response enhancement processing to obtain the hierarchical sub-feature representation corresponding to the target behavior data unit; the hierarchical sub-feature representation corresponding to the target behavior data unit belongs to the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit.

[0059] In a second aspect, embodiments of the present invention provide a server system, including a server, the server being used to execute the method described in the first aspect.

[0060] Compared to existing technologies, the beneficial effects provided by this invention include: Employing an AI-driven LED intelligent dimming method and system that integrates user behavior and ambient light, as disclosed in this invention, the system acquires ambient light data from a target application scenario and extracts its feature representations to characterize the light intensity state and its changing trends. Simultaneously, behavioral representation reconstruction processing is performed on the behavioral data units in the target user behavior data to enhance their core contribution to dimming decisions. Subsequently, based on the reconstructed behavioral data units and ambient light feature representations, corresponding dimming control parameters are generated through fusion decision-making. Finally, the output brightness of the LED light source is adaptively adjusted according to these parameters, thereby achieving intelligent dimming control that deeply integrates user behavioral intentions and ambient light conditions, improving the comfort, personalization, and energy-saving effects of the lighting experience. Attached Figure Description

[0061] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 A flowchart illustrating the steps of the AI-driven LED intelligent dimming method that integrates user behavior and ambient light, as provided in an embodiment of the present invention.

[0063] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0065] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0066] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the AI-driven intelligent LED dimming method that integrates user behavior and ambient light according to an embodiment of this disclosure. The following is a detailed description of the AI-driven intelligent LED dimming method that integrates user behavior and ambient light.

[0067] Step S201: Obtain ambient light data in the target application scenario, wherein the ambient light data includes at least one ambient light acquisition parameter;

[0068] Step S202: Perform ambient light feature extraction processing on the ambient light data to obtain the ambient light feature representation corresponding to the ambient light data; the ambient light feature representation is used to characterize the light intensity state and light change trend in the target application scene;

[0069] Step S203: Perform behavior representation reconstruction processing on the behavior data units in the target user behavior data to obtain the reconstructed behavior data units;

[0070] Step S204: Based on the reconstructed behavior data unit and the ambient light feature representation, generate corresponding dimming control parameters;

[0071] Step S205: Adaptively adjust the output brightness of the LED light source according to the dimming control parameters.

[0072] In this embodiment of the invention, the following example uses a home study as the target application scenario, with the server as the execution entity, to illustrate the scenario-based implementation of the AI-driven LED intelligent dimming method that integrates user behavior and ambient light described in this invention:

[0073] After the server initiates the dimming process for the home study, it first performs an ambient light data acquisition step: using two ambient light sensors deployed above the desk and on the wall near the window, it retrieves ambient light data at a collection cycle of 100 milliseconds. The collected parameters cover four core indicators: light intensity, color temperature, light uniformity, and rate of change. Taking a weekday morning as an example, the server collects data showing that the current light intensity in the desk area is approximately 800 lx and the color temperature is 6000 K. The light intensity in the window area is significantly higher, and due to the gradually increasing sunlight, the overall light intensity is steadily rising at a rate of approximately 500 lx per minute. During the data collection process, the server detects a momentary abnormal light value from the window sensor due to a brief obstruction by the edge of the curtain. It then uses a time series interpolation algorithm to complete the data, ensuring its integrity.

[0074] After acquiring ambient light data, the server performs ambient light feature extraction processing: First, parameters such as light intensity and color temperature are standardized and mapped to the [0,1] range; then, a pre-trained extraction model that integrates static and dynamic features is called to identify that the current study is in a static lighting state of "strong light by the window, medium brightness in the desk area, and average overall lighting uniformity". At the same time, the dynamic trend of "light intensity continuously increasing and approaching 1600 lx in the next 10 minutes" is captured. Finally, an ambient light feature representation covering both static state and dynamic trend is generated to provide a basic reference for subsequent dimming decisions.

[0075] Next, the server performs behavioral representation reconstruction processing on the target user's behavioral data. The user behavioral data comes from the interconnected data of the smart camera in the study, the ergonomic chair sensor, and the desktop electronic devices. The server divides continuous user behavior into three behavioral data units based on behavioral events: "continuous reading of professional books," "writing notes at the desk," and "getting up to retrieve a book." Taking the "continuous reading of professional books" unit as an example, the server initiates the representation reconstruction process: It extracts core features from multiple dimensions using a pre-trained behavioral feature encoder; it identifies from the camera data that the user's head posture is stable and their gaze is highly focused, indicating highly focused static reading behavior; it captures from the chair sensor data that the user's sitting posture is stable and there are no frequent getting-up movements; and it combines this with the electronic device data to determine that the user is in a state of prolonged focused reading. After completing the multi-dimensional feature extraction, the server merges various features and analyzes the correlation between each feature and the dimming requirements, prioritizing the retention of core information such as "focused reading requires a stable color temperature" and "static behavior is sensitive to sudden changes in light." At the same time, combining the correlation between various behavioral units, reading and writing are both focused behaviors, and their lighting requirements are highly correlated. Getting up to pick up a book is a movement behavior, and the lighting requirements are significantly different. After generating a comprehensive behavioral representation, the reconstruction is completed, highlighting the behavioral features that actually contribute to dimming decisions.

[0076] After completing the extraction of ambient light features and the reconstruction of behavioral data units, the server generates dimming control parameters based on both. It calls a pre-trained fusion decision model to cross-analyze the ambient light features and the reconstructed behavioral features, clarifying that the core requirement is "stable lighting to ensure users can concentrate on reading, while offsetting the impact of continuously increasing ambient light." The generated dimming parameters include: adjusting the brightness of the main desk lamp from its current high level to a moderate level suitable for reading, and adjusting the color temperature from a cool 6000K to a more eye-friendly warm white light of 4000K; turning on the auxiliary window light to supplement illumination and improve the uniformity of lighting in the desk area; setting dynamic adjustment rules to automatically accelerate the dimming rate when the rate of change in ambient light exceeds a preset threshold, preventing sudden changes in lighting from interfering with user concentration; and temporarily enhancing corner lighting when the user triggers the "get up to retrieve a book" action, restoring the original lighting state after the user returns to their seat. The server simultaneously verifies whether the parameters are within the hardware capacity of the LED light group to ensure the executability of the instructions.

[0077] Finally, the server converts the dimming control parameters into control commands supported by the LED light group through the home IoT gateway and sends them out. It also receives real-time feedback on the actual operating status of the light group and makes fine adjustments if there are minor deviations to ensure that the lighting effect meets expectations. At the same time, the server records the core process data of this dimming, including ambient light status, user behavior type, dimming parameters, and subsequent user behavior feedback (such as the user not triggering manual dimming after dimming, indicating good lighting adaptability). This type of data is periodically added to the model training set to fine-tune the feature extraction and decision model, continuously optimizing the accuracy and user adaptability of subsequent dimming.

[0078] In this embodiment of the invention, the behavior representation reconstruction process is performed on the behavior data units in the target user behavior data to obtain the reconstructed behavior data units, which can be implemented through the following examples.

[0079] The behavioral data units in the target user behavior data are encoded to obtain at least two behavioral feature representations corresponding to the behavioral data units; the at least two behavioral feature representations are used to reflect feature data of different categories; the behavioral data units are obtained by dividing the target user behavior data into behavioral sequences;

[0080] Perform a feature merging operation on at least two of the behavioral feature representations to obtain a merged feature representation corresponding to the behavioral data unit. Perform feature rearrangement and aggregation on the merged feature representation based on the correlation matrix to obtain a correlation-aggregated feature representation of the merged feature representation used to enhance the dimming decision contribution.

[0081] The behavioral association feature representation corresponding to the target user behavior data and the generated feature representation corresponding to the behavior data unit are subjected to hierarchical weighted synthesis processing to obtain the comprehensive behavioral representation corresponding to the behavior data unit; the generated feature representation is determined by the relevance aggregation feature representation and the merged feature representation; the behavioral association feature representation is used to characterize the association strength between the behavior data units;

[0082] A linear mapping calculation is performed on the comprehensive behavioral representation to obtain the behavioral adjustment matrix corresponding to the behavioral data unit;

[0083] The behavior adjustment matrix is ​​standardized to obtain the standardized adjustment matrix corresponding to the behavior adjustment matrix, and the behavior data unit is reconstructed based on the standardized adjustment matrix to obtain the reconstructed behavior data unit.

[0084] In an embodiment of the present invention, for example, the server initiates a behavior representation reconstruction process for the predefined "continuous reading of professional books" behavior data units in the study room scenario:

[0085] First, encoding processing is performed: behavioral identifier features are extracted from the linked data collected by the smart camera and seat pressure sensor, focusing on core parameters such as user head posture and gaze duration, and encoded into feature vectors reflecting reading focus; at the same time, behavioral type features are extracted, and through behavioral semantic decomposition, it is determined that the unit belongs to "static focused behavior at a desk", and encoded into feature vectors of the corresponding behavioral type.

[0086] Next, feature merging and aggregation are performed: the two types of behavioral feature representations are merged into a unified merged feature representation, and then the correlation matrix is ​​calculated. It is found that the correlation between behavioral identification features and dimming requirements is 0.8 (the higher the focus, the stronger the requirement for light stability), and the correlation between behavioral type features is 0.7. Based on this matrix, the merged features are rearranged, and high-correlation features are retained first, resulting in a correlation-aggregated feature representation that enhances the contribution of dimming decision. The aggregated features and merged features are weighted and fused to generate the generated feature representation of the behavioral data unit.

[0087] Subsequently, a hierarchical weighted synthesis was performed: First, behavioral association feature representations were obtained through a behavioral association analysis model. The association strength between this reading unit and the subsequent "writing notes" unit was 0.95 (both belonging to the behavior of focused attention at a desk), and the association strength with the "getting up to retrieve a book" unit was 0.3. The behavioral association feature representations and generated feature representations were input into a hierarchical weighted network: The first layer performed association analysis on the two types of features, generating hierarchical features that highlight behavioral associations; the second layer, based on the results of the first layer, strengthened the "need for stable lighting for focused behavior," obtaining the final hierarchical feature; this feature was then weighted and integrated with the generated feature to obtain a comprehensive behavioral representation.

[0088] Then, linear mapping and standardization are performed: linear mapping is performed on the comprehensive behavioral representation to obtain the behavioral adjustment matrix used to adjust the weights of behavioral features, clarifying that the "focus" feature needs to be strengthened and the noise feature of occasional head turning needs to be weakened; L2 standardization is performed on the adjustment matrix to obtain the standardized adjustment matrix, and finally the original behavioral data unit is reconstructed based on the matrix, filtering out irrelevant noise and highlighting the behavioral features that are effective for the dimming decision core, resulting in the reconstructed "continuous reading" behavioral data unit.

[0089] In embodiments of the present invention, at least two of the behavioral feature representations include behavioral identification feature representations;

[0090] The process of encoding behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units can be implemented through the following example.

[0091] Load the target user behavior data into the feature encoder in the behavior representation reconstruction model;

[0092] In the feature encoder, at least one behavior identifier unit of the behavior data unit in the target user behavior data is obtained; at least one behavior identifier unit is used to characterize the behavior expression attribute of the behavior data unit;

[0093] Obtain the behavior identifier unit matrix corresponding to at least one of the behavior identifier units, and perform local association calculation on the at least one behavior identifier unit matrix to obtain the local behavior association matrix corresponding to at least one behavior identifier unit matrix.

[0094] At least one of the local behavior association matrices is aggregated and dimensionality reduced to obtain the behavior identifier feature representation corresponding to the behavior data unit.

[0095] In an embodiment of the present invention, for example, the server initiates an encoding process for behavioral identifier features representing the behavioral data unit of "continuous reading of professional books" in a study scenario:

[0096] First, the server loads the coordinated behavioral data collected by the smart camera, ergonomic chair pressure sensor, and desktop e-reader into the feature encoder of the pre-trained behavior representation reconstruction model. This encoder is a lightweight Transformer architecture adapted for indoor behavior analysis optimization.

[0097] Within the feature encoder, the server extracts three core behavioral identifiers from the behavioral data units: first, the head posture unit, which represents the attention-related attributes such as the angle between the user's head and the reader screen and the percentage of time spent looking down; second, the gaze focus unit, which represents the reading engagement attributes such as the duration of time the user's gaze stays on the e-book page and the frequency of page jumps; and third, the blink frequency unit, which represents the visual fatigue-related attributes such as the number of times the user blinks per minute and the duration of continuous eye opening.

[0098] Subsequently, the server obtains the behavior identifier unit matrix corresponding to each behavior identifier unit: the head posture unit matrix contains 12 dimensions of posture parameter values, the gaze focus unit matrix contains 8 dimensions of gaze tracking data, and the blink frequency unit matrix contains 6 dimensions of eye physiological parameters. For each matrix, the server performs local correlation calculations: taking the head posture unit matrix as an example, it calculates the proportion of the head-to-screen angle in the 30°-40° range and its positive correlation coefficient with reading focus, generating a local behavior correlation matrix highlighting the correlation of parameters; similarly, the local behavior correlation matrices for gaze focus and blink frequency are obtained respectively.

[0099] Finally, the server performs PCA aggregation and dimensionality reduction on the three local behavior association matrices, compressing the high-dimensional association parameters into a 32-dimensional low-dimensional space and removing redundant weak association parameters to obtain the behavior identifier feature representation corresponding to the "continuous reading of professional books" behavior data unit. This feature can directly reflect the user's current reading state of high concentration and low visual fatigue, providing a core basis for the need to "stabilize the light color temperature and avoid sudden brightness changes" in subsequent dimming decisions.

[0100] In embodiments of the present invention, at least two of the behavioral feature representations include behavioral appearance feature representations;

[0101] The process of encoding behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units can be implemented through the following example.

[0102] Load the target user behavior data into the feature encoder in the behavior representation reconstruction model;

[0103] In the feature encoder, at least one behavior visual mapping data of the behavior data unit in the target user behavior data is obtained; at least one of the behavior visual mapping data is used to reflect the behavior data unit with different behavior presentation modes;

[0104] Feature representation generation processing is performed on at least one of the behavioral visual mapping data to obtain a performance pattern matrix corresponding to each of the at least one behavioral visual mapping data.

[0105] At least one of the performance pattern matrices is aggregated and dimensionality reduced to obtain the behavioral appearance feature representation corresponding to the behavioral data unit.

[0106] In an embodiment of the present invention, for example, the server initiates an encoding process for representing behavioral appearance features for the behavioral data unit of "continuous reading of professional books" in a study scenario:

[0107] First, the server loads 10 minutes of continuous RGB video frames and depth image data collected by the smart camera, as well as pressure distribution data uploaded in real time by the ergonomic chair, into the feature encoder of the pre-trained behavior representation reconstruction model. This encoder has a built-in visual feature extraction branch optimized for indoor human behavior.

[0108] Within the feature encoder, the server extracts two types of core behavioral visual mapping data from the behavioral data unit: one is posture mapping data, which reflects the static presentation mode of the torso, shoulders and neck when the user is reading; the other is limb movement mapping data, which reflects the dynamic presentation mode of the arms and hands when the user is reading. The two types of data correspond to two different behavioral presentation states: "static sitting" and "micro-movement adjustment".

[0109] Subsequently, the server performs feature representation generation processing on the two types of behavioral visual mapping data: for the sitting posture mapping data, it extracts parameters of 8 dimensions such as shoulder line level, torso curvature, and hip pressure distribution ratio from the depth image, and generates an 8×8 sitting posture performance pattern matrix after quantifying the temporal mean and fluctuation amplitude of each parameter; for the limb movement mapping data, it statistically analyzes parameters of 6 dimensions such as the proportion of arm stillness time and hand contact frequency with the reader from RGB video frames, and generates a 6×6 movement performance pattern matrix. The two matrices quantify the feature details of different presentation modes respectively.

[0110] Finally, the server performs average pooling and PCA aggregation dimensionality reduction on the two types of performance mode matrices of sitting posture and action, compressing the high-dimensional matrix into a 32-dimensional low-dimensional feature space. It removes weakly correlated parameters such as "raising a hand to fix hair" which are unrelated to the core behavior, and obtains the behavioral appearance feature representation corresponding to the behavioral data unit of "continuous reading of professional books". This feature can directly reflect that the user is currently in a highly stable reading state, providing a key basis for the need to "maintain illumination uniformity and avoid color temperature changes" in subsequent dimming decisions.

[0111] In embodiments of the present invention, at least two of the behavioral feature representations include behavioral type feature representations;

[0112] The process of encoding behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units can be implemented through the following example.

[0113] Load the target user behavior data into the feature encoder in the behavior representation reconstruction model;

[0114] In the feature encoder, the target user behavior data is subjected to behavior sequence deconstruction processing to obtain at least one behavior semantic segment in the target user behavior data; at least one behavior semantic segment is used to constitute the target user behavior data, and at least one behavior semantic segment is a behavior semantic segment of different behavior types in the target user behavior data;

[0115] Obtain the behavior type matrix corresponding to the target behavior semantic segment from the behavior type matrices corresponding to at least one of the behavior semantic segments respectively, and use the behavior type matrix corresponding to the target behavior semantic segment as the behavior type feature representation of the behavior data unit in the target user behavior data; the target behavior semantic segment is the behavior semantic segment corresponding to the behavior data unit in at least one of the behavior semantic segments.

[0116] In an embodiment of the present invention, for example, the server initiates an encoding process for behavior type feature representation for behavior data units labeled as "continuous reading of professional books" in a study scenario:

[0117] First, the server loads the smart camera frame data, seat pressure time series data, and e-book operation log associated with the behavior data unit into the feature encoder of the behavior representation reconstruction model. The encoder has a built-in behavior sequence deconstruction module based on a pre-trained semantic segmentation network, which has completed semantic adaptation training for thousands of indoor behavior types.

[0118] Within the feature encoder, the server performs behavioral sequence deconstruction processing on target user behavior data covering 15 minutes: by identifying the boundary nodes of behavioral actions through a semantic segmentation network, the overall behavior is divided into three semantic segments of different behavioral types. The first is the "static reading" segment, accounting for 85%, which corresponds to the core behavior of the user maintaining a stable sitting posture and continuously focusing their gaze on the e-book; the second is the "hand touch operation" segment, accounting for 12%, which corresponds to the auxiliary behavior of the user turning pages and marking electronic notes; and the third is the "head turning" segment, accounting for 3%, which corresponds to the secondary behavior of the user briefly raising their head to stretch. The three segments are spliced ​​together in chronological order to form a complete user behavior sequence.

[0119] Subsequently, the server calls a predefined indoor behavior type matrix library to extract the behavior type matrix corresponding to each semantic fragment. The "static reading" fragment corresponds to a 16-dimensional binary encoding matrix, in which dimensions strongly related to dimming decisions, such as "attention level," "light stability requirement," and "motion amplitude level," are marked with the highest weight. Finally, the server uses the behavior type matrix of the "static reading" semantic fragment, which directly corresponds to the "continuous reading of professional books" behavior data unit, as the behavior type feature representation of that unit. This feature clearly marks the current behavior as a static reading behavior with high attention and high light stability requirements, providing a clear behavior category reference for subsequent dimming decisions.

[0120] In embodiments of the present invention, at least two of the behavioral feature representations include behavioral granularity feature representations;

[0121] The process of encoding behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units can be implemented through the following example.

[0122] Load the target user behavior data into the feature encoder in the behavior representation reconstruction model;

[0123] In the feature encoder, behavioral representation mapping is performed on the behavioral data units in the target user behavior data to obtain the behavioral sub-unit vector corresponding to the behavioral data unit;

[0124] Determine the behavior time-series index of the behavior data unit in the target user behavior data, perform time-series weight mapping on the behavior time-series index of the behavior data unit, and obtain the time-series index vector corresponding to the behavior data unit.

[0125] Obtain the behavior boundary identifier vector corresponding to the behavior data unit, and perform joint feature encoding on the behavior sub-unit vector, the temporal index vector, and the behavior boundary identifier vector corresponding to the behavior data unit to obtain the behavior granularity feature representation corresponding to the behavior data unit.

[0126] In an embodiment of the present invention, for example, the server initiates an encoding process for behavioral granular feature representation for a 15-minute behavioral data unit labeled as "continuous reading of professional books" in a study scenario:

[0127] First, the server loads the smart camera frame sequence, ergonomic chair pressure time series data, and e-book operation log associated with the behavior data unit into the feature encoder of the behavior representation reconstruction model. The encoder has a built-in granular splitting and mapping module optimized for continuous indoor behavior and has completed the adaptation training of nearly 10,000 pieces of family behavior data.

[0128] Within the feature encoder, the server performs behavioral representation mapping on the behavioral data unit: the 15-minute continuous reading behavior is divided into 15 behavioral sub-units with a 1-minute time granularity. Each sub-unit extracts 8 core parameters, such as sitting posture, eye focus rate, and hand operation frequency. These parameters are mapped into 32-dimensional feature vectors through a pre-trained multilayer perceptron. The 15 vectors are then concatenated in chronological order to obtain the 480-dimensional behavioral sub-unit vector corresponding to the behavioral data unit, accurately recording the details of the reading state for each minute.

[0129] Subsequently, the server identifies the reading behavior data unit as the third continuous behavior unit in the overall user behavior sequence of the day (the preceding behaviors are 2 minutes of morning stretching and 1 minute of drinking water). The time sequence index "3" is input into the time sequence weight mapping module. Combining the transience of the preceding behavior with the continuity of the current behavior, a 16-dimensional time sequence index vector reflecting the weight of the core behavior is generated. The weight value of the "core behavior period" dimension is 0.92, highlighting the dominant position of this reading behavior in the daily behavior sequence.

[0130] Meanwhile, the server extracts the boundary attributes of the behavioral data unit: the start time is 8:10 am on weekdays (peak working hours), the duration is 15 minutes (medium to long duration behavior), and the end state is an uninterrupted natural transition. These attributes are encoded into a 16-dimensional behavioral boundary identifier vector to clearly mark the time boundary characteristics of the behavior.

[0131] Finally, the server inputs a 480-dimensional behavior sub-unit vector, a 16-dimensional temporal index vector, and a 16-dimensional behavior boundary marker vector into the joint feature encoding layer. Through linear transformation and attention-weighted fusion, weakly correlated parameters such as brief head tilts and minor posture adjustments are removed, outputting a 64-dimensional behavior-granular feature representation. This feature clearly reflects the duration distribution, temporal importance, and boundary attributes of reading behavior, providing a precise basis for setting a 10-minute long-cycle illumination stability threshold and avoiding frequent dimming interference in subsequent dimming decisions.

[0132] In this embodiment of the invention, the step of performing a feature merging operation on at least two behavioral feature representations to obtain a merged feature representation corresponding to the behavioral data unit, and performing feature rearrangement and aggregation on the merged feature representation based on the correlation matrix to obtain a correlation aggregated feature representation of the merged feature representation used to enhance the contribution of dimming decision, can be implemented through the following example.

[0133] Load at least two of the behavioral feature representations into the feature joint modeling unit in the behavioral representation reconstruction model;

[0134] In the joint feature modeling unit, a feature merging operation is performed on at least two of the behavioral feature representations to obtain the merged feature representation corresponding to the behavioral data unit;

[0135] Perform linear mapping calculation on the merged feature representation to obtain at least one correlation degree calculation matrix corresponding to the merged feature representation;

[0136] Perform correlation degree weighted aggregation processing on at least one of the correlation degree calculation matrices to obtain the correlation degree aggregated feature representation of the merged feature representation used to enhance the contribution of dimming decision.

[0137] In an embodiment of the present invention, for example, the server initiates a joint feature modeling process for the behavioral data unit of "continuous reading of professional books" in a study scenario:

[0138] First, the server synchronously loads the previously encoded 32-dimensional behavior identifier feature representation, 16-dimensional behavior type feature representation, and 64-dimensional behavior granularity feature representation into the feature joint modeling unit built into the behavior representation reconstruction model. This unit has a built-in correlation calculation module optimized for indoor behavior dimming requirements.

[0139] In the joint feature modeling unit, the server performs a feature merging operation: the three types of features are directly concatenated in dimensional order to obtain a merged feature representation of 32+16+64=112 dimensions. This feature includes core attributes such as user reading focus and behavior category, as well as detailed information such as behavior duration distribution and time series weight, thus fully preserving the original information of multi-dimensional behavioral features.

[0140] Subsequently, the server performs linear mapping calculations on the merged feature representations: it calls the pre-trained linear transformation layer within the unit to map the 112-dimensional merged features into three 32×32 correlation calculation matrices. The first matrix corresponds to the correlation between behavioral identification features and core dimming requirements (brightness stability, color temperature comfort), with values ​​ranging from 0 to 1, and a maximum value of 0.91 (correlation between gaze focus and color temperature comfort). The second matrix corresponds to the correlation between behavioral type features and dimming rule matching degree, with a maximum value of 0.87 (correlation between static sitting type and brightness stability). The third matrix corresponds to the correlation between behavioral granularity features and dimming cycle settings, with a maximum value of 0.76 (correlation between long-cycle behavior and dimming gradual change cycle).

[0141] Finally, the server performs a weighted aggregation of the three correlation calculation matrices: it loads pre-trained weight coefficients (0.85 for behavior identifier correlation matrix, 0.78 for behavior type correlation matrix, and 0.62 for behavior granularity correlation matrix), and performs a weighted summation of the matrix elements to obtain a comprehensive correlation matrix; then, it extracts high-correlation feature dimensions with values ​​greater than 0.7 from the matrix through a multi-head attention mechanism, filtering out weak correlation information such as "occasional head turning" that is irrelevant to dimming decisions; finally, it performs aggregation and dimensionality reduction processing on the high-correlation dimensions to obtain a 64-dimensional correlation aggregation feature representation, which specifically strengthens key features that have a core contribution to dimming decisions, such as "high focus requires stable color temperature" and "long-term reading requires gradual dimming".

[0142] In an embodiment of the present invention, at least one of the correlation calculation matrices includes a correlation reference vector, a feature contribution vector, and a correlation matching vector;

[0143] The process of performing correlation-weighted aggregation on at least one of the correlation calculation matrices to obtain the correlation-aggregated feature representation of the merged feature representation used to enhance the contribution of dimming decisions can be implemented through the following example.

[0144] Obtain the dimension alignment vector corresponding to the association reference vector, and use the product of the association matching vector and the dimension alignment vector as the association degree evaluation vector; the association degree evaluation vector is used to characterize the correlation strength between at least two behavioral feature representations.

[0145] The correlation evaluation vector is standardized to obtain the standardized vector corresponding to the correlation evaluation vector. The product of the standardized vector and the feature contribution vector is used as the weighted aggregation vector representing the correlation of at least two behavioral features.

[0146] A linear mapping calculation is performed on the associated aggregated vector to obtain the correlation aggregated feature representation of the merged feature representation used to enhance the contribution of dimming decision.

[0147] In an embodiment of the invention, for example, the server initiates a correlation-weighted aggregation processing flow for the correlation calculation matrix of the "continuous reading of professional books" behavior data unit in the study room scenario. This matrix includes three types of sub-vectors: predefined correlation reference vector, feature contribution vector, and correlation matching vector.

[0148] First, the server obtains the dimension alignment vector corresponding to the associated reference vector. The associated reference vector is a pre-trained 16-dimensional dimming decision core requirement vector, covering dimming dimensions strongly related to reading behavior, such as "color temperature stability," "brightness gradual change rate," and "illuminance uniformity." The dimension alignment vector is used to match the dimensions of behavioral features and dimming requirements. The server generates a 16×32 dimension alignment vector through a pre-trained dimension mapping layer to ensure a one-to-one correspondence between the two types of feature dimensions. Subsequently, the server performs a matrix dot product operation on the associated matching vector and the dimension alignment vector to generate a 16-dimensional correlation evaluation vector. The vector value ranges from 0 to 1, with the correlation strength of the "eye focus - color temperature stability" dimension being 0.91, the "static reading type - brightness gradual change rate" dimension being 0.87, and the "long-cycle reading - dimming cycle adaptation" dimension being 0.76, accurately quantifying the close correlation between behavioral features and dimming decisions.

[0149] Next, the server performs L2 normalization on the correlation evaluation vector, mapping the correlation strength of each dimension to a unified interval of [0,1], eliminating numerical differences between different dimensions, and obtaining a normalized vector. The highest value is maintained at 0.91, and the lowest value is adjusted to 0.12 (corresponding to the "accidental head rotation" feature, which is unrelated to dimming requirements). Subsequently, the server performs element-wise multiplication of the normalized vector with the feature contribution vector. The feature contribution vector is a predefined 32-dimensional vector, where the weight of the behavior identifier feature contribution is 0.85, the weight of the behavior type feature contribution is 0.78, and the weight of the behavior granularity feature contribution is 0.62. After the operation, a 16-dimensional correlation aggregation vector is obtained, which strengthens the weight of high-correlation behavior features and weakens the influence of low-correlation features.

[0150] In this embodiment of the invention, the hierarchical weighted synthesis process of the behavior association feature representation corresponding to the target user behavior data and the generated feature representation corresponding to the behavior data unit to obtain the comprehensive behavior representation corresponding to the behavior data unit can be implemented through the following example.

[0151] The behavioral association feature representation corresponding to the target user behavior data and the generated feature representation corresponding to the behavior data unit are loaded into the hierarchical weighted synthesis processing network in the behavior representation reconstruction model; the hierarchical weighted synthesis processing network includes a feature mapping unit, at least two hierarchical weighted synthesis processing units and a feature synthesis processing unit.

[0152] In the feature mapping unit, the generated feature representation corresponding to the behavior data unit is subjected to feature mapping processing to obtain the mapped feature representation corresponding to the behavior data unit;

[0153] In at least two of the hierarchical weighted synthesis processing units, based on the behavior association feature representation corresponding to the target user behavior data and the mapping feature representation corresponding to the behavior data unit, hierarchical feature representations corresponding to the at least two hierarchical weighted synthesis processing units are generated respectively. The hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit in the at least two hierarchical weighted synthesis processing units is used as the hierarchical weighted synthesis processing vector corresponding to the behavior data unit; the target hierarchical weighted synthesis processing unit is the last hierarchical weighted synthesis processing unit in the at least two hierarchical weighted synthesis processing units.

[0154] In the feature synthesis processing unit, the hierarchical weighted synthesis processing vector corresponding to the behavior data unit and the mapping feature representation corresponding to the behavior data unit are subjected to weight-based result integration processing to obtain the comprehensive behavior representation corresponding to the behavior data unit.

[0155] In an embodiment of the present invention, for example, the server initiates a hierarchical weighted synthesis processing flow for the behavioral data unit of "continuous reading of professional books" in a study scenario:

[0156] First, the pre-obtained behavioral association feature representation and the generated feature representation are loaded into the hierarchical weighted synthesis processing network of the behavioral representation reconstruction model. This network includes a feature mapping unit, two hierarchical weighted synthesis processing units, and one feature synthesis processing unit. The behavioral association feature representation is a 32-dimensional vector representing the association strength (0.95) between the reading unit and the subsequent "writing notes" unit, and the association strength (0.3) between the reading unit and the "getting up to retrieve a book" unit. The generated feature representation is a 112-dimensional vector fusing relevance aggregation features and merged features, covering multiple dimensions of information such as reading focus, behavior type, and behavior cycle.

[0157] In the feature mapping unit, the server maps the 112-dimensional generated features into a 32-dimensional mapped feature representation through a pre-trained linear transformation layer. This is aligned with the dimensions of the behavior-related feature representation, retaining core features that are strongly correlated with dimming decisions, such as "high-focus reading," "static desk work," and "long-term continuous work," while filtering out redundant dimensional information such as "occasionally raising your hand to fix your hair," ensuring that the feature dimensions match and focus on the core needs.

[0158] In the hierarchical weighted synthesis processing stage, the first hierarchical synthesis unit inputs the 32-dimensional behavior-related feature representation and mapping feature representation into the attention mechanism module to calculate the influence weight of behavior-related factors on the current behavior's lighting requirements. Since reading and writing are both focused behaviors, their lighting requirements highly overlap, so the weight of the "stable lighting requirements" dimension is increased to 0.9, generating a 32-dimensional first-layer hierarchical feature representation that highlights the behavior-related attributes. The second hierarchical synthesis unit, as the final target unit, further strengthens the core requirement of "avoiding frequent dimming during long-term reading" based on the first-layer hierarchical feature representation and the behavior-related feature representation, adjusting the weight of the "gradual dimming rate" dimension to 0.85, generating a 32-dimensional hierarchical weighted synthesis processing vector. This vector integrates the behavior-related logic and the current behavior's core requirements.

[0159] In the feature synthesis processing unit, the server loads pre-trained weight coefficients (0.7 for hierarchical weighted synthesis processing vectors and 0.3 for mapped feature representations), performs element-wise weighted summation and non-linear activation processing on the two types of vectors, and outputs a 32-dimensional comprehensive behavioral representation. This representation clearly marks the core lighting requirements of the current behavior: "High-focus reading requires stable warm white light, consistent with the lighting requirements of subsequent writing behavior, requiring a gradually changing lighting rate," providing accurate and scene-aligned behavioral basis for the generation of subsequent lighting parameters.

[0160] In an embodiment of the present invention, at least two of the hierarchical weighted synthesis processing units include a target hierarchical weighted synthesis processing unit, wherein the target hierarchical weighted synthesis processing unit is any one of the hierarchical weighted synthesis processing units in at least one of the hierarchical weighted synthesis processing units.

[0161] The step of generating hierarchical feature representations corresponding to at least two hierarchical weighted synthesis processing units based on the behavior association feature representation corresponding to the target user behavior data and the mapping feature representation corresponding to the behavior data unit can be implemented through the following example.

[0162] If the target hierarchical weighted synthesis processing unit is the first hierarchical weighted synthesis processing unit among at least two hierarchical weighted synthesis processing units, then the target hierarchical weighted synthesis processing unit performs behavior association analysis on the behavior association feature representation corresponding to the target user behavior data and the mapping feature representation corresponding to the behavior data unit to obtain the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit.

[0163] If the target hierarchical weighted synthesis processing unit is not the first hierarchical weighted synthesis processing unit among at least two hierarchical weighted synthesis processing units, then the target hierarchical weighted synthesis processing unit performs behavior association analysis on the behavior association feature representation corresponding to the target user behavior data and the hierarchical feature representation corresponding to the preceding hierarchical weighted synthesis processing unit to obtain the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit; the preceding hierarchical weighted synthesis processing unit is the hierarchical weighted synthesis processing unit preceding the target hierarchical weighted synthesis processing unit.

[0164] In an embodiment of the invention, for example, the server generates a hierarchical feature representation according to rules in two hierarchical weighted synthesis processing units of the hierarchical weighted synthesis processing network for the behavioral data unit of "continuous reading of professional books" in a study scenario:

[0165] First, the server loads the pre-obtained 32-dimensional behavioral association feature representation (the association strength between this reading unit and the subsequent "writing notes" unit is 0.95, and the association strength between this unit and the "getting up to pick up a book" unit is 0.3) and the 32-dimensional mapping feature representation (focusing on the core behavior of "high-concentration reading at a desk") into a hierarchical weighted synthesis processing network. The network contains two cascaded target hierarchical weighted synthesis processing units.

[0166] For the first target hierarchical weighted synthesis processing unit, the server determines it to be the first hierarchical processing unit and then performs behavioral correlation analysis on the behavioral correlation feature representation and the mapping feature representation: It calls the pre-trained multi-head attention mechanism module to match the "behavioral similarity" dimension in the behavioral correlation features with the "lighting stability requirement" dimension in the mapping features. Since reading and writing both belong to the category of focused reading, their lighting requirements highly overlap, so the attention weight of this dimension is increased to 0.92; at the same time, the weight of the low-correlation feature with "getting up to pick up a book" is weakened, and the weight of the "lighting mutation adaptation" dimension is reduced to 0.15. After feature integration, the first 32-dimensional hierarchical feature representation is output, which highlights the attributes of "high-focus reading requires stable color temperature and consistency with the lighting requirements of subsequent writing behavior".

[0167] For the second target hierarchical weighted synthesis processing unit, the server determines that it is not the first hierarchical processing unit, and the preceding unit is the first hierarchical weighted synthesis processing unit. Then, it performs a behavioral correlation analysis on the behavioral association feature representation and the preceding hierarchical feature representation: based on the clearly defined "behavioral similarity requirement" in the preceding features, it further focuses on the continuity requirement of the behavioral sequence, calculating the matching degree between the "behavioral temporal coherence" dimension in the behavioral association features and the "dimming rate" dimension in the preceding features. Because reading and writing behaviors are continuously connected, it is necessary to avoid excessive dimming frequency interfering with focus, so the attention weight of the "dimming rate" dimension is increased from 0.8 to 0.88; at the same time, it strengthens the feature of "long-cycle behavior adapting to long dimming cycles," increasing the weight of the "dimming cycle setting" dimension to 0.85. After feature integration, a 32-dimensional second hierarchical feature representation is output. This feature, based on the first hierarchical feature, further strengthens the core requirement of "continuous focused behavior requires gradual dimming and avoidance of frequent adjustments," providing more logical support for the subsequent generation of comprehensive behavioral representations.

[0168] In this embodiment of the invention, the number of behavioral data units is at least two, and the at least two behavioral data units include target behavioral data units; the behavioral association feature representation corresponding to the target user behavioral data includes behavioral influence weight coefficients of the target behavioral data units for the at least two behavioral data units respectively; the mapping feature representation corresponding to the behavioral data units includes the mapping feature representation corresponding to the target behavioral data units;

[0169] The step of performing behavior association analysis on the behavior association feature representation corresponding to the target user behavior data and the mapping feature representation corresponding to the behavior data unit through the target hierarchical weighted synthesis processing unit to obtain the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit can be implemented through the following example.

[0170] The target hierarchical weighted synthesis processing unit obtains the behavior influence weight matrix of the target behavior data unit for at least two of the behavior data units respectively;

[0171] In the target hierarchical weighted synthesis processing unit, feature merging operations are performed on the mapping feature representation corresponding to the target behavior data unit and at least two behavior influence weight matrices respectively to obtain the merged behavior influence weight matrices corresponding to at least two behavior data units respectively.

[0172] Based on at least two of the said behavior influence weight coefficients, a weighted cumulative calculation is performed on at least two of the said combined behavior influence weight matrices to generate a behavior influence relationship representation corresponding to the target behavior data unit;

[0173] The behavior influence relationship representation corresponding to the target behavior data unit is subjected to response enhancement processing to obtain the hierarchical sub-feature representation corresponding to the target behavior data unit; the hierarchical sub-feature representation corresponding to the target behavior data unit belongs to the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit.

[0174] In an embodiment of the present invention, for example, the server performs behavior correlation analysis in the first target hierarchical weighted synthesis processing unit for the target behavior data unit "continuous reading of professional books" in a study scenario. The specific process is as follows:

[0175] First, the server obtains the behavior influence weight matrices of the target behavior data unit for the other two behavior data units from the pre-trained weight library of the first target hierarchical weighted synthesis processing unit: the weight matrix for the "writing paper notes" unit is 32×32, with the core dimensions focusing on "light uniformity matching and color temperature stability alignment", and the baseline weight of the corresponding dimension in the matrix is ​​0.8; the weight matrix for the "getting up to pick up a book" unit is 32×32, with the core dimensions focusing on "global supplementary lighting adaptation and brightness change tolerance", and the baseline weight of the corresponding dimension in the matrix is ​​0.2.

[0176] Subsequently, the server performs feature merging operations with the 32-dimensional mapping feature representation corresponding to the target behavior data unit (focusing on the core attributes of "high concentration while reading, 4000-4500K color temperature requirement, and low brightness abrupt change tolerance") and the two behavior influence weight matrices respectively: after expanding the mapping feature vector into a 32×32 feature matrix according to the feature dimension, it concatenates it with the corresponding behavior influence weight matrix element by element to obtain two 32×64 merged behavior influence weight matrices for "continuous reading - writing notes" and "continuous reading - getting up to get a book", which fully preserves the core requirements of the target behavior and the rules of its influence on other behaviors.

[0177] Next, the server calls the preset weight coefficients in the behavior association feature representation corresponding to the target user behavior data (the weight coefficient for the target behavior data unit's influence on the behavior of "writing paper notes" is 0.95, and for "getting up to get a book" it is 0.3), and performs a weighted cumulative calculation on the two combined behavior influence weight matrices: using the weight coefficients as weighting factors, the corresponding elements of the two matrices are weighted and then summed to generate a 32×64 behavior influence relationship representation, in which the weighted cumulative value of the dimension of "color temperature stability requirement for continuous reading matches the requirement for writing notes" reaches 0.88, and the weighted cumulative value of the dimension of "low mutation requirement for continuous reading matches the requirement for getting up to get a book" is only 0.06, accurately reflecting the correlation strength of the target behavior with the lighting requirements of different subsequent behaviors.

[0178] Finally, the server performs response enhancement processing on the behavioral impact relationship representation: by strengthening the high correlation dimensions with values ​​greater than 0.7 using a pre-trained ReLU activation function, the feature value of the "color temperature stability alignment" dimension is increased to 0.92; while the low correlation dimensions with values ​​less than 0.2 are weakened, and the feature value of the "brightness mutation adaptation" dimension is reduced to 0.03. After feature compression, a hierarchical sub-feature representation corresponding to the target behavioral data unit is obtained. This sub-feature serves as the core component of the hierarchical feature representation corresponding to the first target hierarchical weighted synthesis processing unit, clearly marking the dimming decision basis of "continuous reading behavior should prioritize matching the lighting requirements of subsequent writing behavior, and weakening the interference of book picking behavior".

[0179] In this embodiment of the invention, the standardized adjustment matrix includes adjustment confidence scores corresponding to at least two behavior optimization processes; the total value of the at least two adjustment confidence scores is a standardized reference threshold.

[0180] The process of reconstructing the behavioral representation of the behavioral data unit based on the standardized adjustment matrix to obtain the reconstructed behavioral data unit can be implemented through the following example.

[0181] Obtain the highest adjustment confidence from at least two adjustment confidences of the standardized adjustment matrix;

[0182] If the highest adjustment confidence exceeds the confidence threshold, then based on the behavior optimization processing corresponding to the highest adjustment confidence, the behavior data unit is reconstructed to obtain the reconstructed behavior data unit.

[0183] If the highest adjustment confidence level does not exceed the confidence level threshold, then the behavioral data unit is used as the behavioral data unit for completing the reconstruction.

[0184] In an embodiment of the present invention, for example, when the server performs a behavior representation reconstruction operation on the target behavior data unit of the study room scenario, it first loads the behavior adjustment matrix that has been standardized, namely the standardized adjustment matrix: the standardized adjustment matrix for the behavior data unit of "continuous reading of professional books" includes the adjustment confidence corresponding to three types of behavior optimization processing: "strengthening focus feature weight" corresponds to 0.82, "filtering local action noise" corresponds to 0.15, and "aligning subsequent behavior association features" corresponds to 0.03. The total value of the three types of adjustment confidence is the standardized reference threshold of 1. All values ​​have been calibrated by L2 standardization to eliminate the numerical deviation between different optimization processing.

[0185] The server extracts the highest value of 0.82 from the three types of adjustment confidence in the standardized adjustment matrix, calls the system's pre-configured confidence threshold of 0.7, and after numerical comparison, 0.82 exceeds the threshold. Then, it performs the "enhanced focus feature weight" behavior optimization processing corresponding to the highest adjustment confidence: the focus feature weights related to light intensity, such as gaze focus duration and posture stability, in the original behavioral data unit are increased from 0.6 to 0.9; at the same time, local motion noise data accounting for 18% in the original data, such as "occasionally looking up to stretch" and "raising hands to fix hair", are filtered out; and the correlation features between this reading behavior and the subsequent "writing notes" behavior are embedded. Finally, the reconstructed behavioral data unit is obtained. This unit only retains the core features of highly correlated focused reading, and the proportion of redundant features is reduced to 3%.

[0186] For another real-world scenario, if the target behavioral data unit is "frequently getting up to browse the bookshelf", its standardized adjustment matrix includes adjustment confidence scores for three types of behavioral optimization processing: "enhancing the lighting adaptation of mobile behavior" corresponds to 0.65, "filtering noise from stationary actions" corresponds to 0.2, and "aligning with preceding reading behavior features" corresponds to 0.15. The highest adjustment confidence score of 0.65 extracted by the server does not exceed the confidence threshold of 0.7. Therefore, the original behavioral data unit is directly used as the behavioral data unit to complete the reconstruction, fully preserving its original behavioral characteristics of strong mobility and discontinuous actions, ensuring that subsequent lighting decisions accurately adapt to the real state of the behavior.

[0187] In this embodiment of the invention, the following implementation methods are also provided.

[0188] Obtain the target user behavior data after reconstruction; the target user behavior data after reconstruction is determined by the behavior data unit after reconstruction.

[0189] If the target user behavior data is consistent with the target user behavior data after reconstruction, then the target user behavior data after reconstruction is used as the behavior representation reconstruction result of the target user behavior data.

[0190] If the target user behavior data and the reconstructed target user behavior data are inconsistent, the step of reconstructing the behavior representation of the reconstructed target user behavior data is repeated.

[0191] In an embodiment of the present invention, for example, after the server completes the representation reconstruction of the behavioral data unit of "continuous reading of professional books" in the study scenario, it initiates the behavioral data consistency verification and iterative reconstruction process:

[0192] First, the server obtains the target user behavior data that has been reconstructed, which is composed of three reconstructed behavioral data units (reconstructed "continuous reading", "writing notes" and "getting up to pick up a book" units). This reconstructed data filters out 18% of the local action noise in the original data, strengthens the weight of behavioral correlation features and focus features, and increases the proportion of core behaviors from 82% to 97%.

[0193] Subsequently, the server compares the original target user behavior data with the reconstructed target user behavior data in terms of feature dimension consistency: by calculating the cosine similarity of the feature vectors of the two types of data, the similarity value is 0.72, which is less than the system's preset consistency threshold of 0.95. Therefore, the two types of data are determined to be inconsistent, and the cyclic reconstruction process is then started.

[0194] The server takes the reconstructed target user behavior data as input and repeatedly executes the behavior representation reconstruction steps: in the behavior data unit encoding stage, a new feature dimension of "behavioral persistence stability" is extracted; in the feature merging and aggregation stage, the weight coefficient of behavior-related features is adjusted to 0.9; in the hierarchical weighted synthesis stage, the feature dimension of "lighting adaptation for long-term focused behavior" is strengthened. After completing the second reconstruction, the server extracts the spliced ​​reconstructed target user behavior data again, calculates its cosine similarity with the original data, and finds it to be 0.96, exceeding the consistency threshold. Therefore, the two types of data are considered consistent, and the cyclic reconstruction is terminated. The target user behavior data from the second reconstruction is used as the final behavior representation reconstruction result.

[0195] In another real-world scenario, if the original behavioral data is "a single short-term act of getting up to retrieve an item", the cosine similarity between the data after the first reconstruction and the original data reaches 0.98, exceeding the consistency threshold. The server determines that the two types of data are consistent and directly uses the target user behavior data from the first reconstruction as the final result, without needing to enter the cyclic reconstruction process.

[0196] In this embodiment of the invention, the generation of corresponding dimming control parameters based on the reconstructed behavioral data unit and the ambient light feature representation can be implemented through the following example.

[0197] The reconstructed behavioral data unit and the ambient light feature representation are loaded into the dimming parameter generation model;

[0198] In the dimming parameter generation model, the behavior dimming requirement is analyzed for the reconstructed behavior data unit to obtain the behavior dimming requirement feature corresponding to the reconstructed behavior data unit.

[0199] Based on the ambient light feature representation, the behavior dimming demand feature is subjected to illumination condition constraint mapping to obtain the environmental constraint feature representation corresponding to the behavior dimming demand feature.

[0200] The dimming control parameters are generated by performing joint decision calculations on the behavioral dimming demand characteristics and the environmental constraint characteristics.

[0201] In an embodiment of the present invention, for example, the server initiates a dimming control parameter generation process to meet the dimming requirements of a study room scenario:

[0202] First, the server loads the reconstructed "continuous reading of professional books" behavioral data unit (32-dimensional feature vector, with the core annotation of highly focused desk state and highly correlated with the lighting requirements of subsequent writing behavior) and the ambient light feature representation (128-dimensional feature vector, annotating the current strong light by the window in the study, the medium brightness in the desk area, and the state of the light intensity continuously increasing at a rate of 500 lx per minute) into the pre-trained Transformer architecture dimming parameter generation model. This model has completed training on 100,000 sets of dimming data in home scenes and has the ability to make joint decisions based on behavioral needs and environmental constraints.

[0203] In the dimming parameter generation model, the server performs behavioral dimming requirement parsing on the reconstructed behavioral data unit: it calls the predefined behavior-requirement mapping rule library, extracts the core attribute of "high-focused reading" from the feature vector, and parses out the corresponding behavioral dimming requirement features, such as warm white light with a color temperature of 4000-4500K, medium intensity with a brightness of 500-600lx, a gradual change rule with a brightness adjustment rate not exceeding 20lx / second, and an area coverage requirement with an illumination uniformity of not less than 0.8.

[0204] Subsequently, the server performs lighting condition constraint mapping processing on the behavioral dimming requirement features based on the ambient light feature representation: combining the current actual brightness of 800lx in the desk area (exceeding the requirement range), the light uniformity deviation caused by strong light from the window (currently 0.72), and the trend of continuously increasing light intensity, the behavioral requirement features are adapted and adjusted to generate an environmental constraint feature representation. The brightness needs to be reduced to the upper limit of the requirement range (600lx) to offset the impact of subsequent increase in ambient light, the auxiliary light from the window needs to be turned on to supplement the light and improve uniformity, and a dynamic adjustment trigger threshold is set (when the ambient light change rate exceeds 100lx / second, the dimming rate is increased to 50lx / second).

[0205] Finally, the server performs joint decision-making calculations on the behavioral dimming requirement features and environmental constraint features: it invokes the multi-head attention mechanism within the model to weightedly integrate the highly correlated dimensions of the two types of features (such as "color temperature requirement - ambient color temperature matching" and "brightness requirement - ambient light change trend"); after linear mapping and activation function processing, specific dimming control parameters are generated. The brightness of the main LED light group on the desk is reduced from 65% to 42% (corresponding to 525 lx), and the color temperature is adjusted from 6200K to 4200K. The brightness of the auxiliary LED light group by the window is increased to 25% (corresponding to 300 lx), and the color temperature is synchronously set to 4200K. The dynamic dimming rule is that when the ambient light change rate exceeds 100 lx / second, the dimming rate is increased to 50 lx / second. After generation, the server verifies that the parameters are within the hardware capacity of the LED light group and then sends them to the light group for execution.

[0206] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned AI-driven LED intelligent dimming method that integrates user behavior with ambient light. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines. For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with different modifications to suit a particular intended application.

Claims

1. An AI-driven intelligent LED dimming method that integrates user behavior and ambient light, characterized in that, include: Acquire ambient light data in the target application scenario, wherein the ambient light data includes at least one ambient light acquisition parameter; The ambient light data is subjected to ambient light feature extraction processing to obtain an ambient light feature representation corresponding to the ambient light data; the ambient light feature representation is used to characterize the light intensity state and light change trend in the target application scene. Perform behavioral representation reconstruction processing on the behavioral data units in the target user behavior data to obtain the reconstructed behavioral data units; Based on the reconstructed behavioral data unit and the ambient light feature representation, corresponding dimming control parameters are generated; The output brightness of the LED light source is adaptively adjusted according to the dimming control parameters.

2. The method according to claim 1, characterized in that, The step of performing behavior representation reconstruction processing on behavior data units in the target user behavior data to obtain reconstructed behavior data units includes: The behavioral data units in the target user behavior data are encoded to obtain at least two behavioral feature representations corresponding to the behavioral data units; the at least two behavioral feature representations are used to reflect feature data of different categories; the behavioral data units are obtained by dividing the target user behavior data into behavioral sequences; Perform a feature merging operation on at least two of the behavioral feature representations to obtain a merged feature representation corresponding to the behavioral data unit. Perform feature rearrangement and aggregation on the merged feature representation based on the correlation matrix to obtain a correlation-aggregated feature representation of the merged feature representation used to enhance the dimming decision contribution. The behavioral association feature representation corresponding to the target user behavior data and the generated feature representation corresponding to the behavior data unit are subjected to hierarchical weighted synthesis processing to obtain the comprehensive behavioral representation corresponding to the behavior data unit; the generated feature representation is determined by the relevance aggregation feature representation and the merged feature representation; the behavioral association feature representation is used to characterize the association strength between the behavior data units; A linear mapping calculation is performed on the comprehensive behavioral representation to obtain the behavioral adjustment matrix corresponding to the behavioral data unit; The behavior adjustment matrix is ​​standardized to obtain the standardized adjustment matrix corresponding to the behavior adjustment matrix, and the behavior data unit is reconstructed based on the standardized adjustment matrix to obtain the reconstructed behavior data unit.

3. The method according to claim 2, characterized in that, At least two of the behavioral feature representations include behavioral identification feature representations; The encoding process of behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units includes: Load the target user behavior data into the feature encoder in the behavior representation reconstruction model; In the feature encoder, at least one behavior identifier unit of the behavior data unit in the target user behavior data is obtained; at least one behavior identifier unit is used to characterize the behavior expression attribute of the behavior data unit; Obtain the behavior identifier unit matrix corresponding to at least one of the behavior identifier units, and perform local association calculation on the at least one behavior identifier unit matrix to obtain the local behavior association matrix corresponding to at least one behavior identifier unit matrix. At least one of the local behavior association matrices is aggregated and dimensionality reduced to obtain the behavior identifier feature representation corresponding to the behavior data unit.

4. The method according to claim 2, characterized in that, At least two of the behavioral feature representations include behavioral appearance feature representations; The encoding process of behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units includes: Load the target user behavior data into the feature encoder in the behavior representation reconstruction model; In the feature encoder, at least one behavior visual mapping data of the behavior data unit in the target user behavior data is obtained; at least one of the behavior visual mapping data is used to reflect the behavior data unit with different behavior presentation modes; Feature representation generation processing is performed on at least one of the behavioral visual mapping data to obtain a performance pattern matrix corresponding to each of the at least one behavioral visual mapping data. At least one of the performance pattern matrices is aggregated and dimensionality reduced to obtain the behavioral appearance feature representation corresponding to the behavioral data unit.

5. The method according to claim 2, characterized in that, At least two of the behavioral feature representations include behavioral type feature representations; The encoding process of behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units includes: Load the target user behavior data into the feature encoder in the behavior representation reconstruction model; In the feature encoder, the target user behavior data is subjected to behavior sequence deconstruction processing to obtain at least one behavior semantic segment in the target user behavior data; at least one behavior semantic segment is used to constitute the target user behavior data, and at least one behavior semantic segment is a behavior semantic segment of different behavior types in the target user behavior data; Obtain the behavior type matrix corresponding to the target behavior semantic segment from the behavior type matrices corresponding to at least one of the behavior semantic segments respectively, and use the behavior type matrix corresponding to the target behavior semantic segment as the behavior type feature representation of the behavior data unit in the target user behavior data; the target behavior semantic segment is the behavior semantic segment corresponding to the behavior data unit in at least one of the behavior semantic segments.

6. The method according to claim 2, characterized in that, At least two of the aforementioned behavioral feature representations include behavioral granularity feature representations; The encoding process of behavioral data units in the target user behavior data to obtain at least two behavioral feature representations corresponding to the behavioral data units includes: Load the target user behavior data into the feature encoder in the behavior representation reconstruction model; In the feature encoder, behavioral representation mapping is performed on the behavioral data units in the target user behavior data to obtain the behavioral sub-unit vector corresponding to the behavioral data unit; Determine the behavior time-series index of the behavior data unit in the target user behavior data, perform time-series weight mapping on the behavior time-series index of the behavior data unit, and obtain the time-series index vector corresponding to the behavior data unit. Obtain the behavior boundary identifier vector corresponding to the behavior data unit, and perform joint feature encoding on the behavior sub-unit vector, the temporal index vector, and the behavior boundary identifier vector corresponding to the behavior data unit to obtain the behavior granularity feature representation corresponding to the behavior data unit.

7. The method according to claim 2, characterized in that, The step of performing a feature merging operation on at least two of the behavioral feature representations to obtain a merged feature representation corresponding to the behavioral data unit, and performing feature rearrangement and aggregation based on the correlation matrix on the merged feature representation to obtain a correlation-aggregated feature representation of the merged feature representation used to enhance the dimming decision contribution, includes: Load at least two of the aforementioned behavioral feature representations into the feature joint modeling unit in the behavioral representation reconstruction model; In the joint feature modeling unit, a feature merging operation is performed on at least two of the behavioral feature representations to obtain the merged feature representation corresponding to the behavioral data unit; A linear mapping calculation is performed on the merged feature representation to obtain at least one correlation calculation matrix corresponding to the merged feature representation; at least one correlation calculation matrix includes a correlation reference vector, a feature contribution vector, and a correlation matching vector. Obtain the dimension alignment vector corresponding to the association reference vector, and use the product of the association matching vector and the dimension alignment vector as the association degree evaluation vector; the association degree evaluation vector is used to characterize the correlation strength between at least two behavioral feature representations. The correlation evaluation vector is standardized to obtain the standardized vector corresponding to the correlation evaluation vector. The product of the standardized vector and the feature contribution vector is used as the weighted aggregation vector representing the correlation of at least two behavioral features. A linear mapping calculation is performed on the associated aggregated vector to obtain the correlation aggregated feature representation of the merged feature representation used to enhance the contribution of dimming decision.

8. The method according to claim 2, characterized in that, The step of performing hierarchical weighted synthesis processing on the behavior association feature representation corresponding to the target user behavior data and the generated feature representation corresponding to the behavior data unit to obtain the comprehensive behavior representation corresponding to the behavior data unit includes: The behavioral association feature representation corresponding to the target user behavior data and the generated feature representation corresponding to the behavior data unit are loaded into the hierarchical weighted synthesis processing network in the behavior representation reconstruction model; the hierarchical weighted synthesis processing network includes a feature mapping unit, at least two hierarchical weighted synthesis processing units, and a feature synthesis processing unit; the at least two hierarchical weighted synthesis processing units include a target hierarchical weighted synthesis processing unit, and the target hierarchical weighted synthesis processing unit is any one of the at least one hierarchical weighted synthesis processing units; In the feature mapping unit, the generated feature representation corresponding to the behavior data unit is subjected to feature mapping processing to obtain the mapped feature representation corresponding to the behavior data unit; If the target hierarchical weighted synthesis processing unit is the first hierarchical weighted synthesis processing unit among at least two hierarchical weighted synthesis processing units, then the target hierarchical weighted synthesis processing unit performs behavior association analysis on the behavior association feature representation corresponding to the target user behavior data and the mapping feature representation corresponding to the behavior data unit to obtain the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit. If the target hierarchical weighted synthesis processing unit is not the first hierarchical weighted synthesis processing unit among at least two hierarchical weighted synthesis processing units, then the target hierarchical weighted synthesis processing unit performs behavior association analysis on the behavior association feature representation corresponding to the target user behavior data and the hierarchical feature representation corresponding to the preceding hierarchical weighted synthesis processing unit to obtain the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit; the preceding hierarchical weighted synthesis processing unit is the hierarchical weighted synthesis processing unit preceding the target hierarchical weighted synthesis processing unit. The hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit in at least two of the hierarchical weighted synthesis processing units is used as the hierarchical weighted synthesis processing vector corresponding to the behavior data unit; the target hierarchical weighted synthesis processing unit is the last hierarchical weighted synthesis processing unit in at least two of the hierarchical weighted synthesis processing units. In the feature synthesis processing unit, the hierarchical weighted synthesis processing vector corresponding to the behavior data unit and the mapping feature representation corresponding to the behavior data unit are subjected to weight-based result integration processing to obtain the comprehensive behavior representation corresponding to the behavior data unit.

9. The method according to claim 8, characterized in that, The number of behavioral data units is at least two, and the at least two behavioral data units include target behavioral data units; the behavioral association feature representation corresponding to the target user behavioral data includes behavioral influence weight coefficients of the target behavioral data units for the at least two behavioral data units respectively; the mapping feature representation corresponding to the behavioral data units includes the mapping feature representation corresponding to the target behavioral data units; The step of performing behavior association analysis on the behavior association feature representation corresponding to the target user behavior data and the mapping feature representation corresponding to the behavior data unit through the target hierarchical weighted synthesis processing unit to obtain the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit includes: The target hierarchical weighted synthesis processing unit obtains the behavior influence weight matrix of the target behavior data unit for at least two of the behavior data units respectively; In the target hierarchical weighted synthesis processing unit, feature merging operations are performed on the mapping feature representation corresponding to the target behavior data unit and at least two behavior influence weight matrices respectively to obtain the merged behavior influence weight matrices corresponding to at least two behavior data units respectively. Based on at least two of the said behavior influence weight coefficients, a weighted cumulative calculation is performed on at least two of the said combined behavior influence weight matrices to generate a behavior influence relationship representation corresponding to the target behavior data unit; The behavior influence relationship representation corresponding to the target behavior data unit is subjected to response enhancement processing to obtain the hierarchical sub-feature representation corresponding to the target behavior data unit; the hierarchical sub-feature representation corresponding to the target behavior data unit belongs to the hierarchical feature representation corresponding to the target hierarchical weighted synthesis processing unit.

10. A server system, characterized in that, Includes a server, said server being used in the method of any one of claims 1-9.

Images