Multi-screen dynamic display control method, device and electronic equipment

Abstract

This application relates to the field of intelligent interactive technology and discloses a multi-screen dynamic display control method, device, and electronic device, including the following steps: constructing a seventh-order behavior tensor through spatiotemporal alignment of heterogeneous sensors, extracting cross-modal intent features using CP-Tucker hybrid decomposition; collaboratively generating resource allocation parameters based on quantum annealing and classical optimization, and dynamically correcting screen geometry mapping in combination with neural radiation fields; and innovatively integrating federated learning noise injection with elastic pixel binding strategies to achieve real-time rendering control under privacy protection. Through the synergistic effects of spatiotemporal synchronization compensation, hybrid tensor decomposition, quantum optimization engine, and differentiable rendering model, dynamic adaptation of display parameters to user behavior and environmental changes is achieved, improving the accuracy and real-time performance of multi-screen collaborative display while ensuring privacy and security.

Description

Technical Field

[0001] The present invention relates to the field of intelligent interaction technology, and specifically to a multi-screen dynamic display control method, device and electronic equipment. Background Art

[0002] In the field of intelligent interactive systems, multi-screen dynamic display control technology faces the dual challenges of multi-dimensional sensory data fusion and real-time resource scheduling. Traditional methods for integrating heterogeneous sensor data lack precise spatiotemporal alignment mechanisms, leading to spatiotemporal misalignment in the behavioral feature extraction process and making it difficult to accurately capture the cross-modal correlation characteristics of user intent.

[0003] When dealing with multi-screen collaborative optimization problems, existing resource allocation algorithms often adopt a static allocation strategy with a fixed threshold. This strategy cannot adapt to network latency fluctuations and dynamic changes in physical space, and can easily cause a spatial mismatch between displayed content and user perception.

[0004] Current display correction technologies rely heavily on preset geometric mapping parameters. Due to the lack of biometric coupling mechanisms, display parameter updates lag behind actual demand when ambient lighting changes suddenly or the user's physiological state fluctuates. Furthermore, the balance between privacy protection and model performance remains unresolved. Traditional federated learning solutions employ fixed noise injection strategies, which can lead to unstable model convergence in dynamic scenarios and make it difficult to maintain sustained multi-device collaborative optimization. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a multi-screen dynamic display control method, device and electronic device, which solves the problems in the existing technology such as inconsistent spatiotemporal benchmarks of multimodal perception data leading to intention recognition deviation, static resource allocation strategies being difficult to adapt to dynamic environmental changes, display correction lagging behind biometric response, and difficulty in coordinated optimization of privacy protection and model performance.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multi-screen dynamic display control method, comprising the following steps:

[0007] S1. Synchronously collect the user's multimodal behavior data and environmental parameters, and generate multi-source perception data with a unified benchmark through spatiotemporal alignment.

[0008] S2. Constructing a seventh-order behavior tensor model including spatiotemporal correlation features based on the multi-source perception data, and performing hybrid tensor decomposition to extract user intention features;

[0009] S3. Solving dynamic resource allocation parameters among multiple screens using a quantum-classical hybrid optimization algorithm based on characteristic parameters output by the behavior tensor model;

[0010] S4. Based on the resource allocation parameters, dynamically correct the geometric parameters of the display device using the neural radiation field to generate a display mapping relationship that conforms to the physical space topology;

[0011] S5. Based on the displayed mapping relationship, a federated learning mechanism is used to update the multimodal intent recognition model, and the updated model parameters are fed back to the tensor decomposition process of step S2;

[0012] S6. Based on the calibrated display parameters and resource allocation plan, drive the multi-screen system to perform dynamic rendering and content output.

[0013] Preferably, step S1 includes the following steps:

[0014] The inertial measurement unit captures the device's motion posture quaternion at a 120Hz sampling rate, a ToF depth camera generates a 640×480 resolution depth map, and an eye tracking module measures the pupil diameter change rate in real time.

[0015] The IEEE1588v2 protocol is used to achieve multi-sensor clock synchronization and construct a spatial transformation matrix chain including translation vectors and rotation matrices.

[0016] Preferably, the mixed tensor decomposition in step S2 satisfies the following formula definition:

[0017] CP decomposition term:

[0018] ;

[0019] in:

[0020] For the eigenvalues;

[0021] Indicates the The modal factor vectors;

[0022] Represents vector outer product operation;

[0023] Tucker decomposition terms:

[0024] ;

[0025] in:

[0026] is the core tensor;

[0027] For the The factor matrix of the modalities;

[0028] Represents the first Modal product.

[0029] Preferably, the quantum optimization process in step S3 constructs the following Hamiltonian:

[0030] ;

[0031] in:

[0032] Display screen and The strength of the association is determined by the Frobenius norm of the corresponding dimension of the behavior tensor;

[0033] is the physical screen spacing;

[0034] It is a dynamic delay threshold that is adaptively adjusted according to the network load rate;

[0035] , For the The Pauli operator for 1 qubit.

[0036] Preferably, step S4 includes the following steps:

[0037] A neural radiation field model is established to generate screen pixel colors through cumulative transmittance calculation, where transmittance is related to the optical properties of the screen material;

[0038] Dynamically adjust the radiation field parameter update rate based on the user's physiological characteristic signals collected in real time, wherein the physiological characteristic signals include at least one of pupil diameter and heart rate variability;

[0039] Screen surface differential constraints are introduced to ensure that the geometric correction process meets the continuity conditions of physical space.

[0040] Preferably, step S5 includes the following steps:

[0041] During on-device training, adaptive noise related to the gradient norm nonlinearity is injected into the model gradient, where the noise intensity decays exponentially with the increase of the gradient amplitude;

[0042] When the edge server aggregates the obfuscated gradients, a periodic decay strategy is used to dynamically adjust the learning rate. The learning rate is kept high in the early stages of training to accelerate convergence, and is gradually reduced in the later stages to improve model stability.

[0043] The updated model parameters are encrypted and transmitted to the tensor decomposition of step S2 for reconstructing the core dimension of the behavior tensor.

[0044] Preferably, step S6 includes the following steps:

[0045] Build a virtual pixel resource pool and define the dynamic mapping relationship between virtual pixels and multiple physical screens. The alpha channel blending weight of each virtual pixel is jointly optimized based on the physical screen spacing and network latency.

[0046] Performs environment-aware dynamic tone compression, jointly adjusting the HDR compression curve based on the ambient light field intensity distribution and user biometric signals, where the compression strength is positively correlated with the mean ambient light intensity;

[0047] Implement physical screen brightness conservation constraints and ensure that the total brightness output of each screen pixel does not exceed a preset threshold through normalization.

[0048] Preferably, the screen correlation The calculation method is:

[0049] ;

[0050] in is the adjustable integration time window, Represents a slice of a tensor in the screen dimension.

[0051] A multi-screen dynamic display control device, comprising:

[0052] A multi-source data synchronization acquisition module is used to perform spatiotemporal alignment processing of user behavior data and environmental parameters through a heterogeneous sensor array;

[0053] Spatiotemporal tensor modeling module, which is used to construct and decompose seventh-order behavioral tensor models based on the hybrid tensor decomposition algorithm;

[0054] Quantum-classical collaborative optimization module, used to solve screen resource allocation parameters with time and space constraints;

[0055] A biophysical coupling correction module is used to dynamically generate a geometric mapping relationship of the display device based on the neural radiation field;

[0056] A private gradient aggregation module for implementing adaptive noise injection and gradient obfuscation in federated learning models;

[0057] Dynamic resource rendering engine module, used to perform elastic pixel mapping and dynamic tone rendering of multi-screen systems.

[0058] An electronic device, comprising:

[0059] processor;

[0060] a memory storing a computer program;

[0061] When the processor executes the program, the above-described method is implemented.

[0062] The present invention provides a multi-screen dynamic display control method, device, and electronic device. It has the following beneficial effects:

[0063] 1. This invention solves the problem of intention recognition bias caused by data fragmentation in existing multi-screen system behavior modeling by building a heterogeneous sensor clock synchronization and spatial transformation chain, combined with CP-Tucker mixed tensor decomposition, and overcomes the defect of traditional single-modal modeling in insufficient capture of cross-dimensional correlation features.

[0064] 2. This invention is based on quantum annealing Hamiltonian construction and classical post-processing fine-tuning, breaking through the bottleneck of traditional heuristic algorithms that are prone to falling into local optimality under complex constraints. It can achieve global optimization of multi-screen resource allocation and avoid the insufficient adaptability of fixed threshold strategies to network delay fluctuations.

[0065] 3. The present invention solves the problem of error accumulation caused by environmental mutations in traditional geometric correction methods by integrating a differentiable rendering model with a pupil diameter-driven parameter update strategy, and overcomes the limitation that static mapping models cannot adapt to dynamic changes in biological features.

[0066] 4. The present invention adopts gradient norm-related adaptive noise and virtual pixel hybrid weight optimization technology to maintain the effectiveness of model updates while ensuring user privacy, eliminating the risk of model performance degradation caused by fixed noise intensity in existing federated learning schemes. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of the method of the present invention;

[0068] Figure 2 Schematic diagram of the device structure of the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0070] Example:

[0071] Please see the attached Figure 1 , an embodiment of the present invention provides a multi-screen dynamic display control method, comprising the following steps:

[0072] S1. Synchronously collect the user's multimodal behavior data and environmental parameters, and generate multi-source perception data with a unified benchmark through spatiotemporal alignment.

[0073] To achieve precise synchronous acquisition and spatiotemporal reference alignment of multimodal data, this step adopts a heterogeneous sensor collaborative perception architecture and implements physical space mapping and temporal consistency control of multi-source information through a layered data processing process. The specific implementation process is as follows:

[0074] First, a multimodal sensor array is configured, including an inertial measurement unit (IMU), a time-of-flight depth camera, and a near-infrared eye tracking module. The IMU preferably uses a nine-axis micro-electromechanical system (MEMS), which calculates the device's attitude quaternion by fusing an accelerometer, gyroscope, and magnetometer. The sampling frequency is set to 120Hz to ensure continuous motion capture. The time-of-flight depth camera is preferably equipped with a 940nm infrared light source, using the phase measurement principle to generate a 640×480 resolution depth map, and achieves millimeter-level ranging accuracy through adaptive modulation frequency adjustment. The eye tracking module integrates a high-frame-rate near-infrared camera, which analyzes the pupil diameter change rate based on the corneal reflection vector. The sampling interval is preferably 33ms to match the human visual persistence characteristic.

[0075] To eliminate differences in spatiotemporal benchmarks among multi-source data, a hierarchical alignment mechanism was established. At the timing synchronization level, a distributed clock network was constructed using the IEEE 1588v2 precision clock protocol. Each sensor node deployed a hardware timestamp module, achieving microsecond-level synchronization accuracy through a master-slave clock architecture. In specific implementation, the master clock node periodically broadcasts synchronization messages. Slave nodes calculate clock offsets based on message transmission delays and compensate for crystal oscillator frequency deviations using a linear regression algorithm. Preferably, the synchronization period is set to 2 seconds to balance network load and synchronization accuracy requirements.

[0076] At the spatial registration level, a rigid transformation relationship chain is constructed between the sensor coordinate systems. Taking the inertial measurement unit coordinate system as the global reference system, the spatial transformation parameters of the depth camera and eye tracking module relative to the inertial unit are obtained through external parameter calibration. A transformation model from the depth camera coordinates to the global reference is established:

[0077] ;

[0078] in Represents the three-dimensional point coordinates in the global coordinate system, is the original coordinate in the depth camera coordinate system, is the rotation matrix, The extrinsic calibration preferably adopts the checkerboard joint calibration method, which solves the least squares optimization problem through the correspondence relationship of feature points under multiple postures.

[0079] For the eye tracking module, establish a mapping relationship between pupil center coordinates and global reference:

[0080] ;

[0081] in is the sight direction vector in the eye movement module coordinate system, and Preferably, a nonlinear optimization algorithm is used to compensate for the coordinate deviation caused by lens distortion and a polynomial distortion model is constructed:

[0082] ;

[0083] in are ideal projection coordinates, is the actual observation coordinate, is the radial distortion coefficient, is the tangential distortion coefficient, .

[0084] In the data fusion stage, the extended Kalman filter is used to achieve the spatiotemporal alignment of multi-source perception data. ,in is the attitude quaternion, is the angular velocity, is the linear acceleration, is the position vector. The prediction equation is derived based on the inertial measurement unit data:

[0085] ;

[0086] The observation equation fuses the depth camera and eye movement data:

[0087] ;

[0088] in is a nonlinear observation function, The covariance matrix is ​​adaptively adjusted to achieve multi-sensor confidence-weighted fusion, and the inertial data weight is automatically increased when the depth data confidence is lower than the threshold.

[0089] Finally, a unified multi-source perception data stream is generated. The data structure includes fields such as timestamp, spatial coordinates, depth information, physiological parameters, and ambient light intensity. The data stream is time-aligned using a circular buffer, and the buffer size is preferably set to 500ms to accommodate the maximum expected transmission delay.

[0090] S2. Build a seventh-order behavior tensor model containing spatiotemporal correlation features based on multi-source perception data, and perform hybrid tensor decomposition to extract user intent features;

[0091] To accurately extract user intent features, this step constructs a behavior tensor model that integrates multi-dimensional spatiotemporal correlations and separates potential behavior patterns using a hybrid tensor decomposition algorithm. The specific implementation process first establishes a seventh-order tensor data structure, then alternately performs CP decomposition and Tucker decomposition to extract local and global features. Finally, the two decomposition results are combined to generate an intent feature vector.

[0092] The dimensionality of the seventh-order behavior tensor is defined based on the multi-source perception data generated in step S1. Each dimension represents spatial coordinates, gaze direction, gesture characteristics, ambient light intensity, sound field distribution, physiological signals, and time series information. Preferably, the time series dimension is constructed using a sliding window mechanism, with the window length adaptively adjusted based on the continuity requirements of the user's behavior. The tensor element values ​​are obtained through multi-source data normalization to ensure dimensional consistency across all modalities.

[0093] The CP decomposition process is used to capture independent behavior patterns, which is mathematically expressed as:

[0094] ;

[0095] in Indicates the The weight coefficient of the characteristic component, For the The modal factor vectors, Represents the vector outer product operation. Decomposition rank The optimal selection is automatically determined based on the Akaike information criterion, which retains important behavior patterns while suppressing noise interference.

[0096] The Tucker decomposition process is used to extract cross-modal correlation features, and its mathematical expression is:

[0097] ;

[0098] in is the core tensor, is the modal factor matrix, Indicates that the tensor is Modal product operation. Core tensor dimensions Preferably set to correspond to the original dimension 30%-50% of the original data, achieving feature dimensionality reduction.

[0099] During the hybrid decomposition process, CP decomposition and Tucker decomposition are performed alternately and iteratively. First, the factor matrix of the CP decomposition is initialized and used as the initial value input of the Tucker decomposition. In each iteration, the CP decomposition result is fixed and the Tucker core tensor is updated. Then, the weight coefficients of the CP decomposition are inversely optimized. Preferably, the alternating least squares method is used to implement parameter update, and the objective function is defined as:

[0100] ;

[0101] in is the regularization coefficient used to control the sparsity of the factor matrix. The objective function is solved by projected gradient descent, and the step size is adaptively adjusted according to the Hessian matrix condition number.

[0102] In the feature fusion stage, the local pattern features of CP decomposition and the global correlation features of Tucker decomposition are linearly combined according to the information entropy weight. The information entropy of each CP component is:

[0103] ;

[0104] Fusion weight coefficient Dynamic adjustment based on information entropy:

[0105] ;

[0106] in is the temperature coefficient, which is preferably set to the inverse of the current iteration number to balance exploration and utilization. The final intention feature vector is generated by weighted summation:

[0107] ;

[0108] in Represents vector concatenation operation, is the expansion matrix of the core tensor in the seventh mode, is the principal component vector of the time series factor matrix.

[0109] The hybrid decomposition algorithm is implemented in a distributed computing framework, preferably using Apache Spark's Resilient Distributed Datasets mechanism. Tensor slices are distributed to different compute nodes, each of which performs local decomposition calculations in parallel, while the driver program periodically performs global parameter aggregation. To reduce communication overhead, a data transmission protocol based on residual compression is designed, which only transmits parameters whose gradient changes exceed a threshold.

[0110] S3. Based on the characteristic parameters output by the behavioral tensor model, the dynamic resource allocation parameters among multiple screens are solved by a quantum-classical hybrid optimization algorithm;

[0111] To achieve dynamic optimization of multi-screen resource allocation, this step constructs an optimization model based on quantum-classical hybrid computing, converting the intent features extracted in step S2 into physical screen parameter configurations. The specific implementation first establishes a discrete optimization problem with spatiotemporal constraints, then uses quantum annealing to search for a near-optimal solution, and finally fine-tunes the parameters using a classical optimization algorithm.

[0112] The objective function of the optimization problem is defined as:

[0113] ;

[0114] in Indicates the The feature vector of each screen (output from step S2), Assign variables to binary, is the content block attribute vector, Characterization screen and content The strength of association, is the regularization coefficient. The correlation strength is calculated using:

[0115] ;

[0116] in For the behavior tensor in screen dimensions and content dimensions slices, is a dynamic integration window whose value is based on the network delay variance Adaptive Adjustment:

[0117] ;

[0118] in is the base integration time, is the attenuation coefficient. The delay variance is obtained by calculating the sample variance of the historical transmission delay series through a sliding window.

[0119] Map the above optimization problem to the quantum annealing model and construct the following Hamiltonian:

[0120] ;

[0121] in and For the The Pauli operator of qubits, Display screen and The physical distance, It is the adaptive delay threshold. The delay threshold is based on the current network load rate. Dynamic Updates:

[0122] ;

[0123] The quantum annealing process is preferably implemented on a D-Wave quantum processor, and the annealing time According to the problem size set up:

[0124] ;

[0125] After annealing is completed and before collection low-energy samples, and generate candidate solution sets through majority voting mechanism The candidate solution is input into the classical optimization module for post-processing, and the objective function is adjusted to:

[0126] ;

[0127] in is the sample weight (negatively correlated with the energy value), is the total variation regularization term. The coordinate descent method is used for iterative optimization, and a single screen allocation variable is updated each time:

[0128] ;

[0129] Optimization results This is output as the final resource allocation parameter to step S4. To improve computational efficiency, a distributed computing framework is implemented, decomposing the qubit mapping problem into multiple subproblems. Each subproblem corresponds to a local screen cluster. Apache Spark performs parallel annealing calculations, and the driver program periodically performs global deaggregation. The data sharding strategy is based on clustering based on the physical location of the screens, ensuring that the coupling strength between subproblems is below a threshold.

[0130] S4. Based on the resource allocation parameters, the neural radiation field is used to dynamically correct the geometric parameters of the display device to generate a display mapping relationship that conforms to the physical space topology;

[0131] To dynamically adjust the display device's geometric parameters, this step constructs a differentiable rendering model based on neural radiation fields, converting the resource allocation parameters generated in step S3 into a geometric mapping of the physical screen. Specifically, this involves first establishing an implicit neural representation of the screen surface, then integrating biophysical features to optimize the radiation field parameters, ultimately generating a display mapping matrix that conforms to spatial continuity.

[0132] The neural radiation field model is implemented by a multi-layer perceptron (MLP), and the network input contains spatial coordinates and sight direction , the output is the volume density and color values Preferably, the network structure contains 8 hidden layers, the number of neurons in each layer decays exponentially with depth, and the activation function uses a combination of ReLU and Softplus:

[0133] ;

[0134] in For the Layer hidden features, , and , is a trainable parameter. When accumulating colors along the light sampling points, the layered volume rendering formula is used:

[0135] ;

[0136] The transmittance and body density and sampling interval satisfy:

[0137] ;

[0138] To integrate biophysical features, define the parameter update rate and physiological signals Nonlinear relationship (such as pupil diameter change rate):

[0139] ;

[0140] in is the baseline learning rate, A small constant to prevent division by zero. When the pupil diameter exceeds the threshold, the sampling point density is automatically increased. , preferably by Dynamic adjustment, is the baseline value of physiological signals.

[0141] The screen surface differential constraint is introduced to ensure that the surface after geometric correction satisfies physical continuity. The curvature constraint loss function is defined as:

[0142] ;

[0143] in is the implicit function of the screen surface, Represents the second-order derivative operator of the surface. This constraint is related to the rendering loss Joint Optimization:

[0144] ;

[0145] in The weight coefficient is determined by grid search on the validation set. The optimization process uses the Adam optimizer, and the momentum parameter , Set to 0.9 and 0.99.

[0146] The dynamic correction process is executed in the following sequence: First, the resource allocation parameters output by step S3 are loaded and the neural radiation field model parameters are initialized. Then, real-time physiological signals are collected and the adaptive learning rate and sampling density are calculated. In each iteration, a set of light rays is randomly sampled from the screen surface and the color accumulation value is calculated. The network parameters are updated by backpropagation. After every 50 iterations, the mean square error of the curvature constraint term is evaluated. If it exceeds the threshold, the error is increased. value.

[0147] In terms of implementation details, the screen surface implicit function The SDF network is represented by a signed distance function (SDF), whose zero-equal-value plane corresponds to the physical screen surface. The SDF network shares underlying features with the radiance field network, and parameter interaction is achieved through a cross-attention mechanism. Model inference is preferably deployed on an NPU accelerator, and the OpenGL ES 3.2 interface is used to generate the geometric mapping matrix during real-time rendering.

[0148] Geometric mapping matrix Each element of stores pixel coordinate transformation parameters, including translation vectors and rotation quaternions. When updating the matrix, regions where the curvature change rate exceeds a threshold are prioritized, and queries are accelerated using a spatial hash table. External parameter calibration uses a checkerboard-assisted offline calibration method to establish a rigid transformation relationship between the screen coordinate system and the global reference.

[0149] S5. Based on the displayed mapping relationship, the multimodal intent recognition model is updated using a federated learning mechanism, and the updated model parameters are fed back to the tensor decomposition process of step S2;

[0150] To achieve privacy protection and dynamic adaptation of model updates, this step constructs a distributed training framework based on federated learning. This framework uses the display mapping generated in step S4 as a supervisory signal to iteratively optimize the multimodal intent recognition model. Specifically, this involves first performing local training on the device and injecting adaptive noise. Next, the edge server aggregates the gradients and adjusts the learning strategy. Finally, the encrypted parameters are transmitted back to the tensor decomposition module in step S2.

[0151] The local training process adopts a differential privacy protection mechanism and defines the gradient perturbation equation as:

[0152] ;

[0153] in For the The device is in The original gradient of the training round, is the baseline noise intensity, is the gradient norm sensitivity parameter. The noise standard deviation is adaptively adjusted with the gradient amplitude: when When the noise intensity decays exponentially Zoom out:

[0154] ;

[0155] Preferably, the attenuation coefficient Set to the current training round The inverse of , ensures that the noise intensity approaches zero in the later stages of training. The sensitivity parameter ϵϵ is dynamically calibrated through the gradient statistical histogram and updated every 5 training rounds.

[0156] When the edge server aggregates the obfuscated gradients, it uses a weighted average strategy with momentum acceleration:

[0157] ;

[0158] in For devices The data weight, represents the number of local samples, is the momentum factor. Adaptive adjustment based on the variance of the aggregated gradient:

[0159] ;

[0160] The learning rate scheduling adopts the periodic cosine annealing strategy, defining the The round learning rate is:

[0161] ;

[0162] in The annealing period is preferably set to 1 / 5 of the total training rounds. When the loss function plateau is detected to exceed 3 periods, the learning rate is automatically reset to To escape from local optimum.

[0163] Updated model parameters The parameters are encrypted and transmitted to the tensor decomposition module via AES-256 to reconstruct the core dimension of the behavior tensor. The parameter injection process is achieved through tensor rank correction, and the update equation of the core tensor GG is defined as:

[0164] ;

[0165] in is the model parameter In the The projection vector of the mode, is the injection intensity coefficient. The projection vector is obtained by singular value decomposition:

[0166] ;

[0167] Preferably, retain the principal components, The value is determined based on the cumulative contribution rate of the characteristic value ≥ 95%. Injection intensity and the corresponding singular values The scaling factor decays linearly with the number of training rounds.

[0168] In terms of implementation details, local training adopts an asynchronous parallel architecture, and each device performs forward inference and backpropagation independently. The gradient aggregation server is deployed on the edge node and receives encrypted gradients through the gRPC protocol. The parameter injection module and the tensor decomposer share memory space and achieve zero-copy data transmission through the DMA channel. The privacy protection strength is achieved through -Differential privacy budget control, privacy loss is calculated in each round of training:

[0169] ;

[0170] in For the Wheel noise standard deviation, is the failure probability. When the preset threshold is exceeded, the training process is automatically terminated and the model reset mechanism is triggered.

[0171] S6. Based on the calibrated display parameters and resource allocation plan, drive the multi-screen system to perform dynamic rendering and content output.

[0172] To achieve dynamic rendering control for multi-screen systems, this step constructs a virtual pixel resource pool and implements environment-aware tone mapping, converting the geometric mapping generated in step S4 and the updated model parameters in step S5 into actual rendering instructions. The specific implementation first establishes a dynamic binding relationship between virtual pixels and the physical screen, then integrates the ambient light field and biometric features to perform tone compression, ultimately ensuring output stability through brightness constraints.

[0173] The construction of the virtual pixel resource pool is based on the resource allocation parameters of step S3, defining the virtual pixel ,in The alpha channel blending weight of each virtual pixel is the candidate physical screen set. Optimize by:

[0174] ;

[0175] in For virtual pixels With physical screen The spacing, is the network transmission delay, is the scaling factor. The exponentially weighted moving average of the historical time delay series is calculated by sliding the window to obtain:

[0176] ;

[0177] in is the current measurement delay, Preferably, the delay window length is set to 3 times the network round trip time to balance the real-time and stability requirements.

[0178] Ambient-aware dynamic tone compression based on ambient light field intensity The user's pupil diameter Joint Adjustment. Define the high dynamic range compression curve:

[0179] ;

[0180] in is the mean ambient light intensity, is the compression strength coefficient. Adaptive adjustment based on biometrics:

[0181] ;

[0182] in is the reference pupil diameter, To adjust the sensitivity parameters. When the pupil diameter is detected to shrink rapidly, it will automatically increase value to enhance shadow details.

[0183] Physical screen brightness conservation constraints are implemented through normalization to ensure that the total output brightness does not exceed the hardware threshold . Define the normalization factor:

[0184] ;

[0185] in For the The original brightness of the screen, A small constant to prevent division by zero. The final output brightness is scaled by:

[0186] ;

[0187] Normalization factor Dynamically updated in the rendering pipeline, when a sudden change in ambient light intensity exceeding a threshold is detected, an emergency brightness limit mechanism is triggered. This mechanism directly writes the brightness cap value through hardware registers, bypassing software layer processing delays.

[0188] The implementation process follows the following sequence: First, the virtual pixel resource pool configuration is loaded, and pixel binding relationships are established based on the geometric mapping matrix from step S4. Next, ambient light sensor data and pupil diameter from the eye tracking module are collected to calculate dynamic compression parameters. In the rendering pipeline, tone compression and alpha blending are performed in parallel on each virtual pixel to generate an intermediate rendering result. Finally, brightness normalization constraints are applied and the result is output to the physical screen via the DisplayPort 1.4a interface.

[0189] In terms of hardware acceleration, alpha blending is performed by the GPU's texture sampling unit, utilizing bilinear interpolation to optimize spatial continuity. Tone compression curves are pre-calculated using a lookup table (LUT) and stored in video memory, allowing for rapid access via the texture buffer during real-time rendering. Luminance normalization factors are calculated using dedicated circuitry implemented in the FPGA, ensuring sub-microsecond response latency.

[0190] Dynamic mapping updates are triggered when the screen's physical position changes exceed a threshold, network latency fluctuations exceed a set value, or the ambient light intensity gradient reaches a critical value. During the update process, a double buffering mechanism is used to prevent rendering tearing, and the new and old mapping matrices are switched atomically.

[0191] Please see the attached Figure 2 The present invention also provides a multi-screen dynamic display control device, comprising:

[0192] A multi-source data synchronization acquisition module is used to perform spatiotemporal alignment processing of user behavior data and environmental parameters through a heterogeneous sensor array;

[0193] Spatiotemporal tensor modeling module, which is used to construct and decompose seventh-order behavioral tensor models based on the hybrid tensor decomposition algorithm;

[0194] Quantum-classical collaborative optimization module, used to solve screen resource allocation parameters with time and space constraints;

[0195] A biophysical coupling correction module is used to dynamically generate a geometric mapping relationship of the display device based on the neural radiation field;

[0196] A private gradient aggregation module for implementing adaptive noise injection and gradient obfuscation in federated learning models;

[0197] Dynamic resource rendering engine module, used to perform elastic pixel mapping and dynamic tone rendering of multi-screen systems.

[0198] The present invention also provides an electronic device, comprising: a processor; and a memory storing a computer program, which executes the above method when processing the computer program.

[0199] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A multi-screen dynamic display control method, characterized in that: The following steps are involved: S1. Synchronously collect the user's multimodal behavior data and environmental parameters, and generate multi-source perception data with a unified benchmark through spatiotemporal alignment. S2. Constructing a seventh-order behavior tensor model including spatiotemporal correlation features based on the multi-source perception data, and performing hybrid tensor decomposition to extract user intention features; S3. Solving dynamic resource allocation parameters among multiple screens using a quantum-classical hybrid optimization algorithm based on characteristic parameters output by the behavior tensor model; S4. Based on the resource allocation parameters, dynamically correct the geometric parameters of the display device using the neural radiation field to generate a display mapping relationship that conforms to the physical space topology; S5. Based on the displayed mapping relationship, a federated learning mechanism is used to update the multimodal intent recognition model, and the updated model parameters are fed back to the tensor decomposition process of step S2; S6. Based on the calibrated display parameters and resource allocation plan, drive the multi-screen system to perform dynamic rendering and content output; The quantum optimization process in step S3 constructs the following Hamiltonian: in: represents the association strength between screens i and j, which is determined by the Frobenius norm of the corresponding dimension of the behavior tensor; τ ij is the physical screen distance; ΔT is the dynamic delay threshold, which is adaptively adjusted according to the network load rate; is the Pauli operator of the i-th quantum bit; The step S4 comprises the following steps: A neural radiation field model is established to generate screen pixel colors through cumulative transmittance calculation, where transmittance is related to the optical properties of the screen material; Dynamically adjust the radiation field parameter update rate based on the user's physiological characteristic signals collected in real time, wherein the physiological characteristic signals include at least one of pupil diameter and heart rate variability; Screen surface differential constraints are introduced to ensure that the geometric correction process meets the continuity conditions of physical space.

2. The multi-screen dynamic display control method according to claim 1, characterized in that: The step S1 comprises the following steps: The inertial measurement unit captures the device motion posture quaternion at the sampling rate, the ToF depth camera generates a high-resolution depth map, and the eye tracking module measures the pupil diameter change rate in real time; The IEEE1588v2 protocol is used to achieve multi-sensor clock synchronization and construct a spatial transformation matrix chain including translation vectors and rotation matrices.

3. The multi-screen dynamic display control method according to claim 1, characterized in that: The mixed tensor decomposition in step S2 satisfies the following formula definition: CP decomposition term: Among them, λ r is the rth eigenvalue, represents the rth factor vector of the kth mode, Represents vector outer product operation; Tucker decomposition terms: in, is the core tensor; is the factor matrix of the kth mode, and ×k represents the kth mode product of the tensor and the matrix.

4. The multi-screen dynamic display control method according to claim 1, characterized in that: The step S5 comprises the following steps: During on-device training, adaptive noise related to the gradient norm nonlinearity is injected into the model gradient, where the noise intensity decays exponentially with the increase of the gradient amplitude; When the edge server aggregates the obfuscated gradients, a periodic decay strategy is used to dynamically adjust the learning rate. The learning rate is kept high in the early stages of training to accelerate convergence, and is gradually reduced in the later stages to improve model stability. The updated model parameters are encrypted and transmitted to the tensor decomposition of step S2 for reconstructing the core dimension of the behavior tensor.

5. The multi-screen dynamic display control method according to claim 1, characterized in that: The step S6 comprises the following steps: Build a virtual pixel resource pool and define the dynamic mapping relationship between virtual pixels and multiple physical screens. The alpha channel blending weight of each virtual pixel is jointly optimized based on the physical screen spacing and network latency. Performs environment-aware dynamic tone compression, jointly adjusting the HDR compression curve based on the ambient light field intensity distribution and user biometric signals, where the compression strength is positively correlated with the mean ambient light intensity; Implement physical screen brightness conservation constraints and ensure that the total brightness output of each screen pixel does not exceed a preset threshold through normalization.

6. The multi-screen dynamic display control method according to claim 1, characterized in that: The J ij The calculation method is: Where T int is the adjustable integration time window, Represents a slice of a tensor in the screen dimension.

7. A multi-screen dynamic display control device, according to the multi-screen dynamic display control method according to any one of claims 1 to 6, characterized in that: include: A multi-source data synchronization acquisition module is used to perform spatiotemporal alignment processing of user behavior data and environmental parameters through a heterogeneous sensor array; Spatiotemporal tensor modeling module, which is used to construct and decompose seventh-order behavioral tensor models based on the hybrid tensor decomposition algorithm; Quantum-classical collaborative optimization module, used to solve screen resource allocation parameters with time and space constraints; A biophysical coupling correction module is used to dynamically generate a geometric mapping relationship of the display device based on the neural radiation field; A private gradient aggregation module for implementing adaptive noise injection and gradient obfuscation in federated learning models; Dynamic resource rendering engine module, used to perform elastic pixel mapping and dynamic tone rendering of multi-screen systems.

8. An electronic device, characterized in that: include: processor; a memory storing a computer program; When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images