Multi-channel signal modulation method for organic display interface

Abstract

The invention relates to the technical field of organic display data processing, in particular to an organic display interface multichannel signal modulation method, which comprises the following steps of: generating a standardized data set through gamma correction and gamut mapping, and inputting the standardized data set into time sequence, amplitude and frequency sub-models deployed in a federal learning framework; and respectively analyzing the pulse width-refresh rate relevance, the driving voltage-brightness nonlinear relationship and the conduction period-mobility dynamic response, generating an optimized weight coefficient, and fusing the optimized weight coefficient into a boundary constraint condition of a cross-sub-model. A migration rate parameter index, amplitude gradient calculation and duty ratio correction service of chain calling is constructed based on a service grid, a reinforcement learning dynamic routing strategy is combined to avoid resource competition nodes, and brightness error characteristics are periodically fed back to reversely optimize gradient updating. According to the method, the conflict between time sequence deviation accumulation and local optimum is effectively suppressed, the color gradation continuity, the dynamic range and the edge sharpness are improved, and efficient cooperative modulation of multi-channel signals is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of organic display data processing, and particularly to a multi-channel signal modulation method for an organic display interface. Background Art

[0002] Multi-channel signal modulation for an organic display interface is an optimization method for display technology based on organic semiconductor materials. Its core lies in dynamically regulating the timing, amplitude, and frequency parameters of different signal channels to achieve precise control over the electroluminescence characteristics of display units. Analyzing from the technical principle, an organic display interface usually consists of multiple independently addressable pixel units, and the light output characteristics of each unit are affected by multi-dimensional signals such as driving voltage, current pulse width, and refresh rate. By introducing a multi-channel modulation algorithm, the system can decouple and reconstruct the driving signals of each pixel in real time according to the gray level, color gamut, and dynamic range requirements of the input image data, thereby improving the display contrast and color restoration while reducing the overall energy consumption. For example, in the case of a high dynamic range scene, the modulation circuit can coordinate the conduction periods of different sub-pixels through a time-division multiplexing strategy, combined with the non-linear response characteristics of the carrier mobility of organic materials, to effectively optimize the edge sharpness and color level continuity during the brightness gradient process.

[0003] The core technical pain point faced by multi-channel signal modulation for an organic display interface in the data processing field is the insufficient efficiency of dynamic collaborative optimization of multi-dimensional signal parameters. The specific manifestations are as follows: During the real-time decoupling process of the gray level, color gamut, and dynamic range of the input image data, due to the coupled non-linear characteristics of the timing, amplitude, and frequency parameters, the traditional single-channel linear driving model is difficult to accurately quantify the mutual modulation effects between multi-channel signals. For example, in a high dynamic range scene, the time-division multiplexing of the sub-pixel conduction period needs to synchronously compensate for the voltage-brightness response delay of the carrier mobility of organic materials, and the existing algorithms are prone to cumulative timing deviation and resource competition due to the lack of a joint optimization architecture for cross-channel non-linear compensation. When reconstructing the driving signal, the pulse width and refresh rate parameters of discrete modulation need to meet the sub-pixel level phase synchronization constraints, but the independent optimization of multi-dimensional parameters will cause conflicts of local optimal solutions, thereby reducing the edge sharpness and color level continuity. This pain point limits the data throughput efficiency and dynamic range expansion ability of high-resolution organic display interfaces in low-power consumption scenarios. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a multi-channel signal modulation method for an organic display interface, which solves the problems of cumulative timing deviation, local optimal solution conflict caused by insufficient efficiency of dynamic collaborative optimization of multi-dimensional signal parameters in the organic display interface, and the resulting problems of decreased display edge sharpness, deteriorated color level continuity, and limited dynamic range.

[0005] To solve the above technical problems, the specific technical solution of the present invention is as follows: The multi-channel signal modulation method for an organic display interface provided by the present invention includes: Obtain the grayscale, color gamut, and dynamic range data of the input image, perform gamma correction and color gamut mapping processing on the grayscale, color gamut, and dynamic range data to generate a standardized data set; Input the standardized data set into the federated learning framework, analyze the correlation characteristics of pulse width and refresh rate through the time series sub-model deployed in the federated learning framework, model the non-linear relationship between driving voltage and brightness through the amplitude sub-model, and analyze the dynamic response of conduction period and mobility through the frequency sub-model to generate optimized weight coefficients for pulse width, refresh rate, and driving voltage; Based on the optimized weight coefficients generated by the federated learning framework, construct a cross-channel compensation strategy in the service mesh, adjust the mobility compensation parameters through the service indexed by the mobility parameter table, update the voltage amplitude correction coefficient through the amplitude gradient calculation service, and reconstruct the pulse timing parameters through the duty cycle correction service to generate a compensated driving signal component; Perform quantization time window segmentation on the compensated driving signal component, trigger the dynamic scaling of microservice instances according to the calculation result of the phase error gradient, and allocate computing resources by combining the time window priority scoring table generated by the edge sharpness index to generate a synchronized and optimized time window scheduling scheme; Based on the synchronized and optimized time window scheduling scheme, perform multi-modal dynamic range expansion through heterogeneous computing nodes, federate and knowledge distill the optimized results of the frequency division strategy generated by the GPU cluster and the optimized physical characteristic data output by the FPGA node to generate a global driving signal and feedback it to the transfer learning model loader of the data acquisition module to form an optimized link with a closed loop of data flow and control flow.

[0006] Further, in the multi-channel signal modulation method for an organic display interface of the present invention, generating a standardized data set includes: Associate the dynamic range data with the temperature parameter collected by the environmental temperature sensor through the timestamp binding module to generate raw data with environmental labels; Parallelly clean the raw data with environmental labels using a containerized microservice architecture, and dynamically scale the number of container instances based on the input image resolution according to the Kubernetes engine; Obtain the electroluminescence response model matching the current organic semiconductor material from the cloud through the transfer learning model dynamic loader, and feedback the feature extraction result output by the pruned electroluminescence response model to the gamma correction module of the data preprocessing pipeline.

[0007] Further, in the multi-channel signal modulation method for an organic display interface of the present invention, the multi-dimensional parameter decoupling in the federated learning framework includes: Deploy the timing sub-model as an independent container, and receive the correlation features of pulse width and refresh rate from the standardized data set through the gRPC protocol; Build a gradient descent optimizer for the driving voltage-luminance non-linear curve in the amplitude sub-model, and use the asynchronous update mechanism to temporarily store the local gradient in the distributed cache, waiting for the gradient synchronization of the timing sub-model and the frequency sub-model; Through the model distillation module, perform attention fusion on the on-time period features of the frequency sub-model and the pulse width features of the timing sub-model, generate cross-sub-model boundary constraint conditions, and input them into the duty cycle correction service interface of the cross-channel compensation strategy.

[0008] Furthermore, for the multi-channel signal modulation method of the organic display interface described in the present invention, constructing the cross-channel compensation strategy includes: Define the chained call order of the mobility parameter table index service, the amplitude gradient calculation service, and the duty cycle correction service in the service mesh, where the duty cycle correction service receives the boundary constraint conditions from the federated learning framework; Through the reinforcement learning routing policy trainer, analyze the real-time load status of the service mesh and the gradient synchronization status of the federated learning sub-model in real time, and generate a dynamic routing priority list to bypass resource competition nodes; Persistently store the luminance error data generated during the compensation process in the service mesh sidecar database, extract periodic error features through the time series analysis engine, and generate a feedback signal to reversely update the gradient descent optimizer of the amplitude sub-model of the federated learning framework.

[0009] Furthermore, for the multi-channel signal modulation method of the organic display interface described in the present invention, the quantization time window segmentation includes: Use qubit encoding to discretize the signal period of the compensated drive signal component into a superposition state time slice combination, and screen the optimal segmentation scheme through the trigger signal of the hardware counter; trigger the elastic scaling of the microservice instance based on the output threshold of the phase error gradient calculation module, and register the new instance in the task assignment queue of the service orchestration engine; Parse the edge sharpness index from the image processing module through the rule engine, and generate a dynamic priority scoring table to drive the time slice task to allocate short time slices in the high sharpness area and merge time slices in the flat area.

[0010] Furthermore, for the multi-channel signal modulation method of the organic display interface described in the present invention, the multi-modal dynamic range expansion includes: Through the heterogeneous computing task dispatcher, allocate the sub-signal components generated by the quantization time window segmentation to the GPU cluster to execute the frequency division strategy optimization, and the FPGA node to execute the physical property optimization calculation based on the simulated annealing algorithm; Construct a cross-architecture feature alignment layer to convert the physically optimized features output by the FPGA nodes into a tensor format compatible with the GPU cluster neural network; Adopt a lightweight encryption aggregation protocol to perform consistency verification on the frequency division strategy optimization results of the GPU cluster and the physical characteristic optimization data of the FPGA nodes, and trigger a sub-signal component weight re-optimization process across nodes when a color level jump anomaly is detected.

[0011] Furthermore, in the multi-channel signal modulation method for an organic display interface according to the present invention, the data cleaning of the containerized microservice architecture includes: Embed a lightweight message broker in the container instance to achieve high-frame-rate and low-latency transmission of raw data with environment tags; Compress the volume of the electroluminescence response model through model pruning technology to adapt to the storage limitations of edge computing nodes; Push the preprocessed standardized data output by the gamma correction module to the message queue to trigger the feature vector update and weight feedback of the transfer learning model dynamic loader.

[0012] Furthermore, in the multi-channel signal modulation method for an organic display interface according to the present invention, the attention fusion of the model distillation module includes: Capture the implicit relationship between the pulse width correlation of the timing sub-model and the on-time period migration rate characteristics of the frequency sub-model through the self-attention mechanism; Use the fused feature vector as the global boundary condition and input it into the interface of the duty cycle correction service; Jointly optimize the compensation coefficient matrix output by the distillation model and the real-time load status of the service mesh to generate a drive signal reconstruction instruction.

[0013] Furthermore, in the multi-channel signal modulation method for an organic display interface according to the present invention, the feedback signal update of the service mesh includes: Extract the time series data of the brightness error from the sidecar database of the service mesh, and identify periodic resource competition events through the sliding window algorithm; Encode the recognition result into a gradient update suggestion and push it to the asynchronous optimizer of the amplitude sub-model; Dynamically adjust the local training rounds of the federated learning framework according to the feedback signal, and synchronously update the pulse width constraint conditions of the timing sub-model.

[0014] Furthermore, in the multi-channel signal modulation method for an organic display interface according to the present invention, the generation of the dynamic priority score table includes: Parse the edge sharpness index generated by the quantization time window segmentation through the rule engine to generate a short time slice task priority score for the high sharpness region; Merging time slices in a flat area and reducing the computational density will reallocate the released computational resources to the contour optimization microservice instances; Jointly analyze the score table with the output data of the phase error gradient calculation module, and dynamically adjust the task weight allocation parameters of the middleware orchestration engine.

[0015] Advantages of the present invention; The present invention realizes dynamic collaborative decoupling of multi-dimensional signal parameters through a federated learning framework. The time sequence, amplitude, and frequency sub-models respectively analyze the pulse width-refresh rate correlation, voltage-brightness non-linear relationship, and on-time-mobility dynamic response, generate optimized weight coefficients and input them into the model distillation module for cross-model attention fusion, generate boundary constraint conditions to reconstruct the duty cycle parameters, and effectively suppress the accumulation of time sequence deviation and the conflict of local optimal solutions; the cross-channel compensation strategy of chained calls in the service mesh combined with reinforcement learning dynamic routing avoids resource competition nodes and periodically feeds back the brightness error characteristics to the federated learning framework, and reversely optimizes the gradient update strategy to maintain color scale continuity; quantization time window segmentation combined with phase error gradient drives microservice elastic scaling and edge sharpness priority scoring, and through the frequency division strategy optimization and physical characteristics co-optimization of heterogeneous computing nodes, realizes dynamic range expansion and edge sharpness improvement, forming a closed-loop optimization link of data cleaning, parameter decoupling, compensation reconstruction, and feedback calibration, and significantly improving the collaborative efficiency and display quality of multi-channel signal modulation of the organic display interface. Description of the drawings

[0016] In order to more clearly illustrate the technical solutions of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the drawings.

[0017] Figure 1 It is a flowchart of the multi-channel signal modulation method for an organic display interface provided by an embodiment of the present invention. Detailed implementation manners

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions of the present invention in conjunction with the specific embodiments and corresponding drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. The following will, in conjunction with the drawings, detail the technical solutions provided by each embodiment of the present invention. To better understand the objectives of the present invention, the following will give a more detailed description of the present invention.

[0019] Please refer to Figure 1, the multi-channel signal modulation method for an organic display interface provided by the present invention includes: Step S101: Obtain the grayscale, color gamut, and dynamic range data of the input image, perform gamma correction and color gamut mapping on the grayscale, color gamut, and dynamic range data to generate a standardized data set; When obtaining the grayscale, color gamut, and dynamic range data of the input image, synchronously collect the temperature parameters output by the environmental temperature sensor through the timestamp binding module, perform millisecond-level time alignment on the dynamic range data frame and the temperature sampling value to generate the original data with environmental labels. The timestamp binding module adopts a hardware clock synchronization mechanism to phase-lock the sampling periods of the image acquisition unit and the temperature sensor, avoiding the timing mismatch between environmental parameters and image data. The original data with environmental labels is transmitted to the containerized microservice architecture through the message bus to trigger the parallel data cleaning process.

[0020] When performing gamma correction on the original data with environmental labels using the containerized microservice architecture, based on the Kubernetes engine, dynamically scale the number of container instances according to the input image resolution. Horizontally expand the container instances in high-resolution scenarios to match the column parallel processing requirements of the pixel matrix. Each container instance deploys a lightweight data verification algorithm to eliminate noise and interpolate and repair abnormal pixel values in the original data, and perform color gamut boundary verification. Realize the intermediate data exchange between containers through shared memory to improve the throughput efficiency of high-frame-rate data.

[0021] During the color gamut mapping process, the transfer learning model dynamic loader obtains the electroluminescence response model matching the current organic semiconductor material from the cloud, and uses the channel pruning technology to remove the redundant feature extraction layers in the model, and retains the convolution kernels strongly related to the carrier mobility and color gamut coverage. The feature vector output by the pruned model is transmitted back to the gamma correction module of the data preprocessing pipeline through the feedback bus, and the slope of the correction curve is driven to be dynamically adjusted according to the non-linear response characteristics of the organic material. The feature feedback mechanism forms a closed-loop adjustment of data cleaning and model optimization, generating a standardized data set adapted to the physical characteristics of the display interface.

[0022] Step S102: Input the standardized data set into the federated learning framework, and analyze the correlation characteristics between the pulse width and the refresh rate through the timing sub-model deployed in the federated learning framework, model the non-linear relationship between the driving voltage and the brightness through the amplitude sub-model, and analyze the dynamic response of the conduction period and the mobility through the frequency sub-model to generate the optimized weight coefficients of the pulse width, refresh rate, and driving voltage; After the standardized data set is input into the federated learning framework, the timing sub-model is deployed as an independent container, and receives the pulse width and refresh rate correlation feature sequence from the data preprocessing pipeline through the gRPC protocol. The gRPC protocol is configured with a bidirectional stream transmission mode to push the feature sequence to the circular buffer in real time for the timing sub-model to perform dynamic time warping analysis, parse the phase synchronization constraints of the signal timing parameters and the refresh rate fluctuation law, and generate the pulse width adjustment coefficient. The amplitude sub-model loads the calibration data of the voltage-luminance response curve of the organic semiconductor material, constructs a piecewise linear interpolation model, calculates the voltage amplitude correction gradient through an asynchronous gradient descent optimizer, and temporarily stores the local gradient in the Redis distributed cache cluster. The cache key value includes the time stamp and the sub-model version identifier. When the frequency sub-model analyzes the dynamic response of the conduction period and the mobility, it adopts a sliding window mechanism to extract the mobility fluctuation characteristics, separates the high-frequency noise components through Fourier transform, and generates the conduction period optimization parameters. The model distillation module fuses the pulse width features of the timing sub-model and the mobility features of the frequency sub-model through a multi-head self-attention mechanism, calculates the similarity matrix in the attention space to generate a weight distribution map, and drives the weighted fusion of the feature vectors to output the boundary constraint condition matrix across sub-models. The matrix is pushed to the duty cycle correction service interface through the service mesh API gateway, and combined with the gradient aggregation result of the amplitude sub-model to generate the optimization weight coefficients of the pulse width, refresh rate and driving voltage, completing the collaborative decoupling and joint optimization of multi-dimensional signal parameters.

[0023] Step S103, based on the optimization weight coefficients generated by the federated learning framework, construct a cross-channel compensation strategy in the service mesh, and adjust the mobility compensation parameters through the mobility parameter table index service, update the voltage amplitude correction coefficient through the amplitude gradient calculation service, and reconstruct the pulse timing parameters through the duty cycle correction service to generate the compensated driving signal component; Deploy the mobility parameter table index service in the service mesh based on the optimization weight coefficients generated by the federated learning framework, and obtain the carrier mobility compensation coefficient table matching the optimization weight by querying the preset material property database. The index service uses a hash mapping algorithm to accelerate the query response, and transmits the compensation coefficient to the amplitude gradient calculation service through the gRPC streaming interface, and calculates the voltage correction gradient in real time in combination with the current driving voltage amplitude. During the calculation process, a sliding window mechanism is used to filter high-frequency noise, generate gradient correction parameters and input them into the duty cycle correction service.

[0024] After receiving the gradient correction parameter, the duty cycle correction service analyzes the timing phase limit and amplitude fluctuation threshold in the boundary constraint condition matrix output by the federated learning framework, and reconstructs the pulse timing waveform parameter. An instruction set for pulse width modulation is generated through a hardware description language. Combining with the historical error data stored in the service mesh sidecar database, the update step size of the duty cycle parameter is dynamically adjusted. The timing deviation data generated during the reconstruction process is collected through the service mesh control plane and persistently stored in the circular buffer of the sidecar database.

[0025] The reinforcement learning routing policy trainer of the service mesh monitors the execution delay and resource occupancy rate of the duty cycle correction service in real time. Combining with the gradient synchronization status identifier of the federated learning sub-model, a dynamic routing priority list is generated. When it is detected that the node queue depth exceeds the threshold, the compensation request is automatically routed to the low-load mirror node to reduce the timing jitter caused by resource competition. The corrected drive signal component is output after multi-channel verification. Its brightness error data extracts periodic features through a time series analysis engine, and reversely updates the gradient descent optimizer of the amplitude sub-model to form an adaptive closed-loop adjustment mechanism for compensation parameters.

[0026] Step S104: Perform quantization time window segmentation on the compensated drive signal component, trigger the dynamic scaling of the microservice instance according to the calculation result of the phase error gradient, allocate computing resources by combining with the time window priority scoring table generated by the edge sharpness index, and generate a synchronized and optimized time window scheduling plan; When performing quantization time window segmentation on the compensated drive signal component, the quantum bit encoding technology is used to discretize the signal period into a superposition state time slice combination, and the optimal segmentation plan that meets the phase synchronization constraint is screened through the trigger signal of the hardware counter. The hardware counter is configured with a noise suppression circuit to eliminate clock jitter interference. The selected time slice interval parameter is transmitted to the microservice orchestration engine through the service bus to trigger the resource scheduling process. The phase error gradient calculation module monitors the change rate of the phase deviation during the segmentation process in real time. When the gradient value exceeds the dynamic threshold range, a microservice instance expansion request is sent to the Kubernetes cluster, and the threshold range is dynamically adjusted according to the thermal stability parameter of the organic material.

[0027] The newly added microservice instance is automatically registered in the task assignment queue of the orchestration engine through the service discovery mechanism. The registration information includes the computing resource capacity and task processing delay index of the instance. The task assignment queue uses the weighted round-robin algorithm to preferentially distribute the time slice tasks in the high phase error gradient area to the newly added instance to accelerate the convergence of the phase error. At the same time, the rule engine analyzes the edge sharpness index from the image processing module, maps the sharpness value of the image contour area to the time slice priority parameter, and generates a dynamic scoring table based on the fuzzy logic rule set.

[0028] The dynamic priority scoring table drives the service orchestration engine to allocate millisecond-level short time slice tasks in high sharpness regions to improve edge resolution, and merge adjacent time slices in flat regions to form batch processing units. The merged task units reduce the computational density through sparse matrix operations, and the released resources are reallocated to the contour optimization microservice instance cluster. The scoring table is jointly analyzed with the phase error gradient data. When the phase error in the high sharpness region exceeds the limit, the service orchestration engine is triggered to dynamically adjust the task weight allocation parameters, achieving a coordinated balance between edge sharpness optimization and timing synchronization, and finally generating a synchronized and optimized time window scheduling scheme.

[0029] Step S105, based on the synchronized and optimized time window scheduling scheme, execute multi-modal dynamic range expansion through heterogeneous computing nodes, federally aggregate and knowledge distill the frequency division strategy optimization results generated by the GPU cluster and the physical characteristic optimization data output by the FPGA node, generate a global driving signal and feedback it to the transfer learning model loader of the data acquisition module, forming an optimized link for the closed-loop of data flow and control flow.

[0030] Based on the synchronized and optimized time window scheduling scheme, the heterogeneous computing task dispatcher distributes sub-signal components to the GPU cluster through the PCIe high-speed bus to execute frequency division strategy optimization, loads a convolutional neural network model to extract frequency domain features and perform non-linear mapping; at the same time, it distributes the physical characteristic optimization task to the FPGA node, and iteratively searches for the optimal driving voltage waveform based on the simulated annealing algorithm. The task distribution instruction contains a time slice index and optimization constraint parameters, achieving load balancing and parallel computing between the GPU and FPGA nodes.

[0031] When constructing the cross-architecture feature alignment layer, deploy a hardware description language conversion interface at the output end of the FPGA node to convert the physical optimization features from the fixed-point number format to the floating-point tensor structure. The interface integrates a bit width matching module to perform dynamic range scaling and sign bit extension on the data output by the FPGA, generating a tensor data stream that matches the neural network input dimension of the GPU cluster. The converted tensor is injected into the GPU batch processing queue through the shared memory channel, and is cascaded with the frequency division strategy optimization result at the feature level, forming a joint input source for multi-modal optimization data.

[0032] When using a lightweight encryption aggregation protocol to perform consistency verification on heterogeneous computing results, an encrypted transmission channel based on elliptic curve cryptography is established. The aggregation server performs color level continuity analysis on the GPU frequency division strategy results and FPGA physical optimization data. When it is detected that the chromaticity jump in adjacent time slices exceeds the gamut tolerance threshold of the organic material, a cross-node re-optimization process is triggered, and an instruction containing the abnormal time slice index and weight correction coefficient is sent to the task dispatcher. The instruction drives the GPU and FPGA to perform collaborative re-optimization on the specified time slice until the chromaticity jump converges to a safe range. The finally generated global drive signal integrates multi-modal optimization results through federated aggregation and knowledge distillation, and feeds back to the transfer learning model loader of the data acquisition module, driving the update of the model pruning and quantization-aware training processes, forming a closed-loop optimization link of data cleaning, parameter decoupling, compensation reconstruction, and feedback calibration.

[0033] The multi-channel signal modulation method for the organic display interface provided by the present invention first associates the dynamic range data with the temperature parameters collected by the environmental temperature sensor through the timestamp binding module to generate the original data with environmental tags. The containerized microservice architecture is used to perform parallel cleaning on the original data, and based on the Kubernetes engine, the number of container instances is dynamically scaled according to the input image resolution to match the data processing requirements in different resolution scenarios. The electro-luminescence response model matching the current organic semiconductor material is obtained from the cloud through the transfer learning model dynamic loader, and the model pruning technology is used to compress the model volume to adapt to the storage limitations of the edge computing nodes, and the feature extraction results output by the pruned model are fed back to the gamma correction module of the data preprocessing pipeline to form an adaptive data normalization processing process. This step provides a basis for highly consistent input data for subsequent signal modulation.

[0034] After the standardized data set is input into the federated learning framework, the timing sub-model deployed in the framework receives the correlation features of the pulse width and refresh rate through the gRPC protocol, and analyzes the dynamic relationship between the signal timing parameters. The amplitude sub-model constructs a gradient descent optimizer for the drive voltage-brightness non-linear curve, and uses an asynchronous update mechanism to temporarily store the local gradient in the distributed cache, waiting for the gradient synchronization of the timing sub-model and the frequency sub-model, avoiding the problem of resource idleness caused by traditional synchronous updates. The frequency sub-model analyzes the dynamic response of the conduction period and mobility, and performs attention fusion on the pulse width feature of the timing sub-model and the mobility feature of the frequency sub-model through the model distillation module to generate cross-sub-model boundary constraint conditions. The boundary constraint conditions are input into the duty cycle correction service interface of the cross-channel compensation strategy as global optimization parameters to achieve collaborative decoupling and joint optimization of multi-dimensional signal parameters.

[0035] Based on the optimized weight coefficients generated by the federated learning framework, construct the chained call order of the mobility parameter table index service, amplitude gradient calculation service, and duty cycle correction service in the service mesh. The duty cycle correction service receives the boundary constraint conditions from the federated learning framework, combines with the reinforcement learning routing policy trainer to analyze the real-time load status and gradient synchronization status of the service mesh in real time, and generates a dynamic routing priority list to bypass resource competition nodes. The luminance error data generated during the compensation process is persistently stored in the service mesh sidecar database, extracts the periodic error characteristics through the time series analysis engine and generates a feedback signal, and reversely updates the amplitude sub-model gradient descent optimizer of the federated learning framework. This process effectively suppresses the accumulation of timing deviation and the conflict of local optimal solutions by dynamically adjusting the compensation parameters and service call paths.

[0036] For the compensated drive signal components, adopt qubit encoding technology to discretize the signal period into a superposition state time slice combination, and screen the optimal segmentation scheme through the trigger signal of the hardware counter. Based on the output threshold of the phase error gradient calculation module, trigger the elastic scaling mechanism of the microservice instance, and register the new instance to the task assignment queue of the service orchestration engine. Parse the edge sharpness index from the image processing module through the rule engine, generate a dynamic priority scoring table, drive the time slice task to allocate short time slices in the high sharpness area to improve the edge sharpness, and merge the time slices in the flat area to reduce the calculation density. The time window scheduling scheme optimizes the response efficiency of the display interface in complex scenarios through dynamic resource allocation.

[0037] Based on the synchronized and optimized time window scheduling scheme, distribute the sub-signal components to the GPU cluster through the heterogeneous computing task dispatcher to execute the frequency division strategy optimization, and the FPGA node to execute the physical characteristic optimization calculation based on the simulated annealing algorithm. Construct a cross-architecture feature alignment layer to convert the physical optimization features output by the FPGA node into a tensor format compatible with the GPU cluster neural network, and realize data intercommunication between heterogeneous computing nodes. Adopt a lightweight encryption aggregation protocol to perform consistency verification on the frequency division strategy optimization results of the GPU cluster and the physical characteristic optimization data of the FPGA node. When a color level jump anomaly is detected, trigger the sub-signal component weight re-optimization process across nodes. The finally generated global drive signal integrates the multi-modal optimization results through federated aggregation and knowledge distillation, and feeds back to the transfer learning model loader of the data acquisition module, forming an optimized link with a closed loop of data flow and control flow. The closed-loop mechanism systematically improves the dynamic range expansion ability and multi-channel signal cooperative modulation efficiency of the organic display interface by continuously iteratively updating the model parameters and compensation strategies.

[0038] Specifically, the multi-channel signal modulation method for the organic display interface described in the present invention generates a standardized data set including: Associate the dynamic range data with the temperature parameters collected by the environmental temperature sensor through the timestamp binding module to generate the original data with environmental tags; Adopt a containerized microservices architecture to perform parallel cleaning on the original data with environmental tags, and dynamically scale the number of container instances based on the Kubernetes engine according to the input image resolution; Obtain the electroluminescence response model matching the current organic semiconductor material from the cloud through the migration learning model dynamic loader, and feedback the feature extraction results output by the pruned electroluminescence response model to the gamma correction module of the data preprocessing pipeline.

[0039] During the generation of the standardized data set, the dynamic range data is synchronously collected with the temperature parameters obtained by the environmental temperature sensor through the timestamp binding module, and the dynamic range data frame is aligned with the temperature sampling value at the millisecond level to generate the original data with environmental tags. The timestamp binding module adopts a hardware clock synchronization mechanism to ensure that the sampling periods of the image acquisition unit and the temperature sensor are phase-locked, avoiding the timing mismatch between the environmental parameters and the image data. The original data with environmental tags is transmitted to the containerized microservices architecture through the message bus to trigger the data cleaning process.

[0040] When performing parallel cleaning on the original data with environmental tags using a containerized microservices architecture, the resolution change characteristics of the input image are monitored in real time based on the Kubernetes engine. When a high-resolution image input is detected, the number of container instances is automatically horizontally scaled to match the column parallel processing requirements of the pixel matrix, and each container instance deploys a lightweight data verification algorithm to eliminate transmission noise. The parallel cleaning process adopts a pipeline architecture, where the front-stage container performs pixel-level outlier detection, and the rear-stage container performs gamut boundary verification. The intermediate data between containers is exchanged through shared memory to improve the throughput efficiency of high-frame-rate data.

[0041] When accessing the cloud model repository through the migration learning model dynamic loader, retrieve the electroluminescence response model that matches the carrier mobility parameters of the current organic semiconductor material. During the retrieval process, a semantic similarity matching algorithm is adopted to encode the material characteristic parameters into feature vectors and calculate the cosine similarity with the cloud model metadata, and the candidate models with similarity thresholds exceeding the preset value are screened. After the download is completed, the channel pruning technology is used to remove the redundant feature extraction layers in the electroluminescence response model, and the convolution kernels strongly related to the gamut coverage rate of the current display interface are retained. The feature vectors output by the pruned model are transmitted back to the data preprocessing pipeline through the feedback bus, driving the gamma correction module to dynamically adjust the slope of the correction curve to make the preprocessing result adapt to the non-linear response characteristics of the organic material. The feature feedback mechanism forms a closed-loop adjustment of data cleaning and model optimization, improving the adaptation accuracy of the standardized data set to the physical characteristics of the display interface.

[0042] Specifically, in the multi-channel signal modulation method for the organic display interface of the present invention, the multi-dimensional parameter decoupling in the federated learning framework includes: Deploy the timing sub-model as an independent container, and receive the pulse width and refresh rate correlation features from the standardized data set through the gRPC protocol; Build a gradient descent optimizer for the driving voltage-luminance non-linear curve in the amplitude sub-model, and use the asynchronous update mechanism to temporarily store the local gradient in the distributed cache, waiting for the gradient synchronization of the timing sub-model and the frequency sub-model; Through the model distillation module, perform attention fusion on the on-time period features of the frequency sub-model and the pulse width features of the timing sub-model, generate cross-sub-model boundary constraint conditions, and input them into the duty cycle correction service interface of the cross-channel compensation strategy.

[0043] During the multi-dimensional parameter decoupling process in the federated learning framework, when deploying the timing sub-model as an independent container, use Docker containerization technology to encapsulate the model inference engine, and establish a message channel with the standardized data set through the gRPC protocol. The gRPC protocol is configured with a bidirectional stream transmission mode to receive the pulse width and refresh rate correlation feature sequence from the data preprocessing pipeline in real time. The feature sequence is sliced by time window and cached in the circular buffer for the timing sub-model to perform dynamic time warping analysis. The pulse width adjustment coefficient and refresh rate optimization parameter output by the timing sub-model are mapped to the distributed message middleware through shared memory to provide input data sources for the amplitude sub-model and the frequency sub-model.

[0044] When building a gradient descent optimizer for the driving voltage-luminance non-linear curve in the amplitude sub-model, calibrate data of the voltage-luminance response curve of the organic semiconductor material is loaded in the initialization stage, and a basic model of the non-linear relationship is constructed using the piecewise linear interpolation algorithm. During the training process, the gradient descent optimizer runs with the asynchronous update mechanism, and the voltage amplitude gradient calculated locally is temporarily stored in the Redis distributed cache cluster. The cache key value includes the time stamp and the sub-model version identifier. The asynchronous update mechanism sets a gradient synchronization waiting queue. When it is detected that the gradient version identifiers of the timing sub-model and the frequency sub-model are consistent, the gradient aggregation service is triggered to perform parameter update to avoid version conflict problems during the training process of multiple sub-models.

[0045] When the model distillation module fuses the on - time characteristics of the frequency sub - model and the pulse - width characteristics of the timing sub - model, it uses the multi - head self - attention mechanism to extract the implicit correlation across sub - models. After the on - time characteristic tensor output by the frequency sub - model is dimension - reduced through the feature projection layer, the similarity matrix is calculated with the pulse - width characteristics of the timing sub - model in the attention space to generate the attention weight distribution map. The distribution map drives the weighted fusion of feature vectors and outputs the boundary constraint condition matrix across sub - models. The boundary constraint condition matrix is pushed to the duty - cycle correction service interface of the cross - channel compensation strategy through the API gateway of the service mesh. The interface parses the timing phase constraint parameters and amplitude fluctuation thresholds in the matrix to reconstruct the duty - cycle correction coefficient of the pulse timing. The compensation parameter deviation value generated during the reconstruction process is transmitted back to the federated learning framework through the closed - loop feedback channel, triggering the online fine - tuning process of the sub - model to form a collaborative optimization mechanism for parameter decoupling and compensation strategy.

[0046] Specifically, for the multi - channel signal modulation method of the organic display interface described in the present invention, constructing the cross - channel compensation strategy includes: Define the chained call order of the mobility parameter table index service, amplitude gradient calculation service, and duty - cycle correction service in the service mesh, where the duty - cycle correction service receives the boundary constraint conditions from the federated learning framework; The reinforcement learning routing policy trainer analyzes the real - time load status of the service mesh and the gradient synchronization status of the federated learning sub - model in real - time to generate a dynamic routing priority list to bypass resource - competing nodes; Persistently store the brightness error data generated during the compensation process into the service mesh sidecar database, extract the periodic error characteristics through the time - series analysis engine and generate a feedback signal to inversely update the amplitude sub - model gradient descent optimizer of the federated learning framework.

[0047] When constructing the cross - channel compensation strategy, define the chained call order of the mobility parameter table index service, amplitude gradient calculation service, and duty - cycle correction service in the service mesh. The chained call realizes the service - to - service communication protocol conversion through the API gateway of the service mesh. The mobility parameter table index service retrieves the matching carrier mobility compensation coefficient table from the pre - set material property database based on the mobility threshold parameter in the boundary constraint conditions output by the federated learning framework, generates the index result and transmits it to the amplitude gradient calculation service through the gRPC streaming interface. After receiving the index result, the amplitude gradient calculation service calculates the voltage compensation gradient in combination with the current drive voltage amplitude, uses the sliding window mechanism to filter high - frequency noise interference, and outputs the gradient correction parameter to the duty - cycle correction service. The duty - cycle correction service reconstructs the pulse timing waveform and generates a duty - cycle adjustment instruction set according to the gradient correction parameter and the timing phase limit in the boundary constraint conditions to complete the iteration of the core parameters of the compensation strategy.

[0048] The reinforcement learning routing policy trainer collects real-time load metrics of each service node through the service mesh control plane, including CPU utilization, request queue depth, and network latency data, and synchronously obtains the gradient synchronization status identifier of the federated learning sub-model. The trainer uses the PPO algorithm to construct a routing decision model. The input layer fuses the multi-dimensional feature vectors of the load metrics and the gradient synchronization status, and the output layer generates a dynamic routing priority list. The routing priority list includes the availability score and path weight coefficient of the service node. When it is detected that the queue depth of the resource competition node exceeds the preset threshold, the compensation request is automatically redirected to the mirror service copy of the low-load node. The routing policy update result dynamically injects traffic management rules through the Envoy proxy of the service mesh to achieve a millisecond-level routing switching response.

[0049] The luminance error data generated during the compensation process is collected through the sidecar proxy of the service mesh and persistently stored in the circular buffer of the sidecar database in a time series format. The time series analysis engine uses the STL decomposition algorithm to separate the trend term, periodic term, and residual term of the error data, and extracts the periodic error characteristics related to the temperature drift characteristics of the organic material. The feature data is encoded to generate a feedback signal, which is pushed to the amplitude sub-model gradient descent optimizer of the federated learning framework through the message bus. After receiving the feedback signal, the gradient descent optimizer dynamically adjusts the learning rate decay strategy based on the periodic error characteristics, increases the update step size of the voltage amplitude correction coefficient during the local training process, and simultaneously triggers the synchronous calibration of the pulse width constraint condition of the timing sub-model. The feedback mechanism effectively suppresses the cumulative deviation in the multi-channel signal modulation process through the recognition of the periodic characteristics of the compensation error and the adaptive adjustment of parameters.

[0050] Specifically, for the multi-channel signal modulation method of the organic display interface described in the present invention, the quantization time window segmentation includes: Using quantum bit encoding to discretize the signal period of the compensated drive signal component into a superposition state time slice combination, and screening the optimal segmentation scheme through the trigger signal of the hardware counter; triggering the elastic scaling of the microservice instance based on the output threshold of the phase error gradient calculation module, and registering the new instance to the task assignment queue of the service orchestration engine; Parsing the edge sharpness index from the image processing module through the rule engine to generate a dynamic priority score table to drive the time slice task to allocate short time slices in the high sharpness area and merge time slices in the flat area.

[0051] During the quantization time window segmentation process, the qubit encoding technology is used to discretize the signal period of the compensated drive signal component, and each time slice corresponds to the phase amplitude parameters of the quantum state superposition. The qubit encoding generates a candidate time slice combination scheme through quantum gate operations, and uses the trigger signal of the hardware counter to traverse and screen the candidate schemes. The hardware counter is configured with a noise suppression circuit to eliminate clock jitter interference, and the optimal time slice segmentation scheme that meets the phase synchronization constraint is selected. The time slice interval parameter of the optimal segmentation scheme is transmitted to the microservice orchestration engine through the service bus to trigger the subsequent resource scheduling process.

[0052] When triggering the elastic scaling of microservice instances based on the output threshold of the phase error gradient calculation module, the phase error gradient value generated during the time window segmentation process is monitored in real time. When the gradient value exceeds the preset dynamic threshold range, an instance expansion request is sent to the Kubernetes cluster, and the dynamic threshold range is adjusted dynamically according to the thermal stability parameters of the organic material. The newly added microservice instances are automatically registered in the task assignment queue of the orchestration engine through the service discovery mechanism, and the registration information includes the computing resource capacity and task processing delay metrics of the instances. The task assignment queue uses the weighted round-robin algorithm to distribute the time slice processing tasks to the newly added instances, and preferentially assigns the time slices in the high-gradient area to accelerate the phase error convergence.

[0053] When parsing the edge sharpness index from the image processing module through the rule engine, the sharpness value of the image contour area is mapped to the time slice processing priority parameter. The rule engine loads the preset fuzzy logic rule set, normalizes the sharpness value and calculates the time slice length weight coefficient to generate a dynamic priority score table. The score table drives the service orchestration engine to allocate millisecond-level short time slice tasks in the high-sharpness area to improve the edge detail resolution, and merges adjacent time slices in the flat area to form a batch processing task unit. The merged task unit reduces the computational density through the pipeline parallel processing mechanism, and the released computing resources are reallocated to the contour optimization microservice instance to achieve the balance optimization of resource utilization efficiency and display quality.

[0054] Specifically, for the multi-channel signal modulation method of the organic display interface described in the present invention, the multi-modal dynamic range expansion includes: The sub-signal components generated by the quantization time window segmentation are distributed to the GPU cluster to execute the frequency division strategy optimization and the FPGA node to execute the physical property optimization calculation based on the simulated annealing algorithm through the heterogeneous computing task dispatcher; Construct a cross-architecture feature alignment layer to convert the physical optimization features output by the FPGA node into a tensor format compatible with the GPU cluster neural network; Use a lightweight encryption aggregation protocol to perform consistency verification on the frequency division strategy optimization results of the GPU cluster and the physical characteristic optimization data of the FPGA node. When a color level jump anomaly is detected, trigger a sub-signal component weight re-optimization process across nodes.

[0055] During the multi-modal dynamic range expansion process, when receiving sub-signal components generated by quantization time window segmentation through a heterogeneous computing task dispatcher, the dynamic scheduling strategy allocates tasks based on the real-time computing load status of the GPU cluster and the FPGA node. The dispatcher deploys a frequency division strategy optimization algorithm in the GPU cluster, loads a convolutional neural network model to perform frequency domain feature extraction and non-linear mapping optimization on the sub-signal components; deploys a simulated annealing algorithm in the FPGA node, and iteratively searches for the optimal driving voltage waveform based on the physical characteristic parameters of the organic material. The task distribution instruction is transmitted to the heterogeneous computing node through the PCIe high-speed bus. Each node receives a task packet containing a time slice index and optimization parameter constraint conditions, realizing load balancing of computing resources and parallel task processing.

[0056] When constructing a cross-architecture feature alignment layer, deploy a data conversion interface at the output end of the FPGA node to convert the physical optimization feature data from the hardware description language format into a floating-point tensor structure that can be parsed by the GPU cluster. The conversion interface integrates a bit width matching module and a normalization processing unit to perform dynamic range scaling and sign bit extension on the fixed-point numbers output by the FPGA, generating a tensor data stream that matches the input dimension of the neural network layer. The converted tensor data is injected into the batch processing queue of the GPU cluster through a shared memory channel, and is cascaded with the frequency division strategy optimization results at the feature level to form a joint input source of multi-modal optimization data.

[0057] When using a lightweight encryption aggregation protocol to perform consistency verification on the heterogeneous computing results, establish a data transmission channel based on elliptic curve cryptography between the GPU cluster and the FPGA node. The aggregation server receives the encrypted frequency division strategy optimization results and physical characteristic optimization data, decrypts them through a key negotiation mechanism, and then performs color level continuity analysis to detect color chroma jump anomaly events between adjacent time slices. When it is detected that the jump amplitude exceeds the color gamut tolerance threshold of the organic material, trigger a cross-node re-optimization process, and send a weight reallocation instruction to the task dispatcher. The instruction contains the abnormal time slice index and the weight correction coefficient, driving the GPU cluster and the FPGA node to perform collaborative re-optimization on the specified time slice until the color level jump value converges to a preset safe interval, completing the dynamic range consistency expansion of the multi-modal data.

[0058] Specifically, for the multi-channel signal modulation method of the organic display interface described in the present invention, the data cleaning of the containerized microservice architecture includes: Embed a lightweight message broker in the container instance to achieve high-frame-rate and low-latency transmission of raw data with environment tags; Compress the volume of the electroluminescence response model through model pruning technology to adapt to the storage limit of edge computing nodes; Push the preprocessed standardized data output by the gamma correction module to the message queue to trigger the feature vector update and weight backpropagation of the transfer learning model dynamic loader.

[0059] During the data cleaning process of the containerized microservice architecture, when embedding a lightweight message broker in the container instance, use the MQTT protocol to establish a communication link with the data acquisition module to achieve high-frame-rate and low-latency transmission of raw data with environmental tags. The message broker configures a message persistent queue and a priority channel to shape the traffic of bursty data streams in high-dynamic-range scenarios, and allocates transmission bandwidth through a time-slicing round-robin mechanism. The raw data with environmental tags is distributed by the message broker to the parallel cleaning container cluster, triggering a multi-instance collaborative processing flow.

[0060] When compressing the volume of the electroluminescence response model through model pruning technology, use the channel pruning technology to remove redundant convolution kernels in the model that are irrelevant to the current organic material characteristics, and retain the feature extraction layers that are strongly correlated with the carrier mobility and color gamut coverage. During the pruning process, load the weight distribution histogram of the transfer learning model, evaluate the weight contribution degree of each layer based on the KL divergence, and iteratively prune the network branches with a contribution degree lower than the threshold. The pruned model is converted to an 8-bit fixed-point number format through quantization-aware training to adapt to the low-precision computing unit of the edge computing node, and the model volume compression rate is dynamically matched with the storage capacity of the edge node.

[0061] When pushing the preprocessed standardized data output by the gamma correction module to the message queue, use Apache Kafka to build a distributed message bus, divide the data batches according to time windows and attach metadata tags. The message consumer listens to the topic partition, and when it detects the arrival of a new batch of data, it triggers the feature extraction service of the transfer learning model dynamic loader. During the feature vector update process, extract the brightness distribution features and color gamut boundary parameters in the standardized data, and fine-tune the weights of the last fully connected layer of the pruned model through the backpropagation algorithm. The updated weight parameters are backpropagated to the gamma correction module of the data preprocessing pipeline through the sidecar proxy, driving the gamma curve parameters to be dynamically calibrated according to the aging characteristics of the organic material, forming a closed-loop iteration of data cleaning and model optimization. The weight backpropagation mechanism manages the parameter update sequence through a version control service to avoid parameter conflicts between multiple nodes.

[0062] Specifically, for the multi-channel signal modulation method of the organic display interface described in the present invention, the attention fusion of the model distillation module includes: Capture the implicit relationship between the pulse width correlation of the timing sub-model and the conduction period mobility characteristics of the frequency sub-model through the self-attention mechanism; Input the fused feature vector as the global boundary condition into the interface of the duty cycle correction service; Jointly optimize the compensation coefficient matrix output by the distillation model with the real-time load status of the service grid to generate a driving signal reconstruction instruction.

[0063] During the attention fusion process of the model distillation module, when capturing the pulse width correlation features of the timing sub-model and the on-time period migration rate features of the frequency sub-model through the self-attention mechanism, a multi-head attention mechanism is used to establish a feature interaction space across sub-models. The pulse width feature vector of the timing sub-model is mapped to the query vector space through a linear projection layer, and the on-time period migration rate feature vector of the frequency sub-model is projected to the key-value vector space, and an attention weight distribution map is generated through similarity matrix calculation. The distribution map drives the dynamic weighted fusion of the timing features and the frequency features, and extracts the implicit association rules for pulse width adjustment and on-time period control.

[0064] After being normalized, the fused feature vector is input as the global boundary condition into the RESTful interface of the duty cycle correction service. The interface parses the timing phase constraint parameters and amplitude fluctuation thresholds in the feature vector, and converts them into an input instruction set for the duty cycle correction service through a parameter mapping table. The correction service calls a pulse waveform generator, and combines the phase synchronization requirements and amplitude tolerance range in the boundary conditions to reconstruct the duty cycle parameter sequence of the driving signal and generate an initial correction instruction.

[0065] When jointly optimizing the compensation coefficient matrix output by the distillation model with the real-time load status of the service grid, a multi-objective optimization algorithm is used to balance the computational resource consumption and signal modulation accuracy. The timing compensation weight and amplitude compensation factor in the compensation coefficient matrix are coupled and analyzed with the node calculation delay and memory occupancy rate collected by the service grid monitoring module to generate a resource-aware optimization objective function. The objective function drives the genetic algorithm to iteratively search for the optimal compensation coefficient combination and outputs a driving signal reconstruction instruction set. The instruction set is injected into the signal modulation pipeline through the sidecar proxy of the service grid, triggering the online parameter update of the driving signal generation module to complete the closed-loop optimization of the multi-dimensional compensation strategy.

[0066] Specifically, for the multi-channel signal modulation method of the organic display interface described in the present invention, the feedback signal update of the service grid includes: Extract the time series data of the brightness error in the sidecar database of the service grid, and identify periodic resource competition events through a sliding window algorithm; Encode the recognition result as a gradient update suggestion and push it to the asynchronous optimizer of the amplitude sub-model; Dynamically adjust the local training rounds of the federated learning framework according to the feedback signal, and synchronously update the pulse width constraint conditions of the timing sub-model.

[0067] During the feedback signal update process of the service mesh, when extracting the time series data of the brightness error from the sidecar database, a time range query instruction is used to filter the error sampling points within a specific time window, and the size of the time window is dynamically adjusted according to the response delay characteristics of the organic material. After the sliding window algorithm loads the error data, the periodic fluctuation components are detected through fast Fourier transform, and combined with the resource allocation records in the service mesh log, the periodic error events caused by the resource competition of the microservice instances are identified. The recognition result marks the event occurrence timestamp and the resource competition node identifier, and generates a structured diagnostic report.

[0068] When encoding the recognition result into a gradient update suggestion, the Protocol Buffers serialization protocol is used to convert the event frequency, node load peak value, and error amplitude in the diagnostic report into a binary data stream. The data stream is pushed to the asynchronous optimizer of the amplitude sub-model through the message middleware, and the optimizer reconstructs the gradient descent direction vector after parsing the data stream. The asynchronous optimizer adopts a delayed update mechanism to accumulate the gradient suggestions of multiple event cycles in the local cache. When the cache queue reaches the batch processing threshold, it triggers the gradient aggregation operation based on momentum acceleration to generate the incremental update parameters of the voltage amplitude correction coefficient.

[0069] When dynamically adjusting the training strategy of the federated learning framework according to the feedback signal, the monitoring module real-time collects the parameter update frequency of the amplitude sub-model and the convergence state of the timing sub-model. When detecting the gradient oscillation phenomenon caused by periodic error events, the local training rounds are automatically reduced and the sensitivity threshold of the early stopping strategy is increased to prevent overfitting. When synchronously updating the pulse width constraint conditions of the timing sub-model, the phase drift error data fed back by the service mesh is input into the constraint generator to reconstruct the upper and lower boundary values of the pulse width. The boundary values are synchronized to all sub-model nodes of the federated learning framework through the distributed lock mechanism to maintain the timing consistency of multi-channel signal modulation.

[0070] Specifically, for the multi-channel signal modulation method of the organic display interface described in the present invention, the generation of the dynamic priority score table includes: Parsing the edge sharpness index generated by the quantization time window segmentation through the rule engine to generate the task priority score of the short time slice in the high sharpness area; Merging the time slices in the flat area and reducing the calculation density, and reallocating the released computing resources to the contour optimization microservice instance; Jointly analyzing the score table and the output data of the phase error gradient calculation module to dynamically adjust the task weight allocation parameters of the middle service orchestration engine.

[0071] During the generation process of the dynamic priority scoring table, when parsing the edge sharpness indicators generated by dividing the quantized time window through the rule engine loading the fuzzy logic rule set, an image segmentation algorithm is used to divide the boundary between the high sharpness area and the flat area. The rule engine maps the edge sharpness value to the time slice processing priority parameter, dynamically adjusts the scoring weight coefficient based on the sharpness gradient change rate, and generates the short time slice task priority scoring table for the high sharpness area. Each time slice task in the scoring table is marked with the area type and priority level to be pushed to the microservice orchestration engine through the service bus.

[0072] When merging time slices in the flat area, a region growing algorithm is used to detect continuous low sharpness pixel blocks, triggering a time slice merging instruction to integrate adjacent time slices into a batch processing unit. The merged time slice tasks reduce the computing density through the resource scheduler and optimize the processing flow using sparse matrix operations. The released computing resources are reallocated to the contour optimization microservice instance cluster according to the load balancing strategy of the service orchestration engine, and the instance cluster receives high priority time slice tasks using the weighted round robin algorithm.

[0073] When jointly analyzing the scoring table and the output data of the phase error gradient calculation module, a multi-objective optimization algorithm is used to balance the requirement for edge sharpness improvement and the phase synchronization accuracy index. An association matrix between the priority parameters of the scoring table and the phase error gradient is established during the analysis process. When it is detected that the phase error of the time slice task in the high sharpness area exceeds the tolerance threshold, the dynamic weight adjustment service of the service orchestration engine is triggered. The adjustment service reallocates the task weight parameters according to the coupling coefficient of the association matrix, driving the computing resources of the microservice instance cluster to tilt towards the phase sensitive tasks, and maintaining the collaborative optimization of the edge sharpness and timing synchronization of the display interface.

[0074] In the specific implementation of the present invention, the grayscale, color gamut, and dynamic range data of the input image are synchronized and aligned at the millisecond level with the temperature parameters collected by the timestamp binding module and the environmental temperature sensor to generate the raw data with environmental labels. The containerized microservice architecture based on the Kubernetes engine dynamically scales the number of instances according to the input image resolution. In high resolution scenarios, the container instances are horizontally extended to 1.5 times the number of pixel matrix columns. Each instance deploys a lightweight noise cancellation algorithm and a color gamut boundary verification module, and realizes data exchange between containers through shared memory. The processed data is pushed to the gamma correction module. The migration learning model dynamic loader retrieves the electroluminescence response model with a matching degree exceeding 85% from the cloud model repository according to the carrier mobility parameters of the current organic semiconductor material, uses the channel pruning technology to remove 40% of the redundant convolution kernels in the model, and feeds the output features of the pruned model back to the gamma correction module to dynamically adjust the slope of the correction curve, generating a standardized data set adapted to the nonlinear characteristics of the organic material.

[0075] After the standardized dataset is input into the federated learning framework, the time series sub-model receives the feature sequence of the correlation between the pulse width and the refresh rate through the gRPC bidirectional stream, performs dynamic time warping analysis after caching in the circular buffer, and generates the pulse width adjustment coefficient; the amplitude sub-model loads the calibration data of the voltage-brightness response curve, constructs a piecewise linear interpolation model and calculates the voltage correction gradient through the asynchronous gradient descent optimizer, and temporarily stores the gradient value in the Redis cache cluster and marks the version identifier; the frequency sub-model uses the sliding window mechanism to extract the mobility fluctuation characteristics, and generates the on-time optimization parameters after separating the noise components through Fourier transform. The model distillation module fuses the pulse width characteristics of the time series sub-model and the mobility characteristics of the frequency sub-model through the multi-head self-attention mechanism, generates the cross-sub-model boundary constraint condition matrix, and pushes this matrix to the duty cycle correction service interface through the service mesh API gateway.

[0076] The mobility parameter table index service deployed in the service mesh accelerates the query of the material property database through the hash mapping algorithm, controls the response time within 5ms, and transmits the matched mobility compensation coefficient to the amplitude gradient calculation service. The amplitude gradient calculation service combines the current driving voltage amplitude, uses a filtering mechanism with a sliding window width of 10ms to generate the gradient correction parameter, and inputs it to the duty cycle correction service to reconstruct the pulse timing waveform. During the reconstruction process, the pulse width modulation instruction set generated by the hardware description language dynamically adjusts the duty cycle update step, with the initial step value of 0.1μs and adaptively adjusted according to the historical error data. The reinforcement learning routing policy trainer of the service mesh monitors the node CPU utilization rate and the queue depth in real time. When it detects that the node load exceeds 75%, it redirects the request to the mirror node through the Envoy proxy to reduce the time series jitter error.

[0077] The compensated drive signal component is discretized into a superposition state time slice combination of 50 - 200μs by quantum bit encoding, and the optimal segmentation scheme that meets the phase synchronization constraint is screened through the hardware counter. The phase error gradient calculation module monitors the gradient change with a period of 100ms. When the gradient value exceeds the dynamic threshold range [0.15, 0.25], it triggers the Kubernetes cluster to expand the number of microservice instances to 1.2 times the original number. After the newly added instances are registered in the task assignment queue, the weighted round-robin algorithm preferentially allocates the time slices in the high-gradient area to the newly added instances, and the phase error convergence speed is increased by 30%. The rule engine parses the edge sharpness index output by the image processing module, allocates 50μs short time slice tasks to the areas with sharpness values higher than 0.8, merges the time slices in the flat areas into the 200μs batch processing unit, and reallocates the released 20% of the computing resources to the contour optimization microservice instances.

[0078] The heterogeneous computing task dispatcher distributes time-slice tasks to the GPU cluster and FPGA nodes through the PCIe 4.0 bus. The GPU cluster loads the frequency division strategy model with the ResNet-18 architecture, and the FPGA node performs simulated annealing optimization calculations with 500 iterations. The cross-architecture feature alignment layer converts the 16-bit fixed-point numbers output by the FPGA into 32-bit floating-point tensors that can be processed by the GPU, and after normalization, performs feature concatenation with the results of the frequency division strategy. The lightweight encryption aggregation protocol uses the ECDH key negotiation mechanism to establish a secure channel. When it is detected that the chromaticity jump of adjacent time slices exceeds ΔE>5, it triggers a cross-node re-optimization process, and the weight correction coefficient is set to 1.3 times the original value for iterative re-optimization. The finally generated global drive signal migrates the physical optimization features of the FPGA to the GPU model through knowledge distillation. The pruning model parameters in the migration learning model loader in the feedback loop are updated every 24 hours, forming a closed-loop optimization link of data preprocessing, parameter decoupling, compensation reconstruction, and feedback calibration, reducing the color level continuity error of the display interface in the 256-level gray scale test to less than 3%, and increasing the edge sharpness MTF value to more than 0.85.

[0079] The explanations of the technical feature terms in the technical solution of the present invention are as follows: Gamma correction and gamut mapping: Gamma correction refers to performing a non-linear transformation on the gray-scale data of the input image to compensate for the non-linear response characteristics of the display device, so that the brightness output is linearly related to the input signal. Gamut mapping matches the gamut coverage rate of the organic display interface by adjusting the color space distribution of the image, eliminating gamut boundary overflow. The two cooperate to generate a standardized data set, enabling the data to adapt to the electroluminescence characteristics of the organic material.

[0080] Federated learning framework: A distributed machine learning architecture composed of time series, amplitude, and frequency sub-models. The time series sub-model generates signal time series optimization parameters by analyzing the time series correlation between the pulse width and the refresh rate; the amplitude sub-model constructs a non-linear mapping relationship between the drive voltage and the brightness and outputs a voltage amplitude correction coefficient; the frequency sub-model analyzes the dynamic response of the conduction period and the carrier mobility and generates a mobility compensation parameter. Each sub-model realizes parameter decoupling through asynchronous gradient update and distributed caching.

[0081] Model distillation module: Adopts a multi-head self-attention mechanism to fuse the pulse width features of the time series sub-model and the mobility features of the frequency sub-model, generates boundary constraint conditions across sub-models through similarity matrix calculation, and uses them as input parameters for the duty cycle correction service to solve the local optimum conflict caused by multi-channel signal coupling.

[0082] Service Mesh: A communication infrastructure based on the microservices architecture, including a mobility parameter table indexing service (rapidly retrieving the material property database), an amplitude gradient calculation service (real-time calculating the voltage correction gradient), and a duty cycle correction service (reconstructing the pulse timing waveform). Dynamically avoiding resource competition nodes through the reinforcement learning routing policy to optimize the service call path.

[0083] Quantized time window segmentation: Discretizing the compensated driving signal period into superposition state time slices, generating candidate segmentation schemes using quantum bit encoding, and screening the optimal solution that meets the phase synchronization constraint through a hardware counter. Triggering the elastic scaling of microservice instances in combination with the phase error gradient to achieve dynamic allocation of computing resources.

[0084] Heterogeneous computing nodes: GPU cluster: Executing frequency division strategy optimization and using a convolutional neural network to extract frequency domain features; FPGA node: Searching for the optimal solution of the driving voltage waveform based on the simulated annealing algorithm. Converting the physical optimization features output by the FPGA into a tensor format that can be processed by the GPU through a cross-architecture feature alignment layer to achieve heterogeneous computing collaboration.

[0085] Federated aggregation and knowledge distillation: Federated aggregation: Performing consistency verification on the GPU frequency division strategy and the FPGA physical optimization results, using a lightweight encryption protocol to detect abnormal color level jumps, and triggering cross-node re-optimization; Knowledge distillation: Transferring the physical optimization knowledge of the FPGA to the GPU neural network model, driving model fine-tuning through a contrastive learning loss function to form a fusion of multi-modal optimization results.

[0086] Closed-loop feedback mechanism: The brightness error data extracts periodic features through time series analysis and reversely updates the gradient descent optimizer of the federated learning sub-model; The global driving signal is fed back to the transfer learning model loader to drive model pruning and quantization-aware training. Forming a double-closed-loop optimization link of data flow (standardization → decoupling → compensation → expansion) and control flow (error feedback → parameter update).

[0087] The models involved in the technical solution of the present invention and their collaborative effects are explained as follows: Federated learning framework sub-model: Timing sub-model: Deployed in an independent container, receiving the pulse width and refresh rate correlation feature sequence of the standardized data set in real time through the gRPC protocol. Using the dynamic time warping algorithm to analyze the phase synchronization constraint and refresh rate fluctuation law of the signal timing, and generating a pulse width adjustment coefficient. For example, analyzing the dynamic correlation of the pulse width in the range of 0.1 - 0.3 ms at a refresh rate of 120 Hz.

[0088] Amplitude sub-model: Load the calibration data of the voltage-luminance response curve of the organic semiconductor material (e.g., the voltage range of 0 - 5V corresponds to the luminance range of 0 - 1000 nit), and construct a piecewise linear interpolation model. Calculate the voltage amplitude correction gradient through the asynchronous gradient descent optimizer, temporarily store the gradient value in the Redis cache cluster, and trigger the aggregation update after the version identifier is synchronized with the frequency sub-model.

[0089] Frequency sub-model: Adopt a sliding window mechanism (window width 10ms) to extract the dynamic response characteristics of the conduction period and carrier mobility, filter out high-frequency noise (>1kHz components) through Fourier transform, and generate optimized parameters for the conduction period. For example, when the mobility fluctuation range is 1.5 - 2.5 cm² / V·s, adjust the conduction period to 5 - 15 μs.

[0090] Model distillation module: Adopt a multi-head self-attention mechanism (8 heads of attention) to fuse the pulse width characteristics (dimension 128) of the timing sub-model and the mobility characteristics (dimension 64) of the frequency sub-model. Calculate the similarity matrix in the attention space, generate a weight distribution map, and then drive the weighted fusion of feature vectors to output a boundary constraint condition matrix across sub-models (dimension 256×256). This matrix contains the timing phase constraint error threshold (e.g., ±0.05 μs) and the amplitude fluctuation tolerance range (e.g., ±0.1V), and is input to the duty cycle correction service to guide the pulse timing reconstruction.

[0091] Core service of the service mesh: Mobility parameter table indexing service: Accelerate the query of the material property database based on the hash mapping algorithm (the number of hash buckets is 1024), and the response time ≤ 5ms. For example, retrieve the compensation coefficient table corresponding to the carrier mobility of 2.0 cm² / V·s, and output the index result to the amplitude gradient calculation service.

[0092] Amplitude gradient calculation service: Combine the current drive voltage amplitude (e.g., 3.2V), and adopt a sliding window filtering (window width 10ms) to generate gradient correction parameters. The initial step size is 0.1 μs, and it is adaptively adjusted to the range of 0.08 - 0.12 μs according to the historical error data.

[0093] Duty cycle correction service: Analyze the timing phase limit in the boundary constraint condition matrix, generate a pulse width modulation instruction set through a hardware description language, and dynamically adjust the duty cycle parameter (e.g., from 50% to 55%).

[0094] Heterogeneous computing node cooperation mechanism: GPU cluster: Deploy a frequency division strategy model with the ResNet-18 architecture. The input is a 100×100 frequency domain feature tensor, and the output is the frequency division optimization result (e.g., decompose the 100kHz fundamental frequency into 10 harmonic components).

[0095] FPGA Node: Perform simulated annealing optimization calculations (500 iterations) to search for the optimal solution of the drive voltage waveform (e.g., the rising edge time of the waveform is optimized from 1.2 μs to 0.9 μs).

[0096] Cross-architecture feature alignment layer: Convert the 16-bit fixed-point number (Q8.8 format) output by the FPGA into a 32-bit floating-point tensor. After normalization (scaling factor 0.01), it is concatenated with the GPU output features to form a combined input source (dimension 200×200).

[0097] Federated aggregation and feedback mechanism: Lightweight encryption aggregation protocol: Use ECDH key agreement to establish a secure channel, detect chromaticity jumps in adjacent time slices (trigger re-optimization when ΔE > 5), and set the weight correction coefficient to 1.3 times the original value for iterative re-optimization.

[0098] Knowledge distillation: Transfer the physical optimization knowledge of the FPGA to the GPU model through a contrastive learning loss function (temperature parameter τ = 0.5), and drive the fine-tuning of the fully connected layer to make the frequency division strategy result approach the physical optimization characteristics.

[0099] Closed-loop feedback: The brightness error data is decomposed by STL to extract periodic features (period 1 s), the learning rate of the amplitude sub-model is updated in the reverse direction (adjusted from 0.001 to 0.0005), and the pruning model parameters of the transfer learning model loader are updated every 24 hours (the pruning rate is dynamically adjusted from 40% to 35% - 45%).

[0100] Example of collaborative logic between models: The pulse width adjustment coefficient (e.g., 0.25 ms) output by the timing sub-model and the on-time period parameter (e.g., 10 μs) of the frequency sub-model are fused by the distillation module to generate boundary constraint conditions to guide the duty cycle correction service to reconstruct the pulse timing (duty cycle from 50% → 55%). When the service mesh dynamic routing strategy detects that the node load > 75%, the request is redirected to a low-load mirror node, reducing the compensation delay from 20 ms to 12 ms. The 50 μs short time slices generated by quantized time window segmentation are optimized by frequency division in the GPU cluster and verified by federated aggregation with the FPGA physical optimization results. The chromaticity jump ΔE is reduced from 6.2 to 3.5, and the edge sharpness MTF value is increased from 0.78 to 0.86, verifying the effectiveness of multi-model collaboration.

[0101] Through the specific implementation and data interaction of the above models, the present invention realizes the decoupling optimization of multi-dimensional signal parameters, the dynamic avoidance of resource competition, and the systematic improvement of display performance.

[0102] The present invention realizes the collaborative decoupling and dynamic optimization of multi-dimensional signal parameters through a federated learning framework. A timing sub-model, an amplitude sub-model, and a frequency sub-model are deployed to analyze the pulse width-refresh rate correlation, the driving voltage-brightness non-linear relationship, and the on-time-mobility dynamic response respectively, and generate optimized weight coefficients. The model distillation module fuses the features of each sub-model through a multi-head attention mechanism, generates cross-sub-model boundary constraint conditions, and inputs them to the duty cycle correction service interface to reconstruct the pulse timing parameters, solving the problems of cumulative timing deviation and conflict of local optimal solutions. The asynchronous gradient update mechanism and the distributed cache design avoid resource competition during the training process of the sub-models, and realize the coordination of parameter decoupling and global optimization.

[0103] The cross-channel compensation strategy based on the service mesh adjusts the mobility compensation parameter and the voltage amplitude correction coefficient in real time by chain-calling the mobility parameter index, the amplitude gradient calculation, and the duty cycle correction service. The reinforcement learning routing policy trainer dynamically analyzes the service load and the gradient synchronization state, generates a routing priority list to bypass high-load nodes, and reduces the delay error caused by resource competition. The brightness error data extracts periodic features through time series analysis, and reversely updates the gradient descent optimizer of the amplitude sub-model, and suppresses the cumulative deviation during the compensation process through a feedback mechanism to maintain the color scale continuity.

[0104] The quantization time window segmentation discretizes the driving signal period into superposition state time slices, and dynamically allocates computing resources in combination with the phase error gradient and the edge sharpness index. The rule engine generates a priority score table to drive the task allocation of short time slices in high sharpness regions, and merges time slices in flat regions to reduce the computing density. The heterogeneous computing task dispatcher distributes sub-signal components to the GPU cluster and the FPGA node, the cross-architecture feature alignment layer integrates the frequency division strategy and the physical optimization result, and the federated aggregation protocol checks the color scale jump anomaly and triggers re-optimization, expanding the dynamic range while improving the edge sharpness, forming a closed-loop optimization link for the data flow and the control flow.

Claims

1. Method for multi-channel signal modulation of an organic display interface, characterized in that, Including: Obtain the grayscale, color gamut, and dynamic range data of the input image, perform gamma correction and color gamut mapping processing on the grayscale, color gamut, and dynamic range data to generate a standardized data set; Input the standardized data set into the federated learning framework, analyze the correlation characteristics of pulse width and refresh rate through the time series sub-model deployed in the federated learning framework, model the non-linear relationship between drive voltage and brightness through the amplitude sub-model, and analyze the dynamic response of conduction period and mobility through the frequency sub-model to generate optimized weight coefficients for pulse width, refresh rate, and drive voltage; Based on the optimized weight coefficients generated by the federated learning framework, construct a cross-channel compensation strategy in the service mesh, adjust the mobility compensation parameters through the service indexed by the mobility parameter table, update the voltage amplitude correction coefficient through the amplitude gradient calculation service, and reconstruct the pulse timing parameters through the duty cycle correction service to generate a compensated drive signal component; Perform quantization time window segmentation on the compensated drive signal component, trigger the dynamic scaling of microservice instances according to the phase error gradient calculation result, and allocate computing resources by combining the time window priority score table generated by the edge sharpness index to generate a synchronized and optimized time window scheduling scheme; Based on the synchronized and optimized time window scheduling scheme, perform multi-modal dynamic range expansion through heterogeneous computing nodes, federate and knowledge distill the optimized results of the frequency division strategy generated by the GPU cluster and the optimized physical characteristic data output by the FPGA node to generate a global drive signal and feedback it to the transfer learning model loader of the data acquisition module to form an optimized link with a closed-loop data flow and control flow.

2. The multi-channel signal modulation method for an organic display interface according to claim 1, wherein Generating a standardized data set includes: Associate the dynamic range data with the temperature parameters collected by the environmental temperature sensor through the timestamp binding module to generate raw data with environmental labels; Adopt a containerized microservice architecture to perform parallel cleaning on the raw data with environmental labels, and dynamically scale the number of container instances based on the input image resolution by the Kubernetes engine; Dynamically load the electroluminescence response model matching the current organic semiconductor material from the cloud through the transfer learning model dynamic loader, and feedback the feature extraction results output by the pruned electroluminescence response model to the gamma correction module of the data preprocessing pipeline.

3. The multi-channel signal modulation method for an organic display interface according to claim 1, wherein Multi-dimensional parameter decoupling in the federated learning framework includes: Deploy the time series sub-model as an independent container and receive the correlation characteristics of pulse width and refresh rate from the standardized data set through the gRPC protocol; Construct a gradient descent optimizer for the drive voltage-brightness non-linear curve in the amplitude sub-model, and use an asynchronous update mechanism to temporarily store the local gradient in the distributed cache and wait for the gradient synchronization of the time series sub-model and the frequency sub-model; Perform attention fusion on the conduction period characteristics of the frequency sub-model and the pulse width characteristics of the time series sub-model through the model distillation module to generate cross-sub-model boundary constraint conditions and input them into the duty cycle correction service interface of the cross-channel compensation strategy.

4. The multi-channel signal modulation method for an organic display interface according to claim 1, characterized in that Constructing a cross-channel compensation strategy includes: Define the chained call order of the mobility parameter table indexing service, amplitude gradient calculation service, and duty cycle correction service in the service mesh, where the duty cycle correction service receives boundary constraint conditions from the federated learning framework; The real-time load status of the service mesh and the gradient synchronization status of the federated learning sub-model are analyzed in real time by the reinforcement learning routing policy trainer to generate a dynamic routing priority list to bypass resource competition nodes; The luminance error data generated during the compensation process is persistently stored in the service mesh sidecar database. The periodic error features are extracted by the time series analysis engine and a feedback signal is generated to inversely update the amplitude sub-model gradient descent optimizer of the federated learning framework.

5. The multi-channel signal modulation method for an organic display interface according to claim 1, wherein Quantized time window segmentation includes: The signal period of the compensated drive signal component is discretized into a superposition state time slice combination by using qubit encoding, and the optimal segmentation scheme is selected through the trigger signal of the hardware counter; Based on the output threshold of the phase error gradient calculation module, the elastic scaling of microservice instances is triggered, and the newly added instances are registered in the task assignment queue of the service orchestration engine; Parse the edge sharpness index from the image processing module through the rule engine to generate a dynamic priority score table to drive the time slice task to allocate short time slices in the high sharpness area and merge time slices in the flat area.

6. The multi-channel signal modulation method for an organic display interface according to claim 1, wherein Multi-modal dynamic range expansion includes: The sub-signal components generated by the quantized time window segmentation are distributed to the GPU cluster to execute the frequency division strategy optimization through the heterogeneous computing task dispatcher, and the FPGA node executes the physical property optimization calculation based on the simulated annealing algorithm; Construct a cross-architecture feature alignment layer to convert the physical optimization features output by the FPGA node into a tensor format compatible with the GPU cluster neural network; Adopt a lightweight encryption aggregation protocol to perform consistency verification on the frequency division strategy optimization results of the GPU cluster and the physical property optimization data of the FPGA node. When a color level jump anomaly is detected, trigger the sub-signal component weight re-optimization process across nodes.

7. The multi-channel signal modulation method for an organic display interface according to claim 2, characterized in that, Data cleaning of the containerized microservice architecture includes: Embed a lightweight message broker in the container instance to achieve high-frame-rate and low-latency transmission of raw data with environment tags; Compress the volume of the electroluminescence response model through model pruning technology to adapt to the storage limit of edge computing nodes; Push the preprocessed standardized data output by the gamma correction module to the message queue to trigger the feature vector update and weight feedback of the transfer learning model dynamic loader.

8. The multi-channel signal modulation method for an organic display interface according to claim 3, wherein Attention fusion of the model distillation module includes: Capture the implicit relationship between the pulse width correlation of the timing sub-model and the conduction cycle mobility characteristics of the frequency sub-model through the self-attention mechanism; Use the fused feature vector as the global boundary condition and input it to the interface of the duty cycle correction service; Jointly optimize the compensation coefficient matrix output by the distillation model and the real-time load status of the service mesh to generate a drive signal reconstruction instruction.

9. The multi-channel signal modulation method for an organic display interface according to claim 4, wherein Feedback signal update of the service mesh includes: Extract the time series data of the luminance error in the service mesh sidecar database, and identify periodic resource competition events through the sliding window algorithm; Encode the recognition result as a gradient update suggestion and push it to the asynchronous optimizer of the amplitude sub-model; Dynamically adjust the local training rounds of the federated learning framework according to the feedback signal, and synchronously update the pulse width constraint conditions of the timing sub-model.

10. The multi-channel signal modulation method for an organic display interface according to claim 5, wherein The generation of the dynamic priority scoring table includes: Parse the edge sharpness index generated by the quantization time window segmentation through the rule engine to generate the short time slice task priority score for the high sharpness area; Merge time slices in the flat area and reduce the computational density, and reallocate the released computational resources to the contour optimization microservice instance; Jointly analyze the scoring table with the output data of the phase error gradient calculation module, and dynamically adjust the task weight allocation parameters of the middle service orchestration engine.

Images