Intelligent household electrical appliance model updating method and device, electronic equipment and medium

Abstract

The invention provides an intelligent household electrical appliance model updating method and device, electronic equipment and a medium, and is applied to the technical field of intelligent home / smart home, and the method comprises the steps: carrying out the incremental updating of an intelligent household electrical appliance model through the obtained environment data and / or user interaction data within a target time, and obtaining a model increment parameter; wherein the intelligent household appliance model is determined based on performance evaluation results of the local model and the global model; and performing performance evaluation on the intelligent household electrical appliance model after incremental learning, and uploading a model increment parameter to a cloud server on the basis that the performance meets a set target performance condition, so that the cloud server performs model maintenance according to the uploaded model increment parameter. According to the method, the defects that centralized training of the cloud server is difficult to adapt to specific family scenes and updating delay exists are overcome, the model can capture subtle changes of user preferences in real time, the learning process is completed on local equipment, the response speed is higher, and user privacy is protected.

Description

Technical Field

[0001] This application relates to the field of smart home / intelligent home control, and in particular to a method, device, electronic device and medium for updating smart home appliance models. Background Technology

[0002] With the deep integration of artificial intelligence and Internet of Things technologies, smart home appliances are undergoing a transformation from simple automation to highly intelligent and personalized services. The core driving force is machine learning models that can continuously learn and optimize. These models enable devices to understand user habits, predict user needs, and make better decisions autonomously, thereby providing a more convenient and comfortable user experience.

[0003] Currently, smart home appliance model update systems generally adopt a centralized training paradigm using cloud servers. In this model, terminal devices deployed in users' homes are responsible for collecting daily usage data and periodically uploading it to the cloud server. The cloud server then aggregates massive amounts of data and uses its powerful computing resources to perform regular batch model training, such as updating daily or weekly. After training is complete, the cloud server pushes the new model to all terminal devices. Once the devices download and load the new model, they can perform more accurate local inference tasks.

[0004] However, the aforementioned method of centralized training using cloud servers results in model update cycles that can last for hours or even days, leading to high model iteration latency and an inability to respond to real-time changes in user behavior. Furthermore, general-purpose models trained on cloud servers struggle to adapt to the unique usage environments and user preferences of different households, resulting in weak personalized service capabilities. In addition, continuously uploading massive amounts of raw data to cloud servers consumes significant network bandwidth, causing network resource inefficiency. More importantly, the centralized storage and processing of user data increases the risk of privacy breaches, raising widespread user concerns about data security. Summary of the Invention

[0005] The purpose of this application is to provide a method, device, electronic device, and medium for updating a smart home appliance model, in order to solve the shortcomings of the prior art, which is that centralized training on cloud servers makes it difficult to apply to specific home scenarios and has update delays. This allows the model to capture subtle changes in user preferences in real time, and the learning process is completed on the local device, avoiding network request delays, resulting in faster response speeds, fundamentally protecting user privacy, and avoiding a large number of invalid model updates, thereby protecting the performance of the global model and greatly saving network bandwidth and cloud computing resources.

[0006] This application provides a method for updating a smart home appliance model, comprising: incrementally updating the smart home appliance model by acquiring environmental data and / or user interaction data within a target time period to obtain model incremental parameters; wherein, the smart home appliance model is determined based on the performance evaluation results of a local model and a global model, the global model being a model maintained by a cloud server, and the model incremental parameters being used to characterize the parameter changes after the incremental update of the smart home appliance model; performing a performance evaluation on the incrementally learned smart home appliance model, and uploading the model incremental parameters to the cloud server based on the requirement that the performance meets a set target performance threshold, so that the cloud server can maintain the model according to the uploaded model incremental parameters.

[0007] According to the present invention, a method for updating a smart home appliance model is provided, which acquires environmental data and / or user interaction data within a target time period and incrementally updates the smart home appliance model, including: incrementally updating the smart home appliance model according to a preset iteration strategy based on the acquired environmental data and / or user interaction data within the target time period; wherein, the preset iteration strategy is used to: for each iteration, construct a first knowledge loss based on the model prediction results obtained from the incremental update and the corresponding determined user satisfaction; construct a second knowledge loss using historical joint representations obtained from the data buffer and the historical user satisfaction corresponding to each historical joint representation; and construct an elastic weight consolidation (EWC) regularization term based on the model parameters obtained after the previous incremental learning and the current model parameters, combined with the preset importance of each model parameter; obtain a total loss function based on the first knowledge loss, the second knowledge loss, and the EWC regularization term; perform back gradient updates on the model parameters based on the total loss function; and perform the next online incremental learning again using the updated model parameters until the preset number of iterations is reached.

[0008] According to the present invention, a method for updating a smart home appliance model includes acquiring environmental data and user interaction data within a target time period and performing online incremental learning on the smart home appliance model. The method further includes: acquiring idle device resources, including CPU utilization, GPU utilization, and available memory; classifying the idle device resources into levels to obtain idle device levels; determining the corresponding batch learning sample quantity based on the idle device level, and acquiring environmental data and user interaction data within the target time period, as well as historical joint representations in the data buffer and historical user satisfaction corresponding to each historical joint representation, according to the batch learning sample quantity, for online incremental learning of the smart home appliance model; determining the gradient adjustment accumulation step number based on the idle device level, and determining the update gradient of the model parameters using the total loss function during each online incremental learning, storing the update gradient in an accumulator and accumulating it with previously stored update gradients until the gradient adjustment accumulation step number is reached, updating the model parameters based on the accumulated update gradient in the accumulator, and then clearing the accumulator.

[0009] According to the present invention, a method for updating a smart home appliance model is provided, which evaluates the performance of an incrementally learned smart home appliance model, including: identifying the model type of the smart home appliance model; determining the corresponding evaluation index, test set generation strategy, stability detection method, and decision threshold based on the model type and a preset model performance evaluation strategy; the preset model performance evaluation strategy is obtained by configuring the corresponding evaluation index, test set generation strategy, stability detection method, and decision threshold based on different model types; generating a corresponding test set according to the test set generation strategy; the test set generation strategy is used to: construct a static test set by fusing the latest target window data and extreme historical data that challenge the model boundary for classification and regression models; and construct a dynamic interactive simulation test environment by replaying key historical scenarios and injecting sensor perturbations for reinforcement learning models; testing the incrementally learned smart home appliance model using the test set, and evaluating it according to the evaluation index and stability detection method respectively, to obtain the corresponding index evaluation results and stability detection results; comparing the index evaluation results with the index evaluation results of the previous version of the smart home appliance model on the same test set to determine the relative improvement; and triggering the uploading of the model incremental parameters when both the relative improvement and the stability detection results meet the corresponding decision threshold.

[0010] According to the present invention, a method for updating a smart home appliance model generates a corresponding test set based on a test set generation strategy. The method further includes: when the model type is determined to be a classification model, using a target window to select current data based on the test set generation strategy, synthesizing joint representation data by combining it with predefined extreme scenarios, and assigning satisfaction labels to the joint representation data to construct a test set; wherein the current data includes joint representations and user satisfaction corresponding to each joint representation; when the model type is determined to be a regression model, using a target window to select current data based on the test set generation strategy, and extracting historical data corresponding to historical extreme events to construct a test set; wherein the historical data includes historical joint representations and user satisfaction corresponding to each historical joint representation; when the model type is determined to be a reinforcement learning model, using pre-labeled key historical scenarios and sensor noise configurations to construct a dynamically interactive simulated test environment; wherein the key historical scenarios include a complete sequence of environmental states that leads to extremely low rewards or policy oscillations, and the sensor noise configuration includes random perturbations injected into the environmental states.

[0011] According to the present invention, a method for updating a smart home appliance model is evaluated using a stability detection method, including: when the model type is determined to be a classification model, based on the stability detection method, resampling the test set with replacement, calculating the performance index difference between the previous version of the smart home appliance model and the incrementally learned smart home appliance model on each resampled subset, constructing an empirical distribution of the performance index difference, and determining the confidence interval of the performance index difference; wherein, the width of the confidence interval is used to evaluate the statistical reliability of the performance improvement; when the model type is determined to be a regression model, based on the stability detection method, calculating the prediction error sequence of the previous version of the smart home appliance model and the incrementally learned smart home appliance model for each sample on the test set, and analyzing the dispersion of the prediction error sequence; wherein, the rolling standard deviation of the prediction error sequence is used to evaluate the volatility of the model's prediction performance at different time points; when the model type is determined to be a reinforcement learning model, based on the stability detection method, independently running the new strategy multiple rounds in a simulated test environment, and recording the cumulative reward sequence for each round; wherein, the mean and standard deviation of the cumulative reward sequence are used to evaluate the average performance and robustness of the strategy when repeatedly facing similar scenarios.

[0012] According to the present invention, a smart home appliance model update method is provided. The global model is obtained by the cloud server based on all received model incremental parameters, allocating aggregate weights according to the data quality uploaded by each device, obtaining global incremental parameters, and then optimizing the global incremental parameters using a generative adversarial network.

[0013] This application also provides a smart home appliance model update device, comprising: a model update module, which incrementally updates the smart home appliance model by acquiring environmental data and / or user interaction data within a target time period to obtain model incremental parameters; wherein, the smart home appliance model is determined based on the performance evaluation results of a local model and a global model, the global model is a model maintained by a cloud server, and the model incremental parameters are used to characterize the parameter changes after the incremental update of the smart home appliance model; and a parameter upload module, which performs performance evaluation on the incrementally learned smart home appliance model, and uploads the model incremental parameters to the cloud server based on the performance meeting the set target performance requirements, so that the cloud server can maintain the model according to the uploaded model incremental parameters.

[0014] This application also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the smart home appliance model update method as described above.

[0015] This application also provides an electronic device, which may be an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the above-described smart home appliance model update methods.

[0016] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the smart home appliance model update method as described above.

[0017] The smart home appliance model update method, device, electronic device, and medium provided in this application acquire environmental data and user interaction data within a target time period and perform online incremental learning on the smart home appliance model. By continuously learning the latest environmental and user interaction data, the model can capture subtle changes in user preferences in real time. The learning process is completed on a local device, using original and highly sensitive user data, which fundamentally protects user privacy. At the same time, local decision-making avoids network request latency, resulting in faster response speeds and a smoother user experience. Furthermore, the performance of the incrementally learned smart home appliance model is evaluated, and if the performance improvement exceeds the target performance threshold, the incremental parameters of the model are uploaded to a cloud server. This avoids a large number of invalid or even negative model updates that could pollute the global model, thereby protecting the performance of the global model. It also greatly saves network bandwidth and cloud computing resources, making the entire system run more efficiently and stably.

[0018] In addition, by conducting performance evaluations on the local model and the global model updated by the cloud server, and selecting the smart home appliance model based on the performance evaluation results, it is ensured that the device uses the most suitable model at any time, avoiding the performance degradation that may be caused by blind updates. Moreover, the global model is obtained by the cloud server based on global aggregation and optimization, so that the global model can learn a wider range of more robust patterns, and its generalization ability far exceeds that of any single local model. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the hardware environment for the smart home appliance model update method provided in this application; Figure 2 This is one of the flowcharts illustrating the smart home appliance model update method provided in this application; Figure 3 This is the second flowchart illustrating the smart home appliance model update method provided in this application; Figure 4 This is a schematic diagram of the structure of the smart home appliance model updating device provided in this application; Figure 5 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] According to one aspect of the embodiments of this application, a method for updating a smart home appliance model is provided. This method is widely applicable to whole-house intelligent digital control application scenarios such as smart homes, smart home ecosystems, smart home device ecosystems, and smart house ecosystems. Optionally, in this embodiment, the above-mentioned smart home appliance model updating method can be applied to, for example... Figure 1 The hardware environment shown consists of terminal device 102 and server 104. For example... Figure 1 As shown, server 104 is connected to terminal device 102 via a network and can be used to provide services (such as application services) to the terminal or clients installed on the terminal. A database can be set up on the server or independently of the server to provide data storage services for server 104. Cloud computing and / or edge computing services can be configured on the server or independently of the server to provide data processing services for server 104.

[0024] The aforementioned network may include, but is not limited to, at least one of the following: wired network, wireless network. The aforementioned wired network may include, but is not limited to, at least one of the following: wide area network, metropolitan area network, local area network. The aforementioned wireless network may include, but is not limited to, at least one of the following: Wi-Fi (Wireless Fidelity), Bluetooth. The terminal device 102 may not be limited to PC, mobile phone, tablet computer, smart air conditioner, smart range hood, smart refrigerator, smart oven, smart stove, smart washing machine, smart water heater, smart washing equipment, smart dishwasher, smart projector, smart TV, smart clothes rack, smart curtains, smart audio-visual equipment, smart socket, smart speaker, smart speaker box, smart fresh air equipment, smart kitchen and bathroom equipment, smart bathroom equipment, smart robot vacuum cleaner, smart window cleaning robot, smart mopping robot, smart air purifier, smart steam oven, smart microwave oven, smart water heater, smart air purifier, smart water dispenser, smart door lock, etc.

[0025] Figure 2 A flowchart illustrating a method for updating a smart home appliance model is shown. The execution entity of this method is the device side, including: S21, by acquiring environmental data and / or user interaction data within the target time period, the smart home appliance model is incrementally updated to obtain model incremental parameters; wherein, the smart home appliance model is determined based on the performance evaluation results of the local model and the global model, the global model is a model maintained by the cloud server, and the model incremental parameters are used to characterize the parameter changes after the incremental update of the smart home appliance model. S22. The performance of the smart home appliance model after incremental learning is evaluated, and if the performance meets the set target performance requirements, the incremental parameters of the model are uploaded to the cloud server so that the cloud server can maintain the model based on the uploaded incremental parameters.

[0026] It should be noted that the step number "S2N" in this manual does not represent the order of the smart home appliance model update methods. The following details will explain further. Figure 3 This application describes a method for updating the smart home appliance model.

[0027] Step S21: The smart home appliance model is incrementally updated using the environmental data and / or user interaction data obtained within the target time period to obtain the model incremental parameters. The smart home appliance model is determined based on the performance evaluation results of the local model and the global model. The global model is a model maintained by the cloud server. The model incremental parameters are used to characterize the parameter changes after the incremental update of the smart home appliance model.

[0028] In this embodiment, acquiring environmental data and / or user interaction data within a target time period and incrementally updating the smart home appliance model includes: incrementally updating the smart home appliance model according to a preset iteration strategy using the acquired environmental data and / or user interaction data within the target time period; wherein, the preset iteration strategy is used to: for each iteration, construct a first knowledge loss based on the model prediction results obtained from the incremental update and the corresponding determined user satisfaction; construct a second knowledge loss using historical joint representations obtained from the data buffer and the historical user satisfaction corresponding to each historical joint representation; and construct an elastic weight consolidation EWC regularization term based on the model parameters obtained after the previous incremental learning and the current model parameters, combined with the preset importance of each model parameter; obtain a total loss function based on the first knowledge loss, the second knowledge loss, and the EWC regularization term; perform back gradient updates on the model parameters based on the total loss function; and use the updated model parameters to perform the next online incremental learning until the preset number of iterations is reached.

[0029] Furthermore, incrementally updating the smart home appliance model according to a preset iterative strategy by acquiring environmental data and / or user interaction data within the target time period also includes: determining a joint representation by combining the acquired environmental data and / or user interaction data within the target time period with a cross-modal attention mechanism, and using the joint representation to incrementally update the smart home appliance model according to a preset iterative strategy.

[0030] Specifically, by acquiring environmental data and / or user interaction data within the target time period, and combining it with a cross-modal attention mechanism, a joint representation is determined, including: extracting features from the acquired environmental data and / or user interaction data within the target time period to obtain corresponding temporal environmental features and user intent features; determining the joint representation based on the temporal environmental features and user intent features, and using the cross-modal attention mechanism, incrementally updating the smart home appliance model according to a preset iterative strategy.

[0031] Furthermore, based on temporal context features and user intent features, and combined with a cross-modal attention mechanism, a joint representation is determined, including: using a multi-head attention mechanism, taking user intent features as the query and temporal context features as the key and value, determining the attention score, and generating a context-aware environment vector; concatenating the context-aware environment vector with the user intent features to obtain the joint representation.

[0032] In addition, before constructing the second loss function using historical joint representations and corresponding historical user satisfaction levels obtained from the data buffer, the process includes: using a K-center greedy algorithm to filter and cache all historical joint representations and corresponding historical user satisfaction levels in the data buffer. Specifically, the joint representations and corresponding historical user satisfaction levels are treated as a set of samples. A sample is randomly selected from the sample pool as the first center and placed in the data buffer. Each sample in the sample pool is traversed, and its distance from the sample in the data buffer is determined to obtain the corresponding sample distance. Based on the distances of each sample, the sample in the sample pool that is furthest away and has not been selected as the center is selected and stored as the new center in the data buffer. Each sample in the sample pool is traversed again, and the distances of the corresponding selected sample and each sample in the data buffer are determined. The minimum distance is selected as the distance between the corresponding selected sample and the sample in the data buffer. Based on the distances between each selected sample and the sample in the data buffer, the sample in the sample pool that is furthest away and has not been selected as the center is selected and stored as the new center in the data buffer. This step is iterated again until the number of samples in the data buffer reaches a preset number, thus obtaining a data buffer that protects a preset number of samples.

[0033] Furthermore, the sample pool can be constructed based on a specific number of historical samples obtained in chronological order from the present to the past, without further limitations here.

[0034] It should be added that temporal environmental features and user intent features are extracted from environmental data and user interaction data within the target time period. This transforms raw and chaotic data such as sensor data and voice commands into structured features that the model can efficiently understand. Temporal environmental data and user intent data are processed separately to better capture their unique patterns, such as trends in environmental data and discreteness in intent data, avoiding information confusion. Cross-modal attention fusion is also performed so that the model's decisions are no longer rigid rules but intelligent responses to the current situation. This allows the smart home appliance model to be trained, and user satisfaction is determined by monitoring whether the user readjusts within a preset time period, thus constructing the first knowledge loss.

[0035] Furthermore, through the model replay mechanism, historical joint representations and corresponding historical user satisfaction levels are obtained from the data buffer using a preset sliding window. This allows the model to learn from representative historical samples stored in the buffer, ensuring that the model does not forget old knowledge while learning new knowledge, thus guaranteeing the stability of the model's individuality and the continuity of knowledge. Based on the model parameters obtained after the previous incremental learning and the current model parameters, combined with the corresponding preset importance, an EWC regularization term is constructed to limit the variation of model parameters during updates. This ensures that the model's core and most general knowledge is not covered by accidental or localized data streams, protecting the model's common sense. This leads to the total loss function, unifying the learning of new knowledge, the retention of old knowledge, and the limitation of model parameter variation under a single framework. This allows for flexible control of the model's learning behavior, enabling the model to continuously learn and evolve.

[0036] In addition, the first knowledge loss is L new = CrossEntropy(Model(z; θ), y), Second knowledge loss L replay = CrossEntropy( Model(z; θ), y ), where the Elastic Weight Consolidation (EWC) regularization term is L ewc = ∑ i (F i / 2) * (θ i - θ* i ) 2 .

[0037] In an optional embodiment, acquiring environmental data and user interaction data within a target time period and performing online incremental learning on the smart home appliance model further includes: acquiring idle device resources, including CPU utilization, GPU utilization, and available memory; classifying idle device resources into levels based on idle device resources to obtain idle device levels; determining the number of batch learning samples for the corresponding number of iterations of online incremental learning based on the idle device levels, and acquiring environmental data and user interaction data within the target time period, as well as historical joint representations in the data buffer and historical user satisfaction corresponding to each historical joint representation, based on the number of batch learning samples, to perform online incremental learning on the smart home appliance model; determining the cumulative gradient adjustment steps based on the idle device levels, and determining the update gradient of the model parameters using the total loss function during each online incremental learning, storing the update gradient in the accumulator and accumulating it with the previously stored update gradient until the cumulative gradient adjustment steps are reached, updating the model parameters based on the update gradient accumulated in the accumulator, and then clearing the accumulator.

[0038] It's important to note that by acquiring available device resources and determining the device's idle level, the corresponding batch learning sample size can be determined based on this level. When the device is busy, the batch learning sample size can be reduced, lowering the memory and peak computing power required for a single computation. This fundamentally avoids the risk of system lag or crashes caused by learning tasks. Conversely, when the device is idle (e.g., late at night), a larger batch learning sample size can be used, utilizing ample resources for more efficient and stable learning. Furthermore, the cumulative step count for gradient adjustment can be determined based on the device's idle level. This allows for simulating a stable learning effect with a corresponding batch learning sample size, using the appropriate cumulative step count based on device activity levels. This ensures learning quality even under resource constraints, making small-sample learning still effective.

[0039] Furthermore, based on the idle resources of the devices, the idle resources are classified into levels to obtain device idle levels. This includes classifying the idle resources using preset resource utilization thresholds. It should be noted that the number and size of preset resource utilization thresholds can be set according to the actual level design requirements and the corresponding device resource types for each level. For example, if the actual level design requirements are low, medium, and high, and the device resource type is GPU utilization, then the preset resource utilization thresholds include 40% and 80%. Idle resources with GPU utilization less than or equal to 40% are classified as low-level, idle resources with GPU utilization greater than 40% and less than or equal to 80% are classified as medium-level, and idle resources with GPU utilization greater than 80% are classified as high-level. No further restrictions are imposed here.

[0040] Step S22: Evaluate the performance of the smart home appliance model after incremental learning, and upload the incremental parameters of the model to the cloud server if the performance meets the set target performance requirements, so that the cloud server can maintain the model based on the uploaded incremental parameters.

[0041] In this embodiment, the performance of the incrementally learned smart home appliance model is evaluated, and the incremental parameters of the model are uploaded to the cloud server based on the performance meeting the set target performance requirements. This includes: identifying the model type of the smart home appliance model; determining the corresponding evaluation indicators, test set generation strategy, stability detection method, and decision threshold based on the model type and a preset model performance evaluation strategy; the preset model performance evaluation strategy is obtained by configuring corresponding evaluation indicators, test set generation strategies, stability detection methods, and decision threshold conditions based on different model types; generating a corresponding test set according to the test set generation strategy; the test set generation strategy is used for classification and regression models, A static test set is constructed by fusing the latest target window data and extreme historical data that challenge the model's boundaries. For the reinforcement learning model, a dynamic interactive simulation test environment is constructed by replaying key historical scenarios and injecting sensor perturbations. The incrementally learned smart home appliance model is tested using the test set, and evaluated according to evaluation metrics and stability detection methods to obtain corresponding metric evaluation results and stability detection results. Based on the metric evaluation results, a baseline comparison is made with the metric evaluation results of the previous version of the smart home appliance model on the same test set to determine the relative improvement. When both the relative improvement and the stability detection results are greater than the corresponding decision threshold, the incremental parameters of the model are triggered for uploading.

[0042] It should be noted that by identifying model types, corresponding evaluation metrics, test set generation strategies, stability testing methods, and decision thresholds are determined based on the model type. This ensures that different evaluation strategies are configured for different model types, and corresponding test sets are generated for each evaluation strategy to evaluate the corresponding model type. Baseline comparison is then performed by combining the evaluation of the previous version of the model, ensuring that each update is a positive and meaningful improvement. This prevents model performance degradation due to catastrophic forgetting or improper training, ensuring that the quality of the local model improves in a spiral manner. Furthermore, the stability test results are used to verify whether the improvement is significant and reliable. Based on the corresponding decision thresholds, automated decision-making is achieved, ensuring that only high-quality model increments are uploaded to the cloud server, avoiding the waste of network bandwidth and cloud computing resources by aggregating useless model updates, thereby guaranteeing the global convergence speed and the final model quality.

[0043] It should be added that when both the relative improvement and stability test results meet the corresponding decision thresholds, the upload of model incremental parameters is triggered. This includes: when the relative improvement is greater than the corresponding decision threshold and the stability test result is greater than the corresponding decision threshold, the upload of model incremental parameters is triggered.

[0044] In addition, performance evaluation of the smart home appliance model after incremental learning also includes: continuing local learning when the relative improvement is less than or equal to the corresponding decision threshold and the stability test result is less than or equal to the corresponding decision threshold.

[0045] Furthermore, the model types include classification models, regression models, and reinforcement learning models. When the model type is a classification model, the corresponding evaluation metric can be a weighted F1 score, the test set generation strategy can be a sliding window or adversarial sampling, etc., the stability detection method can be accuracy and KL divergence, and the decision threshold can be an F1 improvement of >5% and a confidence interval width of <0.1. When the model type is a regression model, the corresponding evaluation metrics can be mean absolute error, root mean square error, and dynamic time warping similarity, the test set generation strategy can be a sliding window or boundary stress test, etc., the stability detection method can be rolling standard deviation analysis, and the decision threshold can be a reduction in MAE of >10% and a reduction in error standard deviation of >5%. When the model type is a reinforcement learning model, the corresponding evaluation metrics can be cumulative reward, policy entropy, and convergence steps, the test set generation strategy can be simulated environment replay or perturbation simulation, the stability detection method can be multi-round independent operation, and the decision threshold can be a cumulative reward improvement of >8% and a reduction in reward standard deviation of >10%.

[0046] Specifically, according to the test set generation strategy, a corresponding test set is generated, including: when the model type is determined to be a classification model, based on the test set generation strategy, using a target window to slide and select the current data, and combining it with predefined extreme scenarios to synthesize joint representation data, and assigning satisfaction labels to the joint representation data to construct the test set; wherein, the current data includes joint representations and user satisfaction corresponding to each joint representation; when the model type is determined to be a regression model, based on the test set generation strategy, using a target window to slide and select the current data, and extracting historical data corresponding to historical extreme events to construct the test set; wherein, the historical data includes historical joint representations and user satisfaction corresponding to each historical joint representation; when the model type is determined to be a reinforcement learning model, using pre-labeled key historical scenarios and sensor noise configurations, a dynamic interactive simulated test environment is constructed; wherein, key historical scenarios include a complete environmental state sequence that leads to extremely low rewards or policy oscillations, and sensor noise configurations include random perturbations injected into the environmental states.

[0047] In addition, the evaluation is conducted using stability testing methods, including: when the model type is determined to be a classification model, based on stability testing methods, the test set is resampled with replacement, the performance index difference between the previous version of the smart home appliance model and the incrementally learned smart home appliance model on each resampled subset is calculated, and an empirical distribution of the performance index difference is constructed to determine the confidence interval of the performance index difference; the width of the confidence interval is used to evaluate the statistical reliability of the performance improvement; when the model type is determined to be a regression model, based on stability testing methods, the prediction error sequence of the previous version of the smart home appliance model and the incrementally learned smart home appliance model on each sample of the test set is calculated, and the dispersion of the prediction error sequence is analyzed; the rolling standard deviation of the prediction error sequence is used to evaluate the volatility of the model's prediction performance at different time points; when the model type is determined to be a reinforcement learning model, based on stability testing methods, the new policy is run independently for multiple rounds in a simulated test environment, and the cumulative reward sequence of each round is recorded; the mean and standard deviation of the cumulative reward sequence are used to evaluate the average performance and robustness of the policy when repeatedly facing similar scenarios.

[0048] It is worth noting that when the model type is a reinforcement learning model, after constructing a dynamic interactive simulation test environment, the previous version of the smart home appliance model and the incrementally learned smart home appliance model are respectively placed into this reconstructed simulation environment, allowing them to start from the beginning of the scene and make autonomous decisions until the end of the scene.

[0049] In one alternative embodiment, reference Figure 3 The global model is obtained by the cloud server based on all received model incremental parameters, allocating aggregate weights according to the data quality uploaded by each device, obtaining global incremental parameters, and then optimizing the global incremental parameters using a generative adversarial network.

[0050] It should be noted that after the cloud server receives the incremental model parameters uploaded by each device, it aggregates the received incremental model data from each device and generates an optimized general model version.

[0051] Furthermore, after receiving the incremental model parameters uploaded by each device, the cloud server normalizes each incremental model parameter to eliminate the impact of hardware differences.

[0052] In addition, if the smart home appliance model structures of different devices are different, parameter mapping (such as low-rank decomposition) can be used to align them to the cloud server model.

[0053] In addition, when uploading incremental model parameters, the device also uploads data quality data, such as signal-to-noise ratio and label confidence, so that the cloud server can allocate aggregation weights to each model's incremental data based on data quality when performing global aggregation.

[0054] In one alternative embodiment, after global aggregation and optimization using generative adversarial networks on the cloud server, the resulting global model undergoes quality verification and version management, and is then sent to the model repository.

[0055] Specifically, quality verification can be achieved by verifying the correctness of the model, testing end-to-end performance in a simulation environment, and conducting adversarial verification, such as injecting anomalous data to test robustness. The specific configuration can be tailored to actual design requirements, and no further limitations are made here.

[0056] In addition, version management uses semantic markup for each version, such as v2.1.3 (major version.feature.fix), and can prevent tampering by hashing the model and putting metadata on the blockchain. The specific configuration can be configured according to actual design requirements, and no further limitations are made here.

[0057] In addition, on the device side, it actively detects whether the model repository has been updated in order to download the model, or receives a global model actively distributed by the cloud server and performs A / B testing to select the model with better effect or performance.

[0058] Furthermore, A / B testing is used to run the old and new models in parallel and automatically switch between them based on user feedback.

[0059] In summary, this invention acquires environmental and user interaction data within a target timeframe and performs online incremental learning on the smart home appliance model. By continuously learning the latest environmental and user interaction data, the model can capture subtle changes in user preferences in real time. The learning process is completed on a local device, using original, highly sensitive user data to fundamentally protect user privacy. Simultaneously, local decision-making avoids network request latency, resulting in faster response times and a smoother user experience. Furthermore, by evaluating the performance of the incrementally learned smart home appliance model and uploading incremental model parameters to a cloud server based on performance improvements exceeding a target performance threshold, a large number of invalid or even negative model updates are avoided, preventing pollution of the global model and protecting its performance. This also significantly saves network bandwidth and cloud computing resources, making the entire system more efficient and robust.

[0060] In addition, by conducting performance evaluations on the local model and the global model updated by the cloud server, and selecting the smart home appliance model based on the performance evaluation results, it is ensured that the device uses the most suitable model at any time, avoiding the performance degradation that may be caused by blind updates. Moreover, the global model is obtained by the cloud server based on global aggregation and optimization, so that the global model can learn a wider range of more robust patterns, and its generalization ability far exceeds that of any single local model.

[0061] The following describes the smart home appliance model updating device provided in this application. The smart home appliance model updating device described below and the smart home appliance model updating method described above can be referred to in correspondence.

[0062] Figure 4 A schematic diagram of a smart home appliance model update device is shown. The device includes: The model update module 41 incrementally updates the smart home appliance model by acquiring environmental data and / or user interaction data within the target time period to obtain model incremental parameters. The smart home appliance model is determined based on the performance evaluation results of the local model and the global model. The global model is a model maintained by the cloud server. The model incremental parameters are used to characterize the parameter changes after the incremental update of the smart home appliance model. The parameter upload module 42 performs performance evaluation on the smart home appliance model after incremental learning, and uploads the incremental parameters of the model to the cloud server based on the performance meeting the set target performance requirements, so that the cloud server can maintain the model according to the uploaded incremental parameters.

[0063] In this embodiment, the model update module 41 is used to: incrementally update the smart home appliance model according to a preset iteration strategy by acquiring environmental data and / or user interaction data within a target time period; wherein, the preset iteration strategy is used to: for each iteration, construct a first knowledge loss based on the model prediction results obtained from the incremental update and the corresponding determined user satisfaction; construct a second knowledge loss using historical joint representations obtained from the data buffer and the historical user satisfaction corresponding to each historical joint representation; and construct an elastic weight consolidation EWC regularization term based on the model parameters obtained after the previous incremental learning and the current model parameters, combined with the preset importance of each model parameter; obtain a total loss function based on the first knowledge loss, the second knowledge loss, and the EWC regularization term; perform back gradient updates on the model parameters based on the total loss function; and use the updated model parameters to perform the next online incremental learning again until the preset number of iterations is reached.

[0064] Furthermore, the model update module 41 is also used to: determine a joint representation by acquiring environmental data and / or user interaction data within a target time period, combined with a cross-modal attention mechanism, and incrementally update the smart home appliance model according to a preset iterative strategy using the joint representation.

[0065] Specifically, the model update module 41 includes: a feature extraction unit, which extracts features from the environmental data and / or user interaction data acquired within the target time period to obtain corresponding temporal environmental features and user intent features; and a cross-modal fusion unit, which determines a joint representation based on the temporal environmental features and user intent features, combined with a cross-modal attention mechanism, and uses the joint representation to incrementally update the smart home appliance model according to a preset iterative strategy.

[0066] Furthermore, the cross-modal fusion unit includes: an attention subunit, which uses a multi-head attention mechanism to take user intent features as the query and temporal environment features as the key and value to determine the attention score and generate a context-aware environment vector; and a fusion subunit, which concatenates the context-aware environment vector with the user intent features to obtain a joint representation.

[0067] In addition, the device also includes a data filtering module, which, before constructing the second loss function by obtaining historical joint representations and corresponding historical user satisfaction from the data buffer, uses a K-center greedy algorithm to filter and cache all historical joint representations and corresponding historical user satisfaction in the data buffer.

[0068] Furthermore, the data filtering module is also used to: take the joint representation and the corresponding historical user satisfaction as a set of samples, randomly select a sample from the sample pool as the first center, and put it into the data buffer; traverse each sample in the sample pool and determine the distance to the sample in the data buffer to obtain the corresponding sample distance; based on the distance of each sample, select the sample in the sample pool that is farthest away and has not been selected as the center, and store it as the new center in the data buffer; traverse each sample in the sample pool again, determine the distance between the corresponding selected sample and each sample in the data buffer, and select the minimum distance as the distance between the corresponding selected sample and the sample in the data buffer; based on the distance between each selected sample and the sample in the data buffer, select the sample in the sample pool that is farthest away and has not been selected as the center, and store it as the new center in the data buffer, and iterate this step again until the number of samples in the data buffer reaches the preset number, thus obtaining a data buffer that protects the preset number of samples.

[0069] In an optional embodiment, the model update module 41 further includes: a resource acquisition unit for acquiring idle device resources, including CPU utilization, GPU utilization, and available memory; a resource rating unit for classifying idle device resources into levels based on the idle device resources to obtain device idle levels; a batch training unit for determining the number of batch learning samples for online incremental learning corresponding to the number of iterations based on the device idle levels, and acquiring environmental data and user interaction data within the corresponding target time period, as well as historical joint representations in the data buffer and historical user satisfaction corresponding to each historical joint representation, based on the number of batch learning samples, to perform online incremental learning on the smart home appliance model; and an accumulated gradient update unit for determining the cumulative gradient adjustment steps based on the device idle levels, and determining the update gradient of the model parameters using the total loss function during each online incremental learning, storing the update gradient in the accumulator and accumulating it with the previously stored update gradient until the cumulative gradient adjustment steps are reached, updating the model parameters based on the update gradient accumulated in the accumulator, and clearing the accumulator.

[0070] Furthermore, the resource rating unit is used to classify idle resources of devices according to preset resource utilization thresholds. It should be noted that the number and size of preset resource utilization thresholds can be set according to the actual level design requirements and the corresponding device resource types for each level. For example, if the actual level design requirements are low, medium, and high, and the device resource type is GPU utilization, then the preset resource utilization thresholds include 40% and 80%. Idle resources of devices with GPU utilization less than or equal to 40% are classified as low-level, idle resources of devices with GPU utilization greater than 40% and less than or equal to 80% are classified as medium-level, and idle resources of devices with GPU utilization greater than 80% are classified as high-level. No further limitations are made here.

[0071] In this embodiment, the parameter uploading module 42 includes: a type identification unit for identifying the model type of the smart home appliance model; a strategy determination unit for determining the corresponding evaluation index, test set generation strategy, stability detection method, and decision threshold based on the model type and a preset model performance evaluation strategy; the preset model performance evaluation strategy is obtained by configuring the corresponding evaluation index, test set generation strategy, stability detection method, and decision threshold based on different model types; and a test set generation unit for generating a corresponding test set according to the test set generation strategy; the test set generation strategy is used for: for classification and regression models, by fusing the latest target window data and extreme historical data that challenge the model boundary. The system uses data to construct a static test set; for reinforcement learning models, it constructs a dynamic interactive simulation test environment by replaying key historical scenarios and injecting sensor perturbations; the evaluation unit uses the test set to test the incrementally learned smart home appliance model and evaluates it according to evaluation indicators and stability detection methods to obtain corresponding indicator evaluation results and stability detection results; the baseline comparison unit compares the indicator evaluation results with the indicator evaluation results of the previous version of the smart home appliance model on the same test set to determine the relative improvement; the parameter upload decision unit triggers the upload of incremental model parameters when the relative improvement and the stability detection results are both greater than the corresponding decision thresholds.

[0072] It should be added that the parameter upload decision unit is also used to: trigger the upload of incremental model parameters when the relative improvement is greater than the corresponding decision threshold and the stability test result is greater than the corresponding decision threshold, based on the decision threshold.

[0073] In addition, the parameter upload module 42 is also used to: continue local learning when the relative improvement is less than or equal to the corresponding decision threshold and the stability detection result is less than or equal to the corresponding decision threshold.

[0074] Specifically, the test set generation unit is used to: when the model type is determined to be a classification model, based on the test set generation strategy, use a target window to slide and select the current data, synthesize joint representation data by combining it with predefined extreme scenarios, and assign satisfaction labels to the joint representation data to construct a test set; wherein, the current data includes joint representations and user satisfaction corresponding to each joint representation; when the model type is determined to be a regression model, based on the test set generation strategy, use a target window to slide and select the current data, and extract historical data corresponding to historical extreme events to construct a test set; wherein, the historical data includes historical joint representations and user satisfaction corresponding to each historical joint representation; when the model type is determined to be a reinforcement learning model, use pre-labeled key historical scenarios and sensor noise configuration to construct a dynamically interactive simulated test environment; wherein, the key historical scenarios include a complete environmental state sequence that leads to extremely low rewards or policy swings, and the sensor noise configuration includes random perturbations injected into the environmental states.

[0075] Additionally, the evaluation unit is used for: when the model type is determined to be a classification model, resampling the test set with replacement based on a stability detection method, calculating the performance index difference between the previous version of the smart home appliance model and the incrementally learned smart home appliance model on each resampled subset, constructing an empirical distribution of the performance index difference, and determining the confidence interval of the performance index difference; wherein, the width of the confidence interval is used to evaluate the statistical reliability of the performance improvement; when the model type is determined to be a regression model, calculating the prediction error sequence of each sample on the test set for the previous version of the smart home appliance model and the incrementally learned smart home appliance model based on a stability detection method, and analyzing the dispersion of the prediction error sequence; wherein, the rolling standard deviation of the prediction error sequence is used to evaluate the volatility of the model's prediction performance at different time points; when the model type is determined to be a reinforcement learning model, independently running the new policy multiple rounds in a simulated test environment based on a stability detection method, and recording the cumulative reward sequence of each round; wherein, the mean and standard deviation of the cumulative reward sequence are used to evaluate the average performance and robustness of the policy when repeatedly facing similar scenarios.

[0076] In one alternative embodiment, the global model is obtained by the cloud server based on all received model incremental parameters, allocating aggregate weights according to the data quality uploaded by each device, obtaining global incremental parameters, and then optimizing the global incremental parameters using a generative adversarial network.

[0077] It should be noted that after the cloud server receives the incremental model parameters uploaded by each device, it aggregates the received incremental model data from each device and generates an optimized general model version.

[0078] Furthermore, after receiving the incremental model parameters uploaded by each device, the cloud server normalizes each incremental model parameter to eliminate the impact of hardware differences.

[0079] In addition, if the smart home appliance model structures of different devices are different, parameter mapping (such as low-rank decomposition) can be used to align them to the cloud server model.

[0080] In addition, when uploading incremental model parameters, the device also uploads data quality data, such as signal-to-noise ratio and label confidence, so that the cloud server can allocate aggregation weights to each model's incremental data based on data quality when performing global aggregation.

[0081] In one alternative embodiment, after global aggregation and optimization using generative adversarial networks on the cloud server, the resulting global model undergoes quality verification and version management, and is then sent to the model repository.

[0082] In summary, this embodiment of the invention acquires environmental and user interaction data within a target time period through a model update module and performs online incremental learning on the smart home appliance model. By continuously learning the latest environmental and user interaction data, the model can capture subtle changes in user preferences in real time. The learning process is completed on a local device, using original, highly sensitive user data, fundamentally protecting user privacy. Simultaneously, local decision-making avoids network request latency, resulting in faster response times and a smoother user experience. Furthermore, a parameter upload module evaluates the performance of the incrementally learned smart home appliance model. If the performance improvement exceeds the target performance threshold, the incremental model parameters are uploaded to the cloud server. This avoids a large number of invalid or even negative model updates that could pollute the global model, thus protecting its performance. It also significantly saves network bandwidth and cloud computing resources, making the entire system run more efficiently and robustly.

[0083] In addition, by conducting performance evaluations on the local model and the global model updated by the cloud server, and selecting the smart home appliance model based on the performance evaluation results, it is ensured that the device uses the most suitable model at any time, avoiding the performance degradation that may be caused by blind updates. Moreover, the global model is obtained by the cloud server based on global aggregation and optimization, so that the global model can learn a wider range of more robust patterns, and its generalization ability far exceeds that of any single local model.

[0084] Figure 5 A schematic diagram of the physical structure of an electronic device is provided. This electronic device may include a processor 510, a communication interface 520, a memory 530, and a communication bus 540. The processor 510, communication interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a smart home appliance model update method. This method includes: incrementally updating the smart home appliance model using environmental data and / or user interaction data acquired within a target time period to obtain model incremental parameters; wherein the smart home appliance model is determined based on the performance evaluation results of a local model and a global model, the global model being a model maintained by a cloud server, and the model incremental parameters characterizing the parameter changes after the incremental update of the smart home appliance model; evaluating the performance of the incrementally learned smart home appliance model, and, based on the performance meeting set target performance conditions, uploading the model incremental parameters to the cloud server for model maintenance by the cloud server.

[0085] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0086] On the other hand, this application also provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can execute the smart home appliance model update method provided by the above methods. The method includes: incrementally updating the smart home appliance model by acquiring environmental data and / or user interaction data within a target time period to obtain model incremental parameters; wherein, the smart home appliance model is determined based on the performance evaluation results of a local model and a global model, the global model is a model maintained by a cloud server, and the model incremental parameters are used to characterize the parameter change after the incremental update of the smart home appliance model; performing performance evaluation on the incrementally learned smart home appliance model, and uploading the model incremental parameters to the cloud server based on the performance meeting the set target performance conditions, so that the cloud server can maintain the model according to the uploaded model incremental parameters.

[0087] In another aspect, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the aforementioned smart home appliance model update methods. The method includes: incrementally updating the smart home appliance model using environmental data and / or user interaction data acquired within a target time period to obtain model incremental parameters; wherein the smart home appliance model is determined based on performance evaluation results of a local model and a global model, the global model being a model maintained by a cloud server, and the model incremental parameters characterizing the parameter changes after the incremental update of the smart home appliance model; evaluating the performance of the incrementally learned smart home appliance model, and, based on the performance meeting set target performance conditions, uploading the model incremental parameters to the cloud server for model maintenance by the cloud server using the uploaded model incremental parameters.

[0088] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0089] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for updating a smart home appliance model, characterized in that, include: By acquiring environmental data and / or user interaction data within the target time period, the smart home appliance model is incrementally updated to obtain model incremental parameters; wherein, the smart home appliance model is determined based on the performance evaluation results of the local model and the global model, the global model is a model maintained by a cloud server, and the model incremental parameters are used to characterize the parameter changes after the incremental update of the smart home appliance model. The performance of the smart home appliance model after incremental learning is evaluated, and if the performance meets the set target performance requirements, the incremental parameters of the model are uploaded to the cloud server so that the cloud server can maintain the model based on the uploaded incremental parameters.

2. The intelligent home appliance model update method according to claim 1, characterized in that, Acquire environmental data and / or user interaction data within a target time period, and incrementally update the smart home appliance model, including: The smart home appliance model is incrementally updated according to a preset iteration strategy based on environmental data and / or user interaction data acquired within a target time period; wherein, the preset iteration strategy is used for: For each iteration, a first knowledge loss is constructed based on the model prediction results obtained from the incremental update and the corresponding determined user satisfaction. A second knowledge loss is constructed using the historical joint representations obtained from the data buffer and the historical user satisfaction corresponding to each historical joint representation. Furthermore, an elastic weight consolidation EWC regularization term is constructed based on the model parameters obtained after the previous incremental learning and the current model parameters, combined with the preset importance of each model parameter. Based on the first knowledge loss, the second knowledge loss, and the EWC regularization term, a total loss function is obtained. Based on the total loss function, the model parameters are updated using backward gradients, and the updated model parameters are used to perform the next online incremental learning until the preset number of iterations is reached.

3. The intelligent home appliance model update method according to claim 2, characterized in that, The process includes acquiring environmental and user interaction data within a target timeframe, performing online incremental learning on the smart home appliance model, and uploading the incremental model parameters to a cloud server based on performance meeting set target requirements. It also includes: Obtain the device's idle resources, including CPU utilization, GPU utilization, and available memory; Based on the idle resources of the equipment, the idle resources of the equipment are classified into different levels to obtain the idle resource levels of the equipment; Based on the device idle level, determine the corresponding batch learning sample quantity, and according to the batch learning sample quantity, obtain the environmental data and user interaction data within the corresponding target time period, as well as the historical joint representation in the data buffer and the historical user satisfaction corresponding to each historical joint representation, so as to perform online incremental learning on the smart home appliance model. Based on the device idle level, the cumulative gradient adjustment steps are determined. During each online incremental learning, the total loss function is used to determine the update gradient of the model parameters. The update gradient is stored in the accumulator and accumulated with the previously stored update gradient until the cumulative gradient adjustment steps are reached. Then, the model parameters are updated based on the update gradient accumulated in the accumulator, and the accumulator is cleared.

4. The intelligent home appliance model update method according to claim 1, characterized in that, The performance of the incrementally learned smart home appliance model is evaluated, including: Identify the model type of smart home appliance models; Based on the model type and in conjunction with the preset model performance evaluation strategy, the corresponding evaluation indicators, test set generation strategy, stability detection method, and decision threshold are determined; the preset model performance evaluation strategy is obtained by configuring the corresponding evaluation indicators, test set generation strategy, stability detection method, and decision threshold based on different model types in advance. According to the test set generation strategy, a corresponding test set is generated; the test set generation strategy is used to: construct a static test set for classification and regression models by fusing the latest target window data and extreme historical data that challenge the model boundary; and construct a dynamic interactive simulation test environment for reinforcement learning models by replaying key historical scenarios and injecting sensor perturbations. The incrementally learned smart home appliance model is tested using the test set, and evaluated according to the evaluation index and the stability detection method respectively, to obtain the corresponding index evaluation results and stability detection results; Based on the evaluation results of the aforementioned indicators, a baseline comparison is made with the evaluation results of the previous version of the smart home appliance model on the same test set to determine the relative improvement. When both the relative improvement and the stability detection results meet the corresponding decision thresholds, the upload of incremental model parameters is triggered.

5. The intelligent home appliance model update method according to claim 4, characterized in that, According to the test set generation strategy, generating the corresponding test set also includes: When the model type is determined to be a classification model, based on the test set generation strategy, the current data is selected by sliding the target window, and joint representation data is synthesized by combining predefined extreme scenarios, and the joint representation data is assigned a satisfaction label to construct a test set; wherein, the current data includes joint representations and user satisfaction corresponding to each joint representation; When the model type is determined to be a regression model, based on the test set generation strategy, the current data is selected by sliding the target window, and historical data corresponding to historical extreme events are extracted to construct the test set; wherein, the historical data includes historical joint representations and user satisfaction corresponding to each historical joint representation; When the model type is determined to be a reinforcement learning model, a dynamic interactive simulation test environment is constructed using pre-labeled key historical scenarios and sensor noise configurations; wherein, the key historical scenarios include a complete sequence of environment states that lead to extremely low rewards or policy swings, and the sensor noise configurations include random perturbations injected into the environment states.

6. The intelligent home appliance model update method according to claim 4, characterized in that, The stability testing method described above is used for evaluation, including: When the model type is determined to be a classification model, based on the stability detection method, the test set is resampled with replacement, the performance index difference between the previous version of the smart home appliance model and the incrementally learned smart home appliance model on each resampled subset is calculated, and an empirical distribution of the performance index difference is constructed to determine the confidence interval of the performance index difference; wherein, the width of the confidence interval is used to evaluate the statistical reliability of the performance improvement. When the model type is determined to be a regression model, based on the stability detection method, the prediction error sequence of the previous version of the smart home appliance model and the incrementally learned smart home appliance model on each sample of the test set is calculated, and the dispersion of the prediction error sequence is analyzed; wherein, the rolling standard deviation of the prediction error sequence is used to evaluate the volatility of the model's prediction performance at different time points. When the model type is determined to be a reinforcement learning model, the new policy is independently run for multiple rounds in the simulation test environment based on the stability detection method, and the cumulative reward sequence of each round is recorded; wherein, the mean and standard deviation of the cumulative reward sequence are used to evaluate the average performance and robustness of the policy when repeatedly facing similar scenarios.

7. The method for updating a smart home appliance model according to any one of claims 1 to 6, characterized in that, The global model is obtained by the cloud server based on all received model incremental parameters, allocating aggregate weights according to the data quality uploaded by each device, obtaining global incremental parameters, and then optimizing the global incremental parameters using a generative adversarial network.

8. A smart home appliance model updating device, characterized in that, include: The model update module incrementally updates the smart home appliance model by acquiring environmental data and / or user interaction data within the target time period to obtain model incremental parameters. The smart home appliance model is determined based on the performance evaluation results of the local model and the global model. The global model is a model maintained by a cloud server. The model incremental parameters are used to characterize the parameter changes after the incremental update of the smart home appliance model. The parameter upload module performs performance evaluation on the incrementally learned smart home appliance model, and uploads the incremental parameters of the model to the cloud server based on the performance meeting the set target performance requirements, so that the cloud server can maintain the model according to the uploaded incremental parameters.

9. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the smart home appliance model update method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the smart home appliance model update method as described in any one of claims 1 to 7.

Images