Remote control method and system for multiple kitchen appliances based on context awareness

Abstract

The invention provides a multi-kitchen-appliance remote control method and system based on context awareness, and relates to the technical field of smart home, and the method comprises the steps: collecting kitchen environment data and user historical use data, and carrying out the wavelet multi-resolution decomposition and density gradient matrix recognition of abnormal data; the method comprises the following steps: extracting data time sequence features by combining a bidirectional recurrent neural network and performing hierarchical coding, inputting a fused feature vector and a density distribution graph into an improved long-short-term memory network, performing feature propagation by combining a Fisher information matrix and a time-varying association graph, and predicting a kitchen appliance use trend by utilizing a prophet model. And based on a risk assessment model and a Monte Carlo tree search method, in combination with an adaptive exploration mechanism and a multi-target evaluation system, generating and assessing a candidate control strategy, and outputting an optimal kitchen electricity cooperative control strategy.

Description

Technical Field

[0001] The present invention relates to the field of smart home technology, and in particular to a scenario-aware multi-kitchen appliance remote control method and system. Background Art

[0002] Kitchen appliance remote control technology has developed rapidly with the rise of the Internet of Things and smart homes. Users can remotely control kitchen appliances through mobile phones or other smart terminals, such as preheating the oven in advance, remotely starting the rice cooker, etc., which greatly improves the convenience and efficiency of cooking. Existing remote control technologies mainly rely on scheduled appointments or online manual control, and lack the ability to perceive and learn the kitchen environment and user habits; The existing remote control technology for kitchen appliances has the following problems: it cannot adapt to complex kitchen environments, lacks understanding of users' personalized needs, and is difficult to achieve coordinated control of multiple kitchen appliances. Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0003] The embodiments of the present invention provide a method and system for remotely controlling multiple kitchen appliances based on situational awareness, which can at least solve some of the problems existing in the prior art.

[0004] According to a first aspect of an embodiment of the present invention, a method for remotely controlling multiple kitchen appliances based on context awareness is provided, comprising: Collecting kitchen environmental data and recording historical usage data of kitchen appliances by users, performing wavelet multi-resolution decomposition on the environmental data, calculating the energy distribution feature matrix under each scale coefficient and performing hierarchical screening on the wavelet coefficients, combining reverse reconstruction to obtain filtered high-quality environmental data, constructing a density distribution map based on the high-quality environmental data and determining a density gradient matrix, setting a density threshold curve based on the density gradient matrix and identifying abnormal data points to obtain purified environmental data, adding the purified environmental data and the historical usage data to a bidirectional recursive neural network, constructing a time series feature correlation heat map and determining a mutual information matrix, hierarchically encoding the historical usage data based on the mutual information matrix, extracting high-value data sequences and performing pyramid pooling operations, constructing a multi-scale feature map and a feature interaction network, and obtaining a fused feature vector; Add the fused feature vector and the density distribution map to a pre-set improved long short-term memory network, construct a parameter sensitivity evaluation module based on the density distribution map and calculate the eigenvalue decomposition of the Fisher information matrix to obtain the parameter importance ranking, use the purification environment data as a constraint condition, construct a time-varying association graph and calculate the degree centrality index of the device node, establish a feature propagation channel and perform information interaction in kitchen appliances to obtain device association features, add the device association features and the density distribution map to a pre-set prophet model to construct a polynomial trend term and predict long-term trends and short-term fluctuations, combine the hierarchical attention mechanism to determine the multi-scale temporal dependency, construct a scenario probability distribution function and determine the prediction confidence interval; The scenario probability distribution function, the prediction confidence interval and the density distribution map are added to a pre-set risk assessment model, a strategy search space based on risk constraints is constructed, and control boundary conditions are set based on the purification environment data. The candidate control strategies are solved by combining the Monte Carlo tree search method and mapped to the device control space, a device collaborative control network is constructed and the attention mechanism is combined to control the information flow, the posterior probability distribution of different control actions in the control strategy is determined, and an error gradient-based priority scoring model is constructed in combination with an adaptive exploration mechanism and Bellman error. The hyperparameters of the priority scoring model are updated in combination with an adaptive learning rate regulator based on the prediction confidence interval and the density distribution map, and the candidate control strategies are evaluated in combination with a pre-set multi-objective evaluation system to obtain the optimal control strategy.

[0005] In an optional embodiment, Collect kitchen environmental data and record the user's historical usage data of kitchen appliances, perform wavelet multi-resolution decomposition on the environmental data, calculate the energy distribution feature matrix under each scale coefficient and perform hierarchical screening on the wavelet coefficients, combine reverse reconstruction to obtain filtered high-quality environmental data, build a density distribution map based on the high-quality environmental data and determine the density gradient matrix, set a density threshold curve based on the density gradient matrix and identify abnormal data points to obtain purified environmental data, add the purified environmental data and the historical usage data to a bidirectional recursive neural network, build a time series feature correlation heat map and determine the mutual information matrix, hierarchically encode the historical usage data based on the mutual information matrix, extract high-value data sequences and perform pyramid pooling operations, build a multi-scale feature map and a feature interaction network, and obtain a fused feature vector including: Collecting kitchen environment data and historical usage data of kitchen appliances by users through pre-deployed sensor nodes, wherein the environment data includes temperature parameters, humidity parameters, light intensity parameters and air quality parameters, and the historical usage data includes device usage time, device usage frequency and device usage power; Perform wavelet multi-resolution decomposition on the environmental data through the Daubey wavelet function to obtain wavelet coefficients at multiple scales, calculate the mean square value of the wavelet coefficients to obtain an energy distribution feature matrix, screen the wavelet coefficients according to a preset first energy threshold and a second energy threshold, delete the wavelet coefficients with energy values ​​lower than the first energy threshold, retain the wavelet coefficients with energy values ​​higher than the second energy threshold, and for the wavelet coefficients with energy values ​​greater than the first energy threshold and less than the second energy threshold, calculate the correlation coefficient between the current wavelet coefficient and the adjacent wavelet coefficients, if the correlation coefficient is greater than the preset correlation threshold, retain the current wavelet coefficient, otherwise delete it, and for the retained wavelet coefficients, obtain the filtered high-quality environmental data through reverse reconstruction; Based on the high-quality environmental data, a density distribution map is constructed by a kernel density estimation method, the density distribution map estimates the probability density function of the filtered environmental data based on a Gaussian kernel function, a density gradient matrix of the density distribution map is calculated by a numerical differentiation method, abnormal point candidates are marked based on a preset threshold, the abnormal point candidates are screened in combination with prior knowledge to determine abnormal data points, and the abnormal data points are removed to obtain purified environmental data; The purified environment data and the historical usage data are aligned according to timestamps to form a multivariate time series, the multivariate time series is input into a bidirectional long short-term memory network, the forward time series features and the reverse time series features are extracted respectively through the forward recursive neural network and the backward recursive neural network of the bidirectional long short-term memory network, the forward time series features and the reverse time series features are spliced ​​to obtain a complete time series feature, the correlation matrix of the complete time series feature is calculated by the Pearson correlation coefficient to construct a time series feature correlation heat map, the feature mutual information matrix is ​​calculated, and the hierarchical clustering method is used to hierarchically encode the complete time series features based on the feature mutual information matrix to extract a high-value data sequence; A time pyramid pooling strategy is adopted to perform multi-scale decomposition on the high-value data sequence, local features are extracted by maximum pooling in each time scale, and the local features are spliced ​​in the time dimension to construct a multi-scale feature map. The multi-scale feature map contains global features and local features. A two-layer fully connected neural network is built as a feature interaction network. The input layer of the feature interaction network receives the multi-scale feature map, the feature interaction correlation relationship is modeled through the hidden layer, and the fused feature vector is generated through the output layer.

[0006] In an optional embodiment, The calculation formula of the Pearson correlation coefficient is as follows: ; in, Represents a time series feature set X The standardized value of each element in Represents a time series feature set Y The standardized value of each element in represents the calculated Pearson correlation coefficient, n Represents a time series feature set X and Y The number of features in P ( x i , y i ) is the joint probability distribution, indicating that x i and y i The probability of P ( x i ) is the time series feature set X The marginal probability distribution of x i The probability of P ( y i ) is the time series feature set Y The marginal probability distribution of y i probability.

[0007] In an optional embodiment, Add the fused feature vector and the density distribution map to a pre-set improved long short-term memory network, construct a parameter sensitivity evaluation module based on the density distribution map and calculate the eigenvalue decomposition of the Fisher information matrix to obtain the parameter importance ranking, use the purification environment data as a constraint condition, construct a time-varying association graph and calculate the degree centrality index of the device node, establish a feature propagation channel and perform information interaction in kitchen appliances to obtain device association features, add the device association features and the density distribution map to a pre-set prophet model to construct a polynomial trend term and predict long-term trends and short-term fluctuations, combine the hierarchical attention mechanism to determine the multi-scale temporal dependency, construct a scenario probability distribution function and determine the prediction confidence interval, including: Input the fused feature vector and the density distribution map into a preset improved long short-term memory network, wherein the improved long short-term memory network includes an encoder and a decoder, wherein the encoder uses a bidirectional gated recurrent unit to encode the fused feature vector to extract high-level semantic information, and the decoder uses an attention mechanism to dynamically weight the output of the encoder to generate a context vector, and the decoder concatenates the output of the previous step, the input at the current moment, and the context vector as the input of the gated recurrent unit, and the output layer of the decoder inserts the density distribution map as prior knowledge; A parameter sensitivity evaluation module is constructed based on the density distribution map, and the environmental data is perturbed and sampled through Monte Carlo simulation, and the Fisher information of each environmental data is calculated. The information matrix formed by the Fisher information is decomposed and sorted based on the eigenvalues, and the importance of the environmental data is determined according to the eigenvalues, and a subset of environmental data whose importance is higher than a preset importance threshold is retained; The purified environment data is used as a constraint condition to construct a device association graph, wherein the device is a node and the environment data is an edge. The value of the environment data determines the weight of the edge. The degree centrality index of each device node is calculated, and the fused feature vector is divided into sub-blocks corresponding to different devices. The sub-blocks are transmitted through physical connections between devices. The receiving end device updates the feature representation according to the received sub-blocks, and the device association feature is formed after multiple rounds of iterations; The device association features and the density distribution map are input into a pre-set prophet model, wherein the prophet model is composed of a plurality of time polynomial basis functions, and the coefficients of the basis functions are fitted by the least squares method to obtain a polynomial trend term, and the probability density of the density distribution map is used as the confidence of the trend term for correction, and a hierarchical attention mechanism is superimposed on the prophet model to divide the time series into time segments of different lengths, and the attention mechanism is applied within the segment to calculate the moment weight distribution to extract local features, and the attention mechanism is applied between segments to calculate the segment weight distribution to extract global features; A scenario probability distribution function is constructed, with the environmental data as the independent variable and the device status and user behavior as the dependent variables. The scenario probability distribution function is sampled through Monte Carlo simulation to obtain scenario samples, and cluster analysis is performed on the scenario samples to identify typical scenarios. The sample mean of each cluster is calculated as the scene center point, and the sample variance is calculated as the uncertainty measure. The scene center point is used as the predicted value, and the uncertainty measure is used as the confidence interval to generate a prediction result.

[0008] In an optional embodiment, The coefficients of the basis function are fitted by the least square method to obtain a polynomial trend term, and the probability density of the density distribution map is used as the confidence of the trend term for correction, including: Constructing an adaptive polynomial basis function group, the adaptive polynomial basis function group includes a combination of Chebyshev polynomials and Legendre polynomials, generating a basis function matrix for each moment in the time series, each column of the basis function matrix represents the value of different basis functions at the moment, and using an elastic network least squares method with a first-order norm regularization term and a second-order norm regularization term to solve a polynomial coefficient vector, wherein the polynomial coefficient vector is obtained by minimizing an objective function including a sample loss term, a sparse term, and a smoothing term; Obtain a density distribution map and adaptively normalize the probability density value of the density distribution map to obtain an initial weight coefficient, introduce a weight calibration mechanism based on relative entropy, calculate the relative entropy between the weight coefficient and the standard normal distribution, construct a penalty term based on the relative entropy, integrate the penalty term into the weight calculation process, and obtain a calibrated weight coefficient matrix; Constructing a multi-scale weighted loss function, the multi-scale weighted loss function includes a local fitting error term, a global trend constraint term and a time series smoothing term, wherein the local fitting error term uses a calibrated weight coefficient to weight the sample error, the global trend constraint term introduces trend monotonicity and convexity constraints, and the time series smoothing term uses total variation regularization; The multi-scale weighted loss function is solved by an alternating direction multiplier method to obtain a plurality of sub-optimization problems, wherein the sub-optimization problems include a weight coefficient optimization sub-problem, a polynomial coefficient optimization sub-problem, and a Lagrange multiplier update sub-problem, and a modified polynomial coefficient vector is obtained by iteratively solving the sub-optimization problems; post-processing the corrected trend term according to an adaptive kernel regression smoother, wherein the window size of the adaptive kernel regression smoother is dynamically adjusted according to the local data density, using a larger window in a data-sparse area to enhance stability, and using a smaller window in a data-dense area to retain local features; Constructing an uncertainty estimation module based on repeated random sampling, generating multiple data subsets through repeated sampling, performing a trend correction process on each data subset, calculating the mean and variance of the trend term based on multiple correction results, and using the variance as an uncertainty measure of the trend term; An adaptive evaluation mechanism for the trend correction effect is established, and a comprehensive evaluation index is constructed based on the fitting error, trend smoothness and computational complexity before and after correction. When the comprehensive evaluation index does not reach the preset threshold, the algorithm parameters are automatically adjusted and the correction process is repeated until the convergence conditions are met.

[0009] In an optional embodiment, Adding the scenario probability distribution function, the prediction confidence interval and the density distribution map to a pre-set risk assessment model, constructing a risk constraint-based strategy search space, setting control boundary conditions based on the purification environment data, and solving the candidate control strategies in combination with the Monte Carlo tree search method and mapping them to the equipment control space include: The scenario probability distribution function, the prediction confidence interval and the density distribution map are input into a pre-set risk assessment model. The risk assessment model performs risk assessment on the control strategy based on the conditional value online risk assessment criterion. By setting the prediction confidence interval as the risk assessment confidence interval and setting the low probability area of ​​the density distribution map as the risk threshold, the potential loss of the control strategy in an abnormal scenario is quantified to generate risk constraint conditions. A multidimensional strategy search space is constructed based on the risk constraint condition, wherein each dimension of the multidimensional strategy search space corresponds to a control variable, and the value range of the control variable is determined by the purification environment data. Each point in the multidimensional strategy search space represents a candidate control strategy. The multidimensional strategy search space is pruned and optimized according to the risk constraint condition, and the search area that does not meet the risk constraint condition is deleted; A Monte Carlo tree search method is used to perform a strategy search in the multidimensional strategy search space, including a selection phase, an expansion phase, a simulation phase, and a back-propagation phase. The selection phase includes starting from a root node and iteratively selecting a child node based on a confidence upper bound formula until reaching an incompletely expanded node. The expansion phase includes performing random sampling in a search area of ​​a selected node, generating a new child node, and initializing the number of visits and cumulative value of the new child node. The simulation phase includes starting from the new child node and simulating to a termination state using a fast walk strategy to obtain a simulated reward value. The back-propagation phase includes back-propagating the simulated reward value from the new child node to the root node, and updating the number of visits and cumulative value of all nodes on the path. The approximate optimal strategy obtained by the Monte Carlo tree search is mapped to the device control space through a predefined state-action mapping table, wherein the state-action mapping table contains a mapping relationship matrix from control variables to device states, and the continuous control variables are converted into a discrete device state sequence according to the mapping relationship matrix to generate a device control instruction sequence.

[0010] In an optional embodiment, Construct a device collaborative control network and combine it with the attention mechanism to control information flow, determine the posterior probability distribution of different control actions in the control strategy, combine the adaptive exploration mechanism and Bellman error to construct a priority scoring model based on error gradient, combine the adaptive learning rate regulator based on the prediction confidence interval and the density distribution map to update the hyperparameters of the priority scoring model, and combine the pre-set multi-objective evaluation system to evaluate the candidate control strategy, and obtain the optimal control strategy including: Constructing a device collaborative control network, wherein the device collaborative control network adopts a directed acyclic graph structure, wherein nodes in the graph represent devices, and edges represent the transmission direction of control information, wherein each node in the directed acyclic graph comprises an attention module and a control module, wherein the attention module receives the current environment state vector and the control information vector of other devices as input, calculates the attention weight distribution through a multi-head attention mechanism, and the control module fuses the input information based on the attention weight distribution, and outputs the device control instruction after being processed by a fully connected neural network; Introducing an adaptive exploration mechanism into the equipment collaborative control network, constructing a priority scoring model based on error gradient, taking the state transfer error of each node as the basis for priority scoring, the state transfer error is calculated by the Bellman equation, including calculating the difference between the current state value function and the next state discount value function, taking the gradient direction of the state transfer error as the priority exploration direction, introducing an adaptive learning rate regulator, dynamically adjusting the learning rate according to the position of the current state in the prediction confidence interval and the density distribution map, the prediction confidence of the current state is inversely proportional to the learning rate, and the probability density of the current state is inversely proportional to the learning rate, and balancing the convergence and exploration of the model by dynamically adjusting the learning rate; The candidate control strategies are evaluated based on a pre-set multi-objective evaluation system, wherein the multi-objective evaluation system includes a user satisfaction index, an energy efficiency index, and an equipment health index. Normalized scores are calculated for the user satisfaction index, the energy efficiency index, and the equipment health index, respectively, and they are combined into a comprehensive evaluation index through weighted summation. The control strategy with the highest comprehensive evaluation index is selected as the optimal control strategy.

[0011] A second aspect of an embodiment of the present invention provides a multi-kitchen appliance remote control system based on context perception, comprising: The first unit is used to collect kitchen environment data and record the user's historical usage data of kitchen appliances, perform wavelet multi-resolution decomposition on the environment data, calculate the energy distribution feature matrix under each scale coefficient and perform hierarchical screening on the wavelet coefficients, obtain filtered high-quality environment data by combining reverse reconstruction, construct a density distribution map based on the high-quality environment data and determine a density gradient matrix, set a density threshold curve based on the density gradient matrix and identify abnormal data points to obtain purified environment data, add the purified environment data and the historical usage data to a bidirectional recursive neural network, construct a time series feature correlation heat map and determine a mutual information matrix, hierarchically encode the historical usage data based on the mutual information matrix, extract high-value data sequences and perform pyramid pooling operations, construct a multi-scale feature map and a feature interaction network, and obtain a fused feature vector; The second unit is used to add the fused feature vector and the density distribution map to a pre-set improved long short-term memory network, construct a parameter sensitivity evaluation module based on the density distribution map and calculate the eigenvalue decomposition of the Fisher information matrix to obtain the parameter importance ranking, use the purification environment data as a constraint condition, construct a time-varying association graph and calculate the degree centrality index of the device node, establish a feature propagation channel and perform information interaction in the kitchen appliance to obtain the device association feature, add the device association feature and the density distribution map to the pre-set prophet model to construct a polynomial trend term and predict the long-term trend and short-term fluctuation, determine the multi-scale time series dependency relationship in combination with the hierarchical attention mechanism, construct a scene probability distribution function and determine the prediction confidence interval; The third unit is used to add the scenario probability distribution function, the prediction confidence interval and the density distribution map to a preset risk assessment model, construct a strategy search space based on risk constraints and set control boundary conditions based on the purification environment data, solve the candidate control strategy in combination with the Monte Carlo tree search method and map it to the device control space, build a device collaborative control network and control information flow in combination with the attention mechanism, determine the posterior probability distribution of different control actions in the control strategy, construct a priority scoring model based on the error gradient in combination with the adaptive exploration mechanism and Bellman error, update the hyperparameters of the priority scoring model in combination with the adaptive learning rate regulator based on the prediction confidence interval and the density distribution map, and evaluate the candidate control strategy in combination with a preset multi-objective evaluation system to obtain the optimal control strategy.

[0012] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0013] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0014] In the present invention, through methods such as wavelet multi-resolution decomposition, density distribution map construction and abnormal data point identification, the noise and interference in the environmental data are effectively filtered out to obtain more accurate and reliable environmental data, providing a better data basis for the subsequent control strategy formulation. By using deep learning technologies such as bidirectional recursive neural networks, improved long short-term memory networks and prophet models, combined with historical usage data, environmental data and equipment association characteristics, multi-scale time series dependencies are constructed, and long-term trends and short-term fluctuations are predicted, so as to formulate more accurate and more practical control strategies. Factors such as parameter sensitivity, control boundary conditions and risk constraints are considered to further improve the reliability and effectiveness of the control strategy. The Monte Carlo tree search method is used to solve the candidate control strategies, and combined with methods such as attention mechanism, adaptive exploration mechanism and priority scoring model based on error gradient, intelligent collaborative control of kitchen appliances is realized. Through adaptive learning rate regulator and multi-objective evaluation system, the control strategy can be dynamically adjusted and optimized according to the actual situation, and finally the optimal control strategy is obtained to improve the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the process of a method for remotely controlling multiple kitchen appliances based on context perception according to an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a multi-kitchen appliance remote control system based on scenario perception according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0018] Figure 1 FIG. 1 is a flow chart of a method for remotely controlling multiple kitchen appliances based on context perception according to an embodiment of the present invention. Figure 1 As shown, the method includes: S1. Collect kitchen environment data and record the user's historical usage data of kitchen appliances, perform wavelet multi-resolution decomposition on the environment data, calculate the energy distribution feature matrix under each scale coefficient and perform hierarchical screening on the wavelet coefficients, combine reverse reconstruction to obtain filtered high-quality environment data, construct a density distribution map based on the high-quality environment data and determine the density gradient matrix, set a density threshold curve based on the density gradient matrix and identify abnormal data points to obtain purified environment data, add the purified environment data and the historical usage data to a bidirectional recursive neural network, construct a time series feature correlation heat map and determine the mutual information matrix, hierarchically encode the historical usage data based on the mutual information matrix, extract high-value data sequences and perform pyramid pooling operations, construct a multi-scale feature map and a feature interaction network, and obtain a fused feature vector; The wavelet multi-resolution decomposition is a signal processing technology that decomposes the signal through wavelet transform so that the signal has different resolutions at different scales and frequencies. The density distribution map is a graphical representation used to describe the distribution of a data set in different areas. The distribution characteristics of the data are displayed by plotting the density values ​​of the data points, which helps to identify the clustered areas, outliers and structural characteristics of the data. The density gradient matrix is ​​a tool used in image processing and signal analysis to represent the local density changes of each pixel or data point in the image. By calculating the density gradient between data points, it helps to identify the edges or change areas in the image. It is commonly used in edge detection and feature extraction. The bidirectional recursive neural network is a neural network structure that processes sequence data in both forward and reverse directions so that the model can consider contextual information. The hierarchical encoding is an encoding method that constructs data representation layer by layer and gradually extracts more abstract features from low to high levels. The pyramid pooling operation is a multi-scale pooling method for image processing that obtains richer feature representations by dividing the image into regions of different scales for pooling.

[0019] In an optional embodiment, Collect kitchen environmental data and record the user's historical usage data of kitchen appliances, perform wavelet multi-resolution decomposition on the environmental data, calculate the energy distribution feature matrix under each scale coefficient and perform hierarchical screening on the wavelet coefficients, combine reverse reconstruction to obtain filtered high-quality environmental data, build a density distribution map based on the high-quality environmental data and determine the density gradient matrix, set a density threshold curve based on the density gradient matrix and identify abnormal data points to obtain purified environmental data, add the purified environmental data and the historical usage data to a bidirectional recursive neural network, build a time series feature correlation heat map and determine the mutual information matrix, hierarchically encode the historical usage data based on the mutual information matrix, extract high-value data sequences and perform pyramid pooling operations, build a multi-scale feature map and a feature interaction network, and obtain a fused feature vector including: Collecting kitchen environment data and historical usage data of kitchen appliances by users through pre-deployed sensor nodes, wherein the environment data includes temperature parameters, humidity parameters, light intensity parameters and air quality parameters, and the historical usage data includes device usage time, device usage frequency and device usage power; Perform wavelet multi-resolution decomposition on the environmental data through the Daubey wavelet function to obtain wavelet coefficients at multiple scales, calculate the mean square value of the wavelet coefficients to obtain an energy distribution feature matrix, screen the wavelet coefficients according to a preset first energy threshold and a second energy threshold, delete the wavelet coefficients with energy values ​​lower than the first energy threshold, retain the wavelet coefficients with energy values ​​higher than the second energy threshold, and for the wavelet coefficients with energy values ​​greater than the first energy threshold and less than the second energy threshold, calculate the correlation coefficient between the current wavelet coefficient and the adjacent wavelet coefficients, if the correlation coefficient is greater than the preset correlation threshold, retain the current wavelet coefficient, otherwise delete it, and for the retained wavelet coefficients, obtain the filtered high-quality environmental data through reverse reconstruction; Based on the high-quality environmental data, a density distribution map is constructed by a kernel density estimation method, the density distribution map estimates the probability density function of the filtered environmental data based on a Gaussian kernel function, a density gradient matrix of the density distribution map is calculated by a numerical differentiation method, abnormal point candidates are marked based on a preset threshold, the abnormal point candidates are screened in combination with prior knowledge to determine abnormal data points, and the abnormal data points are removed to obtain purified environmental data; The purified environment data and the historical usage data are aligned according to timestamps to form a multivariate time series, the multivariate time series is input into a bidirectional long short-term memory network, the forward time series features and the reverse time series features are extracted respectively through the forward recursive neural network and the backward recursive neural network of the bidirectional long short-term memory network, the forward time series features and the reverse time series features are spliced ​​to obtain a complete time series feature, the correlation matrix of the complete time series feature is calculated by the Pearson correlation coefficient to construct a time series feature correlation heat map, the feature mutual information matrix is ​​calculated, and the hierarchical clustering method is used to hierarchically encode the complete time series features based on the feature mutual information matrix to extract a high-value data sequence; A time pyramid pooling strategy is adopted to perform multi-scale decomposition on the high-value data sequence, local features are extracted by maximum pooling in each time scale, and the local features are spliced ​​in the time dimension to construct a multi-scale feature map. The multi-scale feature map contains global features and local features. A two-layer fully connected neural network is built as a feature interaction network. The input layer of the feature interaction network receives the multi-scale feature map, the feature interaction correlation relationship is modeled through the hidden layer, and the fused feature vector is generated through the output layer.

[0020] The numerical differentiation is a numerical method for calculating derivatives by approximately calculating the rate of change of a function at a certain point. The Pearson correlation coefficient is a statistic that measures the linear relationship between two variables. Its value range is between -1 and 1. A value of 1 indicates a perfect positive correlation, a value of -1 indicates a perfect negative correlation, and a value of 0 indicates no linear correlation. The hierarchical clustering method is a clustering algorithm that organizes data objects layer by layer into a tree structure. In hierarchical clustering, data points are recursively divided into subsets, which can be bottom-up (aggregation) or top-down (segmentation). It is often used in classification, pattern recognition and data analysis, and is particularly suitable for situations where there are no predefined categories.

[0021] Collect kitchen environment data and user historical power consumption data. Through various sensor nodes pre-deployed in the kitchen, real-time environmental data is collected, including temperature, humidity, light intensity, and air quality (for example, PM2.5 concentration, CO2 concentration, etc.). At the same time, record the user's historical usage data of kitchen appliances, including the usage time, usage frequency, and power usage of each appliance. For example, record the daily usage time, usage frequency, and power usage of the rice cooker in the past week.

[0022] Perform wavelet multi-resolution decomposition and filtering on environmental data. Perform wavelet multi-resolution decomposition on the collected environmental data using the Daubey wavelet function to decompose the data into wavelet coefficients of different scales. Then, calculate the energy value (e.g., mean square value) of the wavelet coefficients at each scale. Set two energy thresholds: a first energy threshold and a second energy threshold. Delete the wavelet coefficients with energy values ​​lower than the first energy threshold, and retain the wavelet coefficients with energy values ​​higher than the second energy threshold. For the wavelet coefficients with energy values ​​between the two thresholds, calculate the correlation coefficients with the adjacent wavelet coefficients. If the correlation coefficient is greater than the preset correlation threshold, the wavelet coefficient is retained; otherwise, it is deleted. For example, if the energy value of the wavelet coefficient at a certain scale is 0.05, the first energy threshold is 0.01, and the second energy threshold is 0.1, then the coefficient needs to calculate the correlation with the adjacent coefficients. Assuming the correlation threshold is 0.8, if the calculated correlation coefficient is 0.9, the coefficient is retained; if the calculated correlation coefficient is 0.7, the coefficient is deleted. Finally, perform reverse reconstruction on the retained wavelet coefficients to obtain filtered high-quality environmental data.

[0023] Construct a density distribution map based on high-quality environmental data and identify abnormal data. Use the kernel density estimation method to construct a density distribution map of the filtered environmental data based on the Gaussian kernel function and estimate the probability density function of the data. Then, use the numerical differentiation method to calculate the density gradient matrix of the density distribution map. According to the preset density gradient threshold curve, identify the abnormal data point candidates. Combined with prior knowledge, for example, the kitchen temperature usually does not exceed 40 degrees Celsius, screen the abnormal point candidates, determine the final abnormal data points, and remove them to obtain the purified environmental data. For example, if the temperature data at a certain point in time is 50 degrees Celsius, which exceeds the normal kitchen temperature range, the data point will be identified as an abnormal point and removed.

[0024] Input the purified environmental data and historical electricity consumption data into the bidirectional recurrent neural network. Align the purified environmental data and historical electricity consumption data according to the timestamp to form multivariate time series data. Input the multivariate time series data into the bidirectional long short-term memory network. The forward recurrent neural network of the bidirectional long short-term memory network extracts the forward time series features, and the backward recurrent neural network extracts the reverse time series features. Concatenate the forward time series features and the reverse time series features to obtain the complete time series features.

[0025] Construct a time series feature correlation heat map and feature mutual information matrix. Calculate the correlation matrix of the complete time series features (for example, using the Pearson correlation coefficient) and visualize it as a time series feature correlation heat map. Then, calculate the feature mutual information matrix to measure the statistical dependence between different features.

[0026] Hierarchical encoding of historical electricity consumption data and extraction of high-value data sequences. Using the hierarchical clustering method, hierarchical encoding of complete time series features is performed based on the feature mutual information matrix. Based on the encoding results, high-value data sequences are extracted, such as electricity consumption data sequences that are highly correlated with environmental data.

[0027] Construct multi-scale feature maps and feature interaction networks. Use the time pyramid pooling strategy to decompose high-value data sequences at multiple scales. Use maximum pooling to extract local features in each time scale. Concatenate local features of different time scales in the time dimension to construct a multi-scale feature map, which contains global features and local features. Build a two-layer fully connected neural network as a feature interaction network. The input layer of the feature interaction network receives the multi-scale feature map, the hidden layer models the feature interaction correlation, and the output layer generates a fused feature vector.

[0028] In this embodiment, through wavelet multi-resolution decomposition and density gradient outlier detection, noise and outliers in the environmental data are effectively removed, data quality is improved, and a reliable data basis is provided for subsequent analysis. The forward and reverse time series features are extracted through a bidirectional LSTM network, and combined with time pyramid pooling and feature interaction network, multi-scale feature maps and fused feature vectors are constructed, which enhances the expressive power of features and can better capture the complex relationship between environmental data and electricity consumption data. The prediction performance of the model can be effectively improved. For example, the electricity consumption in the future can be predicted, thereby achieving smarter energy management.

[0029] In an optional embodiment, The calculation formula of the Pearson correlation coefficient is as follows: ; in, Represents a time series feature set X The standardized value of each element in Represents a time series feature set Y The standardized value of each element in represents the calculated Pearson correlation coefficient, n Represents a time series feature set X and Y The number of features in P ( x i , y i ) is the joint probability distribution, indicating that x i and y i The probability of P ( x i) is the time series feature set X The marginal probability distribution of x i The probability of P ( y i ) is the time series feature set Y The marginal probability distribution of y i probability.

[0030] In this embodiment, the data is converted to the same scale through standardization processing, which simplifies the correlation calculation process. The Pearson correlation coefficient provides a quantitative indicator that can clearly represent the degree of linear correlation between two time series feature sets. By calculating the Pearson correlation coefficient, it can help users understand the relationship between different features, thereby assisting in decision-making.

[0031] S2. Add the fused feature vector and the density distribution map to a pre-set improved long short-term memory network, construct a parameter sensitivity evaluation module based on the density distribution map and calculate the eigenvalue decomposition of the Fisher information matrix to obtain the parameter importance ranking, use the purification environment data as a constraint condition, construct a time-varying association graph and calculate the degree centrality index of the device node, establish a feature propagation channel and perform information interaction in the kitchen appliance to obtain the device association feature, add the device association feature and the density distribution map to the pre-set prophet model to construct a polynomial trend term and predict the long-term trend and short-term fluctuation, combine the hierarchical attention mechanism to determine the multi-scale temporal dependency, construct the scene probability distribution function and determine the prediction confidence interval; The time-varying association graph is a graph structure used to describe the relationship between elements in a system that changes over time. The degree centrality index is an indicator to measure the importance of nodes in the graph, which indicates the number of edges connected to a node. The feature propagation channel refers to the path or channel through which features propagate from node to node in a graph neural network or other graph analysis methods. It describes how nodes update or transfer features based on the information of adjacent nodes, and is typically used for tasks such as node classification and graph embedding. The prophet model is a model for time series prediction, particularly suitable for data with seasonal and holiday effects. The multi-scale time series dependency refers to the dependency of data at different time scales in time series data. The multi-scale modeling method can capture short-term and long-term dependencies, thereby providing more accurate predictions or analysis.

[0032] In an optional embodiment, Add the fused feature vector and the density distribution map to a pre-set improved long short-term memory network, construct a parameter sensitivity evaluation module based on the density distribution map and calculate the eigenvalue decomposition of the Fisher information matrix to obtain the parameter importance ranking, use the purification environment data as a constraint condition, construct a time-varying association graph and calculate the degree centrality index of the device node, establish a feature propagation channel and perform information interaction in kitchen appliances to obtain device association features, add the device association features and the density distribution map to a pre-set prophet model to construct a polynomial trend term and predict long-term trends and short-term fluctuations, combine the hierarchical attention mechanism to determine the multi-scale temporal dependency, construct a scenario probability distribution function and determine the prediction confidence interval, including: Input the fused feature vector and the density distribution map into a preset improved long short-term memory network, wherein the improved long short-term memory network includes an encoder and a decoder, wherein the encoder uses a bidirectional gated recurrent unit to encode the fused feature vector to extract high-level semantic information, and the decoder uses an attention mechanism to dynamically weight the output of the encoder to generate a context vector, and the decoder concatenates the output of the previous step, the input at the current moment, and the context vector as the input of the gated recurrent unit, and the output layer of the decoder inserts the density distribution map as prior knowledge; A parameter sensitivity evaluation module is constructed based on the density distribution map, and the environmental data is perturbed and sampled through Monte Carlo simulation, and the Fisher information of each environmental data is calculated. The information matrix formed by the Fisher information is decomposed and sorted based on the eigenvalues, and the importance of the environmental data is determined according to the eigenvalues, and a subset of environmental data whose importance is higher than a preset importance threshold is retained; The purified environment data is used as a constraint condition to construct a device association graph, wherein the device is a node and the environment data is an edge. The value of the environment data determines the weight of the edge. The degree centrality index of each device node is calculated, and the fused feature vector is divided into sub-blocks corresponding to different devices. The sub-blocks are transmitted through physical connections between devices. The receiving end device updates the feature representation according to the received sub-blocks, and the device association feature is formed after multiple rounds of iterations; The device association features and the density distribution map are input into a pre-set prophet model, wherein the prophet model is composed of a plurality of time polynomial basis functions, and the coefficients of the basis functions are fitted by the least squares method to obtain a polynomial trend term, and the probability density of the density distribution map is used as the confidence of the trend term for correction, and a hierarchical attention mechanism is superimposed on the prophet model to divide the time series into time segments of different lengths, and the attention mechanism is applied within the segment to calculate the moment weight distribution to extract local features, and the attention mechanism is applied between segments to calculate the segment weight distribution to extract global features; A scenario probability distribution function is constructed, with the environmental data as the independent variable and the device status and user behavior as the dependent variables. The scenario probability distribution function is sampled through Monte Carlo simulation to obtain scenario samples, and cluster analysis is performed on the scenario samples to identify typical scenarios. The sample mean of each cluster is calculated as the scene center point, and the sample variance is calculated as the uncertainty measure. The scene center point is used as the predicted value, and the uncertainty measure is used as the confidence interval to generate a prediction result.

[0033] The Monte Carlo simulation is a method of numerical calculation through random sampling and statistical analysis, which is usually used to solve complex mathematical problems, especially those that cannot be solved by traditional analytical methods. The perturbation sampling is a method of sampling data by introducing small perturbations to a model or system. Perturbation sampling can be used to generate more sample data, thereby helping to improve the robustness and generalization ability of the model. The Fisher information is an important concept in statistics, which is used to measure the amount of information about parameters in sample data. The polynomial basis function is a set of basis functions for function approximation, which are usually used to construct polynomial models.

[0034] Collect operating data of kitchen appliances, such as power consumption, temperature, switch status, etc., as well as environmental data, such as indoor temperature, humidity, light, etc. Clean the collected data to remove outliers and missing values. Normalize the cleaned data, such as Min-Max normalization, to scale the data range to between 0 and 1. Then, extract feature vectors, such as converting time series data into frequency domain features, or statistical features, such as mean, variance, etc. At the same time, construct a density distribution map based on the environmental data, such as using the kernel density estimation method to estimate the probability density function of the environmental data and visualize it as a map. Fuse the extracted feature vector and the constructed density distribution map, such as multiplying the feature vector with the probability density value of the corresponding position in the map to obtain a fused feature vector.

[0035] Suppose the power consumption data collected for a certain oven is [1000, 1200, 1100, 1300, 1200], and the indoor temperature data is [25, 26, 25, 27, 26]. After normalization, the power consumption data becomes [0.0, 0.4, 0.2, 0.6, 0.4], and the indoor temperature data becomes [0.0, 0.5, 0.0, 1.0, 0.5]. Suppose the extracted frequency domain features are [0.5, 0.2, 0.1], and the probability density value of the indoor temperature is 0.8, then the fused feature vector is [0.4, 0.16, 0.08].

[0036] The fused feature vector is input into the pre-set improved long short-term memory network. The network consists of an encoder and a decoder. The encoder uses a bidirectional gated recurrent unit to encode the fused feature vector and extract high-level semantic information. The decoder uses an attention mechanism to dynamically weight the output of the encoder to generate a context vector. The decoder concatenates the output of the previous step, the input of the current moment, and the context vector as the input of the gated recurrent unit. The output layer of the decoder inserts the density distribution map as prior knowledge.

[0037] The parameter sensitivity evaluation module is constructed based on the density distribution map. The environmental data is perturbed and sampled through Monte Carlo simulation, and the Fisher information of each environmental data is calculated. The information matrix composed of the Fisher information is decomposed and sorted based on the eigenvalue, and the importance of the environmental data is determined according to the eigenvalue size. The subset of environmental data whose importance is higher than the preset threshold is retained, that is, the environmental data is purified.

[0038] Assume that the indoor temperature is disturbed and sampled to obtain 100 samples. Calculate the Fisher information of each sample and construct the information matrix. Perform eigenvalue decomposition on the information matrix to obtain the eigenvalue [10, 5, 1]. Set the threshold to 4, and only retain the environmental data subset corresponding to the eigenvalue greater than 4.

[0039] The purified environmental data is used as a constraint to construct a device association graph. The device association graph uses devices as nodes and environmental data as edges. The value of the environmental data determines the weight of the edge. The degree centrality index of each device node is calculated. The fused feature vector is divided into sub-blocks corresponding to different devices. The sub-blocks are transferred through the physical connection between devices. The receiving device updates the feature representation according to the received sub-blocks. After multiple rounds of iterations, the device association feature is formed.

[0040] Assume that there is a physical connection between the oven and the range hood. Divide the fused feature vector of the oven into sub-blocks [0.4, 0.16] and [0.08], and pass them to the range hood respectively. The range hood updates its own feature representation based on the received sub-blocks.

[0041] The device-related features and density distribution map are input into the pre-set prophet model. The prophet model consists of multiple time polynomial basis functions. The coefficients of the basis functions are fitted by the least squares method to obtain the polynomial trend term. The probability density of the density distribution map is used as the confidence of the trend term for correction. The hierarchical attention mechanism is superimposed on the prophet model to divide the time series into time segments of different lengths. The attention mechanism is applied within the segment to calculate the moment weight distribution to extract local features, and the attention mechanism is applied between segments to calculate the segment weight distribution to extract global features. The scene probability distribution function is constructed, with environmental data as the independent variable and device status and user behavior as the dependent variables. The scene probability distribution function is sampled through Monte Carlo simulation to obtain scene samples. Cluster analysis is performed on the scene samples to identify typical scenes. The sample mean of each cluster is calculated as the scene center point, and the sample variance is calculated as the uncertainty measure. The prediction result is generated with the scene center point as the predicted value and the uncertainty measure as the confidence interval.

[0042] In this embodiment, by fusing feature vectors and density distribution maps and introducing device-related features, factors affecting the status of kitchen appliances can be considered more comprehensively, thereby improving prediction accuracy. Parameter sensitivity evaluation and environmental data purification reduce the model's sensitivity to noise data and enhance the robustness of the model. By constructing a scenario probability distribution function, a confidence interval for the prediction result can be provided, thereby improving the credibility of the prediction result.

[0043] In an optional embodiment, The coefficients of the basis function are fitted by the least square method to obtain a polynomial trend term, and the probability density of the density distribution map is used as the confidence of the trend term for correction, including: Constructing an adaptive polynomial basis function group, the adaptive polynomial basis function group includes a combination of Chebyshev polynomials and Legendre polynomials, generating a basis function matrix for each moment in the time series, each column of the basis function matrix represents the value of different basis functions at the moment, and using an elastic network least squares method with a first-order norm regularization term and a second-order norm regularization term to solve a polynomial coefficient vector, wherein the polynomial coefficient vector is obtained by minimizing an objective function including a sample loss term, a sparse term, and a smoothing term; Obtain a density distribution map and adaptively normalize the probability density value of the density distribution map to obtain an initial weight coefficient, introduce a weight calibration mechanism based on relative entropy, calculate the relative entropy between the weight coefficient and the standard normal distribution, construct a penalty term based on the relative entropy, integrate the penalty term into the weight calculation process, and obtain a calibrated weight coefficient matrix; Constructing a multi-scale weighted loss function, the multi-scale weighted loss function includes a local fitting error term, a global trend constraint term and a time series smoothing term, wherein the local fitting error term uses a calibrated weight coefficient to weight the sample error, the global trend constraint term introduces trend monotonicity and convexity constraints, and the time series smoothing term uses total variation regularization; The multi-scale weighted loss function is solved by an alternating direction multiplier method to obtain a plurality of sub-optimization problems, wherein the sub-optimization problems include a weight coefficient optimization sub-problem, a polynomial coefficient optimization sub-problem, and a Lagrange multiplier update sub-problem, and a modified polynomial coefficient vector is obtained by iteratively solving the sub-optimization problems; post-processing the corrected trend term according to an adaptive kernel regression smoother, wherein the window size of the adaptive kernel regression smoother is dynamically adjusted according to the local data density, using a larger window in a data-sparse area to enhance stability, and using a smaller window in a data-dense area to retain local features; Constructing an uncertainty estimation module based on repeated random sampling, generating multiple data subsets through repeated sampling, performing a trend correction process on each data subset, calculating the mean and variance of the trend term based on multiple correction results, and using the variance as an uncertainty measure of the trend term; An adaptive evaluation mechanism for the trend correction effect is established, and a comprehensive evaluation index is constructed based on the fitting error, trend smoothness and computational complexity before and after correction. When the comprehensive evaluation index does not reach the preset threshold, the algorithm parameters are automatically adjusted and the correction process is repeated until the convergence conditions are met.

[0044] The Chebyshev polynomial is a special type of polynomial with the smallest maximum error on the interval [-1, 1]. The Legendre polynomial is a type of orthogonal polynomial widely used in numerical integration (such as Gaussian integral) in numerical analysis. The relative entropy-based weight calibration mechanism is a method for adjusting weights by calculating relative entropy. The alternating direction multiplier method is an algorithm for optimization problems, particularly suitable for large-scale sparse optimization problems. The trend correction is a technique for time series analysis, which aims to adjust the trend part in the sequence to better capture short-term fluctuations or cyclical changes.

[0045] Construct an adaptive polynomial basis function group, which is composed of Chebyshev polynomials and Legendre polynomials, and can better adapt to different types of time series data. For each time point in the time series, the value of each basis function is calculated, and these values ​​are combined into a basis function matrix. Each column of the matrix represents the value of a different basis function at that time point.

[0046] The polynomial coefficients are calculated using the elastic net least squares method. The elastic net least squares method combines the first-order norm regularization term and the second-order norm regularization term, which can effectively control the sparsity and smoothness of the polynomial coefficients. By minimizing the objective function containing the sample loss term, the sparsity term, and the smoothness term, the optimal polynomial coefficient vector can be obtained. For example, for a time series containing 100 time points, assuming that we use the 5th-order Chebyshev polynomial and the 5th-order Legendre polynomial, the size of the basis function matrix is ​​100x10. Through the elastic net least squares method, a polynomial coefficient vector of length 10 can be obtained.

[0047] Obtain the density distribution map of the time series data, and adaptively normalize its probability density value to obtain the initial weight coefficient. In order to further improve the accuracy of the weight coefficient, a weight calibration mechanism based on relative entropy is introduced. Calculate the relative entropy between the weight coefficient and the standard normal distribution, and construct a penalty term based on the relative entropy. Integrate the penalty term into the weight calculation process to obtain a calibrated weight coefficient matrix. For example, assuming that the initial weight coefficient is [0.1, 0.2, 0.3, 0.2, 0.2], after calibration, the weight coefficient may become [0.15, 0.25, 0.3, 0.2, 0.1].

[0048] A multi-scale weighted loss function is constructed. The loss function contains a local fitting error term, a global trend constraint term, and a time series smoothing term. The local fitting error term uses a calibrated weight coefficient to weight the sample error, the global trend constraint term introduces trend monotonicity and convexity constraints, and the time series smoothing term uses total variation regularization.

[0049] The alternating direction multiplier method is used to solve the multi-scale weighted loss function. The problem is decomposed into multiple sub-optimization problems, including the weight coefficient optimization sub-problem, the polynomial coefficient optimization sub-problem and the Lagrange multiplier update sub-problem. By iteratively solving these sub-problems, the corrected polynomial coefficient vector is finally obtained.

[0050] The corrected trend term is post-processed using an adaptive kernel regression smoother. The window size of the smoother is dynamically adjusted based on the local data density. Larger windows are used in data-sparse areas to enhance stability, and smaller windows are used in data-dense areas to preserve local features. For example, suppose the time series has very dense data points in some areas and sparse data points in other areas. In data-dense areas, the smoother can use a smaller window, such as a window size of 3, while in data-sparse areas, the smoother can use a larger window, such as a window size of 7.

[0051] Construct an uncertainty estimation module based on repeated random sampling. Generate multiple data subsets through repeated sampling, perform the above trend correction process on each subset, and calculate the mean and variance of the trend term based on the multiple correction results. Use variance as the uncertainty measure of the trend term. For example, the original time series data is repeatedly sampled 100 times, sampling 80% of the data each time, and trend correction is performed on each subset to obtain 100 corrected trend terms. Then calculate the mean and variance of these 100 trend terms, and the variance is the uncertainty measure of the trend term.

[0052] Establish an adaptive evaluation mechanism for trend correction effect. Construct a comprehensive evaluation index based on the fitting error, trend smoothness and computational complexity before and after correction. If the comprehensive evaluation index does not reach the preset threshold, automatically adjust the algorithm parameters and repeat the correction process until the convergence condition is met.

[0053] In this embodiment, the density distribution information is used to correct the polynomial trend fitting, which can effectively reduce the impact of noise, thereby more accurately extracting the real trend component in the time series data. The use of technologies such as adaptive polynomial basis function group, elastic network least squares method and multi-scale weighted loss function can better adapt to different types of time series data and improve the robustness of trend extraction. Through repeated random sampling and uncertainty estimation modules, the uncertainty measurement of trend items can be provided, providing a more reliable basis for subsequent analysis and decision-making.

[0054] S3. Add the scenario probability distribution function, the prediction confidence interval and the density distribution map to a pre-set risk assessment model, construct a strategy search space based on risk constraints and set control boundary conditions based on the purification environment data, solve the candidate control strategies by combining the Monte Carlo tree search method and map them to the device control space, build a device collaborative control network and combine the attention mechanism to control information flow, determine the posterior probability distribution of different control actions in the control strategy, combine the adaptive exploration mechanism and Bellman error to construct a priority scoring model based on error gradient, combine the adaptive learning rate regulator based on the prediction confidence interval and the density distribution map to update the hyperparameters of the priority scoring model, and evaluate the candidate control strategies in combination with the pre-set multi-objective evaluation system to obtain the optimal control strategy.

[0055] The strategy search space refers to the set of all possible strategies in reinforcement learning or optimization problems. The Monte Carlo tree search method is a search algorithm for decision-making processes, commonly used in game AI, planning problems and reinforcement learning. The device control space refers to the set of all possible control commands or operations in the device or system control process. The Bellman error refers to the difference between the estimate and the actual value of the state value function or action value function in reinforcement learning and dynamic programming.

[0056] In an optional embodiment, Adding the scenario probability distribution function, the prediction confidence interval and the density distribution map to a pre-set risk assessment model, constructing a risk constraint-based strategy search space, setting control boundary conditions based on the purification environment data, and solving the candidate control strategies in combination with the Monte Carlo tree search method and mapping them to the equipment control space include: The scenario probability distribution function, the prediction confidence interval and the density distribution map are input into a pre-set risk assessment model. The risk assessment model performs risk assessment on the control strategy based on the conditional value online risk assessment criterion. By setting the prediction confidence interval as the risk assessment confidence interval and setting the low probability area of ​​the density distribution map as the risk threshold, the potential loss of the control strategy in an abnormal scenario is quantified to generate risk constraint conditions. A multidimensional strategy search space is constructed based on the risk constraint condition, wherein each dimension of the multidimensional strategy search space corresponds to a control variable, and the value range of the control variable is determined by the purification environment data. Each point in the multidimensional strategy search space represents a candidate control strategy. The multidimensional strategy search space is pruned and optimized according to the risk constraint condition, and the search area that does not meet the risk constraint condition is deleted; A Monte Carlo tree search method is used to perform a strategy search in the multidimensional strategy search space, including a selection phase, an expansion phase, a simulation phase, and a back-propagation phase. The selection phase includes starting from a root node and iteratively selecting a child node based on a confidence upper bound formula until reaching an incompletely expanded node. The expansion phase includes performing random sampling in a search area of ​​a selected node, generating a new child node, and initializing the number of visits and cumulative value of the new child node. The simulation phase includes starting from the new child node and simulating to a termination state using a fast walk strategy to obtain a simulated reward value. The back-propagation phase includes back-propagating the simulated reward value from the new child node to the root node, and updating the number of visits and cumulative value of all nodes on the path. The approximate optimal strategy obtained by the Monte Carlo tree search is mapped to the device control space through a predefined state-action mapping table, wherein the state-action mapping table contains a mapping relationship matrix from control variables to device states, and the continuous control variables are converted into a discrete device state sequence according to the mapping relationship matrix to generate a device control instruction sequence.

[0057] The conditional value-based online risk assessment criterion is a risk assessment method, which is mainly used to measure the potential loss in the worst case under given conditions. The confidence upper limit formula is a strategy for balancing exploration and utilization, especially in the multi-armed bandit problem. By assigning a confidence upper limit to each choice, it encourages the exploration of options with greater uncertainty. It is suitable for reinforcement learning, constrained optimization problems and multi-armed bandit algorithms. The fast move strategy refers to improving the efficiency and response speed of the algorithm by quickly evaluating and selecting actions in reinforcement learning or game problems.

[0058] Construct a risk assessment model. The model receives scenario probability distribution functions, prediction confidence intervals, and density distribution maps as input. The scenario probability distribution function describes the likelihood of different scenarios. The prediction confidence interval quantifies the uncertainty of the prediction results. The density distribution map shows the distribution density of system states in different regions. The risk assessment model is based on the conditional value risk assessment criterion, uses the prediction confidence interval as the risk assessment confidence interval, and sets the low probability area in the density distribution map as the risk threshold. By evaluating the potential loss of the control strategy under abnormal scenarios, the model generates a series of risk constraints.

[0059] Assume a scenario of controlling a sewage treatment plant. The scenario probability distribution function can describe the probability of occurrence of different pollutant concentrations. The prediction confidence interval can reflect the confidence in the prediction of future pollutant concentrations. The density distribution map can show the distribution of water quality parameters after treatment. If the low probability area of ​​a certain water quality parameter corresponds to the risk of exceeding the discharge standard, the risk assessment model will generate corresponding risk constraints, such as limiting the value range of certain control variables.

[0060] A multidimensional strategy search space is constructed based on the generated risk constraints. Each dimension of the search space corresponds to a control variable, such as the dosage and aeration volume in a sewage treatment plant. The value range of each control variable is determined by the purification environment data, such as historical operation data, environmental regulations, etc. Each point in the search space represents a candidate control strategy. The search space is pruned and optimized according to the risk constraints, and areas that do not meet the risk constraints are removed. For example, if the risk constraint stipulates that the dosage cannot exceed a certain threshold, the areas in the search space that exceed the threshold will be removed.

[0061] The Monte Carlo tree search method is used to search for strategies in the optimized search space. Monte Carlo tree search is an iterative process that includes four stages: selection, expansion, simulation, and backpropagation. In the selection stage, starting from the root node, child nodes are iteratively selected according to the confidence upper bound formula until the node that is not fully expanded is reached. The confidence upper bound formula balances exploration and utilization, and tends to select nodes with fewer visits or higher cumulative values. In the expansion stage, random sampling is performed in the search area of ​​the selected node to generate new child nodes, and their visit counts and cumulative values ​​are initialized. In the simulation stage, starting from the newly generated child node, the fast walk strategy is used to simulate to the termination state to obtain the simulated reward value. For example, in the example of a sewage treatment plant, the water quality compliance rate and operating cost under different control strategies can be simulated. In the backpropagation stage, the simulated reward value is backpropagated from the new node to the root node, and the visit counts and cumulative values ​​of all nodes on the path are updated.

[0062] The approximate optimal strategy obtained by Monte Carlo tree search is mapped to the device control space. Through the pre-defined state-action mapping table, the continuous control variables are converted into a discrete device state sequence to generate a device control instruction sequence. The state-action mapping table contains a mapping relationship matrix from control variables to device states. For example, the continuous value of the dosing amount can be converted into the on / off state and frequency setting value of the dosing pump.

[0063] In this embodiment, by integrating risk assessment into the strategy search process, high-risk control strategies can be effectively avoided to ensure that the equipment can operate safely in various scenarios. Through the Monte Carlo tree search algorithm, the approximate optimal strategy that meets the risk constraints can be quickly found to improve control efficiency and reduce resource consumption. By considering the scenario probability distribution and prediction confidence interval, the control strategy can be made more robust to environmental changes and prediction errors, thereby improving the overall stability of the control system.

[0064] In an optional embodiment, Construct a device collaborative control network and combine it with the attention mechanism to control information flow, determine the posterior probability distribution of different control actions in the control strategy, combine the adaptive exploration mechanism and Bellman error to construct a priority scoring model based on error gradient, combine the adaptive learning rate regulator based on the prediction confidence interval and the density distribution map to update the hyperparameters of the priority scoring model, and combine the pre-set multi-objective evaluation system to evaluate the candidate control strategy, and obtain the optimal control strategy including: Constructing a device collaborative control network, wherein the device collaborative control network adopts a directed acyclic graph structure, wherein nodes in the graph represent devices, and edges represent the transmission direction of control information, wherein each node in the directed acyclic graph comprises an attention module and a control module, wherein the attention module receives the current environment state vector and the control information vector of other devices as input, calculates the attention weight distribution through a multi-head attention mechanism, and the control module fuses the input information based on the attention weight distribution, and outputs the device control instruction after being processed by a fully connected neural network; Introducing an adaptive exploration mechanism into the equipment collaborative control network, constructing a priority scoring model based on error gradient, taking the state transfer error of each node as the basis for priority scoring, the state transfer error is calculated by the Bellman equation, including calculating the difference between the current state value function and the next state discount value function, taking the gradient direction of the state transfer error as the priority exploration direction, introducing an adaptive learning rate regulator, dynamically adjusting the learning rate according to the position of the current state in the prediction confidence interval and the density distribution map, the prediction confidence of the current state is inversely proportional to the learning rate, and the probability density of the current state is inversely proportional to the learning rate, and balancing the convergence and exploration of the model by dynamically adjusting the learning rate; The candidate control strategies are evaluated based on a pre-set multi-objective evaluation system, wherein the multi-objective evaluation system includes a user satisfaction index, an energy efficiency index, and an equipment health index. Normalized scores are calculated for the user satisfaction index, the energy efficiency index, and the equipment health index, respectively, and they are combined into a comprehensive evaluation index through weighted summation. The control strategy with the highest comprehensive evaluation index is selected as the optimal control strategy.

[0065] The Bellman equation is a recursive equation that describes state transitions in dynamic programming or reinforcement learning, and expresses the relationship between the expected value of future rewards and the current rewards and the expected value of the next state given the current state and action.

[0066] Construct a device collaborative control network. The network adopts a directed acyclic graph structure. The nodes in the graph represent each device, and the edges represent the transmission direction of the control information. Each node contains an attention module and a control module. For example, assume a smart home environment, which includes three devices: air conditioning, lighting, and curtains. The air conditioning node can receive control information from the lighting node, such as lighting intensity, to adjust its own cooling strategy. The attention module of each node receives the current environment state vector and the control information vector of other devices as input. For example, the attention module of the air conditioning node receives the current indoor temperature, humidity, and lighting intensity information sent by the lighting node. The attention module calculates the attention weight distribution through a multi-head attention mechanism. For example, if the current indoor temperature is very high, the attention module will give a higher weight to the temperature information. The control module fuses the input information based on the attention weight distribution, and outputs the device control instruction after processing by the fully connected neural network. For example, the control module of the air conditioning node calculates the appropriate cooling temperature setting value based on the weighted temperature, humidity, and lighting intensity information.

[0067] An adaptive exploration mechanism is introduced into the equipment cooperative control network. A priority scoring model based on error gradient is constructed, and the state transition error of each node is used as the basis for priority scoring. The state transition error is obtained by comparing the current state value function with the next state discount value function. For example, if the indoor temperature decreases slower than expected after the air conditioner is turned on, the state transition error will be larger. The gradient direction of the state transition error is used as the priority exploration direction. For example, if increasing the cooling power can effectively reduce the state transition error, the direction of increasing the cooling power is preferentially explored. At the same time, an adaptive learning rate regulator is introduced. The regulator dynamically adjusts the learning rate according to the position of the current state in the prediction confidence interval and density distribution map. The prediction confidence of the current state is inversely proportional to the learning rate, and the probability density of the current state is inversely proportional to the learning rate. For example, if the prediction confidence of the current state is high and the probability density is also high, the learning rate will be adjusted down to avoid over-exploration. Conversely, if the prediction confidence and probability density of the current state are both low, the learning rate will be increased to encourage the exploration of new control strategies. Assuming that the current indoor temperature is 28 degrees Celsius, the predicted confidence interval is 27-29 degrees Celsius, and the probability density is 0.8, the learning rate will be set to a lower value, such as 0.01.

[0068] The candidate control strategies are evaluated based on a pre-set multi-objective evaluation system. The evaluation system includes user satisfaction index, energy efficiency index and equipment health index. For example, user satisfaction can be obtained through user feedback questionnaires, energy efficiency can be obtained through equipment power consumption calculations, and equipment health can be obtained through equipment operation status monitoring. The normalized scores of these indicators are calculated separately, for example, the user satisfaction score is normalized to between 0 and 1. Then, they are combined into a comprehensive evaluation index by weighted summation. For example, assuming that the user satisfaction weight is 0.5, the energy efficiency weight is 0.3, and the equipment health weight is 0.2, the comprehensive evaluation index is 0.5*user satisfaction score+0.3*energy efficiency score+0.2*equipment health score. The control strategy with the highest comprehensive evaluation index is selected as the optimal control strategy. For example, assuming that the user satisfaction score of a control strategy is 0.9, the energy efficiency score is 0.8, and the equipment health score is 0.7, then its comprehensive evaluation index is 0.83. If this is the highest comprehensive evaluation index among all candidate strategies, then the strategy is selected as the optimal control strategy.

[0069] In this embodiment, through the attention mechanism, the device can more effectively select and utilize information from other devices, thereby improving control efficiency. The adaptive exploration mechanism can help the system better cope with environmental changes and uncertainties and enhance the robustness of the system. The multi-objective evaluation system can comprehensively consider multiple factors such as user needs, energy efficiency, and equipment health, thereby optimizing the overall control performance.

[0070] Figure 2 FIG. 1 is a schematic diagram of a multi-kitchen appliance remote control system based on context perception according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to collect kitchen environment data and record the user's historical usage data of kitchen appliances, perform wavelet multi-resolution decomposition on the environment data, calculate the energy distribution feature matrix under each scale coefficient and perform hierarchical screening on the wavelet coefficients, obtain filtered high-quality environment data by combining reverse reconstruction, construct a density distribution map based on the high-quality environment data and determine a density gradient matrix, set a density threshold curve based on the density gradient matrix and identify abnormal data points to obtain purified environment data, add the purified environment data and the historical usage data to a bidirectional recursive neural network, construct a time series feature correlation heat map and determine a mutual information matrix, hierarchically encode the historical usage data based on the mutual information matrix, extract high-value data sequences and perform pyramid pooling operations, construct a multi-scale feature map and a feature interaction network, and obtain a fused feature vector; The second unit is used to add the fused feature vector and the density distribution map to a pre-set improved long short-term memory network, construct a parameter sensitivity evaluation module based on the density distribution map and calculate the eigenvalue decomposition of the Fisher information matrix to obtain the parameter importance ranking, use the purification environment data as a constraint condition, construct a time-varying association graph and calculate the degree centrality index of the device node, establish a feature propagation channel and perform information interaction in the kitchen appliance to obtain the device association feature, add the device association feature and the density distribution map to the pre-set prophet model to construct a polynomial trend term and predict the long-term trend and short-term fluctuation, determine the multi-scale time series dependency relationship in combination with the hierarchical attention mechanism, construct a scene probability distribution function and determine the prediction confidence interval; The third unit is used to add the scenario probability distribution function, the prediction confidence interval and the density distribution map to a preset risk assessment model, construct a strategy search space based on risk constraints and set control boundary conditions based on the purification environment data, solve the candidate control strategy in combination with the Monte Carlo tree search method and map it to the device control space, build a device collaborative control network and control information flow in combination with the attention mechanism, determine the posterior probability distribution of different control actions in the control strategy, construct a priority scoring model based on the error gradient in combination with the adaptive exploration mechanism and Bellman error, update the hyperparameters of the priority scoring model in combination with the adaptive learning rate regulator based on the prediction confidence interval and the density distribution map, and evaluate the candidate control strategy in combination with a preset multi-objective evaluation system to obtain the optimal control strategy.

[0071] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0072] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0073] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-kitchen appliance remote control method based on situational awareness, characterized in that: include: Collecting kitchen environmental data and recording historical usage data of kitchen appliances by users, performing wavelet multi-resolution decomposition on the environmental data, calculating the energy distribution feature matrix under each scale coefficient and performing hierarchical screening on the wavelet coefficients, combining reverse reconstruction to obtain filtered high-quality environmental data, constructing a density distribution map based on the high-quality environmental data and determining a density gradient matrix, setting a density threshold curve based on the density gradient matrix and identifying abnormal data points to obtain purified environmental data, adding the purified environmental data and the historical usage data to a bidirectional recursive neural network, constructing a time series feature correlation heat map and determining a mutual information matrix, hierarchically encoding the historical usage data based on the mutual information matrix, extracting high-value data sequences and performing pyramid pooling operations, constructing a multi-scale feature map and a feature interaction network, and obtaining a fused feature vector; Add the fused feature vector and the density distribution map to a pre-set improved long short-term memory network, construct a parameter sensitivity evaluation module based on the density distribution map and calculate the eigenvalue decomposition of the Fisher information matrix to obtain the parameter importance ranking, use the purification environment data as a constraint condition, construct a time-varying association graph and calculate the degree centrality index of the device node, establish a feature propagation channel and perform information interaction in kitchen appliances to obtain device association features, add the device association features and the density distribution map to a pre-set prophet model to construct a polynomial trend term and predict long-term trends and short-term fluctuations, combine the hierarchical attention mechanism to determine the multi-scale temporal dependency, construct a scenario probability distribution function and determine the prediction confidence interval; The scenario probability distribution function, the prediction confidence interval and the density distribution map are added to a pre-set risk assessment model, a strategy search space based on risk constraints is constructed, and control boundary conditions are set based on the purification environment data. The candidate control strategies are solved by combining the Monte Carlo tree search method and mapped to the device control space, a device collaborative control network is constructed and the attention mechanism is combined to control the information flow, the posterior probability distribution of different control actions in the control strategy is determined, and an error gradient-based priority scoring model is constructed in combination with an adaptive exploration mechanism and Bellman error. The hyperparameters of the priority scoring model are updated in combination with an adaptive learning rate regulator based on the prediction confidence interval and the density distribution map, and the candidate control strategies are evaluated in combination with a pre-set multi-objective evaluation system to obtain the optimal control strategy.

2. The method according to claim 1, characterized in that Collect kitchen environmental data and record the user's historical usage data of kitchen appliances, perform wavelet multi-resolution decomposition on the environmental data, calculate the energy distribution feature matrix under each scale coefficient and perform hierarchical screening on the wavelet coefficients, combine reverse reconstruction to obtain filtered high-quality environmental data, build a density distribution map based on the high-quality environmental data and determine the density gradient matrix, set a density threshold curve based on the density gradient matrix and identify abnormal data points to obtain purified environmental data, add the purified environmental data and the historical usage data to a bidirectional recursive neural network, build a time series feature correlation heat map and determine the mutual information matrix, hierarchically encode the historical usage data based on the mutual information matrix, extract high-value data sequences and perform pyramid pooling operations, build a multi-scale feature map and a feature interaction network, and obtain a fused feature vector including: Collecting kitchen environment data and historical usage data of kitchen appliances by users through pre-deployed sensor nodes, wherein the environment data includes temperature parameters, humidity parameters, light intensity parameters and air quality parameters, and the historical usage data includes device usage time, device usage frequency and device usage power; Perform wavelet multi-resolution decomposition on the environmental data through the Daubey wavelet function to obtain wavelet coefficients at multiple scales, calculate the mean square value of the wavelet coefficients to obtain an energy distribution feature matrix, screen the wavelet coefficients according to a preset first energy threshold and a second energy threshold, delete the wavelet coefficients with energy values ​​lower than the first energy threshold, retain the wavelet coefficients with energy values ​​higher than the second energy threshold, and for the wavelet coefficients with energy values ​​greater than the first energy threshold and less than the second energy threshold, calculate the correlation coefficient between the current wavelet coefficient and the adjacent wavelet coefficients, if the correlation coefficient is greater than the preset correlation threshold, retain the current wavelet coefficient, otherwise delete it, and for the retained wavelet coefficients, obtain the filtered high-quality environmental data through reverse reconstruction; Based on the high-quality environmental data, a density distribution map is constructed by a kernel density estimation method, the density distribution map estimates the probability density function of the filtered environmental data based on a Gaussian kernel function, a density gradient matrix of the density distribution map is calculated by a numerical differentiation method, abnormal point candidates are marked based on a preset threshold, the abnormal point candidates are screened in combination with prior knowledge to determine abnormal data points, and the abnormal data points are removed to obtain purified environmental data; The purified environment data and the historical usage data are aligned according to timestamps to form a multivariate time series, the multivariate time series is input into a bidirectional long short-term memory network, the forward time series features and the reverse time series features are extracted respectively through the forward recursive neural network and the backward recursive neural network of the bidirectional long short-term memory network, the forward time series features and the reverse time series features are spliced ​​to obtain a complete time series feature, the correlation matrix of the complete time series feature is calculated by the Pearson correlation coefficient to construct a time series feature correlation heat map, the feature mutual information matrix is ​​calculated, and the hierarchical clustering method is used to hierarchically encode the complete time series features based on the feature mutual information matrix to extract a high-value data sequence; A time pyramid pooling strategy is adopted to perform multi-scale decomposition on the high-value data sequence, local features are extracted by maximum pooling in each time scale, and the local features are spliced ​​in the time dimension to construct a multi-scale feature map. The multi-scale feature map contains global features and local features. A two-layer fully connected neural network is built as a feature interaction network. The input layer of the feature interaction network receives the multi-scale feature map, the feature interaction correlation relationship is modeled through the hidden layer, and the fused feature vector is generated through the output layer.

3. The method according to claim 2, characterized in that The calculation formula of the Pearson correlation coefficient is as follows: ； in, Represents a time series feature set X The standardized value of each element in Represents a time series feature set Y The standardized value of each element in represents the calculated Pearson correlation coefficient, n Represents a time series feature set X and Y The number of features in P ( x i , y i ) is the joint probability distribution, indicating that x i and y i The probability of P ( x i ) is the time series feature set X The marginal probability distribution of x i The probability of P ( y i ) is the time series feature set Y The marginal probability distribution of y i probability.

4. The method according to claim 1, characterized in that: Add the fused feature vector and the density distribution map to a pre-set improved long short-term memory network, construct a parameter sensitivity evaluation module based on the density distribution map and calculate the eigenvalue decomposition of the Fisher information matrix to obtain the parameter importance ranking, use the purification environment data as a constraint condition, construct a time-varying association graph and calculate the degree centrality index of the device node, establish a feature propagation channel and perform information interaction in kitchen appliances to obtain device association features, add the device association features and the density distribution map to a pre-set prophet model to construct a polynomial trend term and predict long-term trends and short-term fluctuations, combine the hierarchical attention mechanism to determine the multi-scale temporal dependency, construct a scenario probability distribution function and determine the prediction confidence interval, including: Input the fused feature vector and the density distribution map into a preset improved long short-term memory network, wherein the improved long short-term memory network includes an encoder and a decoder, wherein the encoder uses a bidirectional gated recurrent unit to encode the fused feature vector to extract high-level semantic information, and the decoder uses an attention mechanism to dynamically weight the output of the encoder to generate a context vector, and the decoder concatenates the output of the previous step, the input at the current moment, and the context vector as the input of the gated recurrent unit, and the output layer of the decoder inserts the density distribution map as prior knowledge; A parameter sensitivity evaluation module is constructed based on the density distribution map, and the environmental data is perturbed and sampled through Monte Carlo simulation, and the Fisher information of each environmental data is calculated. The information matrix formed by the Fisher information is decomposed and sorted based on the eigenvalues, and the importance of the environmental data is determined according to the eigenvalues, and a subset of environmental data whose importance is higher than a preset importance threshold is retained; The purified environment data is used as a constraint condition to construct a device association graph, wherein the device is a node and the environment data is an edge. The value of the environment data determines the weight of the edge. The degree centrality index of each device node is calculated, and the fused feature vector is divided into sub-blocks corresponding to different devices. The sub-blocks are transmitted through physical connections between devices. The receiving end device updates the feature representation according to the received sub-blocks, and the device association feature is formed after multiple rounds of iterations; The device association features and the density distribution map are input into a pre-set prophet model, wherein the prophet model is composed of a plurality of time polynomial basis functions, and the coefficients of the basis functions are fitted by the least squares method to obtain a polynomial trend term, and the probability density of the density distribution map is used as the confidence of the trend term for correction, and a hierarchical attention mechanism is superimposed on the prophet model to divide the time series into time segments of different lengths, and the attention mechanism is applied within the segment to calculate the moment weight distribution to extract local features, and the attention mechanism is applied between segments to calculate the segment weight distribution to extract global features; A scenario probability distribution function is constructed, with the environmental data as the independent variable and the device status and user behavior as the dependent variables. The scenario probability distribution function is sampled through Monte Carlo simulation to obtain scenario samples, and cluster analysis is performed on the scenario samples to identify typical scenarios. The sample mean of each cluster is calculated as the scene center point, and the sample variance is calculated as the uncertainty measure. The scene center point is used as the predicted value, and the uncertainty measure is used as the confidence interval to generate a prediction result.

5. The method according to claim 4, characterized in that The coefficients of the basis function are fitted by the least square method to obtain a polynomial trend term, and the probability density of the density distribution map is used as the confidence of the trend term for correction, including: Constructing an adaptive polynomial basis function group, the adaptive polynomial basis function group includes a combination of Chebyshev polynomials and Legendre polynomials, generating a basis function matrix for each moment in the time series, each column of the basis function matrix represents the value of different basis functions at the moment, and using an elastic network least squares method with a first-order norm regularization term and a second-order norm regularization term to solve a polynomial coefficient vector, wherein the polynomial coefficient vector is obtained by minimizing an objective function including a sample loss term, a sparse term, and a smoothing term; Obtain a density distribution map and adaptively normalize the probability density value of the density distribution map to obtain an initial weight coefficient, introduce a weight calibration mechanism based on relative entropy, calculate the relative entropy between the weight coefficient and the standard normal distribution, construct a penalty term based on the relative entropy, integrate the penalty term into the weight calculation process, and obtain a calibrated weight coefficient matrix; Constructing a multi-scale weighted loss function, the multi-scale weighted loss function includes a local fitting error term, a global trend constraint term and a time series smoothing term, wherein the local fitting error term uses a calibrated weight coefficient to weight the sample error, the global trend constraint term introduces trend monotonicity and convexity constraints, and the time series smoothing term uses total variation regularization; The multi-scale weighted loss function is solved by an alternating direction multiplier method to obtain a plurality of sub-optimization problems, wherein the sub-optimization problems include a weight coefficient optimization sub-problem, a polynomial coefficient optimization sub-problem, and a Lagrange multiplier update sub-problem, and a modified polynomial coefficient vector is obtained by iteratively solving the sub-optimization problems; post-processing the corrected trend term according to an adaptive kernel regression smoother, wherein the window size of the adaptive kernel regression smoother is dynamically adjusted according to the local data density, using a larger window in a data-sparse area to enhance stability, and using a smaller window in a data-dense area to retain local features; Constructing an uncertainty estimation module based on repeated random sampling, generating multiple data subsets through repeated sampling, performing a trend correction process on each data subset, calculating the mean and variance of the trend term based on multiple correction results, and using the variance as an uncertainty measure of the trend term; An adaptive evaluation mechanism for the trend correction effect is established, and a comprehensive evaluation index is constructed based on the fitting error, trend smoothness and computational complexity before and after correction. When the comprehensive evaluation index does not reach the preset threshold, the algorithm parameters are automatically adjusted and the correction process is repeated until the convergence conditions are met.

6. The method according to claim 1, characterized in that Adding the scenario probability distribution function, the prediction confidence interval and the density distribution map to a pre-set risk assessment model, constructing a risk constraint-based strategy search space, setting control boundary conditions based on the purification environment data, and solving the candidate control strategies in combination with the Monte Carlo tree search method and mapping them to the equipment control space include: The scenario probability distribution function, the prediction confidence interval and the density distribution map are input into a pre-set risk assessment model. The risk assessment model performs risk assessment on the control strategy based on the conditional value online risk assessment criterion. By setting the prediction confidence interval as the risk assessment confidence interval and setting the low probability area of ​​the density distribution map as the risk threshold, the potential loss of the control strategy in an abnormal scenario is quantified to generate risk constraint conditions. A multidimensional strategy search space is constructed based on the risk constraint condition, wherein each dimension of the multidimensional strategy search space corresponds to a control variable, and the value range of the control variable is determined by the purification environment data. Each point in the multidimensional strategy search space represents a candidate control strategy. The multidimensional strategy search space is pruned and optimized according to the risk constraint condition, and the search area that does not meet the risk constraint condition is deleted; A Monte Carlo tree search method is used to perform a strategy search in the multidimensional strategy search space, including a selection phase, an expansion phase, a simulation phase, and a back-propagation phase. The selection phase includes starting from a root node and iteratively selecting a child node based on a confidence upper bound formula until reaching an incompletely expanded node. The expansion phase includes performing random sampling in a search area of ​​a selected node, generating a new child node, and initializing the number of visits and cumulative value of the new child node. The simulation phase includes starting from the new child node and simulating to a termination state using a fast walk strategy to obtain a simulated reward value. The back-propagation phase includes back-propagating the simulated reward value from the new child node to the root node, and updating the number of visits and cumulative value of all nodes on the path. The approximate optimal strategy obtained by the Monte Carlo tree search is mapped to the device control space through a predefined state-action mapping table, wherein the state-action mapping table contains a mapping relationship matrix from control variables to device states, and the continuous control variables are converted into a discrete device state sequence according to the mapping relationship matrix to generate a device control instruction sequence.

7. The method according to claim 1, characterized in that Construct a device collaborative control network and combine it with the attention mechanism to control information flow, determine the posterior probability distribution of different control actions in the control strategy, combine the adaptive exploration mechanism and Bellman error to construct a priority scoring model based on error gradient, combine the adaptive learning rate regulator based on the prediction confidence interval and the density distribution map to update the hyperparameters of the priority scoring model, and combine the pre-set multi-objective evaluation system to evaluate the candidate control strategy, and obtain the optimal control strategy including: Constructing a device collaborative control network, wherein the device collaborative control network adopts a directed acyclic graph structure, wherein nodes in the graph represent devices, and edges represent the transmission direction of control information, wherein each node in the directed acyclic graph comprises an attention module and a control module, wherein the attention module receives the current environment state vector and the control information vector of other devices as input, calculates the attention weight distribution through a multi-head attention mechanism, and the control module fuses the input information based on the attention weight distribution, and outputs the device control instruction after being processed by a fully connected neural network; Introducing an adaptive exploration mechanism into the equipment collaborative control network, constructing a priority scoring model based on error gradient, taking the state transfer error of each node as the basis for priority scoring, the state transfer error is calculated by the Bellman equation, including calculating the difference between the current state value function and the next state discount value function, taking the gradient direction of the state transfer error as the priority exploration direction, introducing an adaptive learning rate regulator, dynamically adjusting the learning rate according to the position of the current state in the prediction confidence interval and the density distribution map, the prediction confidence of the current state is inversely proportional to the learning rate, and the probability density of the current state is inversely proportional to the learning rate, and balancing the convergence and exploration of the model by dynamically adjusting the learning rate; The candidate control strategies are evaluated based on a pre-set multi-objective evaluation system, wherein the multi-objective evaluation system includes a user satisfaction index, an energy efficiency index, and an equipment health index. Normalized scores are calculated for the user satisfaction index, the energy efficiency index, and the equipment health index, respectively, and they are combined into a comprehensive evaluation index through weighted summation. The control strategy with the highest comprehensive evaluation index is selected as the optimal control strategy.

8. A multi-kitchen appliance remote control system based on situational awareness, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect kitchen environment data and record the user's historical usage data of kitchen appliances, perform wavelet multi-resolution decomposition on the environment data, calculate the energy distribution feature matrix under each scale coefficient and perform hierarchical screening on the wavelet coefficients, obtain filtered high-quality environment data by combining reverse reconstruction, construct a density distribution map based on the high-quality environment data and determine a density gradient matrix, set a density threshold curve based on the density gradient matrix and identify abnormal data points to obtain purified environment data, add the purified environment data and the historical usage data to a bidirectional recursive neural network, construct a time series feature correlation heat map and determine a mutual information matrix, hierarchically encode the historical usage data based on the mutual information matrix, extract high-value data sequences and perform pyramid pooling operations, construct a multi-scale feature map and a feature interaction network, and obtain a fused feature vector; The second unit is used to add the fused feature vector and the density distribution map to a pre-set improved long short-term memory network, construct a parameter sensitivity evaluation module based on the density distribution map and calculate the eigenvalue decomposition of the Fisher information matrix to obtain the parameter importance ranking, use the purification environment data as a constraint condition, construct a time-varying association graph and calculate the degree centrality index of the device node, establish a feature propagation channel and perform information interaction in the kitchen appliance to obtain the device association feature, add the device association feature and the density distribution map to the pre-set prophet model to construct a polynomial trend term and predict the long-term trend and short-term fluctuation, determine the multi-scale time series dependency relationship in combination with the hierarchical attention mechanism, construct a scene probability distribution function and determine the prediction confidence interval; The third unit is used to add the scenario probability distribution function, the prediction confidence interval and the density distribution map to a preset risk assessment model, construct a strategy search space based on risk constraints and set control boundary conditions based on the purification environment data, solve the candidate control strategy in combination with the Monte Carlo tree search method and map it to the device control space, build a device collaborative control network and control information flow in combination with the attention mechanism, determine the posterior probability distribution of different control actions in the control strategy, construct a priority scoring model based on the error gradient in combination with the adaptive exploration mechanism and Bellman error, update the hyperparameters of the priority scoring model in combination with the adaptive learning rate regulator based on the prediction confidence interval and the density distribution map, and evaluate the candidate control strategy in combination with a preset multi-objective evaluation system to obtain the optimal control strategy.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images