Washing machine remote control method and system based on internet of things

Abstract

The present application relates to the technical field of smart home, especially a washing machine remote control method and system based on the Internet of Things, comprising: image capture of clothes by an image acquisition module, and input of the captured image into a clothes recognition module; output of an image feature vector, a clothes type distribution vector and a clothes volume estimate value by the clothes recognition module, collection of environmental state data by a local sensor and an edge device interface, and formation of an environmental state vector; input of the collected data into a perception enhancement function, and output of a context enhancement feature vector by the perception enhancement function; input of the context enhancement feature vector into a neural network model, and acquisition of a recommended parameter set; and sending of the recommended parameter set to a washing machine terminal after confirmation by a user to achieve remote control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart home, in particular to a washing machine remote control method and system based on Internet of Things. BACKGROUND

[0002] With the development of smart home and Internet of Things technology, as a household appliance with high frequency of daily use, washing machines have gradually evolved from traditional mechanical control to digital, remote and intelligent. At present, many high-end washing machines have the ability to connect with mobile phone Apps or home intelligent control systems (such as voice assistants), and users can start, pause, set washing mode, and schedule time through remote operation without touching the device. These functions have brought high convenience to users when they are out, busy or have difficulty in operation. However, the remote control technology currently used on the market still has significant limitations. The main problem is the lack of perception ability of the "on-site state" of the washing task. In other words, although this system expands the physical distance of control, it does not significantly improve the intelligence level of control.

[0003] In the remote operation scenario, users usually cannot directly observe the details of the actual type, quantity, material structure, color depth, etc. of the clothes, nor can they real-time understand the changes of the environment where the family is located, such as whether the current temperature and humidity are suitable for airing, whether the electricity price is in the peak period, whether the power grid load is high, etc. Due to the lack of these key information, users often choose the washing program based on experience or preset preferences, resulting in inaccurate program selection, and further problems such as incomplete washing of clothes, cross-staining, and excessive energy consumption. At the same time, the existing remote control system basically belongs to the execution structure of "static parameters + fixed process", and the system cannot dynamically adapt to the changes of the state of the clothes or the adjustment of the environmental conditions, nor can it optimize or feedback correct in the middle according to the washing effect, lacking the intelligent ability of "process perception-strategy reasoning-action adjustment". In addition, although some manufacturers have introduced basic AI algorithms to try to control energy saving or stain recognition, these functions are mostly limited to local control, and cannot be effectively activated in remote operation mode, nor can they form a closed-loop interaction with the user's remote control needs. Therefore, how to overcome the technical shortcomings of "can't see, can't judge, can't control" in the remote operation scenario, so that the washing machine can not only be remotely started, but also make intelligent response based on the actual washing task and environmental state, is a key technical problem that needs to be broken through in the current Internet of Things washing machine control field. SUMMARY

[0004] To solve the above problems, the present application provides a washing machine remote control method and system based on Internet of Things, which can solve the problems in the above background technology.

[0005] To achieve the above object, the technical scheme adopted by the present application is:

[0006] The washing machine remote control method based on the Internet of Things comprises:

[0007] S1: image grabbing of clothes is performed by an image acquisition module, and the grabbed image is input into a clothes recognition module;

[0008] S2: an image feature vector, a clothes type distribution vector and a clothes volume estimation value are output by the clothes recognition module;

[0009] S3: while clothes recognition is being performed, environmental state data is synchronously collected by a local sensor and an edge device interface to form an environmental state vector;

[0010] S4: the clothes type distribution vector, the clothes volume estimation value and the environmental state vector are input into a perception enhancement function, and a context enhancement feature vector is output by the perception enhancement function;

[0011] S5: the context enhancement feature vector is input into a neural network model, and a recommended parameter set is obtained by the neural network model;

[0012] S6: the recommended parameter set is confirmed by a user, and the recommended parameter set is sent to a washing machine terminal to realize remote control.

[0013] As a preferred, the S2 comprises:

[0014] S2.1: the grabbed image is acquired, and the image is preprocessed;

[0015] S2.2: the preprocessed image is acquired, and the image is input into a clothes recognition module;

[0016] S2.3: an image feature vector is output by a lightweight convolutional neural network of the clothes recognition module;

[0017] S2.4: the image feature vector is input into a sub-module group for multi-target feature decoding, and a clothes type distribution vector and a volume estimation value are output by the sub-module group.

[0018] As a preferred, the S2.4 comprises:

[0019] S2.4.1: the image feature vector is input into a first sub-module and a second sub-module respectively for multi-target feature decoding;

[0020] S2.4.2: a type distribution vector is output by a full connection classifier of the first sub-module, and a volume estimation value is output by a linear regression model of the second sub-module.

[0021] As a preferred, the calculation formula of the volume estimation value is:

[0022]

[0023] wherein V c is the volume estimate, w is the linear regression model weight, b is the bias term, and z is the image feature vector;

[0024] The formula for calculating the context-enhanced feature vector is:

[0025] X f = X input + λ (T c · θ1) (E env · θ2) ;

[0026] wherein X f is the context-enhanced feature vector, λ is the adjustable fusion weight coefficient, θ1 and θ2 are the learnable scale factors learned during training, non-fixed constants, · is the Hadamard element-wise product, i.e., multiplication of corresponding dimensions, and X input is the comprehensive input vector.

[0027] As a preference, the expression of the recommended parameter set is:

[0028] P wash = [T wash , M mode , R spin , D dry ] ;

[0029] wherein P wash is the recommended parameter set, T wash is the washing water temperature, M mode is the washing mode number, R spin is the spin speed ratio, and D dry is the drying start suggestion.

[0030] As a preference, the S6 comprises:

[0031] S6.1: The control system visualizes the recommended parameter set, and the user confirms whether to accept the recommended parameter set according to the visualization result;

[0032] If yes, the control system defaults the recommended parameter set as the output parameter set;

[0033] If no, the control system adjusts the output parameter set, and the output parameter set is the sum of the recommended parameter set and the modified parameter set at this time;

[0034] S6.2: The recommended parameter set is sent to the washing machine terminal to realize remote control.

[0035] As a preference, the S6.1 comprises:

[0036] S6.1.1: After the user confirms the recommended parameter set, the control system sends the recommended parameter set to the washing machine terminal in a structured JSON format through a standard Internet of Things communication protocol;

[0037] S6.1.2: After receiving the recommended parameter set, the main control chip of the washing machine terminal performs structural verification, parses the control parameters involved, and calls the hardware driver to complete task configuration.

[0038] Preferably, the preprocessing of S2.1 includes size scaling, color balance processing and edge enhancement.

[0039] The Internet of Things-based remote control system for washing machines comprises:

[0040] An image acquisition module for acquiring images inside the washing machine drum;

[0041] A clothing recognition module comprising a lightweight convolutional neural network module, a first sub-module and a second sub-module;

[0042] The lightweight convolutional neural network is used to extract local texture and edge information of the image, and expand convolution and global average pooling to aggregate spatial semantics, and output an image feature vector;

[0043] The first sub-module is used to predict the material type of the clothing and output a type distribution vector through a fully connected classifier;

[0044] The second sub-module is used to estimate the overall volume of the clothing and output a volume estimate value through a linear regression model;

[0045] A sensor module and an edge device are used to collect environmental state data and form an environmental state vector.

[0046] Preferably, the image feature vector includes the structure, color distribution and local texture state in the clothing image.

[0047] The present application has the following advantages:

[0048] The clothing recognition module directly extracts two control parameters that play a key role in washing decisions: type distribution vector and volume estimate, providing a theoretical basis for the generation of structured decision information;

[0049] By introducing a perception enhancement function, a complete washing context feature expression is constructed, and an enhanced mechanism with coupled interaction terms is further introduced to enhance the expression ability of the model for "clothing-environment linkage influence". This context perception feature fusion method not only meets the needs of adaptive intelligent control, but also has high practicality and deployment feasibility, clear data source, simple and efficient computing structure, ensuring that the washing machine can still make accurate strategy judgments based on complete information when operated remotely by the user.

[0050] The closed-loop inference logic of image recognition, environment perception and intelligent parameter recommendation is realized by the application, the risk of error setting of the user in remote operation due to lack of on-site perception is avoided, and compared with the traditional fixed program control mode, the application has the advantages of dynamic adaptation, scene self-judgment, controllable energy consumption and stable strategy. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The flow chart of the remote control method of the washing machine in the embodiment of the application is shown in the figure.

[0052] Figure 2 The engineering flow chart of the clothes recognition module in the embodiment of the application is shown in the figure.

[0053] Figure 3 The block diagram of the remote control system of the washing machine in the embodiment of the application is shown in the figure.

[0054] Figure 4 The composition block diagram of the image feature vector in the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0055] Please refer to Figures 1-4 The application relates to a washing machine remote control method and system based on the Internet of Things, and specifically comprises the following steps:

[0056] Step 1: clothes image recognition and feature extraction

[0057] The input of this step is the image in the washing machine cylinder after the user finishes putting in clothes, and the system automatically activates the image acquisition module to complete image capture before the user issues a start request through a remote App. The image is denoted as is a color RGB image data, the acquisition device is a fixed-focus wide-angle camera module installed at the top center position of the washing machine cylinder cover, the resolution is 1280x720, and near-infrared night vision light supplement is supported. Image acquisition is performed in a clothes static state, so as to ensure that the image has structural stability and contour boundary definition. The acquisition is automatically triggered after the detection of the hatch door closure or the completion of the putting-in operation.

[0058] After image acquisition is completed, the first step is to perform geometric and illumination standardization processing on the image. Specifically, the processing includes size scaling (uniform to 128x128 resolution), color balance processing (using histogram stretching to improve color contrast), edge enhancement (using Laplacian filtering to improve structural definition), etc. The image after standardization processing is sent to the clothes recognition module as the model input. The module adopts a lightweight convolutional neural network Model vision, with 13 convolutional layers, the first 7 of which are used to extract local texture and edge information, and the last 6 of which aggregate spatial semantics through step-by-step expansion convolution and global average pooling, finally outputting a fixed-length feature vector This vector encodes the structure, color distribution, and local texture state in the entire clothing image.

[0059] The feature vector z is simultaneously sent to a sub-module group for multi-target feature decoding, where the sub-module group includes a first sub-module and a second sub-module.

[0060] Specifically, the first sub-module is used to predict the clothing material type, and a 2-layer fully connected classifier is used to achieve this, outputting where each dimension represents the existence probability of a different material category (such as cotton, blended, wool, silk). The classifier uses a softmax function to normalize the output, so that ∑T c = 1. For example, for a scenario where a light gray sweater and a white T-shirt are put in, the system may output T c = [0.65, 0.30, 0.05, 0.00], indicating that the system identifies 65% as cotton, 30% as blended, 5% as wool, and essentially no silk clothing.

[0061] The second sub-module is used to estimate the overall volume of the clothing, and a linear regression model is used:

[0062]

[0063] where is the model regression weight, b is the bias term, and the max function is used to constrain the predicted value to be non-negative. Here V c is the normalized volume estimate, which is used to measure the degree of clothing load, and will be used as the basis for determining whether to use large water volume, extend the washing time, or trigger the energy-saving mode in subsequent control logic. For example, if the system identifies that the clothing distribution in the image is dense and the texture features are significantly superimposed, it may output V c ≈ 0.83, indicating that the current washing machine load is high; if the clothing is sparse or thin, it may output V c ≈ 0.25.

[0064] z: 256-dimensional clothing image deep feature vector extracted by Model vision , obtained by forward propagation of the standardized image I through MobileNetV2.

[0065] T c : 4-dimensional clothing type vector output by the classification branch, representing the confidence distribution of the four material types of cotton, blended, wool, and silk, normalized to probability form.

[0066] V c: The estimated volume of the clothes, which is a numerical indicator output by the regression branch after the regression, reflects the size of the clothes occupying space, and has been normalized.

[0067] w: The weight vector of the linear regression model, obtained by model training.

[0068] b: The regression bias term, which is a constant.

[0069] The core innovation of this step is that the image recognition not only outputs the traditional "image label", but directly extracts two control parameters that play a key role in laundry decision-making: the type distribution vector T c and the volume estimate V c . This design breaks the traditional pattern of "image recognition results still need user judgment to control the way", making the visual module for the first time in the washing system to assume the responsibility of generating structured decision-making information. At the same time, the model structure has deployability, supporting real-time running in low-power chips (such as SoC based on ARM Cortex A53), without relying on external computing resources.

[0070] Output:

[0071] 1. The clothes type distribution vector Output by the Model vision classification branch, representing the proportion of each type of material in the current clothes;

[0072] 2. The estimated volume of clothes V c Output by the regression function f v (z), indicating the overall load state of the clothes, which is a normalized volume indicator.

[0073] Through the image recognition module, the washing machine completes self-sensing of the clothes content before the user remotely controls the start: including material classification and quantity volume estimation. Compared with the traditional mode that requires the user to actively set the washing program, this step provides the core input basis for the subsequent automatic control recommendation of the system. This method uses the achievable embedded image processing + lightweight network structure to directly convert the clothes image information into two structure variables that can be used for reasoning, and realizes automatic recognition and feedback without relying on external server support, laying a solid foundation for remote intelligent washing control.

[0074] Step 2: Environment state perception and feature fusion

[0075] The system synchronously collects environment state data through local sensors and edge device interfaces to form the environment state vector Specifically, it includes the following 7 dimensions:

[0076] Indoor humidity h (from built-in DHT22 temperature and humidity sensor in washing machine);

[0077] Current grid load g (obtained by smart meter interface via Modbus protocol);

[0078] Current electricity price p (from local API or pre-downloaded time-of-use price table);

[0079] Three-dimensional weather features w1, w2, w3 (queried from weather service and mapped to temperature, rainfall probability, and wind speed level);

[0080] Whether the user's region allows the washing time period flag a (0 or 1, given by the home energy scheduling platform or local time period restriction table).

[0081] Each of the above variables is normalized to convert its value to the [0, 1] interval to be compatible with the input feature range of the neural network model. For example, if the humidity h is 63%, the normalized value is 0.63; if the electricity price p is 0.89 yuan / kWh during the peak period of the day, it is normalized to 0.742 under the condition of setting the upper limit to 1.2 yuan. This part of the processing is completed by the local computing module in advance and does not require user intervention.

[0082] Next, the system will concatenate T c (4 dimensions), V c (1 dimension), and E env (7 dimensions) to form a comprehensive input vector X At the same time as constructing X input , to further improve the environmental adaptability and scenario understanding ability of the control strategy generation, a context perception enhancement function is introduced to guide the learning of the nonlinear interaction between "clothes-environment" in the feature space in a weakly supervised manner.

[0083] This enhancement function is reflected through the coupling term between features and is constructed as follows:

[0084] X f = X input + λ(T c · θ1)(E env · θ2);

[0085] λ is the adjustable fusion weight coefficient, which controls the influence strength of the coupling term;

[0086] is the learnable scaling factor obtained during training, which is not a fixed constant;

[0087] · denotes the Hadamard element-wise product (i.e., multiplication of corresponding dimensions);

[0088] For the final fused context-aware feature, it is used as the input of the subsequent control policy recommendation model.

[0089] The creativity of this fusion term is reflected in the following points: (1) It not only splices internal and external features, but also introduces a soft interaction mechanism, so that there is a dynamic weight between a certain type of clothing and a certain type of environment. For example, when the proportion of "wool" is high and the current h is high (high humidity), the system will automatically emphasize the policy attention on the spin-drying intensity and drying time; (2) Compared with the traditional splicing input method, this fusion structure with cross terms can capture the complex environment-dependent relationship in the remote laundry scene; (3) λ can be dynamically adjusted according to actual deployment, such as using different values in energy-saving priority or washing degree priority scenarios. c The proportion of "wool" is high and the current h is high (high humidity), the system will automatically emphasize the policy attention on the spin-drying intensity and drying time; (2) Compared with the traditional splicing input method, this fusion structure with cross terms can capture the complex environment-dependent relationship in the remote laundry scene; (3) λ can be dynamically adjusted according to actual deployment, such as using different values in energy-saving priority or washing degree priority scenarios.

[0090] In order to avoid X f Out of control in extreme scenarios (such as all 0 or 1), a stability regularization constraint term is added:

[0091]

[0092] The regularization term is added to the loss function when training the control model, and β is the regularization weight constant. The innovation of this structure is that it is not a traditional parameter penalty, but introduces a regulation of the fluctuation degree of the fused feature, ensuring that the policy model has sufficient adaptability to external feature changes while not overfitting.

[0093] X input : Basic spliced features, obtained by splicing T c , V c and E env , a total of 12 dimensions.

[0094] X f : Enhanced context-aware features, after the interaction term is processed to form a structure input, which is used for policy recommendation module

[0095] λ: Weight adjustment parameter of the fusion coupling term.

[0096] θ1, θ2: Feature field scaling factor, obtained by self-learning during training.

[0097] Regularization constraint, used to control the stability of the fused feature change.

[0098] β: Regularization term coefficient, set to 10 -3 magnitude to limit the deviation of the optimization process.

[0099] Output:

[0100] 1. Context-enhanced feature vector as the core input of the subsequent control model;

[0101] 2. Original feature set X input As an alternative interpretable input to support interpretable tracking of parameter recommendation results.

[0102] This step involves collecting key physical variables from the home environment and combining them with the clothing features identified in step 1 to construct a complete laundry context feature representation X. input Furthermore, an enhancement mechanism X with coupled interaction terms is introduced. f This enhances the model's ability to express the "linkage between clothing and the environment," and also has high practicality and deployment feasibility. The data sources are clear, and the calculation structure is simple and efficient, ensuring that the washing machine can still make accurate policy judgments based on complete information when the user operates it remotely.

[0103] Step 3: Intelligent Recommendation of Control Parameters

[0104] The input for this step is the context-enhanced feature vector output from step 2. It consists of the following parts: the clothing type distribution vector identified in step 1. and clothing volume estimates And the environmental state vector collected in step 2 by local temperature and humidity sensors (such as DHT22), smart meter modules, weather APIs, and time period management systems. These features are processed by the context coupling function defined in step 2 to generate the final input vector X. f This reflects the overall status perception result of the current laundry task.

[0105] In remote control scenarios, users often cannot accurately set washing parameters. Therefore, this step designs a lightweight multi-head neural network model that can be deployed on embedded devices. control This allows for the automatic recommendation of washing strategies. The model receives X... f As input, it outputs a control vector containing four core parameters: washing water temperature T. wash Washing mode number M mode Dehydration speed ratio R spin And drying start-up recommendation D dry The model structure employs a three-layer linear transformation, where the input layer maps the 12-dimensional input to a 32-dimensional hidden space, the intermediate layers are 24-dimensional and 16-dimensional respectively, and the final output controls the parameters through four independent linear heads. LeakyReLU is used as the activation function between each layer.

[0106] For example, when a user puts a batch of heavy cotton and blended clothing into the washing machine, and at the same time the external detector shows that the current humidity is 85%, the electricity price is at its peak (e.g., 0.98 yuan / kWh), and the temperature is below 10°C, X fthe clothes load V c , material weight (e.g., high cotton content) and environmental conditions (e.g., cold and humid, expensive electricity) are explicitly expressed in the input vector. In this case, the system should tend to recommend a medium water temperature (e.g., T wash = 0.6, corresponding to about 36°C), a mode number that is biased towards energy saving (e.g., M mode = 1 indicates fast washing), a medium dehydration intensity (R spin = 0.5), and to start the drying program (D dry = 1) to reduce the burden of airing.

[0107] The final output of the model is as follows:

[0108] P wash = [T wash , M mode , R spin , D dry ];

[0109] where T wash ∈ [0, 1] is the normalized washing water temperature, and the actual temperature can be converted by T actual = 30 + 30 · T wash ; is the mode number, such as standard washing (0), fast washing (1), wool (2), etc.; R spin ∈ [0, 1] represents the dehydration speed intensity ratio; D dry ∈ {0, 1} indicates whether to recommend starting the drying program.

[0110] To make the recommended parameters more in line with the household energy use strategy and the user's long-term expectations, two patented control mechanisms are designed in this step. The first is the energy consumption penalty regularization term, which is used to suppress the frequent occurrence of high energy consumption configurations during the training phase:

[0111]

[0112] where η1, η2, η3 are the energy consumption penalty coefficients brought by water temperature rise, speed increase and drying enablement, respectively, derived from the average energy consumption model provided by the washing machine manufacturer (e.g., η1 = 2.0, indicating that the temperature increases by 2 units of energy consumption for every 10°C increase; η2 = 1.5, η3 = 3.2); α is the regularization weight hyperparameter, used to balance the tension between energy consumption and performance goals. This design not only constrains the behavior of the model from a cost perspective, but also improves the practicality of the recommended strategy, especially for users who do not set energy saving as a priority.

[0113] The second is the strategy robustness regularization term, which is used to limit the excessive sensitivity of the model output to environmental micro-variations:

[0114]

[0115] where represents the gradient of control output to environmental variable, λ is the adjustment factor. For example, if the system jumps the dehydration intensity from 80% to 40% when the electricity price rises from 0.90 yuan to 0.91 yuan, the gradient is large and the system will be punished. Through this mechanism, the model tends to generate a control strategy that is stable in response to continuous changes in the environment, improving the credibility and user acceptance of the strategy.

[0116] The final output of this step is the recommended control parameter set P wash , which contains four control variables that will be displayed on the user's remote interface in the next step for user confirmation and generation of the final control instruction P exec .

[0117] The above regularization term is only used in the model training stage to guide the model to find a balance between energy saving and robustness. The entire control parameter recommendation system is based on X f to achieve direct mapping from input to control output, which constitutes the core decision component in the remote intelligent control path of this patent.

[0118] This step realizes the closed-loop reasoning logic of "image recognition, environmental perception and intelligent parameter recommendation" for the first time in the remote control scenario of household washing machines, avoiding the risk of incorrect settings by users due to lack of on-site perception when operating remotely. Compared to traditional fixed program control methods, this scheme has four key advantages: dynamic adaptation, scene self-judgment, controllable energy consumption, and stable strategy. Through fine feature fusion and regularization mechanism design, it forms a complete, self-consistent and engineering implementable intelligent control decision module.

[0119] Step 4: Remote user confirmation and control instruction issuance

[0120] The input of this step is the recommended control parameter set output by step 3:

[0121] P wash = [T wash , M mode , R spin , D dry ];

[0122] where T wash is the normalized recommended washing water temperature (e.g. 0.6 corresponds to 36℃), M mode is the recommended washing program number (e.g. 1 represents fast wash), R spin represents the recommended dehydration rotation speed ratio (e.g. 0.7 is 70% of the maximum rotation speed), and D dry indicates whether to recommend enabling the drying function. All these parameters are generated by Model control based on X fThe reasoning leads to the conclusion that these are control recommendations based on a combined optimization of clothing contents and environmental conditions.

[0123] The goal of this step is to allow the user to confirm the system-recommended control parameters remotely, make fine-tuning adjustments if necessary, and ultimately generate structured control commands P. exec This information is then sent to the washing machine for execution. Since users are not physically present in remote control scenarios and cannot observe the clothes or intuitively assess the washing environment, the system needs to provide sufficient explanation and flexibility to ensure that users can understand the recommended results and easily intervene in certain parameters, thereby improving control reliability.

[0124] The system first displays P in the user's remote App interface using both graphical and textual methods. wash For example, T wash =0.6 will be achieved through the mapping function T actual =30+30·T wash The actual water temperature was determined to be 36℃, and it was displayed as "Recommended water temperature: 36℃". Similarly, M... mode =2 is mapped to "wool pattern", R spin =0.8 is displayed as "Dehydration Strength: Strong", D dry =1 indicates "Drying is recommended". Each item is accompanied by visual controls (such as sliders, switches, and drop-down menus) for user adjustment.

[0125] If users have no special needs, they can directly click the "Start" button to accept the recommendation; the system will then default to P. exec =P wash If the user makes any minor adjustments to the parameters, the system records the user's adjustment value ΔP. user The final control command is then synthesized according to the following formula:

[0126] P exec =P wash +ΔP user ;

[0127] Where P exec Represents a set of structured control instructions, and is related to P. wash Having the same structure, the values ​​of elements may change depending on the user's intent. For example, if the user recommends T... wash If the value is increased from 0.6 (36℃) to 0.8 (48℃), then ΔT = +0.2. In the new control command, T... wash The value is 0.8.

[0128] P wash These are the system-recommended control parameters, derived by the model based on current clothing and environmental perception; ΔP user It is an incremental change manually made by the user through the App's user interface; P execis the final control execution parameter, which combines the results of system intelligent recommendation and user individual adjustment.

[0129] After the user confirms the control scheme, the system sends the P exec to the washing machine terminal in a structured JSON format. After receiving the instruction, the master chip first performs structure verification on the fields, parses the control parameters, and calls the corresponding hardware drivers to complete the task configuration. For example, T wash controls the target temperature of the heating module, R spin controls the motor speed, M mode is mapped to the preset program pipeline, D dry controls whether the drying module is activated.

[0130] To ensure the controllability and user experience of this process in a remote environment, the system also designs a recommendation strategy credibility scoring mechanism to help users determine whether to accept the recommendation:

[0131]

[0132] where h is the hidden feature vector of the intermediate layer in the control model (taken from the hidden layer output), w and b are the trained weights and biases, and σ is the sigmoid function. S conf reflects the credibility of the current recommendation and is used for interface prompts such as "System Confidence: High / Medium / Low".

[0133] h is the intermediate vector in the model, representing the abstract representation of the current washing scenario; w and b are the weight and bias parameters obtained through training; S conf represents the system recommendation credibility score, which is an auxiliary index used to enhance user trust in remote decision-making.

[0134] For example, if the system output S conf = 0.85, the user interface will mark "The system has high confidence in this recommendation, and it is recommended to adopt it directly"; if S conf = 0.32, the system will remind the user "The current recommendation strategy has low confidence, and it is recommended to adjust according to individual needs".

[0135] The final output of this step is a structured control instruction set P exec , which includes T wash , M mode , R spin , and D dry four parameters, the values of which are consistent with the system recommendation or generated after user fine-tuning. The instruction will be sent to the washing machine terminal through a secure communication link to drive the control module to perform the corresponding operation.

[0136] The above embodiments are only used to describe the preferred embodiments of the present application, and are not used to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements made by those skilled in the art to the technical solutions of the present application shall fall within the protection scope of the claims of the present application.

Claims

1. A method for remote control of a washing machine based on Internet of Things, characterized in that, Comprise: S1: image acquisition module for clothes image capture, the image input clothes recognition module after capture; S2: through the clothes recognition module output image feature vector, clothes type distribution vector and clothes volume estimate value; S2.1: get the image after capture, the image is pretreated; S2.2: get the pretreated image, input the image to clothes recognition module; S2.3: through the light weight convolutional neural network of clothes recognition module output image feature vector; S2.4: the image feature vector is input to the sub module group for multi target feature decoding, and the clothes type distribution vector and volume estimate value are output through the sub module group; S2.4.1: the image feature vector is input to the first sub module and the second sub module respectively for multi target feature decoding; S2.4.2: the type distribution vector is output through the full connection classifier of the first sub module, and the volume estimate value is output through the linear regression model of the second sub module; S3: while clothes recognition is carried out, the environment state data is synchronously collected through local sensor and edge device interface, and the environment state vector is formed; S4: the clothes type distribution vector, clothes volume estimate value and environment state vector are input into the perception enhancement function, and the context enhanced feature vector is output through the perception enhancement function; S5: the context enhanced feature vector is input into the neural network model, and the recommended parameter set is obtained through the neural network model; S6: the user confirms the recommended parameter set, and sends the recommended parameter set to the washing machine terminal to realize remote control; The calculation formula of the volume estimate value is: where V c is the volume estimate, w are linear regression model weights, b is a bias term, and z is an image feature vector; The calculation formula of the context enhanced feature vector is: X f = X input + λ(T c · θ1)(E env · θ2); where X f is the contextual enhanced feature vector, λ is the adjustable fusion weight coefficient, θ1, θ2 are the learnable scale factors during training, non-fixed constants, · is the Hadamard element-wise product, i.e., multiplication of corresponding dimensions, X input is the comprehensive input vector. 2.The IoT-based remote control method of a washing machine according to claim 1, wherein, The expression of the recommended parameter set is: P wash = [T wash , M mode , R spin , D dry ] ; where P wash is the recommended parameter set, T wash is the wash water temperature, M mode is the wash mode number, R spin is the spin speed ratio, and D dry is the drying start recommendation. 3.The IoT-based remote control method of a washing machine according to claim 1, wherein, The S6 comprises: S6.1: the control system carries out visual processing on the recommended parameter set, and the user confirms whether to accept the recommended parameter set according to the visual processing result; If yes, the control system defaults the recommended parameter set as the output parameter set; If no, the control system adjusts the recommended parameter set, and at this time the recommended parameter set is the sum of the control system recommended parameter set and the modified parameter set; S6.2: the recommended parameter set is sent to the washing machine terminal to realize remote control. 4.The IoT-based remote control method of a washing machine according to claim 3, characterized in that, The S6.1 comprises: S6.1.1: after the user confirms the recommended parameter set, the control system sends the recommended parameter set to the washing machine terminal in structured JSON format through the standard Internet of Things communication protocol; S6.1.2: the main control chip of the washing machine terminal receives it, carries out structure checking on the recommended parameter set, analyzes the control parameters involved, and calls the hardware driver to complete task configuration. 5.The IoT-based remote control method of a washing machine according to claim 1, wherein, The pretreatment of S2.1 comprises size scaling, color balance processing and edge enhancement.

6. A remote control system for a washing machine based on the Internet of Things, characterized in that, Comprise: Image acquisition module, for collecting images in the washing machine cylinder; Clothes recognition module, comprising light weight convolutional neural network module, first sub module and second sub module; Light weight convolutional neural network, for extracting local texture and edge information of image, and expanding convolution and global average pooling aggregation space semantics, output image feature vector; First sub module, for predicting clothes material type, output type distribution vector through full connection classifier; The second sub-module is configured to estimate the total volume of the clothes and output a volume estimation value through a linear regression model; The sensor module and the edge device are configured to collect environmental state data and form an environmental state vector.

7. The IoT-based remote control system for washing machines as claimed in claim 6 wherein, The image feature vector includes structure, color distribution and local texture state in the clothes image.

Images