A smart toilet control method, system and storage medium based on natural language interaction

Abstract

The present invention relates to the field of smart toilet technology, and more specifically, to a smart toilet control method, system, and storage medium based on natural language interaction, comprising the following steps: receiving instructions input by a user via natural language; determining the user's identity; initializing a user mode and setting parameters; diverting natural language instructions; if the natural language instruction is determined to be a simple instruction, the local edge layer will parse and directly execute the control operation; if the natural language instruction is determined to be a complex instruction, a large cloud-based model will parse and generate a local fixed control instruction; the local fixed control instruction will be saved in the corresponding user personalized mode; and the control instruction will be executed. In the present invention, interactive control using natural language is implemented, and model distillation technology is used through a large cloud-based language model to reduce computing power requirements and improve the overall user experience and intelligence level of the smart toilet.

Description

Technical Field

[0001] The present invention relates to the technical field of smart toilets, and in particular to a smart toilet control method, system and storage medium based on natural language interaction. Background Art

[0002] With the continuous advancement of science and technology, smart toilets have gradually become popular in people's lives; traditional smart toilet control methods mostly rely on physical buttons or simple remote control operations, which have certain limitations; on the one hand, for users who are not familiar with the layout of the operation panel, especially the elderly or children, operation may be more difficult, reducing the user experience; on the other hand, when the user's hands are inconvenient or in a specific scenario, it is difficult to perform precise operations.

[0003] In addition, as people's requirements for the convenience and intelligence of smart homes continue to increase, simple control methods can no longer meet user needs; existing smart toilet control technologies lack personalization and cannot be customized according to the habits and preferences of different users, making it difficult to provide more humane services; moreover, when faced with complex control needs, existing control methods are inefficient and cannot quickly and accurately parse and execute user commands, limiting the further expansion and optimization of smart toilet functions. Summary of the Invention

[0004] The purpose of the present invention is to provide a smart toilet control method, system and storage medium based on natural language interaction to solve the problems raised in the above background technology.

[0005] To solve the above technical problems, one of the objectives of the present invention is to provide a smart toilet control method based on natural language interaction, comprising the following steps:

[0006] S1, receiving the user's instructions input through natural language and obtaining the current status data of the smart toilet;

[0007] S2. Perform multimodal identity recognition on the user to determine the user's identity;

[0008] S3, user mode initialization and parameter setting, users can customize multiple modes;

[0009] S4. Diverting natural language instructions based on preset instruction complexity judgment conditions and real-time network status;

[0010] S5. When the natural language instruction is determined to be a simple instruction, the local edge layer parses it and directly executes the control operation;

[0011] S6. When the natural language instruction is determined to be a complex instruction, the natural language instruction and the current state data of the smart toilet are sent to the cloud. The cloud-based large model parses and generates a local fixed control instruction, and returns the local fixed control instruction to the smart toilet for execution.

[0012] S7. Saving the local fixed control instruction to the corresponding user personalized mode according to the user identity;

[0013] S8. Execute the control instructions to complete the interactive operation of the smart toilet.

[0014] As a further improvement of the present technical solution, the multimodal identity recognition in S2 includes at least one or more combinations of voiceprint recognition, fingerprint recognition and WeChat identity recognition, and the user identity is bound to the unique ID of the smart toilet.

[0015] As a further improvement of this technical solution, the user mode initialization in S3 can be achieved by editing text and transmitting it to the smart toilet, which will then parse and initialize it. The user mode parameter setting can utilize the roaming function of WeChat to migrate the user's mode parameters from one smart toilet to another.

[0016] As a further improvement of this technical solution, the instruction complexity judgment condition in S4 includes at least one of the following:

[0017] The length of the natural language instruction exceeds the preset threshold;

[0018] The natural language instruction contains more parameters than the preset number;

[0019] Natural language instructions are not stored in the local instruction library;

[0020] Real-time network latency or bandwidth meets cloud processing conditions.

[0021] As a further improvement of this technical solution, the specific steps of building and deploying the cloud-based large model in S6 include:

[0022] A1. Use the full DeepSeek-V3 teacher model to generate a 10k-sample smart toilet command dataset, including the mapping between natural language commands and local fixed control commands.

[0023] A2. Use a three-stage distillation training strategy to train the smart toilet natural language parsing model.

[0024] As a further improvement to this technical solution, the cloud-based smart toilet natural language parsing model in A2 is trained using a three-stage distillation training strategy. The specific steps include:

[0025] B1, Logits distillation stage, using The divergence loss function transfers the output probability distribution of the teacher model to the student model, using the temperature coefficient Control the smoothness of the probability distribution and set a higher temperature value in the initial stage Soften the probability distribution and gradually reduce it to By hardening the distribution, the model complexity is reduced while maintaining the semantic understanding ability of the student large model; the teacher model is used to generate instruction variants, expand the diversity of training samples, and achieve data enhancement; using temperature coefficients of Divergence loss function, the teacher model uses the training data to generate soft labels, and the student model learns both the soft labels and the original hard labels. The corresponding formula is:

[0026] ;

[0027] in, Represents knowledge distillation Layer loss function, represents the normalization function, represents the raw output of the student model, represents the original output of the teacher model, Represents the probability of comparing the output of the student model and the teacher model, represents the information loss when using the student model to approximate the teacher model;

[0028] B2, feature distillation stage, aligning the intermediate layer feature expressions of the teacher model and the student model;

[0029] B3. Construct a joint loss function to jointly optimize knowledge transfer and task performance;

[0030] B4. Parameter freezing and fine-tuning phase: freeze the underlying parameters of the student model and only fine-tune the top-level classifier and attention layer. Use hard-label cross-entropy loss for enhanced training, introduce domain data augmentation, and enable dynamic quantization to reduce computational overhead during deployment.

[0031] As a further improvement of this technical solution, the intermediate layer feature expressions of the teacher model and the student model are forced to be aligned in B2, and the shadow space knowledge transfer is achieved through the mean square error loss. The corresponding formula is:

[0032] ;

[0033] in, represents the intermediate layer loss function in knowledge distillation, represents the feature dimension of the middle layer, Represents the dimension index of the intermediate layer feature vector, Indicates that the intermediate layer features of the teacher model are The output of dimension, Indicates that the intermediate layer features of the student model are The output of dimension, represents a learnable linear projection matrix that adapts to feature dimension differences, Represents the square of the L2 norm.

[0034] As a further improvement of this technical solution, the formula for constructing the joint loss function in B3 is:

[0035] ;

[0036] ;

[0037] in, represents the joint loss function, and represents the weight coefficient, Indicates the number of classification categories, Represents the classification category index, represents the true label, represents the predicted probability.

[0038] As a further improvement of this technical solution, the specific steps of the cloud-based large model parsing and generating local fixed control instructions in S6 include:

[0039] C1, based on Few-shot learning technology, maps natural language commands into local fixed control commands through a preset command example library;

[0040] C2. Dynamically adjust the output value based on the current status data of the smart toilet.

[0041] A second object of the present invention is to provide an intelligent toilet control system based on natural language interaction, which is used to implement the steps of the above-mentioned intelligent toilet control method based on natural language interaction, including:

[0042] Data acquisition module: used to receive instructions input by the user through natural language and obtain the current status data of the smart toilet;

[0043] Identity recognition module: used to perform multimodal identity recognition on the user, determine the user's identity, and bind the user's identity to the unique ID of the smart toilet;

[0044] Instruction diversion module: used to divert natural language instructions based on preset instruction complexity judgment conditions and real-time network status;

[0045] Local voice module: used to assist the local edge layer in parsing instructions;

[0046] Cloud parsing module: used to receive natural language commands and the current status data of the smart toilet, parse the cloud-based large model and generate local fixed control commands, and return the local fixed control commands to the smart toilet for execution;

[0047] User mode management module: used to save local fixed control instructions to the corresponding user personalized mode according to user identity;

[0048] Local execution module: used to receive control instructions from the local edge layer and the cloud, execute control operations, and complete interactive operations on the smart toilet.

[0049] The third object of the present invention is to provide a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the steps of the above-mentioned smart toilet control method based on natural language interaction.

[0050] Compared with the existing technology, the beneficial effects of the present invention are as follows: the intelligent toilet control method, system and storage medium based on natural language interaction realize interactive control of natural language, improve user experience, integrate intelligent toilet data through the large language model in the cloud, simplify the complexity of the terminal side, adopt Few-shot technology to achieve the accuracy of data output, adopt model distillation technology to reduce the demand for computing power, and can deploy the model to the edge, thereby improving the overall user experience and intelligence level of the intelligent toilet. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of the method of the present invention;

[0052] Figure 2 It is a system structure diagram of the present invention;

[0053] The meanings of the various marks in the figure are: 100, data acquisition module; 200, identity recognition module; 300, instruction diversion module; 400, local voice module; 500, cloud analysis module; 600, user mode management module; 700, local execution module. DETAILED DESCRIPTION

[0054] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0055] like Figure 1As shown, this embodiment provides a smart toilet control method based on natural language interaction, including the following steps:

[0056] S1, receiving the user's instructions input through natural language and obtaining the current status data of the smart toilet;

[0057] S2. Perform multimodal identification on users to confirm their identities. For individuals who have not entered user authentication, the mode adapts to seasonal changes and is automatically set to spring, summer, autumn, and winter modes to meet customer needs. Authenticated users can use the cloud-based large model to parse natural language, providing a better user experience. Unauthenticated individuals can use the local voice service module or use WeChat scanning and a one-time payment method to use the cloud-based large model voice recognition service.

[0058] The multimodal identity recognition in S2 includes at least one or more combinations of voiceprint recognition, fingerprint recognition and WeChat identity recognition, and the user identity is bound to the unique ID of the smart toilet. First, the user's voiceprint, fingerprint and WeChat OpenID and UnionID are collected to perform user settings, and the identity information is bound to the unique ID of the smart toilet. You can choose one, two combinations or three of these identity information for identity recognition settings. While ensuring the security of the user's identity, select the identity authentication settings suitable for the toilet scene. The WeChat identity has a roaming function, which can migrate the user's mode parameters from one smart toilet to another, such as migrating the user's comfort mode and its parameters from toilet A to toilet B;

[0059] S3. User mode initialization and parameter setting. Users can customize multiple modes. Each user can set multiple modes, such as comfort mode, clean mode, menstrual mode, quick wash mode, soft wash mode, defecation-aid mode, etc. Each mode contains a set of controllable parameters. The control parameters include parameters of the washing stage and parameters of the drying stage; the parameters of the washing stage include at least one of the following: water splash, water temperature, water pressure, nozzle direction and duration; the parameters of the drying stage include: wind temperature, duration, and the mode can be set by editing a text and sending it to the smart toilet. The toilet will analyze and set it by itself without manual tedious settings. Some common modes can be set, such as child washing mode, menstrual protection mode, etc. These modes can synchronize data with the smart toilet, and you can also set some personalized modes, such as the men's clean mode. After the user enters these modes, he can control the toilet through relatively simple modes. For example, issuing more casual instructions such as "dry it", "dry it", and "dry it at 25 degrees" can dry it. The user mode initialization in S3 can be sent to the smart toilet by editing text, and the smart toilet will parse and initialize it; user mode parameter setting, using the roaming function of WeChat, you can migrate the user's mode parameters from one smart toilet to another;

[0060] S4. Divert natural language commands based on preset command complexity judgment conditions and real-time network status. The existing technical solutions for controlling toilets with fixed vocabulary natural language still face multiple challenges in practical applications: First, it is limited to fixed vocabulary. If the scope is exceeded, the toilet cannot be effectively controlled, affecting the user experience. For example, a smart toilet can respond to basic commands (flushing, heating, water temperature, seat temperature, air temperature, drying, hip washing, feminine washing, defecation assistance, water pressure adjustment, position adjustment, and deodorization), but it is difficult for the local voice module to process voice commands such as playing music, adjusting the water temperature to 38 degrees, and turning on the massage mode. Secondly, for fuzzy commands (such as "flush cleanly", "higher water temperature", "a little forward"), the semantic parsing model is difficult to convert into For specific parameters, the user selects a mode after identity verification. If the mode is not selected within the specified time, the default mode will be automatically selected. In user mode, the user can interact with the smart toilet through voice. The system splits the processing based on the complexity of the command (entry length, number of parameters, or whether it is stored locally) and the real-time network status (latency, bandwidth): simple commands that the system can recognize (such as "flush", "splash", "water temperature", "water pressure", "nozzle direction", "duration", "wind temperature", etc.) are executed by the local edge layer. Commands that the system considers to be complex (such as "adjust the water temperature to 38 degrees and turn on the massage mode", "play the hero song for a while") are parsed by the cloud-based large model to generate local fixed control commands to realize the control of the smart toilet.

[0061] The instruction complexity judgment conditions in S4 include at least one of the following:

[0062] The length of the natural language instruction exceeds the preset threshold;

[0063] The natural language instruction contains more parameters than the preset number;

[0064] Natural language instructions are not stored in the local instruction library;

[0065] Real-time network latency or bandwidth meets cloud processing requirements;

[0066] S5. When the natural language instruction is determined to be a simple instruction, the local edge layer parses it and directly executes the control operation;

[0067] S6. When the natural language instruction is determined to be a complex instruction, the natural language instruction and the current status data of the smart toilet are sent to the cloud. The large cloud model parses and generates local fixed control instructions, and returns the local fixed control instructions to the smart toilet for execution. The natural language parsing model of the smart toilet is constructed through distillation, and the knowledge in a large and complex model ("teacher model") is "refined" or "distilled" into a more compact and simple model (i.e., "student model"), so that the student model can significantly reduce the consumption of computing resources and time while maintaining a high level of performance. The teacher model uses a relatively high-performance model such as the full DeepSeek-V3 as the teacher model to generate 10k inference data samples (natural language parsing task). The student model transfers knowledge to the student model through supervised fine-tuning (SFT), focusing on retaining the core reasoning ability.

[0068] The specific steps for building and deploying large cloud models in S6 include:

[0069] A1. Use the full DeepSeek-V3 teacher model to generate a 10k-sample dataset of smart toilet commands, including the mapping of natural language commands to local fixed control commands. We also use dialect conversion, fuzzy expression clarification, and teacher model construction to enhance the data. For example:

[0070] {"text": "Increase the flushing force", "label": "Water pressure adjustment +3 levels"};

[0071] {"text": "The seat is too cold", "label": "Seat temperature +5℃"};

[0072] {"text": "My butt is a little cold", "label": "Seat temperature +2℃"};

[0073] Generate augmented data through the teacher model:

[0074] "Please express 'increase the flushing water temperature of the smart toilet' in different ways, as many as possible, and in a colloquial manner."

[0075] Such as enhancing data through dialect conversion and clarifying vague expressions:

[0076] Dialect conversion (such as converting "buttocks" to "hips");

[0077] Clarifying vague expressions (such as converting "medium strength" to "gear 3");

[0078] A2. Train the natural language parsing model of the smart toilet using a three-stage distillation training strategy to meet the natural language control accuracy of the smart toilet and ensure the response speed of the model;

[0079] In A2, the natural language parsing model of the cloud smart toilet is trained using a three-stage distillation training strategy. The specific steps include:

[0080] B1. Logits distillation stage, using the divergence loss function to transfer the output probability distribution of the teacher model to the student model, and using the temperature coefficient to control the smoothness of the probability distribution. Set a higher temperature value in the initial stage to soften the probability distribution, and gradually reduce it to to harden the distribution, realizing the transition from fuzzy learning to exact matching, achieving dynamic temperature adjustment, reducing the model complexity while maintaining the semantic understanding ability. This stage focuses on the knowledge transfer of the output of the classification layer and is applicable to scenarios such as instruction recognition that require retaining semantic discrimination ability; generate instruction variants using the teacher model (such as "increase the water temperature" varying to "make the hot water hotter"), expand the diversity of training samples, and achieve data augmentation; use the divergence loss function with the temperature coefficient to enhance the learning of the relative relationship between categories. The teacher model generates soft labels for the training data (such as the probability distribution of "seat temperature adjustment"), and the student model simultaneously learns the soft labels and the original hard labels. The corresponding formula is:

[0081] ;

[0082] Among them, the temperature coefficient is used to soften the probability distribution, and the smoothness of the probability distribution is adjusted through the temperature coefficient . When , the distribution is smoother, and when , it degenerates into a standard classification task. The scaling factor is used to balance the loss magnitude, prevent gradient disappearance at high temperatures, and soften the probability distribution (the larger the value, the smoother), represents the layer loss function in knowledge distillation, represents the normalization function, represents the raw output of the student model, represents the original output of the teacher model, Represents the comparison of the output of the student model and the teacher model , Represents the information loss when using the student model to approximate the teacher model;

[0083] B2. In the feature distillation stage, the intermediate layer feature representations of the teacher model and the student model are aligned, and shadow space knowledge transfer is achieved through mean squared error loss. This strategy can capture deep semantic association features and effectively improve the small model's ability to generalize and parse ambiguous instructions (such as "turn up the heat" corresponds to "water temperature +2°C"). The corresponding formula is:

[0084] ;

[0085] in, represents the intermediate layer loss function in knowledge distillation, represents the feature dimension of the middle layer, Represents the dimension index of the intermediate layer feature vector, Indicates that the intermediate layer features of the teacher model are The output of dimension, Indicates that the intermediate layer features of the student model are The output of dimension, represents a learnable linear projection matrix that adapts to feature dimension differences, represents the square of L2 norm;

[0086] B3. Construct a joint loss function to jointly optimize knowledge transfer and task performance, achieving a balance between accuracy and efficiency in model compression scenarios. The formula of the joint loss function is:

[0087] ;

[0088] ;

[0089] in, represents the joint loss function, and represents the weight coefficient, Indicates the number of classification categories, Represents the classification category index, represents the true label, represents the predicted probability;

[0090] B4. To improve the accuracy of the student model in specific scenarios (such as smart toilet command parsing), the underlying parameters such as the embedding layer are frozen, and only the top-level classifier is fine-tuned. Hard-label cross-entropy loss is used for reinforcement training, and domain data enhancement (such as dialect conversion and parameter fuzzy expression) is introduced. Dynamic quantization (4-bit) is enabled to reduce the computational overhead during deployment. For example, layer-by-layer parameter freezing is implemented, the parameters of the embedding layer and the first four layers of the student model are fixed, and only the last two Transformer layers and the classification head are fine-tuned. In terms of hardware selection, 4-bit quantization-aware training is used to optimize memory usage. The student model is quantized and compressed. The PaddleSlim and BitsAndBytes compression libraries are used to quantize the model and compress the model size. The model supports servers with low computing power and even devices such as Jetson. The model is deployed using the Ollama framework.

[0091] The specific steps in S6 to parse the cloud-based large model and generate local fixed control instructions include:

[0092] C1. Based on Few-shot learning technology, natural language instructions are mapped into local fixed control instructions through a preset instruction example library. In the cloud, accurate parsing of natural language is achieved through the Few-Shot learning technology of a large language model. Few-Shot Learning is a machine learning method whose core idea is to learn from a small number of examples and complete specific tasks or identify new instances. Few-Shot learning aims to train the model by using only a few labeled samples (usually only a few), thereby achieving the ability to recognize and classify unseen instances. This method simulates the ability of humans to quickly master new things with only a small number of examples.

[0093] In the natural language command parsing scenario of smart toilets, Few-Shot technology uses a small number of annotated examples (e.g., "Raise seat temperature" corresponds to "seat temperature +" and "Water flow is too strong" corresponds to "water pressure -") to guide a large language model to understand the mapping relationship between user intent and fixed commands, without relying on large-scale training data. For example, the user input "My butt is a little cold" can be parsed into the command "seat temperature +" by matching the example "The seat is cold, heat it up"; while "Wash the female parts with gentle water flow" can be broken down into the combined command "female wash + water pressure -". The technology adapts to scenarios such as formulas and fuzzy expressions through a dynamic example library. For example, "End deodorization" is mapped to the command "deodorization", accurately covering the core functions of smart toilets such as seat temperature adjustment, water pressure control, and cleaning modes (female wash / female wash).

[0094] C2. Dynamically adjust the output value based on the current status data of the smart toilet. The output command samples of the large model in different scenarios are shown below. The output is in JSON format.

[0095] Example 1:

[0096] Natural language text: Please adjust the water temperature to 24℃, the water pressure to 0.5, the air temperature to 40℃, and the seat temperature to 20℃.

[0097] Model output:

[0098] { "Water Temperature": 24, "Water Pressure": 0.5, "Air Temperature": 40, "Seat Temperature": 20};

[0099] Example 2:

[0100] Natural language text: Adjust the water temperature to 38 degrees and turn on massage mode.

[0101] Model output:

[0102] {"water temperature": 24, "mode": massage};

[0103] For fuzzy instructions, data from the IoT side is needed, or when transmitting fuzzy instructions, relevant data of the smart toilet is also sent, and the fuzzy instructions and relevant data of the smart toilet are submitted to the big model together.

[0104] Example 3:

[0105] Natural language text: Please increase the water temperature a little.

[0106] The smart toilet data obtained through the Internet of Things or mixed data is:

[0107] {"waterTemp":24};

[0108] Model output:

[0109] {"waterTemp":26};

[0110] S7. Saving the local fixed control instruction to the corresponding user personalized mode according to the user identity;

[0111] S8. Execute the control instructions to complete the interactive operation of the smart toilet.

[0112] like Figure 2 As shown, this embodiment also provides an intelligent toilet control system based on natural language interaction, which is used to implement the steps of the above-mentioned intelligent toilet control method based on natural language interaction, including:

[0113] Data acquisition module 100: used to receive instructions input by the user through natural language and obtain the current status data of the smart toilet;

[0114] Identity recognition module 200: used to perform multimodal identity recognition on the user, determine the user's identity, and bind the user's identity to the unique ID of the smart toilet;

[0115] Instruction diversion module 300: used to divert natural language instructions based on preset instruction complexity judgment conditions and real-time network status;

[0116] Local voice module 400: used to assist the local edge layer in parsing instructions;

[0117] Cloud parsing module 500: used to receive natural language instructions and the current status data of the smart toilet, parse and generate local fixed control instructions using a large cloud model, and return the local fixed control instructions to the smart toilet for execution;

[0118] User mode management module 600: used to save local fixed control instructions to the corresponding user personalized mode according to the user identity;

[0119] Local execution module 700: used to receive control instructions from the local edge layer and the cloud, execute control operations, and complete interactive operations on the smart toilet.

[0120] In addition, the present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the above-mentioned smart toilet control method based on natural language interaction.

[0121] Optionally, the present invention also provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the steps of the above-mentioned smart toilet control method based on natural language interaction.

[0122] Those skilled in the art will appreciate that the process of implementing all or part of the steps of the above embodiments may be accomplished by hardware, or may be accomplished by instructing the relevant hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk, or an optical disk, etc.

[0123] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart toilet control method based on natural language interaction, characterized in that: The steps include: S1, receiving the user's instructions input through natural language and obtaining the current status data of the smart toilet; S2. Perform multimodal identity recognition on the user to determine the user's identity; S3, user mode initialization and parameter setting, users can customize multiple modes; S4. Diverting natural language instructions based on preset instruction complexity judgment conditions and real-time network status; S5. When the natural language instruction is determined to be a simple instruction, the local edge layer parses it and directly executes the control operation; S6. When the natural language instruction is determined to be a complex instruction, the natural language instruction and the current state data of the smart toilet are sent to the cloud. The cloud-based large model parses and generates a local fixed control instruction, and returns the local fixed control instruction to the smart toilet for execution. S7. Saving the local fixed control instruction to the corresponding user personalized mode according to the user identity; S8, executing the control instruction to complete the interactive operation of the smart toilet; The specific steps for building and deploying the cloud-based large model in S6 include: A1. Use the full DeepSeek-V3 teacher model to generate a 10k-sample smart toilet command dataset, including the mapping between natural language commands and local fixed control commands. A2. Use a three-stage distillation training strategy to train the smart toilet natural language parsing model; The cloud-based smart toilet natural language parsing model in A2 is trained using a three-stage distillation training strategy. The specific steps include: B1, Logits distillation stage, using The divergence loss function transfers the output probability distribution of the teacher model to the student model, using the temperature coefficient Control the smoothness of the probability distribution and set a higher temperature value in the initial stage Soften the probability distribution and gradually reduce it to By hardening the distribution, the model complexity is reduced while maintaining the semantic understanding ability of the student large model; the teacher model is used to generate instruction variants, expand the diversity of training samples, and achieve data enhancement; using temperature coefficients of Divergence loss function, the teacher model uses the training data to generate soft labels, and the student model learns both the soft labels and the original hard labels. The corresponding formula is: ； in, Represents knowledge distillation Layer loss function, represents the normalization function, represents the raw output of the student model, represents the original output of the teacher model, Represents the information loss when using the student model to approximate the teacher model; B2, feature distillation stage, aligning the intermediate layer feature expressions of the teacher model and the student model; B3. Construct a joint loss function to jointly optimize knowledge transfer and task performance; B4, parameter freezing and fine-tuning phase: freeze the underlying parameters of the student model and fine-tune only the top-level classifier and attention layer. Strengthen training with hard-label cross-entropy loss, introduce domain data augmentation, and enable dynamic quantization to reduce computational overhead during deployment. In B2, the intermediate layer feature expressions of the teacher model and the student model are forced to be aligned, and the shadow space knowledge transfer is achieved through the mean square error loss. The corresponding formula is: ； in, represents the intermediate layer loss function in knowledge distillation, represents the feature dimension of the middle layer, Represents the dimension index of the intermediate layer feature vector, Indicates that the intermediate layer features of the teacher model are The output of dimension, Indicates that the intermediate layer features of the student model are The output of dimension, represents a learnable linear projection matrix that adapts to feature dimension differences, represents the square of L2 norm; The formula for constructing the joint loss function in B3 is: ； ； in, represents the joint loss function, and represents the weight coefficient, Indicates the number of classification categories, Represents the classification category index, represents the true label, represents the predicted probability.

2. The intelligent toilet control method based on natural language interaction according to claim 1, characterized in that: The multimodal identity recognition in S2 includes at least one or more combinations of voiceprint recognition, fingerprint recognition and WeChat identity recognition, and the user identity is bound to the unique ID of the smart toilet; the user mode initialization in S3 can be achieved by editing text, transmitting it to the smart toilet, and initializing it by parsing it; user mode parameter setting, using the roaming function of WeChat, can migrate the user's mode parameters from one smart toilet to another.

3. The intelligent toilet control method based on natural language interaction according to claim 1, characterized in that: The instruction complexity judgment condition in S4 includes at least one of the following: The length of the natural language instruction exceeds the preset threshold; The natural language instruction contains more parameters than the preset number; Natural language instructions are not stored in the local instruction library; Real-time network latency or bandwidth meets cloud processing conditions.

4. The intelligent toilet control method based on natural language interaction according to claim 1, characterized in that: The specific steps of the cloud-based large model parsing and generating local fixed control instructions in S6 include: C1, based on Few-shot learning technology, maps natural language commands into local fixed control commands through a preset command example library; C2. Dynamically adjust the output value based on the current status data of the smart toilet.

5. A smart toilet control system based on natural language interaction, used to implement the steps of the smart toilet control method based on natural language interaction described in any one of claims 1 to 4, characterized in that: include: Data acquisition module (100): used to receive instructions input by the user through natural language and obtain current status data of the smart toilet; Identity recognition module (200): used to perform multimodal identity recognition on the user, determine the user's identity, and bind the user's identity to the unique ID of the smart toilet; Instruction diversion module (300): used for diverting natural language instructions according to preset instruction complexity judgment conditions and real-time network status; Local voice module (400): used to assist the local edge layer in parsing instructions; Cloud parsing module (500): used to receive natural language instructions and current status data of the smart toilet, parse and generate local fixed control instructions using a cloud-based large model, and return the local fixed control instructions to the smart toilet for execution; User mode management module (600): used to save local fixed control instructions to the corresponding user personalized mode according to the user identity; Local execution module (700): used to receive control instructions from the local edge layer and the cloud, execute control operations, and complete interactive operations on the smart toilet.

6. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the smart toilet control method based on natural language interaction described in any one of claims 1 to 4 are implemented.

Images