Intelligent liquid blending device

Abstract

The invention relates to the field of ornamental green plant maintenance, and provides an intelligent liquid blending device which comprises a mixing cavity used for collecting stock solution and water blending feed liquid; the stock solution bin is used for storing stock solution and is communicated with the mixing cavity through a solution adding device, and the solution adding device is provided with a piezoelectric atomization module; the water inlet pipeline is used for conveying water to the mixing cavity, and the water inlet pipeline is provided with a water inlet valve; the discharging pipeline is used for discharging the feed liquid in the mixing cavity, and the discharging pipeline is provided with a flow meter; the interaction module is used for acquiring set concentration and set flow input by a user; the control module is used for controlling the opening degree of the water inlet valve according to the set flow and collecting the feedback flow of the flow meter; inputting the set concentration, the feedback flow and the stock solution concentration into an atomization amount calculation model, and calculating a target set atomization amount after local fine adjustment and global adjustment of the atomization amount calculation model; and generating a corresponding piezoelectric control parameter based on the target set atomization amount so as to indicate the piezoelectric atomization module to execute an atomization action.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ornamental plant maintenance, and particularly to an intelligent liquid preparation device. BACKGROUND

[0002] Fertilization is a necessary means for daily maintenance of garden and courtyard plants. Liquid fertilizer has the characteristics of easy absorption and high utilization rate, and is a commonly used fertilizer type in the field of garden and courtyard plant maintenance. With the existing water and fertilizer integrated irrigation system, intelligent preparation and application of liquid fertilizer can be realized. However, the above irrigation system is only suitable for commercial fields, and is suitable for large parks and urban green belts. The equipment is large in size and high in cost, and is not suitable for civilian fields. On the other hand, the maintenance of ornamental plants planted in residential courtyards and terraces is relatively more difficult and more sensitive to the concentration of liquid fertilizer, requiring high preparation accuracy. At present, there is no intelligent equipment on the market that can prepare and apply liquid fertilizer for such scenes. SUMMARY

[0003] The main purpose of the present application is to provide an intelligent liquid preparation device that can meet the requirement of high preparation accuracy.

[0004] To achieve the above purpose, the present application provides an intelligent liquid preparation device, comprising: a mixing chamber for collecting raw liquid and water to prepare liquid; a raw liquid bin for storing raw liquid and communicating with the mixing chamber through a liquid adding device, the liquid adding device having a piezoelectric atomization module; a water inlet pipeline for conveying water to the mixing chamber, the water inlet pipeline being provided with a water inlet valve; a discharge pipeline for discharging the liquid in the mixing chamber, the discharge pipeline being provided with a flow meter; an interactive module for obtaining a set concentration and a set flow rate input by a user; a control module for controlling the opening degree of the water inlet valve according to the set flow rate, collecting the feedback flow rate of the flow meter, inputting the set concentration, the feedback flow rate and the raw liquid concentration into an atomization amount calculation model, calculating the target set atomization amount after local fine tuning and global adjustment of the atomization amount calculation model, generating corresponding piezoelectric control parameters based on the target set atomization amount to instruct the piezoelectric atomization module to perform atomization action, constructing a multi-modal perception system by adding an electric conductivity sensor, a pH sensor and a camera, identifying plant varieties and growth states through a convolutional neural network, and real-time inferring the current optimal nutrient solution concentration in combination with the corresponding sensor data of the electric conductivity sensor and the pH sensor; the growth state includes leaf color and lesion.

[0005] In some embodiments, the set concentration, feedback flow rate and stock solution concentration are input into the atomization amount calculation model, and after local fine tuning and global adjustment of the atomization amount calculation model, the target set atomization amount is calculated, including: normalizing the set concentration, feedback flow rate and stock solution concentration according to the input layer of the preset atomization amount calculation model to obtain numerical type variables corresponding to the normalized set concentration, feedback flow rate and stock solution concentration; performing feature processing on the normalized numerical type variables according to the middle layer of the preset atomization amount calculation model to obtain a feature vector; mapping the feature vector to an initial set atomization amount according to the output layer of the preset atomization amount calculation model; based on the initial set atomization amount, locally fine tuning the output layer of the preset atomization amount calculation model according to a pre-constructed reward function, remapping the feature vector according to the locally fine tuned output layer to obtain a candidate set atomization amount; based on the local fine tuning value, globally adjusting the output layer of the preset atomization amount calculation model according to a pre-constructed federated learning algorithm, and remapping the feature vector according to the globally fine tuned output layer to obtain the target set atomization amount.

[0006] In some embodiments, based on the initial set atomization amount, the output layer of the preset atomization amount calculation model is locally fine tuned according to a pre-constructed reward function, and the feature vector is remapped according to the locally fine tuned output layer to obtain a candidate set atomization amount, including: evaluating the effect of the initial atomization amount by the pre-constructed reward function, and calculating the gradient feedback of the reward function to the initial atomization amount to obtain the improvement direction of the current output quality; based on the gradient feedback of the reward function, the weight and bias parameters of the output layer are locally fine tuned using a local learning rate, and the purpose of the adjustment is to quickly improve the direction of the current output without affecting the global parameters of the entire model; the feature vector is input into the locally fine tuned output layer and forward propagated again to obtain an updated candidate set atomization amount.

[0007] In some embodiments, based on the local fine tuning value, the output layer of the preset atomization amount calculation model is globally adjusted according to a pre-constructed federated learning algorithm, and the feature vector is remapped by the locally fine tuned output layer to obtain the target set atomization amount, including: each participating device locally completes the output layer local fine tuning operation based on the reward function, and calculates the local parameter update amount relative to the original output layer parameters; all clients upload their respective local parameter update amounts to the server, and the server aggregates all client local parameter update amounts by a federated averaging algorithm to generate a set of globally updated output layer global update parameters, so that the model can fuse the optimization experience of multiple devices while maintaining privacy; the updated global output layer parameters of the server are used to forward calculate the feature vector again to output the target set atomization amount.

[0008] In some embodiments, when the server aggregates the local parameter update amount of all clients through a federated averaging algorithm, the server is further configured to perform the following steps: obtaining the local parameter update amount uploaded by each client and the corresponding device runtime length data; calculating the weight coefficient of each client according to the device runtime length data, wherein the longer the runtime length, the higher the weight coefficient corresponding to the client; performing weighted average on the local parameter update amount of each client according to the weight coefficient to generate a global updated output layer global update parameter; before issuing the global update parameter to each client, performing encryption processing on the global update parameter, the encryption processing including hierarchical encryption of the parameter using a homomorphic encryption algorithm, so that each client can only obtain the parameter part matching the local model structure when decrypting.

[0009] In some embodiments, generating the corresponding piezoelectric control parameter based on the target setting atomization amount includes: obtaining a parameter set reflecting the atomization characteristics of the raw liquid at the working temperature; performing function fitting between the parameter set and the piezoelectric control parameter and the atomization amount to establish an atomization parameter control model; and calculating the set piezoelectric control parameter corresponding to the target setting atomization amount according to the atomization parameter control model, the piezoelectric control parameter including a driving voltage, and the parameter set including the viscosity and surface tension of the raw liquid.

[0010] In some embodiments, further comprising: a concentration detection sensor configured in the discharge pipeline, the control module further collects the feedback concentration of the concentration detection sensor, calculates the deviation value between the feedback concentration and the set concentration, and compares it with the set range, the control module issues a raw liquid abnormal prompt information through the interaction module when the deviation value exceeds the set range, and the concentration detection sensor includes a conductivity sensor.

[0011] In some embodiments, the discharge pipeline has a discharge direction, the mixing chamber has an upper end face located on the upstream side of the discharge direction, a lower end face located on the downstream side of the discharge direction, and a side wall surface connecting the upper end face and the lower end face, the projection area of the upper end face along the discharge direction is greater than the projection area of the lower end face along the discharge direction, the water inlet pipeline communicates with the mixing chamber at the upper end face, the liquid feeder communicates with the mixing chamber at the side wall surface, and the discharge pipeline communicates with the mixing chamber at the lower end face.

[0012] In some embodiments, the flow meter includes an electromagnetic coupling rotating body and a detection module, the rotating body is placed in the discharge pipeline and can rotate along a plane perpendicular to the discharge direction, and the detection module includes a Hall sensor.

[0013] In some embodiments, the rotating body comprises a driving member and a driven member provided with a magnet, the driving member comprises a rotating shaft part parallel to the discharging direction and a plurality of fan parts arranged around the outside of the rotating shaft part, the fan parts have concave arc surfaces intersecting the discharging direction, the driven member is connected to the discharging direction end of the rotating shaft part, and the flow meter further comprises a guide plate located on the side of the driving member opposite to the discharging direction, the guide plate is provided with a plurality of guide holes for guiding the liquid to contact the concave arc surfaces.

[0014] As can be seen from the above scheme, the above-mentioned dispensing device uses a piezoelectric atomization module for liquid dispensing, which helps to accurately control the concentration of the liquid and improve the preparation efficiency of the liquid. The control module can control the dilution ratio of the original liquid, i.e. the concentration of the liquid, by adjusting the opening degree of the water inlet valve and the atomization amount. The feedback flow collected by the flow meter can reflect the real flow of the liquid. The control module can calculate the target set atomization amount based on the atomization amount calculation model of the feedback flow and the set concentration and set flow collected by the interaction module, so as to accurately control the concentration of the liquid. Further, based on the target set atomization amount, the corresponding piezoelectric control parameters can be calculated to accurately control the atomization action.

[0015] In view of the high sensitivity of liquid fertilizer concentration of ornamental plants in residential courtyards and terraces, by using the atomization amount calculation model (local fine tuning + global adjustment) of the control module and the micron-level atomization precision of the piezoelectric atomization module, the concentration control error can be controlled within ±0.3%, so as to avoid the problems of root burning and malnutrition of plants caused by concentration deviation, and significantly improve the safety of family green plant maintenance.

[0016] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating any creative labor.

[0018] In order to more completely understand the present application and its beneficial effects, the following will be described in conjunction with the drawings, wherein the same reference numerals in the following description represent the same parts.

[0019] Figure 1 is the external structure diagram of the intelligent liquid dispensing device embodiment of the present application.

[0020] Figure 2 is the internal structure diagram of the intelligent liquid dispensing device embodiment of the present application.

[0021] Figure 3 is a sectional structure diagram of the present application. Figure 2

[0022] Figure 4 is an enlarged structural diagram of A in the present application. Figure 3

[0023] Figure 5 is a partial structural diagram of a flow meter in an embodiment of the intelligent liquid dispensing device of the present application.

[0024] Figure 6 is a structural exploded view of the present application. Figure 5

[0025] Reference signs: 1, mixing cavity; 11, upper end face; 12, lower end face; 13, side wall face; 2, raw liquid bin; 3, liquid adding device; 31, piezoelectric atomization module; 4, water inlet pipeline; 41, water inlet valve; 5, discharge pipeline; 6, flow meter; 61, rotating body; 611, driving piece; 611a, rotating shaft part; 611b, fan blade part; 611c, concave arc face; 612, driven piece; 612a, magnet; 62, detection module; 63, flow guide plate; 631, guide hole; 64, support part; 7, interactive module; 71, display screen; 72, rotary knob; 73, button; 8, solar cell module; 9, concentration detection sensor.

[0026] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. DETAILED DESCRIPTION

[0027] Hereinafter, the preferred embodiments of the present application will be described in detail with reference to the accompanying drawings. In the following description, the same parts are given the same reference numerals, and repeated description is omitted. In addition, the drawings are only schematic views, and the ratio of the dimensions between the parts or the shape of the parts and the like can be different from the actual ones. It should be noted that all the directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition and the like between the parts in a certain specific posture, and if the specific posture changes, the directional indications also change accordingly.

[0028] It should also be noted that when an element is referred to as being “fixed to” or “disposed on” another element, it can be directly on the other element or a middle element can be present at the same time. When an element is referred to as being “connected” to another element, it can be directly connected to the other element or a middle element can be present at the same time.

[0029] ​​​Fertilization is a necessary means for daily maintenance of garden and courtyard green plants. Liquid fertilizer has the characteristics of easy absorption and high utilization rate, and is a commonly used fertilizer type in the field of garden and courtyard green plant maintenance. With the existing water and fertilizer integrated irrigation system, intelligent deployment and application of liquid fertilizer can be realized. However, the above irrigation system is only suitable for commercial field, and is suitable for large parks and urban green belts. The equipment is large in size and high in cost, and is not suitable for civilian field. On the other hand, the maintenance of ornamental plants planted in residential courtyards and terraces is relatively more difficult and more sensitive to the concentration of liquid fertilizer, requiring high deployment accuracy. At present, there is no intelligent device on the market suitable for deploying and applying liquid fertilizer in such scenarios.

[0030] Referring to Figures 1-6 The intelligent liquid deployment device provided by the application comprises a mixing cavity 1 for collecting raw liquid and water to deploy liquid; a raw liquid bin 2 for storing raw liquid and communicating with the mixing cavity 1 through a liquid adding device 3, the liquid adding device 3 having a piezoelectric atomization module 31; a water inlet pipeline 4 for conveying water to the mixing cavity 1, the water inlet pipeline 4 being provided with a water inlet valve 41; a discharge pipeline 5 for discharging the liquid in the mixing cavity 1, the discharge pipeline 5 being provided with a flow meter 6; an interactive module 7 for obtaining a set concentration and a set flow rate input by a user; a control module for controlling the opening degree of the water inlet valve 41 according to the set flow rate, collecting the feedback flow rate of the flow meter 6; inputting the set concentration, the feedback flow rate and the raw liquid concentration into an atomization amount calculation model, calculating the target set atomization amount after local fine tuning and global adjustment of the atomization amount calculation model; generating corresponding piezoelectric control parameters based on the target set atomization amount to instruct the piezoelectric atomization module to perform atomization action; constructing a multi-modal perception system by adding an electric conductivity sensor, a pH sensor and a camera, identifying plant varieties and growth states through a convolutional neural network, and real-time inferring the current optimal nutrient solution concentration in combination with the corresponding sensor data of the electric conductivity sensor and the pH sensor; the growth state includes leaf color and lesion.

[0031] The deployment device uses the piezoelectric atomization module 31 to atomize the raw liquid and input it into the mixing cavity 1 to mix with the conveyed water, which helps to accurately control the concentration of the liquid and improve the preparation efficiency of the liquid. The control module can control the dilution ratio of the raw liquid, i.e. the concentration of the liquid, by adjusting the opening degree of the water inlet valve 41 and the atomization amount. The feedback flow rate collected by the flow meter 6 can reflect the real flow rate of the liquid. The control module can calculate the target set atomization amount based on the feedback flow rate, the set concentration and the set flow rate collected by the interactive module 7 through the atomization amount calculation model, so as to accurately control the concentration of the liquid. Further based on the target set atomization amount, corresponding piezoelectric control parameters can be calculated to accurately control the atomization action.

[0032] The intelligent liquid preparation device provided by the application is a mechatronic scheme, has a casing, the display screen 71, the knob 72 and the button 73 which are parts of the interactive module 7 are all arranged on the casing. In addition, the interface part of the water inlet pipeline 4 and the discharge pipeline 5 is also exposed outside the casing, the water inlet pipeline 4 can be directly communicated with a faucet, the discharge pipeline 5 can be communicated with a yard spraying device, the raw liquid bin 2 is arranged on the upper side of the casing, the main part of the water inlet pipeline 4 and the discharge pipeline 5, the water inlet valve 41, the liquid adding device 3 and the flow meter 6 are all arranged inside the casing.

[0033] The discharge pipeline 5 has a discharge direction D1, the mixing cavity 1 has an upper end face 11 located on the upstream side of the discharge direction D1, a lower end face 12 located on the downstream side of the discharge direction D1 and a side wall face 13 connecting the upper end face 11 and the lower end face 12, the projection area of the upper end face 11 along the discharge direction D1 is greater than the projection area of the lower end face 12 along the discharge direction D1, the water inlet pipeline 4 is communicated with the mixing cavity 1 at the upper end face 11, the liquid adding device 3 is communicated with the mixing cavity 1 at the side wall face 13, and the discharge pipeline 5 is communicated with the mixing cavity 1 at the lower end face 12.

[0034] The mixing cavity 1 is roughly funnel-shaped, the water inlet pipeline 4 and the discharge pipeline 5 are respectively communicated with the upper and lower ends of the mixing cavity 1, and the communication part of the water inlet pipeline 4 with the mixing cavity 1 is located on the upstream side of the communication part of the liquid adding device 3 with the mixing cavity 1, so that the mixing efficiency of the raw liquid and water can be improved by using turbulent flow and the concentration uniformity of the liquid can be improved.

[0035] The flow meter 6 includes a rotating body 61 and a detection module 62 which are electromagnetically coupled, the rotating body 61 is arranged in the discharge pipeline 5 and can rotate along a plane perpendicular to the discharge direction D1, and the detection module 62 includes a Hall sensor.

[0036] The flow meter 6 belongs to a turbine flow meter and has high detection precision, the rotating body 61 arranged in the discharge pipeline 5 can promote the mixing of the raw liquid and water, so that the preparation of the liquid can be assisted to a certain extent.

[0037] The rotating body 61 includes a driving part 611 and a driven part 612 provided with a magnet 612a, the driving part 611 includes a rotating shaft part 611a parallel to the discharge direction D1 and a plurality of fan blade parts 611b arranged outside the rotating shaft part 611a, the fan blade part 611b has a concave arc surface 611c intersecting the discharge direction D1, the driven part 612 is connected to the discharge direction D1 end of the rotating shaft part 611a, and the flow meter 6 further includes a flow guide plate 63 located on the side of the driving part 611 opposite to the discharge direction D1, and the flow guide plate 63 is provided with a plurality of guide holes 631 guiding the liquid to contact the concave arc surface 611c.

[0038] The flow guide plate with multiple guide holes 631 can guide the liquid to fully act on the driving member 611, so that the liquid can better drive the rotating body 61, and the rotating speed of the rotating body 61 can better match the flow rate of the liquid, thereby helping to improve the detection accuracy of the flowmeter 6. The cooperation of the flow guide plate 63 and the rotating body 61 can also promote the mixing of the raw liquid and water, thereby helping to improve the uniformity of the concentration of the liquid.

[0039] In addition, the flowmeter further comprises a support part 64, the support part 64 comprises a horizontal rod and vertical rods arranged at both ends of the horizontal rod, the support part 64 is connected with the flow guide plate 63 through the vertical rod part, the lower end of the rotating shaft part 611a is connected with the horizontal rod part of the support part 64 through the follower 612, and the upper end of the rotating shaft part 611a is connected with the flow guide plate 63. The two magnets 612a are symmetrically arranged in the two grooves of the follower.

[0040] The intelligent liquid dispensing device further comprises a power supply module, and the power supply module comprises a solar cell module 8. Considering that the sunlight is sufficient in outdoor environments such as courtyards and terraces, but it is inconvenient to take electricity, the solar cell module 8 can be used as a supplement to the conventional power supply mode of the intelligent liquid dispensing device, and in the case that the power is sufficient, the solar cell module 8 can also be used as the main power supply means of the intelligent liquid dispensing device. In the embodiment, the solar cell module 8 is arranged on the top surface of the housing.

[0041] The intelligent liquid dispensing device further comprises a concentration detection sensor 9 arranged on the discharge pipeline 5, and the control module further collects the feedback concentration of the concentration detection sensor 9, calculates the deviation value between the feedback concentration and the set concentration, and compares it with the set range. When the deviation value exceeds the set range, the control module sends an abnormal raw liquid prompt information through the interaction module 7. The concentration detection sensor comprises an electric conductivity sensor.

[0042] The raw liquid may be deteriorated due to its own quality problems or improper storage, and the fertilizer components in the raw liquid may be partially degraded and invalid. Specifically, there is a significant deviation between the actual concentration of the fertilizer components in the liquid and the set concentration. The concentration detection sensor 9 arranged in the discharge pipeline 5 can introduce a feedback mechanism for the concentration of the liquid into the intelligent liquid dispensing device. When the deviation value between the feedback concentration and the set concentration exceeds the set range, the control module can determine that the raw liquid is deteriorated and prompt the user to replace the liquid through the interaction module 7, so as to better cope with the liquid problems caused by the deterioration of the raw liquid. Considering that most of the fertilizer components of the raw liquid are ionic compounds, an electric conductivity sensor can be selected as the concentration detection sensor 9.

[0043] By adding EC (electric conductivity) sensors, pH sensors and cameras in the device, a multi-modal perception system is constructed. Through a convolutional neural network (CNN), the plant variety and growth state (such as leaf color, disease spot) are recognized, and the current optimal nutrient solution concentration is inferred in real time combined with sensor data, forming a “perception-decision-execution” closed-loop self-learning system.

[0044] Multi-modal data acquisition Real-time monitoring of conductivity (reflecting ion concentration) and pH of the outflow liquid in the mixing chamber by EC / pH sensor, periodic image capture of plant leaf by camera.

[0045] Data fusion model includes: design of multi-branch neural network: image branch: CNN extracts leaf features (such as chlorophyll content, leaf area); sensor branch: fully connected layer processes EC / pH / flow data; fusion layer: weighted fusion of two types of features through attention mechanism, output recommended concentration Crecommended.

[0046] Self-learning mechanism records the growth feedback of the plant after each fertilization (through image comparison of leaf state changes), updates the fusion model parameters using backpropagation, and forms a closed loop of "fertilization effect-model optimization".

[0047] No need for users to input plant varieties, the device automatically identifies and matches the best concentration (such as automatically increasing the iron element fertilizer concentration when identifying that the rose leaves are yellow), reduces the maintenance threshold, and is suitable for gardening beginners. Combined with physiological indicators (EC value reflects ion concentration) and apparent characteristics (leaf state), dynamically adjust the formula, solve the blindness of traditional "fixed concentration + experience fertilization", promote the healthy growth of plants (such as reducing the problem of root burning caused by improper concentration). Through continuous learning of plant maintenance data of different users (under the premise of protecting privacy through federated learning), the model can identify rare varieties and generate adaptive solutions, building an evolving plant nutrition knowledge base.

[0048] In some embodiments, based on the target setting, the corresponding piezoelectric control parameters are generated, including: obtaining a parameter set reflecting the atomization characteristics of the original liquid at the working temperature; function fitting between the parameter set and the piezoelectric control parameters and the atomization amount, establishing an atomization parameter control model; calculating the set piezoelectric control parameters corresponding to the target set atomization amount according to the atomization parameter control model, the piezoelectric control parameters including the driving voltage, and the parameter set including the viscosity and surface tension of the original liquid.

[0049] The atomization amount of the piezoelectric atomization module has a certain relationship with the piezoelectric control parameters and the specific physicochemical parameters of the original liquid at the working temperature, which can be understood with reference to the following formula: Q=(C*V 2 *D*A / γ) / (f*n) 0.5 ; Wherein Q is the piezoelectric atomization amount, V is the driving voltage, D is the diameter of the piezoelectric element, A is the nozzle area, f is the vibration frequency, η is the viscosity of the original liquid, γ is the surface tension of the original liquid, and C is the device characteristic constant.

[0050] To accurately control the atomization amount Q, it is necessary to establish an atomization parameter control model based on piezoelectric control parameters and specific physicochemical parameters of the original liquid at the working temperature. Referring to the above formula, the driving voltage V is taken as the piezoelectric control parameter, the viscosity η and the surface tension γ of the original liquid are taken as the physicochemical parameters of the original liquid, and the piezoelectric element diameter D, the nozzle area A and the equipment characteristic constant C can be combined into a constant. Among them, the physicochemical parameters of the original liquid are a set of data sets related to the type, concentration and working temperature of the original liquid. The amount of related data required for function fitting is large, and specific related databases or data models can be established locally for storage, or they can also be retrieved from the cloud as needed.

[0051] In some embodiments, the set concentration, feedback flow and original liquid concentration are input into the atomization amount calculation model, and after local fine tuning and global adjustment of the atomization amount calculation model, the target set atomization amount is calculated, including: normalizing the set concentration, feedback flow and original liquid concentration according to the input layer of the preset atomization amount calculation model, to obtain numerical type variables corresponding to the normalized set concentration, feedback flow and original liquid concentration; according to the intermediate layer of the preset atomization amount calculation model, the normalized numerical type variables are processed to obtain a feature vector; according to the output layer of the preset atomization amount calculation model, the feature vector is mapped to an initial set atomization amount; according to the pre-constructed reward function, the output layer of the preset atomization amount calculation model is locally fine tuned based on the initial set atomization amount, the feature vector is remapped according to the output layer after local fine tuning, to obtain a candidate set atomization amount; according to the pre-constructed federated learning algorithm, the output layer of the preset atomization amount calculation model is globally adjusted based on the local fine tuning value, and the feature vector is mapped again according to the output layer after global fine tuning, to obtain the target set atomization amount.

[0052] The core processing flow of the atomization amount calculation model includes input layer normalization, intermediate layer feature processing, output layer initial mapping, and local fine tuning based on the reward function and global adjustment based on the federated learning algorithm. The input layer processing converts the set concentration, feedback flow and original liquid concentration into numerical type variables by normalizing them to adapt to the model input. The intermediate layer feature processing extracts and processes the normalized variables through the intermediate layer (such as the neural network hidden layer) of the preset model to generate a feature vector. The output layer initial mapping maps the feature vector to an initial set atomization amount using the output layer. The local fine tuning evaluates the effect of the initial atomization amount based on the reward function, adjusts the output layer parameters, and generates a candidate set atomization amount. The global adjustment aggregates the local adjustment experience of multiple devices through the federated learning algorithm to optimize the output layer parameters, and finally obtains the target set atomization amount.

[0053] The normalization process employs minimum-maximum normalization or Z-score normalization methods to convert the set concentration (range 0-100%), feedback flow rate (unit L / min), and stock solution concentration (percentage) into numerical variables within the range of [0,1] or [-1,1], thus eliminating the influence of dimensions.

[0054] Feature processing can be designed as a multilayer perceptron (MLP) or convolutional neural network (CNN) through intermediate layers. Nonlinear features are extracted through weighted summation and activation functions (such as ReLU) to generate feature vectors containing the concentration-flow correlation.

[0055] The initial mapping involves a linear output layer, where the feature vector is mapped to the initial atomization level using a weight matrix. The formula is: Initial atomization level = Feature vector × W + b; where W is the weight and b is the bias. Local fine-tuning is achieved by defining the reward function as a negative function of the concentration error (e.g., ...). Where R is the value of the reward function, C is the set concentration, such as the target liquid concentration input through the interactive module (such as the display screen, knob, etc.), representing the concentration value of the nutrient solution or preparation solution expected by the user (usually expressed as a percentage or conductivity, etc.). C is actually the actual concentration of the liquid detected in real time by the concentration detection sensor (such as the conductivity sensor) in the device, reflecting the true concentration value of the current preparation result. It is collected in real time by the concentration detection sensor in the discharge pipeline and directly feeds back the actual ion concentration or other concentration indicators of the liquid discharged from the current mixing chamber. The output layer parameters are adjusted by gradient descent to quickly optimize the atomization calculation accuracy of the current device.

[0056] Global adjustments are made by uploading the differences in parameters after local adjustments from each device. The server then aggregates data from multiple devices through federated learning to update the global model parameters, thereby improving the universality across devices.

[0057] By normalizing input parameters of different dimensions, the model ensures consistent processing capabilities for data such as concentration and flow rate. Feature processing in the intermediate layers uncovers the nonlinear correlation between concentration and flow rate, improving the theoretical accuracy of atomization calculations. Combining local fine-tuning and global adjustment satisfies the real-time optimization needs of single devices while leveraging multi-device data to enhance the model's generalization ability, avoiding overfitting or underfitting. Through reward functions and federated learning, the model can dynamically adjust to the high-precision requirements of small scenarios such as residential courtyards, solving the "one-size-fits-all" adaptation problem of traditional devices.

[0058] In some embodiments, according to the pre-constructed reward function, the output layer of the preset atomization amount calculation model is locally fine-tuned based on the initial set atomization amount, the feature vector is remapped according to the locally fine-tuned output layer, and a candidate set atomization amount is obtained, comprising: evaluating the effect of the initial atomization amount by the pre-constructed reward function, and calculating the gradient feedback of the reward function to the initial atomization amount to obtain the improvement direction of the current output quality; based on the gradient feedback of the reward function, the weight and bias parameters of the output layer are locally fine-tuned using a local learning rate, and the purpose of the adjustment is to quickly improve the direction of the current output without affecting the global parameters of the entire model; the feature vector is input into the locally fine-tuned output layer, and forward propagation is performed again to obtain an updated candidate set atomization amount.

[0059] The effect of the initial atomization amount is evaluated by the reward function, the gradient feedback is calculated to adjust the weight and bias parameters of the output layer, and the atomization amount calculation accuracy of the current device is quickly optimized. The reward function evaluation generates gradient feedback to indicate the parameter adjustment direction by quantifying the concentration error corresponding to the initial atomization amount. Local parameter adjustment adjusts only the output layer parameters by using a learning rate independent of the global model, avoiding affecting the general feature extraction capability of the intermediate layer. Forward propagation update realizes real-time optimization by recalculating the candidate set atomization amount based on the adjusted output layer parameters. The reward function design defines the square loss function of the concentration error , where R is the value of the reward function, Cset is the set concentration, such as the target feed liquid concentration input through the interaction module (such as a display screen, a knob, etc.), representing the concentration value of the nutrient solution or the blending solution that the user expects (usually expressed in percentage or conductivity, etc.), Cactual is the actual concentration of the feed liquid detected in real time by the concentration detection sensor (such as a conductivity sensor) in the device, reflecting the true concentration value of the current blending result, which is collected by the concentration detection sensor in the discharge pipeline and directly feeds back the actual ion concentration or other concentration indicators of the mixed cavity discharge liquid.

[0060] , where (suppose the atomization amount is much smaller than the water flow, which can be simplified as ).

[0061] The gradient calculation calculates the gradient of the reward function to the weight W and bias b of the output layer: ; ; The update formula is: ; , where alocal is a learning rate independent of the global model, usually set to a large value to speed up local optimization, R is the value of the reward function, WR is the gradient of the reward function with respect to the output layer weights W, W' is the updated output layer weight matrix for the next forward pass calculation of the atomization amount, W is the output layer weight matrix of the current iteration (value before updating), a is the local W R is the direction and magnitude of weight update, adjusted in the opposite direction of the reward function gradient (gradient descent) to minimize the concentration error (maximize the reward R), b' is the updated output layer bias vector. b is the output layer bias vector of the current iteration (value before updating), similar to weight update, the deviation of the initial atomization amount is compensated by the adjustment of the bias, and the concentration error is quickly corrected, b R is the gradient of the reward function with respect to the output layer bias vector b of the current iteration.

[0062] The candidate atomization amount is generated by inputting the feature vector into the updated output layer and calculating the candidate set atomization amount by forward propagation: .

[0063] The concentration error is directly related to the reward function, so that the device can quickly adjust the atomization amount for the current irrigation scene, solving the problem of concentration sensitivity in small-batch fertilization (such as poor tolerance of high-concentration fertilizer for balcony plants). Only the output layer parameters are adjusted to avoid retraining the entire model, reducing the amount of calculation and energy consumption, and adapting to the low computing power requirements of consumer devices. Each device can be optimized independently to adapt to the differentiated needs of different users' plant varieties (such as succulents and flowers) for concentration, improving the deployment accuracy of a single device.

[0064] In some embodiments, according to a pre-constructed federated learning algorithm, the output layer of the preset atomization amount calculation model is globally adjusted based on the local fine-tuning value, and the output layer after local fine-tuning is mapped again to the feature vector to obtain a target set atomization amount, including: each participating device locally completes the output layer local fine-tuning operation based on the reward function, and calculates the local parameter update amount relative to the original output layer parameters; all clients upload their respective local parameter update amounts to the server, and the server aggregates all client local parameter update amounts through a federated averaging algorithm to generate a set of global updated output layer global update parameters, so that the model can integrate the optimization experience of multiple devices while maintaining privacy; the updated global output layer parameters of the server are used to re-calculate the feature vector forwardly, and output the target set atomization amount.

[0065] By aggregating the local parameter update amounts of multiple devices using a federated learning algorithm, the global model parameters are optimized while protecting user privacy, so that the atomization amount calculation model can integrate the maintenance experience of different scenarios.

[0066] Local parameter update is completed by each device locally to fine-tune the output layer, and the parameter update amount (such as , W is an initial weight matrix of a preset atomization amount calculation model output layer (initial parameters issued by a global model or parameters updated after a previous global update), which is a basic parameter for linear mapping of the output layer and maps the feature vector generated by the intermediate layer to an initial set atomization amount, b is an initial bias vector of the output layer, which determines the linear mapping result of the initial atomization amount together with the weight matrix W, W' is an output layer weight matrix adjusted by each device locally through gradient descent of a reward function, b' is an output layer bias vector after local fine-tuning, AW is an update increment of the weight matrix, which is equal to the difference between the fine-tuned weight and the original weight, and Ab is an update increment of the bias vector, which is equal to the difference between the fine-tuned bias and the original bias). Parameter aggregation is to collect local update amounts of all devices by the server, and generate global update parameters through a federated average algorithm (FedAvg). Global model update is to update the output layer weight and bias by issuing global parameters to each device, and to improve the cross-device generalization ability of the model.

[0067] Local update amount calculation records the change amount AWi, Abi of the output layer parameters after local fine-tuning by each device (i is the device number). The federated average algorithm calculates global update parameters where N is the number of participating devices.

[0068] The privacy protection mechanism avoids user planting data (such as plant varieties and fertilization frequency) from being leaked by uploading only parameter update amounts instead of original data or complete models. Global parameter issuance updates the local output layer after each device receives the global update parameters.

[0069] Federal learning avoids uploading user data to the server, and only exchanges parameter update amounts, which meets the privacy security requirements of civil devices (such as family gardening data not being leaked externally). Aggregating the maintenance experience of multiple users (such as fertilization habits of different regions and plant types) enables the model to adapt to more complex scenarios and improve the global deployment accuracy. Without centralized data training, the model is continuously optimized through real-time feedback of distributed devices, solving the problem of traditional intelligent devices relying on fixed algorithms and being difficult to update.

[0070] In some embodiments, when the server aggregates the local parameter update amounts of all clients using a federated averaging algorithm, it is further configured to perform the following steps: obtain the local parameter update amount uploaded by each client and its corresponding device runtime data; calculate the weight coefficient of each client based on the device runtime data, wherein the client with a longer runtime corresponds to a higher weight coefficient; perform a weighted average of the local parameter update amounts of each client according to the weight coefficient to generate the globally updated output layer global update parameters; before distributing the global update parameters to each client, encrypt the global update parameters, wherein the encryption process includes using a homomorphic encryption algorithm to perform layered encryption of the parameters, so that each client can only obtain the parameter portion that matches the local model structure when decrypting.

[0071] By incorporating device runtime as a weighting factor into the federated averaging algorithm and adding homomorphic encryption, the rationality of weight allocation and the security of parameter transmission are ensured. Weighting factor calculation allocates weights based on device runtime (reflecting the amount of data accumulated), with devices running longer contributing more to the global model. Weighted parameter aggregation uses the weighting factor to perform a weighted average of local update amounts, increasing the influence of data-rich devices. Layered encryption ensures that devices can only decrypt the parameter portions matching their local model by applying homomorphic encryption to global update parameters, preventing information leakage.

[0072] The weighting coefficient is calculated by defining the runtime of device i as Ti, and the total runtime as Ti. The weighting coefficients are: ωi = Ti / Ttotal. The weighted average aggregation is performed by globally updating the parameters to a weighted sum: ; Homomorphic encryption uses homomorphic encryption algorithms (such as BFV and CKKS) on the server to encrypt global parameters in layers. Each device only has the private key to decrypt the corresponding layer of the local model (such as the output layer) and cannot obtain the parameters of other layers or update information from other devices.

[0073] After receiving the encrypted global parameters through the device, the parameter distribution and decryption process only decrypts the weights and biases of the output layer, ensuring the privacy of the model structure and parameters.

[0074] Weighting devices based on runtime allows those accumulating more valuable data (such as yards with consistently applied fertilizer) to contribute more significantly to model optimization, improving the overall model's reliability and robustness. Homomorphic encryption ensures that model details and update information from user devices are not leaked during parameter transmission, meeting the high data security requirements of civilian scenarios (e.g., preventing malicious tampering with model parameters). This incentive mechanism encourages long-term device use, creating a positive cycle of "data contribution - model optimization - experience improvement," thus enhancing user engagement.

[0075] In some embodiments, by modeling the liquid preparation process as a Markov Decision Process (MDP), a reinforcement learning agent (RL Agent) is designed to optimize concentration accuracy, fertilization speed, and stock solution consumption efficiency simultaneously. Through Deep Q Network (DQN) or policy gradient algorithm, the device autonomously selects the optimal preparation strategy under different plant types (such as succulents requiring low concentration and slow application, shrubs requiring high concentration and fast application).

[0076] The state space definition includes: state S={Cset, Creal-time, Qfeedback, Vstock solution remaining, Tambient temperature}, corresponding to set concentration, real-time concentration, flow rate, stock solution remaining, and ambient temperature, respectively.

[0077] The action space design includes: action A={Δvalve opening, Δatomization frequency}, with discrete steps (such as ±5% opening, ±10% atomization frequency) as selectable actions.

[0078] The reward function construction includes: multi-objective reward: R=α( |Cerror|) + β(fertilization speed) + γ(1 / stock solution consumption); where α / β / γ are weight coefficients (users can customize through interactive modules, such as α=0.6, β=0.2, γ=0.2 for succulent mode), where Cerror is the value of the set concentration and the real-time detected concentration, and the absolute value is taken negatively to ensure that the smaller the error, the higher the reward, the fertilization speed is the effective fertilization amount completed per unit time, and the reciprocal of the stock solution consumption reflects the effective utilization rate of unit stock solution (the less consumed, the higher the reward), and R is the value of the reward function.

[0079] Training and execution include: offline stage: pre-train DQN network in simulation environment using historical fertilization data; online stage: balance exploration and utilization through Epsilon-Greedy strategy, and update Q value table in real time.

[0080] For different plant fertilization sensitivity characteristics (such as orchids that dislike concentrated fertilizer and require low rate and long time application), the optimal preparation strategy is automatically generated to avoid plant damage caused by manual trial and error. Under the premise of ensuring concentration accuracy, by reducing stock solution waste (such as using up the current batch when there are 50ml of stock solution remaining) and accelerating fertilization speed (non-sensitive plant scenarios), the consumable utilization rate is improved. Adjust the strategy in combination with the ambient temperature (plant absorption efficiency changes at high temperatures) to address the shortcomings of traditional algorithms that ignore environmental variables, making it particularly suitable for seasonal maintenance needs in outdoor gardens.

[0081] In some embodiments, a model initialization method based on transfer learning is designed to solve the problem of data scarcity in new user scenarios. A pre-trained "general plant fertilization model" is used as the basis, and the output layer parameters are fine-tuned using a small amount of target user data (such as 3-5 fertilization records) to quickly adapt to different garden plant species and soil conditions, solving the problem of poor precision in traditional model "cold start".

[0082] The pre-trained model is constructed by training a general model on the cloud using millions of plant maintenance data (covering multiple common varieties), including the input layer (concentration / flow / stock solution parameters), the intermediate layer (deep neural network to extract general fertilization rules), and the output layer (atomization amount prediction).

[0083] The transfer learning process includes: feature transfer: preserve the input layer and intermediate layer parameters (frozen weights) of the pre-trained model, and replace the output layer with a randomly initialized fully connected layer; fine-tuning: use a small amount of data from the new user (such as the plant type and successful concentration case entered by the user at the first fertilization), fine-tune the output layer using the mean square error loss function, the pre-trained model is compressed to ≤10MB, suitable for low-power MCUs (such as STM32) of civilian devices, and the fine-tuning process is completed locally without the need for continuous networking.

[0084] When a new user uses it for the first time, the device can achieve an initial concentration accuracy of more than 80% without long training, especially suitable for quick maintenance of temporary purchased potted plants. Only 3-5 successful fertilization records are needed to complete the adaptation, solving the problem of sparse data in civilian scenarios (families usually maintain 5-10 plants), and reducing the user's participation cost. The pre-trained model covers the common fertilization rules of common plants (such as nitrogen-loving foliage plants and phosphorus-loving flowering plants), and the transferred model can quickly adapt to unrecorded varieties (such as inferring concentration requirements through similar genus characteristics).

[0085] In some embodiments, a fuzzy logic controller is designed to address the problem of ambiguous concentration description by non-professional users (such as "a little lighter" and "a little thinner than last time"). The user's natural language input is converted into fuzzy concentration instructions, which are mapped to precise atomization amount and valve opening control parameters through a fuzzy rule base, reducing the complexity of operation.

[0086] The fuzzy set definition includes: the language variable "concentration preference" is defined as {extremely light, relatively light, moderate, relatively thick, extremely thick}, corresponding to the domain [0, 100%], and the membership function adopts triangular or trapezoidal distribution (such as "relatively light" corresponding to 15%-30% concentration interval, and the membership degree decreases linearly with the increase of concentration).

[0087] The fuzzy rule base is constructed by establishing "user instruction-concentration range-control parameter" rules, such as: IF concentration preference = lighter AND plant type = succulent THEN atomization amount = basic value x 0.8, valve opening = 80%; IF concentration preference = thicker AND plant type = shrub THEN atomization amount = basic value x 1.2, valve opening = 90%.

[0088] The fuzzy inference and defuzzification are calculated by the Mamdani inference method to obtain the fuzzy set of the control amount, and then the centroid method is used to defuzzify to obtain the accurate value, the formula is: u = ∑μ(ui)*ui / ∑μ(ui); wherein μ(ui) is the membership degree of the control amount, and ui is the control amount. The user does not need to understand the concentration percentage, and directly issues an instruction through natural language (such as "give the Chinese rose light fertilizer"), and the device automatically matches the plant database (with the default concentration range of 50+ common plants), to solve the barrier of professional terms. The user is allowed to input inaccurate instructions (such as "reduce the concentration by one point from the last time"), and the historical data memory and fuzzy logic are automatically adjusted to avoid concentration deviation caused by operation errors, which is especially suitable for middle-aged and elderly users or gardening enthusiasts. Combined with the user's historical instruction habits (such as the user-defined "light fertilizer" actually corresponding to 20% concentration), the fuzzy rule base is dynamically corrected to form a special operation habit, and the interactive friendliness is improved.

[0089] It should be understood that the terms used herein in this application are merely for the purpose of describing particular embodiments and are not intended to limit the present application. It should be understood that when an element or layer is referred to as "on", "adjacent to", "connected to", or "coupled to" another element or layer, it can be directly on, adjacent to, connected or coupled to the other element or layer, or there can be an intervening element or layer. In contrast, when an element is referred to as "directly on", "directly adjacent to", "directly connected to" or "directly coupled to" another element or layer, there is no intervening element or layer. It should be understood that although the terms first, second, third, etc. are used to describe various elements, components, regions, layers and / or parts, these elements, components, regions, layers and / or parts should not be limited by these terms. These terms are only used to distinguish one element, component, region, layer or part from another element, component, region, layer or part. Therefore, the first element, component, region, layer or part discussed below can be represented as the second element, component, region, layer or part without departing from the teachings of the present application.

[0090] Spatially relative terms, such as "beneath", "below", "lower", "under", "above", "upper" and the like, can be used herein for ease of description to describe one element or feature's relationship to another element(s) or feature(s) as illustrated in the figures. It will be understood that the spatially relative terms are intended to encompass different orientations of the device in use and / or operation in addition to the orientations depicted in the figures. For example, if a device in the figures is inverted, then a dependent- element described as "below" or "beneath" another element or feature would then be oriented "above" and "over" the other element or feature. Thus, the exemplary term "below" can encompass both an orientation of above and below. The device can be otherwise oriented (rotated 90 degrees or at other orientations) and the spatially relative descriptors used herein interpreted accordingly.

[0091] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising", when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0092] It will be further understood that the terms "comprises" and / or "comprising", when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0093] The above description is only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An intelligent liquid preparation device, characterized in that, include: The mixing chamber is used to collect the original solution and water to prepare the mixture; The raw liquid chamber is used to store the raw liquid and is connected to the mixing chamber through a liquid dispenser, the liquid dispenser having a piezoelectric atomization module; A water inlet pipe is used to supply water to the mixing chamber, and the water inlet pipe is equipped with a water inlet valve; A discharge pipe is used to discharge the liquid material in the mixing chamber, and the discharge pipe is equipped with a flow meter; The interactive module is used to obtain the user's input of the set concentration and set flow rate; The control module is used to control the opening degree of the inlet valve according to the set flow rate and to collect the feedback flow rate of the flow meter. The set concentration, feedback flow rate, and original solution concentration are input into the atomization volume calculation model. After local fine-tuning and global adjustment of the atomization volume calculation model, the target set atomization volume is calculated. Based on the target, the atomization amount is set to generate corresponding piezoelectric control parameters to instruct the piezoelectric atomization module to perform atomization actions; by adding a conductivity sensor, a pH sensor, and a camera, a multimodal perception system is constructed, and a convolutional neural network is used to identify plant varieties and growth status, and the current optimal nutrient solution concentration is inferred in real time by combining the sensor data corresponding to the conductivity sensor and pH sensor; the growth status includes leaf color and lesions.

2. The intelligent liquid preparation device as described in claim 1, characterized in that, The set concentration, feedback flow rate, and stock solution concentration are input into the atomization volume calculation model. After local fine-tuning and global adjustment of the atomization volume calculation model, the target set atomization volume is calculated, including: The set concentration, feedback flow rate and stock solution concentration are normalized according to the input layer of the preset atomization amount calculation model to obtain the numerical variables corresponding to the normalized set concentration, feedback flow rate and stock solution concentration. Based on the intermediate layer of the preset atomization amount calculation model, the normalized numerical variables are processed to obtain feature vectors; Based on the output layer of the preset atomization amount calculation model, the feature vector is mapped to the initially set atomization amount; Based on the pre-constructed reward function, the output layer of the preset atomization amount calculation model is locally fine-tuned based on the initial set atomization amount, and the feature vector is remapped based on the locally fine-tuned output layer to obtain the candidate set atomization amount; Based on the pre-built federated learning algorithm, the output layer of the preset atomization amount calculation model is globally adjusted based on the local fine-tuning value. The feature vector is then re-mapped based on the globally fine-tuned output layer to obtain the target atomization amount.

3. The intelligent liquid preparation device as described in claim 2, characterized in that, Based on the pre-constructed reward function, the output layer of the preset atomization amount calculation model is locally fine-tuned based on the initial set atomization amount. The feature vector is then remapped based on the locally fine-tuned output layer to obtain candidate set atomization amounts, including: The effect of the initial atomization amount is evaluated by a pre-built reward function, and the gradient feedback of the reward function with respect to the initial atomization amount is calculated to obtain the direction for improving the current output quality. Based on gradient feedback of the reward function, the weights and bias parameters of the output layer are locally fine-tuned using a local learning rate. The purpose of this adjustment is to quickly improve the direction of the current output without affecting the global parameters of the entire model. The feature vector is input into the locally fine-tuned output layer, and forward propagation is performed again to obtain the updated candidate atomization amount.

4. The intelligent liquid preparation device as described in claim 2, characterized in that, Based on a pre-built federated learning algorithm, the output layer of a preset atomization calculation model is globally adjusted based on local fine-tuning values. The locally fine-tuned output layer then remaps the feature vector to obtain the target atomization amount, including: Each participating device locally performs a local fine-tuning operation of the output layer based on the reward function and calculates its local parameter update amount relative to the original output layer parameters; All clients upload their local parameter updates to the server. The server aggregates the local parameter updates of all clients using a federated averaging algorithm to generate a set of globally updated output layer global update parameters, enabling the model to integrate optimization experience from multiple devices while maintaining privacy. The updated global output layer parameters from the server are used to recalculate the feature vectors, and the output target is set with atomization amount.

5. The intelligent liquid preparation device as described in claim 4, characterized in that, When the server aggregates the local parameter updates of all clients using a federated averaging algorithm, it is also configured to perform the following steps: Obtain the local parameter update amount uploaded by each client and its corresponding device runtime data; The weight coefficient of each client is calculated based on the device runtime data, where the client with a longer runtime corresponds to a higher weight coefficient. The local parameter update amounts of each client are weighted and averaged according to the weighting coefficients to generate the globally updated output layer global update parameters. Before the global update parameters are sent to each client, the global update parameters are encrypted. The encryption process includes using a homomorphic encryption algorithm to perform layered encryption on the parameters, so that each client can only obtain the parameter part that matches the local model structure when decrypting.

6. The intelligent liquid preparation device as described in claim 1, characterized in that, Based on the target, the atomization amount is set to generate corresponding piezoelectric control parameters, including: Obtain a set of parameters reflecting the atomization characteristics of the original solution at the operating temperature; The parameter set is fitted with a function between the piezoelectric control parameters and the atomization amount to establish an atomization parameter control model. The piezoelectric control parameters corresponding to the target atomization amount are calculated based on the atomization parameter control model. The piezoelectric control parameters include the driving voltage, and the parameter set includes the viscosity and surface tension of the original liquid.

7. The intelligent liquid preparation device according to any one of claims 1-5, characterized in that, Also includes: The control module also collects the feedback concentration from the concentration detection sensor configured in the discharge pipeline, calculates the deviation between the feedback concentration and the set concentration, and compares it with the set range. When the deviation exceeds the set range, the control module issues an abnormality prompt message for the raw liquid through the interactive module. The concentration detection sensor includes a conductivity sensor.

8. The intelligent liquid preparation device according to any one of claims 1-5, characterized in that: The discharge pipe has a discharge direction, and the mixing chamber has an upper end face located upstream of the discharge direction, a lower end face located downstream of the discharge direction, and a side wall connecting the upper end face and the lower end face. The projected area of ​​the upper end face along the discharge direction is larger than the projected area of ​​the lower end face along the discharge direction. The water inlet pipe is connected to the mixing chamber at the upper end face, the liquid dispenser is connected to the mixing chamber at the side wall, and the discharge pipe is connected to the mixing chamber at the lower end face.

9. The intelligent liquid preparation device according to any one of claims 1-5, characterized in that: The flow meter includes an electromagnetically coupled rotating body and a detection module. The rotating body is placed inside the discharge pipe and can rotate along a plane perpendicular to the discharge direction. The detection module includes a Hall sensor.

10. The intelligent liquid preparation device as described in claim 9, characterized in that: The rotating body includes a driving member and a driven member equipped with a magnet. The driving member includes a rotating shaft portion parallel to the discharge direction and a plurality of fan blade portions surrounding the outside of the rotating shaft portion. The fan blade portions have concave arc surfaces intersecting the discharge direction. The driven member is connected to the discharge direction end of the rotating shaft portion. The flow meter also includes a guide plate located on the side of the driving member opposite to the discharge direction. The guide plate is provided with a plurality of guide holes for guiding the liquid material to contact the concave arc surface.

Images