A smart coffee brewing machine

By using a multi-channel sensor array and a multi-dimensional spatial boundary convergence adaptive extraction control module, the problem of uneven extraction caused by the internal channel effect of the coffee grounds was solved, and real-time dynamic adjustment and uniformity of the coffee extraction process were achieved.

CN123100975APending Publication Date: 2026-11-17SPICE & BEVERAGE RES INST CHINESE ACAD OF TROPICAL AGRI SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611201625.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing smart coffee brewing equipment suffers from uneven extraction because a single sensor cannot acquire information about the internal spatial distribution of the coffee grounds. Furthermore, data from multiple sensors cannot be fused in real time, making it impossible to dynamically adjust the pump drive parameters and creating a channel effect.

Method used

The system employs a multi-channel fluid pressure sensor array, an ultrasonic concentration detection unit, a water pump drive module, a heating water tank, an outlet flow rate sensor, a spectral data acquisition unit, and a control chip. Combined with a multi-dimensional spatial boundary convergence adaptive extraction control module, it uses a multi-dimensional spatial boundary convergence mechanism network to invert the internal fluid dynamics of the coffee powder bed in real time, dynamically adjusts the water pump drive parameters, suppresses channel effects, and achieves uniform extraction.

Benefits of technology

It achieves uniformity and real-time control of the coffee extraction process, eliminates the long-tail delay of multi-channel sensor data, and ensures the consistency and efficiency of extraction quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123100975A_ABST
    Figure CN123100975A_ABST
Patent Text Reader

Abstract

This invention provides an intelligent coffee brewing machine, belonging to the technical field of coffee brewing machines. It includes: a multi-channel fluid pressure sensor array, an ultrasonic concentration detection unit, a water pump drive module, a heating water tank, a coffee grounds bed support cavity, a water flow rate sensor, a spectral data acquisition unit, an edge microcontroller unit, and a control chip. The control chip is equipped with a multi-dimensional spatial boundary convergence adaptive extraction control module, used to fuse the multi-source heterogeneous data streams from the multi-channel fluid pressure sensor array, the ultrasonic concentration detection unit, the water flow rate sensor, and the spectral data acquisition unit. Through a multi-dimensional spatial boundary convergence mechanism network, it performs real-time inversion of the internal fluid dynamics state of the coffee grounds bed, dynamically adjusts the instantaneous pulse frequency of the water pump drive module, suppresses channel effects, and achieves uniform extraction. This solves the technical problems of uneven extraction caused by channel effects within the coffee grounds bed and the inability to fuse multi-source sensor data in real time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of coffee brewing machine technology, and more specifically, relates to an intelligent coffee brewing machine. Background Technology

[0002] In the field of automated coffee brewing, existing intelligent coffee brewing equipment typically relies on a single pressure sensor and a fixed water pump frequency to control the extraction process. Some solutions introduce temperature sensors or simple flow meters as auxiliary parameters, driving the water pump to output a constant pressure through preset open-loop control logic. In actual brewing conditions, when the coffee grounds are unevenly packed, have varying particle sizes, or experience localized compaction and migration after being impacted by water, significant differences in flow resistance occur in different spatial regions within the coffee grounds. Water preferentially penetrates the coffee grounds along low-resistance paths, creating a channel effect. In existing technologies, because a single sensor cannot acquire information about the spatial distribution within the coffee grounds, and because there are temporal inconsistencies between data streams when multiple sensors sample concurrently, fixed control strategies cannot detect the channel effect or dynamically adjust the water pump drive parameters based on the real-time state of the coffee grounds' fluid dynamics. This results in a significant deviation of the soluble component distribution in the extract from the standard state. In other words, existing technologies suffer from technical problems such as uneven extraction due to the channel effect within the coffee grounds and the inability to fuse data from multiple sensors in real time. Summary of the Invention

[0003] In view of this, the present invention provides an intelligent coffee brewing machine that can solve the technical problems in the prior art, such as uneven extraction caused by the internal channel effect of the coffee powder bed and the inability to fuse data from multiple sources in real time.

[0004] The present invention is implemented as follows: The present invention provides an intelligent coffee brewing machine, comprising: a multi-channel fluid pressure sensor array, an ultrasonic concentration detection unit, a water pump drive module, a heating water tank, a coffee grounds bed support cavity, a water flow rate sensor, a spectral data acquisition unit, an edge microcontroller unit, and a control chip. The heating water tank and the coffee grounds bed support cavity are connected in series via an inlet pipe. The multi-channel fluid pressure sensor array is embedded in the sidewall and bottom of the coffee grounds bed support cavity in an array configuration. The ultrasonic concentration detection unit is fixedly installed on the coffee grounds bed support cavity. Below the outlet of the coffee grounds bed carrier chamber, the water flow rate sensor is installed on the pipeline between the outlet of the coffee grounds bed carrier chamber and the ultrasonic concentration detection unit. The spectral data acquisition unit is fixed on both sides of the outlet pipeline in a through-beam manner. The edge microcontroller unit is fixedly installed on the inner wall of the housing. The control chip is embedded on the main board of the edge microcontroller unit. The control chip is electrically connected to and interacts with the multi-channel fluid pressure sensor array, the ultrasonic concentration detection unit, the water pump drive module, the heating water tank, the water flow rate sensor, and the spectral data acquisition unit.

[0005] The multi-channel fluid pressure sensor array is used to collect fluid pressure distribution data at various spatial locations inside the coffee powder bed support cavity in real time; the ultrasonic concentration detection unit is used to continuously measure the concentration of coffee solids in the extract by means of ultrasonic attenuation coefficient; the water pump drive module is used to receive the pulse frequency modulation signal output by the control chip and output the corresponding instantaneous water flow pressure; the heating water tank is used to heat the injected water to the set temperature and supply water stably; the coffee powder bed support cavity is a cylindrical cavity structure with a powder bed support filter inside, used to support coffee powder and make the water flow uniformly through the powder bed; the outlet flow rate sensor is used to collect the liquid volume flow rate through the outlet in real time; the spectral data acquisition unit is used to collect the absorbance data of the extract in the visible and near-infrared bands to characterize the characteristic component distribution of soluble substances in the extract; the edge microcontroller unit is used to carry the control chip and complete the concurrent convergence and buffering processing of multi-channel sensor data.

[0006] The control chip is equipped with a multi-dimensional spatial boundary convergence adaptive extraction control module, which is used to fuse the multi-source heterogeneous data streams from the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, water flow rate sensor, and spectral data acquisition unit. Through the multi-dimensional spatial boundary convergence mechanism network, the internal fluid dynamics state of the coffee grounds is inverted in real time, and the instantaneous pulse frequency of the water pump drive module is dynamically adjusted to suppress channel effects and achieve uniform extraction. At the same time, a dual-buffered queue mechanism is used to eliminate the long-tail delay of multi-channel concurrent data, achieving zero-latency multi-channel data fusion.

[0007] The multidimensional space boundary convergence adaptive extraction control module includes the following steps:

[0008] The data streams from the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, effluent flow rate sensor and spectral data acquisition unit are independently buffered and received using a dual-buffered queue. The thread suspension waiting time of each data stream is dynamically adjusted, and the fused data frame is output after timestamp alignment is completed.

[0009] Orthogonal projection processing is performed on the multi-channel pressure distribution data in the fused data frame to eliminate noise coupling interference between channels, construct the pressure distribution tensor and extract the pressure gradient feature vector of each spatial node.

[0010] The pressure distribution tensor and ultrasonic concentration data are input into the multidimensional spatial boundary convergence mechanism network. The conservation of mass and momentum is transformed into multiplier constraints through the conservation law constraint layer. The flow resistance spatial distribution inside the powder bed is inverted to obtain the flow resistance spatial distribution matrix.

[0011] Calculate the deviation coefficient between the flow resistance spatial distribution matrix and the standard uniform flow resistance vector, and trigger three strategies respectively: maintain the current frequency, non-steady-state compensation fine-tuning, or forced frequency reduction according to the interval to which the deviation coefficient belongs.

[0012] Based on the pulse frequency adjustment amount output by the deviation coefficient, the updated instantaneous pulse frequency control signal is output to the water pump drive module through the weight update mechanism of the multidimensional spatial boundary convergence mechanism network.

[0013] Collect the values ​​from the outflow velocity sensor and the absorbance feature vector from the spectral data acquisition unit, calculate the similarity value between the absorbance feature vector and the pre-stored standard extraction feature vector, and perform the following actions based on the similarity value range: extend the extraction time, maintain the current parameters, or reduce the subsequent water supply.

[0014] The flow resistance spatial distribution matrix sequence and ultrasonic concentration data sequence during the extraction process are written into the storage area of ​​the edge microcontroller unit. The jump matrix parameters of the multidimensional spatial boundary convergence mechanism network are updated to complete the adaptive learning closed loop and output the extraction completion signal.

[0015] Specifically, the dual-buffered queue refers to maintaining two independent first-in-first-out buffers for each sensor data stream within the edge microcontroller unit. One buffer is in a write state, and the other is in a read state. The two buffers automatically switch roles when they are full or empty.

[0016] Specifically, the dynamic adjustment of the thread suspension waiting time for each data stream is determined dynamically based on the characteristic variance of temperature, flow rate, and spectral data.

[0017] Specifically, the orthogonal projection process refers to mapping the original numerical vectors of the multi-channel pressure sensor onto a space composed of a set of mutually orthogonal basis vectors, so that the common-mode error components caused by the spatial coupling of different sensor channels cancel each other out in the coordinates of each dimension after projection.

[0018] The multidimensional spatial boundary convergence mechanism network consists of a multi-source data heterogeneous input layer, a graph theory topology mapping layer, a mechanism conservation law constraint layer, a self-circulating feature extraction layer, an unsteady compensation branch, and a jump decision output layer cascaded together. The graph theory topology mapping layer internally nests the Kruskal minimum spanning tree algorithm, which abstracts the physical nodes of the powder bed space into a weighted undirected graph and uses the flow resistance difference between nodes as the edge weight to generate the topological connectivity matrix in real time.

[0019] Specifically, the conservation law constraint layer refers to the constraint condition that transforms the continuity equation and momentum conservation equation of incompressible fluid into the multiplication operation of the activation values ​​of each node in the network, so that the internal numerical calculation process of the network always satisfies the conservation relationship specified by the fluid physics mechanism when inverting the flow resistance of the powder bed.

[0020] Specifically, the flow resistance spatial distribution matrix is ​​a two-dimensional numerical matrix composed of the spatial coordinate positions of each sensor in the multi-channel fluid pressure sensor array as row and column indices, and the local flow resistance values ​​of each spatial node obtained by inversion of the multi-dimensional spatial boundary convergence mechanism network as matrix elements.

[0021] Specifically, the deviation coefficient is obtained by dividing the norm of the difference between each element of the flow resistance spatial distribution matrix and the corresponding element of the standard uniform flow resistance vector by the norm of the standard uniform flow resistance vector. When the deviation coefficient is lower than the first deviation threshold, the current pump pulse frequency is maintained. When the deviation coefficient is between the first and second deviation thresholds, the unsteady compensation module is triggered to output a fine-tuning amount. When the deviation coefficient is higher than the second deviation threshold, the pulse frequency of the pump drive module is reduced to the frequency reduction ratio and the flow resistance inversion is re-executed.

[0022] Specifically, the unsteady-state compensation branch is a set of independent hidden layer branches in the multidimensional spatial boundary convergence mechanism network. When the network's error trigger detects a sudden change in the outflow velocity, the main propagation path jumps to the unsteady-state compensation branch through the conditional branch to perform multiple rounds of iterative correction on the current hidden layer state. After the correction is completed, the result is re-integrated into the deep decision output of the main network.

[0023] Specifically, the jump matrix refers to the additional linear mapping matrix established between adjacent and cross-layer hidden layers in the multidimensional spatial boundary convergence mechanism network. This matrix enables the original local features of the graph theory topology mapping layer to be directly weighted and superimposed onto the input of the deep nodes, ensuring that the local pressure detail features are not lost due to multi-layer nonlinear transformations when they are transmitted to the final decision output.

[0024] Specifically, the similarity value is calculated by normalizing the absorbance feature vector output by the spectral data acquisition unit in the current extraction stage and the pre-stored standard extraction feature vector, and then calculating the cosine similarity value obtained by the dot product of the two. When the similarity value is lower than the first similarity threshold, the duration of the current extraction stage is increased; when the similarity value is between the first and second similarity thresholds, the current extraction parameters are maintained; when the similarity value is higher than the second similarity threshold, the water supply in subsequent extraction stages is reduced.

[0025] The self-looping feature extraction layer integrates a resource allocation mechanism, which is directly linked to the deviation coefficient. This mechanism dynamically controls the proportion of activated neurons in each hidden layer, the number of computation steps in the self-looping iteration, and the size of the dynamic cache block in the video memory.

[0026] Among them, the resource allocation mechanism adjusts the exponential function through resource allocation. accomplish, Based on deviation coefficient With the characteristic variance of spectral data Calculate, when When in a steady-state low-load range, a low-energy computing power maintenance strategy is adopted; when When the disturbance range is slightly in the range, a dynamic balance adaptive strategy is adopted; when When the high load range is reached, a full computing power boundary limit convergence strategy is adopted.

[0027] The training dataset for the multidimensional spatial boundary convergence mechanism network was established in the following way: virtual samples were generated by simulating the unsteady fluid dynamics behavior of the coffee bed under different compaction, particle size distribution and channel effect severity using computational fluid dynamics software; real heterogeneous physical data streams were recorded by conducting multiple rounds of standard brewing tests using a smart coffee machine prototype; and paired training datasets were constructed by normalizing the spatiotemporal dimensions of the simulation data and the experimental data and using the spatial distribution of flow resistance of the coffee bed and brewing quality indicators as supervision labels.

[0028] Among them, the first deviation threshold is 0.35, the second deviation threshold is 0.70, and the frequency reduction ratio is 60%; the first similarity threshold is 0.80, and the second similarity threshold is 0.95; resource allocation adjustment index function. ,when Time-activated neuron ratio parameter 40%, number of self-looping iterations For 2 times; when hour 75% For 5 times; when hour 100% It is 12 times.

[0029] This invention combines a multi-channel fluid pressure sensor array with a multi-dimensional spatial boundary convergence mechanism network. The pressure distribution tensor of each spatial node within the coffee bed is used as the network input. A conservation law constraint layer transforms fluid mass and momentum conservation into hard boundary constraints for the network weights, ensuring that the network's inversion results always conform to fluid physics mechanisms. This accurately identifies the spatially non-uniform distribution of flow resistance within the coffee bed. Furthermore, a deviation coefficient is used to trigger a strategy that adjusts the frequency of water pump pulses of varying intensities, dynamically converging the flow resistance distribution towards a uniform state and fundamentally suppressing the development of channel effects. Simultaneously, a dual-buffered queue mechanism eliminates the long-tail delay of concurrent sampling from multiple sensors, ensuring the temporal consistency of data fusion frames and ensuring that control decisions are always based on the actual coffee bed state at the current moment. In summary, this invention solves the technical problems mentioned in the background art, such as uneven extraction caused by channel effects within the coffee bed and the inability to fuse multi-source sensor data in real time. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the components of a smart coffee maker.

[0031] Figure 2 A flowchart illustrating the execution steps of the adaptive extraction control module for multidimensional space boundary convergence.

[0032] Figure 3 This is a schematic diagram of the time-series variation curve of the deviation coefficient during the extraction process of an intelligent coffee brewing machine.

[0033] The reference numerals in the attached drawings are explained as follows: 01, water inlet pipe; 02, water pump drive module; 03, coffee powder bed support cavity; 04, powder bed support filter; 05, multi-channel fluid pressure sensor array; 06, liquid outlet; 07, liquid outlet pipe; 08, water flow rate sensor; 09, spectral data acquisition unit; 10, ultrasonic concentration detection unit; 11, heating water tank; 12, edge microcontroller unit; 13, control chip. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0035] like Figure 1The diagram shows the composition of an intelligent coffee brewing machine provided by the present invention, including: a multi-channel fluid pressure sensor array 05, an ultrasonic concentration detection unit 10, a water pump drive module 02, a heating water tank 11, a coffee grounds bed support cavity 03, a water flow rate sensor 08, a spectral data acquisition unit 09, an edge microcontroller unit 12, and a control chip 13. The heating water tank and the coffee grounds bed support cavity 03 are connected in series via a water inlet pipe 01. The multi-channel fluid pressure sensor array is embedded in the sidewalls and bottom of the coffee grounds bed support cavity in an array configuration. The ultrasonic concentration detection unit 10... The measuring unit is fixedly installed below the outlet 06 of the coffee powder bed bearing cavity. The water flow rate sensor is installed on the pipeline between the outlet of the coffee powder bed bearing cavity and the ultrasonic concentration detection unit. The spectral data acquisition unit is fixed on both sides of the outlet pipeline 07 in a through-beam manner. The edge microcontroller unit is fixedly installed on the inner wall of the housing. The control chip is embedded on the main board of the edge microcontroller unit. The control chip is electrically connected to and interacts with the multi-channel fluid pressure sensor array, the ultrasonic concentration detection unit, the water pump drive module, the heating water tank, the water flow rate sensor, and the spectral data acquisition unit.

[0036] The multi-channel fluid pressure sensor array is used to collect fluid pressure distribution data at various spatial locations inside the coffee powder bed support cavity in real time; the ultrasonic concentration detection unit is used to continuously measure the concentration of coffee solids in the extract by means of ultrasonic attenuation coefficient; the water pump drive module is used to receive the pulse frequency modulation signal output by the control chip and output the corresponding instantaneous water flow pressure; the heating water tank is used to heat the injected water to the set temperature and supply water stably; the coffee powder bed support cavity is a cylindrical cavity structure with a powder bed support filter 04 inside, which is used to support coffee powder and make the water flow uniformly through the powder bed; the outlet flow rate sensor is used to collect the liquid volume flow rate through the outlet in real time; the spectral data acquisition unit is used to collect the absorbance data of the extract in the visible and near-infrared bands to characterize the characteristic component distribution of soluble substances in the extract; the edge microcontroller unit is used to carry the control chip and complete the concurrent convergence and buffering processing of multi-channel sensor data.

[0037] The control chip is equipped with a multi-dimensional spatial boundary convergence adaptive extraction control module, which is used to fuse the multi-source heterogeneous data streams from the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, water flow rate sensor, and spectral data acquisition unit. Through the multi-dimensional spatial boundary convergence mechanism network, the internal fluid dynamics state of the coffee grounds is inverted in real time, and the instantaneous pulse frequency of the water pump drive module is dynamically adjusted to suppress channel effects and achieve uniform extraction. At the same time, a dual-buffered queue mechanism is used to eliminate the long-tail delay of multi-channel concurrent data, achieving zero-latency multi-channel data fusion.

[0038] like Figure 2 The diagram shows the execution steps of the multidimensional space boundary convergence adaptive extraction control module, including the following steps:

[0039] S01. The dual-buffered queue of the edge microcontroller unit independently buffers and receives the data streams from the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, effluent flow rate sensor and spectral data acquisition unit, respectively. It dynamically adjusts the thread suspension waiting time of each data stream according to the characteristic variance of temperature, flow rate and spectral data, and outputs the fused data frame after completing the timestamp alignment of each data stream.

[0040] S02. Perform orthogonal projection processing on the multi-channel pressure distribution data in the fused data frame to eliminate noise coupling interference between sensor channels, construct a pressure distribution tensor with spatial location as the dimension coordinate, and extract the pressure gradient feature vector of each spatial node.

[0041] S03. Input the pressure distribution tensor and ultrasonic concentration data into the multidimensional spatial boundary convergence mechanism network. The conservation law constraint layer of the network transforms the mass conservation and momentum conservation of water flow into multiplier constraints in matrix multiplication. The spatial distribution of flow resistance inside the powder bed is inverted through the multi-level self-circulating feedback matrix to obtain the flow resistance spatial distribution matrix.

[0042] S04. Calculate the deviation coefficient between the flow resistance spatial distribution matrix and the standard uniform flow resistance vector. When the deviation coefficient is in the range of 0 to 0.35, the fluid dynamics of the powder bed is determined to be normal, and the current pump pulse frequency is maintained unchanged. When the deviation coefficient is in the range of 0.35 to 0.70, the powder bed is determined to have a slight channel effect, triggering the unsteady compensation module of the multidimensional spatial boundary convergence mechanism network, and the output pulse frequency is fine-tuned by iterative correction of the hidden layer state. When the deviation coefficient is greater than 0.70, the powder bed is determined to have a severe channel effect, and the pulse frequency of the pump drive module is directly reduced to 60% of the current value and S03 is re-executed.

[0043] S05. Based on the pulse frequency adjustment amount output by S04, the dynamic convergence term of the flow resistance boundary is incorporated into the loss function through the weight update mechanism of the multidimensional spatial boundary convergence mechanism network. The network weight update is directly linked to the pulse modulation coefficient of the water pump drive module, and the updated instantaneous pulse frequency control signal is output to the water pump drive module.

[0044] S06: Collect the effluent flow rate sensor value and the absorbance feature vector of the spectral data acquisition unit at the current extraction stage. Calculate the similarity value between the absorbance feature vector and the pre-stored standard extraction feature vector. If the similarity value is below 0.80, increase the duration of the current extraction stage and re-trigger S03. If the similarity value is between 0.80 and 0.95, maintain the current extraction parameters and continue execution. If the similarity value is above 0.95, determine that the extraction quality meets the standard, reduce the water supply for subsequent extraction stages, and proceed to S07.

[0045] S07. Write the flow resistance spatial distribution matrix sequence and ultrasonic concentration data sequence of the entire extraction process into the storage area of ​​the edge microcontroller unit, update the jump matrix parameters of the multidimensional spatial boundary convergence mechanism network, complete the adaptive learning closed loop of this extraction, and output the extraction completion signal to the control chip.

[0046] Specifically, the dual-buffered queue refers to maintaining two independent first-in-first-out buffers for each sensor data stream within the edge microcontroller unit. One buffer is in a write state to receive newly arrived sensor sampling data, while the other buffer is in a read state to output the received data to subsequent processing threads. The two buffers automatically switch roles when they are full or empty, thereby avoiding long-tail latency caused by mutual blocking of read and write operations.

[0047] Specifically, orthogonal projection processing refers to mapping the original numerical vectors of multi-channel pressure sensors onto a space composed of a set of mutually orthogonal basis vectors, so that the common-mode error components caused by spatial coupling of different sensor channels cancel each other out in the coordinates of each dimension after projection, thereby retaining only the pressure characteristic information of each spatial location.

[0048] Specifically, the conservation law constraint layer refers to a computational layer within the multidimensional spatial boundary convergence mechanism network. This layer transforms the mathematical relationships corresponding to the mass conservation law and momentum conservation law in fluid mechanics into constraint conditions for the multiplication operation of the activation values ​​of each node in the network. This ensures that when the network inverts the flow resistance of the powder bed, its internal numerical calculation process always satisfies the conservation relationship specified by the fluid physics mechanism, preventing the network from outputting inversion results that violate physical laws.

[0049] Specifically, the flow resistance spatial distribution matrix refers to a two-dimensional numerical matrix composed of the spatial coordinate positions of each sensor in the multi-channel fluid pressure sensor array as row and column indices, and the local flow resistance values ​​of each spatial node obtained by inversion of the multi-dimensional spatial boundary convergence mechanism network as matrix elements. This matrix intuitively reflects the non-uniform distribution of resistance to water flow in different spatial regions inside the coffee powder bed.

[0050] Specifically, the standard uniform flow resistance vector refers to a one-dimensional vector composed of the flow resistance reference values ​​of each spatial node collected and stored in the control chip under ideal conditions where the coffee grounds are uniformly filled and there is no channel effect. The deviation coefficient is obtained by dividing the norm of the difference between each element of the flow resistance spatial distribution matrix and the corresponding element of the standard uniform flow resistance vector by the norm of the standard uniform flow resistance vector. The larger the value, the more uneven the spatial distribution of flow resistance in the coffee grounds.

[0051] Specifically, the unsteady-state compensation module refers to a set of independent hidden layer branches in the multidimensional spatial boundary convergence mechanism network. When the network's error trigger detects a sudden change in the outflow velocity, the main propagation path jumps to this module through a conditional branch. This module performs multiple rounds of iterative correction on the current hidden layer state to compensate for the network output deviation caused by the sudden change in flow state. After the correction is completed, the result is re-integrated into the deep decision output of the main network.

[0052] Specifically, the jump matrix refers to the additional linear mapping matrix established between adjacent and cross-layer hidden layers in the multidimensional spatial boundary convergence mechanism network. This matrix enables the micro-flow features extracted by the shallow network to be directly weighted and superimposed on the input of the deep nodes without going through all the intermediate layers of transformation. This ensures that the local pressure details perceived by the shallow layer are not lost due to multi-layer nonlinear transformation when they are transmitted to the final decision output.

[0053] Specifically, the similarity value refers to the cosine similarity value obtained by normalizing the absorbance feature vector output by the spectral data acquisition unit of the current extraction stage and the pre-stored standard extraction feature vector, and then calculating the dot product of the two. The value ranges from 0 to 1. The closer the value is to 1, the closer the component distribution of the current extract is to the standard extraction state.

[0054] The specific structure of the multidimensional spatial boundary convergence mechanism network is as follows: the network consists of a cascaded multi-source data heterogeneous input layer, a graph theory topology mapping layer, a mechanism conservation law constraint layer, a self-looping feature extraction layer, an unsteady-state compensation branch, and a jump decision output layer. The input layer receives aligned pressure distribution tensor and concentration scalar data. The graph theory topology mapping layer internally incorporates the Kruskal minimum spanning tree algorithm, abstracting the spatial physical nodes of the coffee grounds bed as vertices of a weighted undirected graph. Using the flow resistance difference between nodes as edge weights, it generates a topological connectivity matrix representing the core conduction path of the fluid in real time, completing the physical spatial topology modeling of the non-neural network default graph algorithm. The mechanism conservation law constraint layer follows immediately, transforming the continuity equation and momentum conservation equation of the incompressible fluid into hard boundary constraints for the network weights through a multiplier constraint matrix. The self-recurrent feature extraction layer consists of three deep recurrent units with hidden state self-feedback, and integrates a resource allocation mechanism. This mechanism is directly linked to the deviation coefficient in step S04, used to dynamically control the proportion of activated neurons in each hidden layer, the number of computation steps in the self-recurrent iteration, and the size of the dynamic memory cache block, avoiding wasted computing power and storage caused by idle neurons. The jump matrix spans the intermediate layers, intuitively superimposing the original local features of the graph theory topology mapping layer onto the output. The non-steady-state compensation branch is controlled by a mutation trigger and contains two densely connected nonlinear correction networks. The final decision output layer outputs the pulse frequency control signal and the flow resistance spatial distribution matrix through a fully connected structure.

[0055] The specific steps for establishing the training dataset for the multidimensional spatial boundary convergence mechanism network include:

[0056] Physical simulation data generation: Using computational fluid dynamics (CFD) software, the unsteady fluid dynamics behavior of coffee powder bed under different compaction, particle size distribution and channel effect severity is simulated, and the pressure, flow rate and soluble solids concentration of each spatial coordinate are collected as virtual samples.

[0057] Real-world physical calibration test: Multiple rounds of standard brewing tests were conducted using a prototype smart coffee machine, and heterogeneous physical data streams under real-world conditions were recorded in real time using a multi-channel fluid pressure sensor array and an ultrasonic concentration detection unit.

[0058] Heterogeneous data alignment and labeling: Simulation data and experimental data are normalized in terms of spatiotemporal dimensions. The measured spatial distribution of powder bed flow resistance and the final brewing quality index are used as supervision labels to construct a paired training dataset containing input tensors and target matrices.

[0059] The specific steps for training the multidimensional space boundary convergence mechanism network include:

[0060] Graph topology and physical boundary initialization: Input the training samples into the graph theory topology mapping layer, initialize the spatial topology matrix using the Kruskal algorithm, and inject physical prior parameters into the conservation law constraint layer.

[0061] Forward propagation and mechanism loss calculation: Perform forward computation of the network, check the residuals of the output flow resistance matrix and fluid dynamics equations, and construct a composite loss function that integrates traditional cross-entropy and fluid conservation law penalty terms.

[0062] Error backpropagation and jump parameter update: The Adam optimizer is used to iteratively update the linear mapping parameters between the hidden layer weights and the jump matrix, so that the flow resistance inversion network converges to the global optimum while keeping the physical constraints intact.

[0063] This multidimensional spatial boundary convergence mechanism network, by introducing the Kruskal minimum spanning tree algorithm and a fluid dynamics conservation law constraint layer, ensures that the feature extraction process of the deep learning model is no longer limited to purely data-driven approaches, but is built upon solid fluid physics mechanisms. This structural design effectively avoids irrational predictions that violate physical laws when the network inverts complex flow resistance distributions, significantly improving the model's generalization ability and noise resistance. Simultaneously, through the synergy of the jump matrix and dynamic computational resource allocation mechanism, the model can seamlessly transmit subtle local pressure changes to the decision-making end and dynamically schedule microcontroller hardware resources according to the fluid state, eliminating long-tail latency while ensuring real-time control of channel effects and efficient uniform extraction.

[0064] To precisely control the allocation of computational resources within a neural network (including the proportion of activated neurons and the number of self-looping iterations), a resource allocation adjustment exponential function is designed. This function is based on the deviation coefficient in step S04. The characteristic variance of the spectral data in step S01 Perform real-time calculations:

[0065] ;

[0066] in, and This is a preset weight adjustment coefficient. It is obtained through calculation. The control range to which the value belongs is used to dynamically adjust the proportion of activated neurons in the neural network using different strategies. and the self-loop iteration step parameter :

[0067] when When the system is determined to be in a steady-state, low-load operation period, a low-energy computing power maintenance strategy is adopted: the proportion of activated neurons in the neural network is configured as follows: The number of iterations in the self-loop is limited to 1. This is done to minimize the power consumption of video memory and the processor.

[0068] when When a slight flow disturbance is detected in the powder bed, a dynamic equilibrium adaptive strategy is adopted: the proportion of activated neurons in the neural network is increased to [value missing]. The number of iteration steps in the self-loop is adjusted to Secondly, it balances computational accuracy with real-time response speed.

[0069] when When a severe channel effect or spectral abrupt change is detected in the system, a full-computational-power boundary-limit convergence strategy is adopted: the proportion of activated neurons in the neural network is forcibly locked to a certain value. The number of iteration steps in the self-loop has been increased to Next, the largest dynamic memory cache block is allocated simultaneously to suppress fluid unsteady distortion with the highest computing power precision.

[0070] The specific implementation of step S01 is as follows: The edge microcontroller unit allocates independent dual-buffer queues to the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, effluent flow rate sensor, and spectral data acquisition unit. Each dual-buffer queue consists of two independent first-in-first-out (FIFO) buffers. One buffer is always in a write state, receiving the currently arriving sensor sampling data, while the other buffer is in a read state, outputting the buffered data to subsequent processing threads. When the write buffer reaches its capacity limit or the read buffer is emptied, the two buffers automatically switch read / write roles, ensuring that read and write operations do not block each other and eliminating long-tail latency when data arrives concurrently. During the timestamp alignment stage, the system extracts the characteristic variance of each data stream. When the variance of temperature data is small, the thread suspension waiting time is short; when the variance of spectral data is large, the waiting window is extended accordingly, ensuring that the data streams are accurately aligned in the time dimension before being merged and output as a fused data frame.

[0071] The specific implementation of step S02 is as follows: Using the original multi-channel pressure numerical vectors from the fused data frame as input, orthogonal projection processing is performed on them. The core principle of orthogonal projection processing is to map the original numerical vectors onto a space spanned by a set of mutually orthogonal basis vectors constructed by the Gram-Schmidt orthogonalization process. Since the projections of common-mode error components introduced by the physical location coupling of different sensor channels onto the orthogonal basis directions cancel each other out, the coordinates of each dimension after projection only retain the independent pressure characteristics of the corresponding spatial location, thereby eliminating noise coupling interference between channels. Using the spatial coordinates of each sensor as the dimension index, the pressure feature values ​​after orthogonal projection are organized into a pressure distribution tensor, and the pressure difference between adjacent spatial nodes in the tensor is calculated to extract the pressure gradient feature vector of each node for use by subsequent network layers.

[0072] The specific implementation of step S03 is as follows: The pressure distribution tensor and the concentration scalar data output by the ultrasonic concentration detection unit are jointly input into the multidimensional spatial boundary convergence mechanism network. The network's graph theory topology mapping layer first uses each sensor spatial node as a vertex and the estimated flow resistance difference between nodes as edge weights. It then calls the Kruskal minimum spanning tree algorithm to add edges that do not form loops in ascending order of edge weight, constructing a topological connectivity matrix representing the core fluid conduction path. The Kruskal minimum spanning tree algorithm is a greedy algorithm that connects all nodes with the lowest total weight, thus ensuring that the topological connectivity matrix accurately reflects the fluid priority path with the least resistance. The mechanism conservation law constraint layer then applies the continuity equation of the incompressible fluid. The momentum conservation term in the Navier-Stokes equations is transformed into a multiplier constraint matrix for the multiplication of activation values ​​at each node. This forces the network to satisfy the fluid mass and momentum conservation relationships during forward propagation, preventing the output from violating physical laws. The self-looping feature extraction layer consists of three deep looping units with hidden state self-feedback. It performs multi-step iterative extraction of the temporal evolution features of flow resistance, and finally outputs the spatial distribution matrix of flow resistance through a jump decision output layer. The matrix elements are the local flow resistance inversion values ​​of each spatial node.

[0073] The specific implementation of step S04 is as follows: A difference vector is constructed by the differences between each element of the flow resistance spatial distribution matrix and the corresponding element of the standard uniform flow resistance vector. The L2 norm of this difference vector is calculated, and then divided by the L2 norm of the standard uniform flow resistance vector to obtain the deviation coefficient. The standard uniform flow resistance vector is acquired and stored within the control chip under ideal conditions of uniform powder bed filling and no channel effect, through pre-calibration experiments. When the value is below 0.35, the fluid dynamics of the powder bed are considered normal, and the current pump pulse frequency is maintained unchanged. When the value is between 0.35 and 0.70, a slight channeling effect is determined to exist in the powder bed, triggering an unsteady-state compensation branch. This branch contains two densely connected nonlinear correction networks, which perform multiple rounds of iterative correction on the current hidden layer state before outputting a fine-tuning amount of the pulse frequency. When the value is greater than 0.70, it is determined that there is a serious channel effect in the powder bed. The pulse frequency of the water pump drive module is directly reduced to 60% of the current value, and S03 is re-triggered to perform flow resistance inversion.

[0074] The specific implementation of step S05 is as follows: using the pulse frequency adjustment output in S04 as a monitoring signal, the jump matrix parameters are updated online through the weight update mechanism of the multidimensional spatial boundary convergence mechanism network. The jump matrix is ​​established on an additional linear mapping path between the graph theory topology mapping layer and the jump decision output layer, so that the local flow resistance details extracted by the graph theory topology mapping layer can be directly weighted and superimposed to the output without going through the layer-by-layer nonlinear transformation of all intermediate layers, avoiding the loss of microscopic pressure change information in the multi-layer transmission. The weight update uses the Adam optimizer, and the loss function integrates the flow resistance inversion error term and the fluid conservation law penalty term to ensure that the network does not violate physical constraints during the convergence process. After the update is completed, the corresponding instantaneous pulse frequency control signal is output to the water pump drive module through pulse frequency modulation.

[0075] The specific implementation of step S06 is as follows: The current volumetric flow rate value of the effluent flow rate sensor is read, and simultaneously, the absorbance feature vectors of the extract in the visible and near-infrared bands are obtained from the spectral data acquisition unit. After L2 normalization of the absorbance feature vector and the pre-stored standard extraction feature vector, their dot product is calculated to obtain a cosine similarity value, ranging from 0 to 1. The closer the value is to 1, the better the component distribution of the extract matches the standard state. When the similarity value is below 0.80, the system determines that the extraction has not yet met the standard, increases the duration of the current extraction stage, and re-triggers S03; when the similarity value is between 0.80 and 0.95, the current extraction parameters are maintained and execution continues; when the similarity value is above 0.95, the extraction quality is determined to meet the standard, the water supply for subsequent extraction stages is reduced, and the process proceeds to S07.

[0076] The specific implementation of step S07 is as follows: The spatial distribution matrix of flow resistance at each moment during the entire extraction process is organized according to the time sequence and written together with the ultrasonic concentration data sequence into the non-volatile storage area of ​​the edge microcontroller unit as reference historical data for the next extraction. Subsequently, based on the flow resistance matrix sequence and ultrasonic concentration sequence of this extraction, the jump matrix parameters of the multidimensional spatial boundary convergence mechanism network are incrementally updated online, so that the flow resistance inversion capability of the network is continuously accumulated and optimized with each brewing experience. Finally, an extraction completion signal is output to the control chip, the water pump drive module is turned off, and the current control cycle ends.

[0077] The specific structure of the multidimensional spatial boundary convergence mechanism network is as follows: the network consists of a cascaded multi-source data heterogeneous input layer, a graph theory topology mapping layer, a mechanism conservation law constraint layer, a self-circulating feature extraction layer, an unsteady-state compensation branch, and a jump decision output layer. The input layer receives time-stamped pressure distribution tensor and ultrasonic concentration scalar data, uniformly encodes the heterogeneous data, and sends it to subsequent computational layers. The graph theory topology mapping layer uses each spatial sensor node in the powder bed as a vertex and the flow resistance difference between nodes as the edge weight, embedding the Kruskal minimum spanning tree algorithm to generate a topological connectivity matrix in real time, completing the physical spatial topology modeling. This algorithm constructs a minimum spanning tree in ascending order of edge weights using a greedy strategy, which can restore the fluid's preferred conduction path with the lowest computational cost. It is an explicit algorithm used by default in non-neural networks in graph theory, giving the network interpretable physical topological priors. The mechanism conservation law constraint layer follows immediately, using a multiplier constraint matrix to transform the continuity equation and momentum conservation equation of incompressible fluids into hard boundary constraints on the network weights, preventing the network from outputting inversion results that violate the fluid physics mechanism. The self-recurrent feature extraction layer consists of three layers of deep recurrent units with hidden state self-feedback. At each time step, the hidden state of the recurrent unit re-injects the output of the previous step into the current input, enabling the network to extract the dynamic features of flow resistance evolving over time. The self-recurrent feature extraction layer integrates a resource allocation mechanism, which differs from the deviation coefficient calculated in step S04. Directly linked, adjusting the exponential function through resource allocation. Dynamically control the allocation of the following three resources: the proportion of activated neurons participating in forward computation in each hidden layer out of the total number of neurons. The number of computational steps each loop unit performs in a single inference iteration. And the size of the video memory cache block dynamically allocated for storing the activation value of the intermediate layer. When When smaller, Falling into the low load range, Reduced to 40% Set to 2 times, minimize the allocation of video memory cache blocks to meet steady-state operation requirements with the lowest computing power; when At medium time, It falls within the range of slight disturbances. Increased to 75% Adjusted to 5 times, striking a balance between accuracy and real-time performance; when When the high load threshold is exceeded, Reaching the full computing power range, Forced lock to 100% The computational power was increased to 12 times, and the largest memory cache block was allocated to suppress severe channeling effects with maximum computing power. The jump matrix establishes a cross-layer linear mapping path between the graph theory topology mapping layer and the jump decision output layer, allowing shallow local pressure details to be directly superimposed to the output, avoiding information loss during multi-layer nonlinear transformations. The unsteady-state compensation branch is controlled by a mutation trigger, activated when the outflow velocity changes abruptly. It contains two densely connected nonlinear correction networks that iteratively correct the hidden layer state multiple times before re-integrating into the main network's decision output. The jump decision output layer simultaneously outputs the instantaneous pulse frequency control signal and the flow resistance spatial distribution matrix through a fully connected structure.

[0078] The steps for establishing the training dataset of the multidimensional spatial boundary convergence mechanism network specifically include: using computational fluid dynamics software to numerically simulate the unsteady fluid dynamics behavior of coffee grounds under different compaction, different grinding particle size distributions, and different channel effect severity, and collecting the pressure field, velocity field, and soluble solids concentration field of each spatial coordinate as virtual samples; conducting multiple rounds of standard brewing tests through a smart coffee machine prototype, and synchronously recording heterogeneous physical data streams under real-world conditions using a multi-channel fluid pressure sensor array and an ultrasonic concentration detection unit; normalizing the simulation data and experimental data in terms of spatiotemporal dimensions, and using the actually measured spatial distribution matrix of coffee grounds flow resistance and the final brewing quality index as supervision labels to construct a paired training dataset containing input tensors and target matrices.

[0079] The specific steps of training the multidimensional spatial boundary convergence mechanism network include: inputting training samples into the graph theory topology mapping layer, initializing the spatial topology matrix using the Kruskal minimum spanning tree algorithm, and simultaneously injecting prior physical parameters of mass and momentum conservation into the conservation law constraint layer; performing forward computation of the network, performing residual verification between the output flow resistance spatial distribution matrix and the fluid dynamics equations, and constructing a composite loss function that integrates the mean square error of flow resistance inversion and the fluid conservation law penalty term; and using the Adam optimizer to alternately iteratively update the hidden layer weights of the self-loop feature extraction layer and the linear mapping parameters of the jump matrix, so as to converge the flow resistance inversion network to the global optimum while keeping the conservation law constraints intact.

[0080] This multidimensional spatial boundary convergence mechanism network injects the topological structure of the fluid-priority conduction path into the network in an interpretable physical prior form through the Kruskal minimum spanning tree algorithm of the graph theory topology mapping layer, making feature extraction independent of purely data-driven approaches. The conservation law constraint layer embeds the fluid physics mechanism into the network weight space in the form of hard multipliers, fundamentally eliminating inversion results that violate physical laws. The jump matrix seamlessly transmits shallow local pressure details to the output, ensuring that microscopic flow resistance mutation information is fully reflected in deep decision-making. The resource allocation mechanism dynamically binds the deviation coefficient to the proportion of activated neurons, the number of self-looping iterations, and the size of the GPU cache block. This reduces computational power consumption when the powder bed is stable and mobilizes all computational resources when channel effects are severe, ensuring precise matching between the microcontroller's hardware computing power and the actual fluid disturbance intensity. This eliminates long-tail latency while guaranteeing real-time control of channel effects.

[0081] Resource allocation adjustment index function The definition is as follows. This function is based on the deviation coefficient calculated in step S04. The characteristic variance of the spectral data in step S01 Perform real-time calculations: ,in and This is the preset weight adjustment coefficient. When... When the system is determined to be in a steady-state, low-load operation period, the activation neuron ratio parameter is configured as follows: The number of iterations in the self-loop is limited to 1. Next, maintain network inference with minimal computing power. At that time, if a slight flow disturbance is detected in the powder bed, the proportion parameter of activated neurons is increased to [value missing]. The number of iteration steps in the self-loop is adjusted to Next, a balance is struck between computational accuracy and real-time response speed. When a severe channel effect or spectral abrupt change is detected in the system, the proportional parameter of the activated neurons is forcibly locked to a certain value. The number of iteration steps in the self-loop has been increased to Next, the largest dynamic memory cache block is allocated simultaneously to suppress fluid unsteady distortion with the highest computing power precision.

[0082] It should be noted that the key technologies of this invention include: First, the orthogonal projection and flow resistance spatial inversion technology based on a multi-channel pressure array. By canceling the common-mode coupling errors of each sensor channel in the orthogonal basis space, an accurate pressure distribution tensor is constructed. Combined with the topological modeling of the Kruskal minimum spanning tree algorithm and the physical boundary constraints of the conservation law constraint layer, the network can accurately invert the spatial non-uniform distribution of flow resistance inside the powder bed. Traditional single-point pressure measurement cannot obtain spatial distribution information, nor can it apply physical mechanism constraints, thus failing to identify the location and degree of channel effects. Second, the graded closed-loop pump control technology driven by deviation coefficient. By quantifying the norm deviation between the flow resistance spatial distribution matrix and the standard uniform flow resistance vector into a deviation coefficient, and dividing it into three control intervals to trigger frequency adjustment strategies of different intensities, the control response is precisely matched with the severity of the channel effect. Traditional fixed-pressure open-loop control cannot adaptively adjust the output according to the real-time state of the powder bed, resulting in the neglect of mild channel effects and over-response to severe channel effects. Thirdly, the dynamic computing power allocation technology driven by the resource allocation adjustment exponential function maps the deviation coefficient and the spectral characteristic variance into an adjustment exponent, dynamically controlling the proportion of activated neurons, the number of self-loop iteration steps, and the size of the video memory cache block in real time. This reduces computing power consumption when the powder bed is stable and mobilizes all computing resources when channel effects are severe, ensuring precise matching between the microcontroller's hardware computing power and the actual disturbance intensity. These three key technologies work synergistically: multi-channel array inversion provides the basis for accurate calculation of the deviation coefficient, the deviation coefficient provides the driving input for the resource allocation function, and dynamic computing power allocation ensures that flow resistance inversion and control decisions can be completed in a timely manner under different disturbance intensities. These three elements form a tightly linked technical loop, jointly achieving continuous, accurate, and real-time suppression of powder bed channel effects.

[0083] It should be noted that under conditions of concurrent high-frequency sampling of multiple sensor data streams, the arrival time of each data stream is often non-uniform in the time dimension due to differences in sensor response time, inconsistent communication bus bandwidth, and differences in processing thread scheduling priorities. This results in the inaccurate alignment of multiple sampled values ​​corresponding to the same physical moment, a problem that is particularly prominent when the data stream density is high or the system load fluctuates greatly. The reason for the above technical problem is that the measurement cycle of the ultrasonic concentration detection unit is usually on the order of milliseconds, while the spectral data acquisition unit needs to scan multiple bands successively, and its single frame acquisition time may reach tens of milliseconds. The inherent time difference between the two will evolve into a randomly distributed long-tailed delay when the concurrent buffer competes for write rights. If the latest timestamp value is directly aligned, the fast sensor data will frequently wait for the slow sensor data, resulting in a decrease in the real-time performance of control decisions. If the time difference is ignored and the data is directly fused, the values ​​of different channels in the fused data frame correspond to different physical moments, destroying the multidimensional data correlation on which the flow resistance inversion depends, and severely reducing the inversion accuracy. To address the aforementioned technical issues, common solutions include setting a fixed sampling period for all sensors and using an interrupt synchronization mechanism to force alignment, or employing a first-in-first-out (FIFO) buffer to queue and wait for the slowest sensor to arrive before reading it all at once. However, the fixed sampling period scheme requires consistent response times for all sensors, a requirement that cannot be met due to the differences in the physical measurement mechanisms of spectral, ultrasonic, and pressure sensors. While the unified queuing scheme can ensure timing alignment, the process of waiting for the slowest sensor introduces a fixed delay equivalent to the acquisition time of the slowest sensor. When a sudden change occurs in the powder bed flow pattern, there is an unacceptable time difference between the moment the control chip receives the fused data frame and the moment the change occurs, thus compressing the response window for channel effects. This invention effectively solves this technical problem. The dual-buffered queue mechanism allocates two independent first-in-first-out (FIFO) buffers to each sensor data stream. Write and read operations are executed in parallel on different buffers without blocking each other, eliminating random queuing delays caused by bus contention. Simultaneously, the system extracts the characteristic variance of each data stream as a rate feature. Temperature data threads with smaller variances have shorter suspension times, while spectral data threads with larger variances have appropriately longer suspension times. This differentiated thread scheduling strategy ensures precise temporal alignment of the data streams, rather than passively waiting for the slowest sensor. This mechanism, while maintaining timing alignment accuracy, compresses the maximum waiting delay to the extremely short computation time required for single-stream characteristic variance estimation. This ensures that the fused data frame always reflects the true powder bed state at the current physical moment, providing a fundamental guarantee for the real-time performance of subsequent flow resistance inversion and control decisions.

[0084] Specifically, the principle of this invention is as follows: The core reason why this invention can solve the above-mentioned technical problems lies in the following aspects. First, the multi-channel fluid pressure sensor array is embedded in the sidewall and bottom of the powder bed bearing cavity in an array form, which can obtain pressure distribution information at different spatial locations inside the powder bed, rather than relying on the scalar value of a single measurement point. This provides the necessary multi-dimensional input for subsequent spatial flow resistance inversion. Second, the orthogonal projection processing cancels out the common-mode error components caused by the spatial coupling of each sensor channel in the projection space, retaining only the independent pressure characteristic information of each spatial node, thereby constructing an accurate pressure distribution tensor. Third, the conservation law constraint layer embedded in the multi-dimensional spatial boundary convergence mechanism network transforms the continuity equation and momentum conservation equation of incompressible fluid into constraint conditions for node activation value multiplication operations, so that the network will not output results that violate physical laws when inverting the flow resistance distribution, significantly improving the reliability and generalization ability of the inversion. Fourth, a graph theory topology mapping layer nested with the Kruskal minimum spanning tree algorithm abstracts the physical nodes of the powder bed space into a weighted undirected graph. The topological connectivity matrix is ​​generated in real time using the flow resistance difference between nodes as edge weights, intuitively depicting the spatial morphology of the core fluid conduction path and providing a topological basis for identifying the location and extent of channel effects. Fifth, the deviation coefficient, as a quantitative representation of the spatial non-uniformity of flow resistance, is divided into three control intervals, corresponding to three response strategies: maintaining the current frequency, fine-tuning the frequency, and forced frequency reduction. This forms a closed-loop control logic from perception to decision-making, ensuring that channel effects can be targeted and suppressed at different severity levels. Sixth, a dual-buffered queue mechanism maintains two independent first-in-first-out buffers for each sensor data stream, ensuring that read and write operations do not block each other. This eliminates the long-tail delay during concurrent sampling of multiple data streams, guarantees the consistency of timestamps in the fused data frames, and ensures that subsequent flow resistance inversion and control decisions are based on the real physical state at the same moment. The above mechanisms work together to form a complete technical chain from multi-dimensional perception, physical constraint inversion, topological path identification to dynamic closed-loop control, enabling the present invention to continuously and accurately suppress the powder bed channel effect and achieve uniform extraction under actual brewing conditions.

[0085] The following is a specific embodiment 1 of the multidimensional space boundary convergence adaptive extraction control module. The specific implementation of each step in this embodiment 1 is described in detail below.

[0086] The specific implementation of step S01 is as follows: The edge microcontroller unit internally maintains two independent first-in-first-out (FIFO) buffers for each of the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, effluent flow rate sensor, and spectral data acquisition unit. One buffer is in a write state, receiving newly arrived sensor sampling data, while the other buffer is in a read state, outputting received data to subsequent processing threads. The two buffers automatically switch roles when full or empty, thereby eliminating long-tail latency caused by read / write blocking. (Thread suspension waiting time for each data path is also mentioned.) Based on the characteristic variance of temperature, flow rate, and spectral data Dynamic adjustment, the formula is expressed as follows:

[0087] ;

[0088] In the formula, For the first The thread suspension and waiting time for the data stream, in milliseconds. For the first The basic hang time for the data stream, measured in milliseconds (ms), is empirically estimated at 5ms. For the first The variance sensitivity coefficient of the data stream is dimensionless and obtained through pre-calibrated experimental fitting; its value is typically between 0.5 and 2.0. For the first The characteristic variance of the road sensor data within the current sampling window, with units consistent with the square dimension of the corresponding sensor physical quantity. For sensor channel index, It is an exponential function with the natural constant as its base. Because... Dimensionless Since it is dimensionless, the dimensions on both sides of the equation are milliseconds (ms). After aligning the timestamps of each data stream, the data streams are concatenated and output as a fused data frame.

[0089] The specific implementation of step S02 is to process the original numerical vector of multi-channel pressure distribution in the fused data frame. Orthogonal projection processing is performed to eliminate noise coupling interference between sensor channels. The formula is as follows:

[0090] ;

[0091] In the formula, This is the raw numerical vector from the multi-channel pressure sensor, in Pa. This represents the total number of sensor channels. The orthogonal projection matrix, composed of orthogonal basis vectors arranged in columns, is dimensionless and is pre-obtained by principal component decomposition of the covariance matrix of historical pressure data. This is the pressure eigenvector after projection to eliminate common-mode error, in Pa, with superscript. This represents the matrix transpose. Then, the spatial coordinates of each sensor are used. For dimension indexing, Reorganized into pressure distribution tensor The unit is Pa, where , , They are respectively , , The spatial node indexes are determined by orientation, and the pressure gradient feature vectors of each spatial node are extracted. The formula for the directional gradient components is expressed as follows:

[0092] ;

[0093] In the formula, For nodes along The pressure gradient component in the direction, in Pa / m. and These represent the projected pressure values ​​of adjacent spatial nodes, in Pa. For adjacent sensors in Spacing in directions, in meters. Directional gradient components and Directional gradient components The calculation method and Similar.

[0094] The specific implementation of step S03 is to convert the pressure distribution tensor into a single value. and ultrasonic concentration data Input a multidimensional spatial boundary convergence mechanism network. A graph theory topology mapping layer abstracts the spatial physical nodes of the coffee grounds into a weighted undirected graph. The vertex, For the set of vertices, Let the set of edges be defined, with the difference in flow resistance between nodes as the edge weight. The formula is expressed as follows:

[0095] ;

[0096] In the formula, For nodes With nodes The edge weights between them are expressed in Pa·s / m. , and They are nodes With nodes The estimated local flow resistance, in Pa·s / m , Represents absolute value. The Kruskal minimum spanning tree algorithm is called to... Processing is performed to generate a topological connectivity matrix in real time. ,in The total number of nodes. Middle elements If the edge It belongs to the minimum spanning tree, otherwise The Kruskal's minimum spanning tree algorithm proceeds as follows: ... Sort all edges in ascending order of their weights; then, take the edge with the smallest weight in sequence. If adding this edge does not create a cycle, add it to the spanning tree; otherwise, skip it. Repeat this process until the spanning tree contains all edges. By adding edges, we obtain the topological connectivity matrix representing the conduction path of the fluid core. The conservation law constraint layer transforms the continuity equation and momentum conservation equation of incompressible fluids into a network weight multiplier constraint matrix. Activation value of each node Apply constraints, as expressed in the formula below:

[0097] ;

[0098] In the formula, This is the activation value vector for each node. This is the conservation law constraint coefficient matrix. The total number of conservation law equations, Each row corresponds to the discretized linear coefficients of the continuity equation and the momentum conservation equation at each node, ensuring that the numerical values ​​within the network always satisfy the constraints of the fluid physics mechanism. The dimensions of are consistent with the dimensions of the left-hand side of the corresponding conservation law equation, and the constraint value is a zero vector. The self-looping feature extraction layer contains three deep recurrent units, the first of which is... Hidden state of the next iteration The updated formula is expressed as follows:

[0099] ;

[0100] In the formula, For the first The hidden state vector of the next iteration. This represents the total number of hidden layer neurons. The self-circulating weight matrix is ​​dimensionless. The input weight matrix is ​​dimensionless. For the input feature dimension, The input feature vector at the current time step. The bias vector is dimensionless. It is a non-linear activation function. This is the iteration step number. After... After the iteration, the final output flow resistance spatial distribution matrix is... Its elements For the first Local flow resistance inversion values ​​at spatial nodes, in Pa·s / m , and They are respectively direction and Number of spatial nodes in a direction.

[0101] The specific implementation of step S04 is to calculate the spatial distribution matrix of flow resistance. Compared with standard uniform flow resistance vector Deviation coefficient between The formula is expressed as follows:

[0102] ;

[0103] In the formula, The deviation coefficient is dimensionless. To make the matrix The column vector obtained by expanding by column, with units of Pa·s / m , This is a standard uniform current resistance reference vector, pre-acquired through calibration experiments and stored in the control chip, with units of Pa·s / m. , This represents the Euclidean norm, where the numerator and denominator have the same dimensions, and the quotient is dimensionless. When When, maintain the current water pump pulse frequency unchanged; when At that time, the unsteady-state compensation module is triggered, and the output pulse frequency is fine-tuned by iterating through multiple rounds of hidden layer states. The unit is Hz; when When this happens, the pulse frequency of the water pump drive module is directly reduced to 60% of the current value and step S03 is executed again.

[0104] The specific implementation of step S05 is to adjust the pulse frequency based on the pulse frequency output in step S04. By incorporating the dynamic convergence term of the flow resistance boundary into the loss function through the weight update mechanism of the multidimensional spatial boundary convergence mechanism network, a composite loss function is formed. The formula is expressed as follows:

[0105] ;

[0106] In the formula, The total value of the composite loss function is dimensionless. The traditional cross-entropy loss term is dimensionless. The fluid conservation law penalty term is calculated from the residual norm of the non-constrained elements in the conservation law constraint layer in step S03. It is dimensionless. This is the conservation law penalty coefficient, dimensionless, with an empirical value of 0.1 to 1.0. This is the convergence weighting coefficient for the flow resistance boundary, dimensionless, with an empirical value of 0.05 to 0.5. Here, the mean square deviation term between the flow resistance distribution and the standard distribution is... Dimensional Pa ·s ·m To ensure that the term is dimensionless, The overall dimensionless property. Network weights and pump pulse modulation coefficients. The update formula is expressed as follows:

[0107] ;

[0108] In the formula, The updated pulse modulation coefficients are in Hz. The pulse modulation coefficients before the update are in Hz. The pulse frequency learning rate is dimensionless, with an empirical value of 0.01 to 0.1. The output from the unsteady-state compensation module in step S04, in Hz, will The corresponding instantaneous pulse frequency control signal is output to the water pump drive module.

[0109] The specific implementation of step S06 is to acquire the absorbance feature vector output by the spectral data acquisition unit in the current extraction stage. Extract feature vectors from pre-stored standards Calculate cosine similarity The formula is expressed as follows:

[0110] ;

[0111] In the formula, This is a similarity value, dimensionless, ranging from 0 to 1. This is the absorbance feature vector output by the spectral data acquisition unit in the current extraction stage; it is dimensionless. The pre-stored standard extracted feature vector is dimensionless. and They are respectively and The normalized vectors obtained by dividing each vector by its Euclidean norm are dimensionless. This represents the number of wavelength channels for spectral sampling. This represents the vector dot product, where the numerator and denominator have the same dimensions, and the quotient is dimensionless. When When this happens, increase the duration of the current extraction stage and re-trigger step S03; when When, maintain the current extraction parameters and continue execution; when At this time, reduce the water supply for subsequent extraction stages and proceed to step S07.

[0112] The specific implementation of step S07 is to sequence the spatial distribution matrix of the flow resistance throughout the extraction process. With ultrasonic concentration data sequence The data is written to the storage area of ​​the edge microcontroller unit, and the jump matrix of the multidimensional spatial boundary convergence mechanism network is updated based on the extracted data. The formula is expressed as follows:

[0113] ;

[0114] In the formula, This is the updated jump matrix. This is the jump matrix before the update. The learning rate for the jump matrix is ​​dimensionless, with an empirical value of 0.001 to 0.01. The error gradient vector at the output of the deep decision-making process is dimensionless. The shallow local feature vectors output by the graph theory topological mapping layer are dimensionless. This represents the number of deep nodes. This represents the number of shallow nodes. It is an outer product matrix, dimensionless, and after the update is completed, it outputs an extraction completion signal to the control chip.

[0115] Resource allocation adjustment index function The specific implementation method is based on the deviation coefficient in step S04. The characteristic variance of the spectral data in step S01 Real-time calculations are performed, and the formula is expressed as follows:

[0116] ;

[0117] In the formula, This is a resource allocation and adjustment index, dimensionless. and The preset weighting adjustment coefficients are dimensionless and obtained through pre-calibrated experimental fitting; they are empirical values. The value ranges from 0.6 to 1.0. It is 0.2 to 0.5. The deviation coefficient in step S04 is dimensionless. The characteristic variance of the spectral data in step S01 within the current sampling window is here. It is made dimensionless after normalization. It is the natural logarithm function. Dimensionless, therefore The whole is dimensionless. When At that time, the proportional parameter of activated neurons Number of self-loop iterations times; when hour, , times; when hour, , Next, the largest dynamic memory cache block is allocated simultaneously to suppress fluid unsteady distortion with the highest computing power precision.

[0118] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0119] This embodiment uses a standard espresso brewing process as the verification scenario. The coffee powder filling amount is 18g, the median particle size distribution is 350μm, and 16 pressure sensors are embedded in the coffee bed support cavity, distributed as follows: 12 channels on 4 layers of the side walls and 4 channels on the bottom, with a sampling frequency of 100Hz. The heating water tank stabilizes the water temperature at 93℃, and the initial pulse frequency of the water pump drive module is set to the frequency value corresponding to a 9bar water pressure.

[0120] Once brewing begins, the dual-buffer queue of the edge microcontroller unit is immediately activated, assigning independent dual-buffer queues to the pressure array, ultrasonic concentration detection unit, outlet flow rate sensor, and spectral data acquisition unit, respectively. During the first data acquisition cycle, the spectral data acquisition unit needs to scan band by band, and its characteristic variance estimate is... The corresponding thread suspension time is dynamically extended to 28ms by the system; the characteristic variance of the stress array is... The thread suspension time is only 4ms. Data from each channel is timestamped under a differentiated waiting strategy, and the first frame of fused data is output. The timing deviation between different sensor channels is controlled within 1ms.

[0121] After obtaining the fused data frame, the system performs orthogonal projection processing on the 16 original pressure numerical vectors. A 16-dimensional orthogonal basis is constructed using the Gram-Schmidt orthogonalization process, mapping the original vectors to the orthogonal basis space and eliminating common-mode errors introduced by the physical coupling of adjacent sensors. The independent pressure characteristic values ​​of each spatial node in the constructed pressure distribution tensor after projection are shown in Table 1 (taking the bottom 4 sensors as an example).

[0122] Table 1. Characteristic values ​​of the bottom pressure sensor after orthogonal projection

[0123]

[0124] As shown in Table 1, the original pressure value of sensor B03 is significantly lower than that of other bottom sensors. After orthogonal projection, the characteristic value of this node retains the local low pressure characteristics, while the common mode error components of other nodes are effectively canceled. This indicates that orthogonal projection processing can retain the true local pressure anomaly information while eliminating coupling noise.

[0125] The pressure distribution tensor and the concentration scalar data output by the ultrasonic concentration detection unit (current value) The data is then input into a multidimensional spatial boundary convergence mechanism network, which sequentially passes through a graph theory topology mapping layer, a mechanism conservation law constraint layer, and a self-circulating feature extraction layer. The graph theory topology mapping layer uses 16 sensor nodes as vertices and the difference in flow resistance between adjacent nodes as edge weights, applying the Kruskal minimum spanning tree algorithm to generate a topological connectivity matrix. Because the pressure at node B03 is relatively low, the corresponding edge weight is significantly larger. The minimum spanning tree bypasses the high-resistance path near node B03 in this region, making the topological connectivity matrix clearly show a low-resistance conduction path converging from directions B01, B02, and B04 towards region B03, accurately locating the location of the channel effect. The conservation law constraint layer embeds the continuity equation and momentum conservation constraints into the forward calculation in the form of a multiplier constraint matrix, ensuring that the inverted flow resistance spatial distribution matrix satisfies the laws of fluid physics. The three-layer loop unit of the self-circulating feature extraction layer iteratively extracts the temporal features of flow resistance, ultimately outputting the flow resistance spatial distribution matrix.

[0126] When calculating the deviation coefficient, the difference vector is constructed by the difference between the inverted flow resistance values ​​of the 16 spatial nodes and the corresponding elements of the standard uniform flow resistance vector. The deviation coefficient is obtained by dividing the L2 norm of the difference vector by the L2 norm of the standard uniform flow resistance vector. The value was in the range of 0.35 to 0.70, indicating a slight channeling effect in the powder bed. Meanwhile, the characteristic variance of the spectral data... Resource allocation and adjustment index ,Pick , Calculated ,fall into At the critical interval, the system will increase the proportion of activated neurons to [a certain value]. The number of iteration steps in the self-loop is adjusted to Second-rate.

[0127] The unsteady-state compensation branch is triggered, and after five rounds of iterative correction of the current hidden layer state, the output pulse frequency is fine-tuned. The pulse frequency of the water pump drive module is reduced by approximately 8% from its original value, corresponding to an adjustment of the outlet water pressure from 9 bar to approximately 8.3 bar. This reduces the impact of water flow on the low-resistivity channel, promoting a redistribution of water flow across the powder bed cross-section. After adjustment, S03 is re-executed, and the deviation coefficient obtained from the second flow resistance inversion decreases to... If the concentration is below 0.35, the fluid dynamics of the powder bed are considered to have returned to normal, and extraction should continue at the current frequency.

[0128] The trend of the deviation coefficient at different times during the extraction process is as follows: Figure 3 As shown. By Figure 3 As can be seen, before the pump pulse frequency was adjusted, the deviation coefficient reached a peak of 0.47 at 8 seconds. As the frequency adjustment took effect, the deviation coefficient continued to decrease and leveled off between 12 and 18 seconds, eventually stabilizing at around 0.25. The physical reason for the gradual leveling off of the deviation coefficient is that reducing the pump pulse frequency lowers the outlet water pressure, weakens the preferential penetration driving force of the water flow into the low-resistivity region of the powder bed, reduces the difference in water flow velocity in different regions of the powder bed, and gradually converges the spatial non-uniformity of flow resistance. This convergence process is the direct physical manifestation of the suppression of the channel effect.

[0129] At the 20s mark of extraction, the absorbance feature vector output by the spectral data acquisition unit, after L2 normalization, showed a cosine similarity of 0.83 with the pre-stored standard extraction feature vector, falling between 0.80 and 0.95. The current extraction parameters were maintained, and the process continued. At the 28s mark, the similarity value rose to 0.97, exceeding 0.95, indicating that the extraction quality met the standards. The system then reduced the water supply for subsequent extraction stages and entered step S07. The similarity values ​​and pump adjustment status at each key moment are shown in Table 2.

[0130] Table 2 Extraction of State Parameters at Key Moments

[0131]

[0132] As shown in Table 2, as the deviation coefficient continuously decreased from 0.47, the similarity value steadily increased from 0.61 to 0.97, showing an inverse correlation between the two trends. The decrease in the deviation coefficient reflects the gradual homogenization of the flow resistance distribution in the powder bed and the gradual equalization of the fluid's permeation path across the powder bed cross-section. This leads to the gradual convergence of the proportion of soluble substances in the extract towards the standard state, resulting in an increase in the similarity value. This pattern, from a physical mechanism perspective, explains that real-time suppression of channel effects to homogenize the flow resistance in the powder bed is the fundamental reason driving the convergence of the extract component distribution towards the standard state, rather than relying on simply extending the time or increasing the water volume to compensate for extraction deviations.

[0133] In stage S07, the system writes the 16-frame flow resistance spatial distribution matrix sequence and the ultrasonic concentration data sequence at 28 time points of the entire extraction process into the non-volatile storage area of ​​the edge microcontroller unit, and performs incremental updates on the jump matrix parameters of the multidimensional spatial boundary convergence mechanism network based on the data, completes the adaptive learning closed loop, and outputs the extraction completion signal.

[0134] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0135] Table 3. Variable Explanation Table (Part 1)

[0136]

[0137] Table 4. Variable Explanation Table (Part Two)

[0138]

[0139] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart coffee brewing machine, characterized in that, include: Multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, water pump drive module, heating water tank, coffee grounds support cavity, outflow velocity sensor, spectral data acquisition unit, edge microcontroller unit and control chip; The control chip is equipped with a multi-dimensional spatial boundary convergence adaptive extraction control module, which is used to fuse the multi-source heterogeneous data streams from the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, water flow rate sensor, and spectral data acquisition unit. Through the multi-dimensional spatial boundary convergence mechanism network, the internal fluid dynamics state of the coffee grounds is inverted in real time, and the instantaneous pulse frequency of the water pump drive module is dynamically adjusted to suppress channel effects and achieve uniform extraction. At the same time, a dual-buffered queue mechanism is used to eliminate the long-tail delay of multi-channel concurrent data, achieving zero-latency multi-channel data fusion.

2. The intelligent coffee maker according to claim 1, characterized in that, The water pump drive module is connected in series with the heating water tank and the coffee powder bed support cavity through a water inlet pipe. The multi-channel fluid pressure sensor array is embedded in the side wall and bottom of the coffee powder bed support cavity in an array form.

3. The intelligent coffee maker according to claim 2, characterized in that, The ultrasonic concentration detection unit is fixedly installed below the liquid outlet of the coffee powder bed bearing cavity, and the water flow rate sensor is installed on the pipeline between the liquid outlet of the coffee powder bed bearing cavity and the ultrasonic concentration detection unit.

4. The intelligent coffee maker according to claim 3, characterized in that, The spectral data acquisition unit is fixed on both sides of the liquid outlet pipe in an through-beam manner, the edge microcontroller unit is fixedly installed on the inner wall of the housing, and the control chip is embedded on the main board of the edge microcontroller unit.

5. The intelligent coffee maker according to claim 4, characterized in that, The control chip is electrically connected to and interacts with the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, water pump drive module, heating water tank, water flow rate sensor, and spectral data acquisition unit.

6. The intelligent coffee maker according to claim 5, characterized in that, The multi-channel fluid pressure sensor array is used to collect fluid pressure distribution data at various spatial locations inside the coffee powder bed bearing cavity in real time; the ultrasonic concentration detection unit is used to continuously measure the concentration of coffee solids in the extract through ultrasonic attenuation coefficient.

7. The intelligent coffee maker according to claim 6, characterized in that, The water pump drive module is used to receive the pulse frequency modulation signal output by the control chip and output the corresponding instantaneous water flow pressure; the heating water tank is used to heat the injected water to the set temperature and supply water stably; the coffee powder bed bearing cavity is a cylindrical cavity structure with a powder bed support filter screen inside, which is used to carry coffee powder and make the water flow evenly through the powder bed.

8. The intelligent coffee maker according to claim 7, characterized in that, The water flow rate sensor is used to collect the volumetric flow rate of the liquid flowing through the outlet in real time; the spectral data acquisition unit is used to collect the absorbance data of the extract in the visible and near-infrared bands to characterize the distribution of characteristic components of soluble substances in the extract; the edge microcontroller unit is used to carry the control chip and complete the concurrent convergence and buffering of multi-channel sensor data.

9. The intelligent coffee maker according to claim 8, characterized in that, The control chip is equipped with a multi-dimensional spatial boundary convergence adaptive extraction control module, which is used to fuse the multi-source heterogeneous data streams from the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, water flow rate sensor, and spectral data acquisition unit. Through the multi-dimensional spatial boundary convergence mechanism network, the internal fluid dynamics state of the coffee grounds is inverted in real time, and the instantaneous pulse frequency of the water pump drive module is dynamically adjusted to suppress channel effects and achieve uniform extraction. At the same time, a dual-buffered queue mechanism is used to eliminate the long-tail delay of multi-channel concurrent data, achieving zero-latency multi-channel data fusion.

10. The intelligent coffee maker according to claim 1, characterized in that, The multidimensional space boundary convergence adaptive extraction control module specifically performs the following steps: The data streams from the multi-channel fluid pressure sensor array, ultrasonic concentration detection unit, effluent flow rate sensor and spectral data acquisition unit are independently buffered and received using a dual-buffered queue. The thread suspension waiting time of each data stream is dynamically adjusted, and the fused data frame is output after timestamp alignment is completed. Orthogonal projection processing is performed on the multi-channel pressure distribution data in the fused data frame to eliminate noise coupling interference between channels, construct the pressure distribution tensor and extract the pressure gradient feature vector of each spatial node. The pressure distribution tensor and ultrasonic concentration data are input into the multidimensional spatial boundary convergence mechanism network. The conservation of mass and momentum is transformed into multiplier constraints through the conservation law constraint layer. The flow resistance spatial distribution inside the powder bed is inverted to obtain the flow resistance spatial distribution matrix. Calculate the deviation coefficient between the flow resistance spatial distribution matrix and the standard uniform flow resistance vector, and trigger three strategies respectively: maintain the current frequency, non-steady-state compensation fine-tuning, or forced frequency reduction according to the interval to which the deviation coefficient belongs. Based on the pulse frequency adjustment amount output by the deviation coefficient, the updated instantaneous pulse frequency control signal is output to the water pump drive module through the weight update mechanism of the multidimensional spatial boundary convergence mechanism network. Collect the values ​​from the outflow velocity sensor and the absorbance feature vector from the spectral data acquisition unit, calculate the similarity value between the absorbance feature vector and the pre-stored standard extraction feature vector, and perform the following actions based on the similarity value range: extend the extraction time, maintain the current parameters, or reduce the subsequent water supply. The flow resistance spatial distribution matrix sequence and ultrasonic concentration data sequence during the extraction process are written into the storage area of ​​the edge microcontroller unit. The jump matrix parameters of the multidimensional spatial boundary convergence mechanism network are updated to complete the adaptive learning closed loop and output the extraction completion signal.