A kitchen scene customization networking method and system

Abstract

The application discloses a kitchen scene customized networking method, the steps are as follows: controlling the wireless communication module to broadcast the beacon frame, the new equipment judges whether to initiate the access request according to the signal strength, and the network dynamically receives according to the load and the equipment priority; the weighted model is used to calculate the equipment priority and allocate resources; the input voice signal is extracted and processed by the lightweight neural network; based on the pre-training model fine-tuning, the kitchen instruction intention is recognized and the slot is filled, and the instruction is analyzed; the hardware performance index is collected in real time, the load trend is predicted, and the voltage frequency is adjusted; the data is encrypted in the wireless communication link layer, and the session key is managed by the key distribution protocol. The method can improve the network access and resource utilization efficiency, reduce congestion, ensure the access and resource supply of key equipment, improve the voice processing precision and instruction recognition accuracy, reduce the energy consumption of the equipment, improve the operation stability, and protect the data security.

Description

Technical Field

[0001] This invention belongs to the field of smart home and Internet of Things technology, specifically a customized networking method and system for a kitchen scene. Background Technology

[0002] With the rapid development of smart home technology, the number of smart devices in the kitchen is constantly increasing, with products such as smart refrigerators, smart ovens, smart range hoods, and voice assistants gradually becoming standard in households. These devices need to rely on networks to achieve interconnectivity in order to truly meet users' core needs for a smart and convenient kitchen environment.

[0003] However, current kitchen equipment networking technology still has many significant shortcomings and is difficult to adapt to the characteristics of different scenarios:

[0004] At the device access control level, traditional networking methods often employ fixed access strategies, failing to fully consider the diverse types of devices and varying degrees of urgency in a kitchen setting. This results in an inability to dynamically adjust based on real-time network load and device priorities, easily leading to network congestion or delays in accessing critical devices.

[0005] The resource allocation method also lacks flexibility. Because it does not take into account core factors such as the real-time data requirements of kitchen equipment and the impact of failures, the goal of efficient resource utilization is difficult to achieve.

[0006] For voice command processing in a kitchen environment, existing technologies have significant bottlenecks: on the one hand, complex neural network models result in excessive computational complexity, making it difficult to achieve real-time operation on edge devices; on the other hand, there is still considerable room for improvement in the recognition accuracy of kitchen-specific commands.

[0007] Kitchen equipment experiences significant load fluctuations during operation, but current technologies lack the ability to monitor hardware performance indicators in real time and predict load trends, making it impossible to adjust voltage and frequency in a timely manner. This can lead to excessive energy consumption and decreased performance stability.

[0008] In terms of data security, traditional networking solutions have significant deficiencies in encryption measures and key management mechanisms at the wireless communication link layer, making it difficult to meet the rigid requirements for user privacy protection and data security in kitchen scenarios. Summary of the Invention

[0009] The purpose of this invention is to provide a customized networking method and system for kitchen scenarios, so as to solve the problems of inflexible access control, low resource allocation efficiency, poor real-time voice processing, unintelligent load adjustment, and insufficient data security in the existing kitchen equipment networking technology.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0011] A customized networking method for a kitchen scene includes the following steps:

[0012] Step S1: Control the wireless communication module to broadcast beacon frames. New devices determine whether to initiate an access request based on the signal strength of the received beacon frames. The network performs dynamic admission control based on load and device priority.

[0013] Step S2: Calculate device priorities using a weighted model and allocate resources based on those priorities;

[0014] Step S3: Perform feature extraction and lightweight neural network processing on the input speech signal;

[0015] Step S4: Fine-tune the pre-trained model to achieve intent recognition and slot filling for kitchen instructions, and parse the kitchen instructions;

[0016] Step S5: Collect hardware performance indicators in real time, predict load trends, and adjust voltage frequency.

[0017] Step S6: Data encryption is performed at the wireless communication link layer, and a key distribution protocol is used to manage the session key.

[0018] According to the above technical solution, the wireless communication module broadcasts beacon frames, and the new device determines whether to initiate an access request based on the signal strength of the received beacon frames. Specifically:

[0019] Control the wireless communication module to broadcast at a preset period. Broadcast beacon frames carrying device identifiers, device types, and initial priority information; new devices receive beacon frames and calculate signal strength using the following formula:

[0020]

[0021] Wherein, the path loss index n ranges from 3.0 to 4.0, and the distance between devices is d. The reference distance is defined as the range of signal strength A at that distance, from [value to be filled in]. When the signal strength is greater than the preset access threshold (The range of values ​​is from) When a new device initiates an access request to the network, it does so.

[0022] Based on the above technical solution, the weighted model is used to calculate the device priority P, and the calculation formula is as follows:

[0023]

[0024] In the formula, the weighting coefficient , The value range is 0.3-0.5; The value range is 0.2-0.4; The value range is 0.2-0.4, where F represents the urgency of the function, R represents the real-time nature of the data, and I represents the impact of the failure; each update cycle... (Value range: 5s-15s) Recalculate priorities based on device status, and allocate resources using a priority queue algorithm based on these priorities. Resource allocation amount... The calculation formula is:

[0025]

[0026] in, Let N be the priority of device i, and N be the total number of devices in the network. This represents the total amount of resources.

[0027] Based on the above technical solution, feature extraction is performed on the input speech signal, including: pre-emphasis processing, the formula of which is:

[0028]

[0029] in, The preemphasis coefficient represents the signal value at time t. The value range is 0.9-1.0; The signal value at time (t - 1) is represented by the Mel spectrum calculated through frame-by-frame windowing; Principal component analysis (PCA) is used for feature dimensionality reduction.

[0030] According to the above technical solution, frame-by-frame windowing processing divides the speech signal into frames of length L and frame shift M, and multiplies them by a window function. The power spectrum is calculated using the following formula:

[0031]

[0032] In the formula, This represents the power spectral density value at a discrete frequency k; It is the processed discrete-time signal, where n is the discrete-time index; L represents the signal. The length of , i.e., the number of sample points in the discrete signal;

[0033] The Mel filter bank contains M filters (ranging from 30 to 50), and the formula for calculating the Mel spectrum is:

[0034]

[0035] In the formula, k represents a discrete frequency index value, used to traverse the frequency range from 0 to (L / 2 - 1) to sum the correlation quantities; This represents the frequency response value of the m-th filter in the Mel filter bank at frequency index k;

[0036] Principal component analysis (PCA), assuming the original feature matrix is ​​X, calculates the covariance matrix as follows:

[0037]

[0038] Where C represents the covariance matrix; n represents the sample size, n-1 is used to obtain the covariance matrix for the unbiased estimate; X is the data matrix. Denotes the transpose of X;

[0039] Perform eigenvalue decomposition on C, select the eigenvectors corresponding to the top k largest eigenvalues ​​(ranging from 10 to 20) to form the transformation matrix U, and then perform dimensionality reduction on the eigenvalues. .

[0040] According to the above technical solution, a lightweight neural network is used to process speech signals. Specifically, a lightweight self-attention neural network architecture is adopted, a sparse attention mechanism is introduced to reduce computational complexity, and the model is optimized by combining knowledge distillation technology.

[0041] Based on the above technical solution, fine-tuning is performed on a pre-trained model, specifically a natural language processing pre-trained model, on a dataset containing no fewer than 15,000 kitchen instructions. A multi-task learning framework is used to jointly optimize intent recognition and slot filling, with the loss function being:

[0042]

[0043] Where L represents the total loss; Represents the weighting coefficients of the loss function; weighting coefficients The value range is 0.4-0.6. The cross-entropy loss is calculated using the following formula:

[0044]

[0045] in, The CRF loss is filled to fit the slots, and the slot label sequence is decoded using the Viterbi algorithm to parse kitchen instructions. N is the number of samples, and C is the number of classes. For real labels, Predict probabilities for the model.

[0046] Based on the above technical solution, hardware performance indicators are collected in real time, and an LSTM neural network is used to predict future load trends. The cell state update formula of the LSTM neural network is as follows:

[0047]

[0048] in, This represents the cell state at the current time t; This represents the output of the forget gate at the previous time t - 1; It represents the cell state at the previous time step t-1; This indicates element-wise multiplication; It is usually the input gate at the current time t; Represents the candidate cell state at the current time t;

[0049] The formula for updating the hidden state is: Map the prediction results to the load level; adjust the voltage frequency using a proportional-integral controller based on the load level, using the following formula:

[0050]

[0051] in, The proportional coefficient represents the amount of voltage frequency adjustment. The value ranges from 0.05 to 0.15, and the integral coefficient is... The value range is 0.005-0.015. This represents the error between the current load and the target load.

[0052] According to the above technical solution, data encryption is performed at the wireless communication link layer using an AES-256 encryption module in CBC (Cipher Block Link) mode. The encryption formula is as follows:

[0053]

[0054] in, This represents the i-th encrypted data block. Represents Advanced Encryption Standard (AES) encryption operations; It is the i-th plaintext data block; This represents the XOR operation. Let K be the (i-1)th encrypted data block, and K be the encryption key.

[0055] A key distribution protocol based on a security association is adopted, and the session key is negotiated using the elliptic curve Diffie-Hellman (ECDH) algorithm. Two-way authentication is performed upon device access, and every [time period missing]. (Value range is 20min−40min) Automatically update the key.

[0056] According to the above technical solution, the networking system includes:

[0057] The beacon broadcasting and access control unit is used to control the wireless communication module to broadcast beacon frames, receive new device access requests, and perform dynamic admission control based on load and priority.

[0058] The priority management and resource allocation unit is used to calculate device priorities using a weighted model and allocate resources based on those priorities.

[0059] The speech signal processing unit is used to perform feature extraction and lightweight neural network processing on the input speech signal;

[0060] The semantic understanding unit is used to fine-tune based on the pre-trained model and parse kitchen instructions;

[0061] The load monitoring and voltage / frequency regulation unit is used to collect hardware performance indicators, predict load trends, and regulate voltage and frequency.

[0062] The data encryption and key management unit is used to encrypt data at the wireless communication link layer and manage session keys using a key distribution protocol.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] In this invention, by judging the signal strength of beacon frames and using dynamic admission control based on load and priority, combined with weighted model calculation of device priority and priority queue algorithm for resource allocation, it can adapt to the characteristics of diverse equipment types and different functional urgency in kitchen scenarios, improve network access efficiency and resource utilization efficiency, reduce network congestion, and ensure priority access and resource supply for critical equipment.

[0065] The system performs pre-emphasis, frame-by-frame windowing, power spectrum calculation, Mel spectrum extraction, and PCA feature dimensionality reduction on the speech signal. Combined with a lightweight self-attention neural network architecture, sparse attention mechanism, and knowledge distillation technology, it reduces computational complexity while ensuring speech processing accuracy, enabling real-time speech processing on edge devices and improving the recognition accuracy and response speed of kitchen instructions.

[0066] By collecting hardware performance indicators in real time, using an LSTM neural network to predict load trends, and adjusting the voltage frequency through a proportional-integral controller, the system can make timely adjustments based on load changes during the operation of kitchen equipment, thereby reducing equipment energy consumption and improving equipment stability and performance. Attached Figure Description

[0067] Figure 1 This is a flowchart of the customized networking method of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example 1

[0070] like Figure 1 As shown, a customized networking method for a kitchen scene includes the following steps:

[0071] Step S1: Control the wireless communication module to broadcast beacon frames. New devices determine whether to initiate an access request based on the signal strength of the received beacon frames. The network performs dynamic admission control based on load and device priority.

[0072] In the dynamic admission control of step S1, multi-dimensional status data of the access device must first be collected, specifically including:

[0073] Basic device status parameters: Bandwidth utilization B (unit: Mbps), Remaining power E (unit: %), Storage capacity utilization S (unit: %), CPU utilization C (unit: %).

[0074] Business type parameters: Business urgency U (quantified as 0-1, 1 is the most urgent, such as the alarm signal of a gas leak alarm; 0 is not urgent, such as food inventory reporting), business real-time level T (quantified as 0-1, 1 is strong real-time, such as voice control commands; 0 is not real-time, such as equipment log synchronization).

[0075] Business requirement parameters: Data transmission priority (0-1, emergency control commands take 1, non-real-time data takes 0.2), latency tolerance D (unit: ms, such as real-time video stream (D≤50ms), sensor data (D≤10000ms), redundancy space R (0-1, 1 indicates tolerable service degradation, such as reduced camera resolution; 0 indicates no degradation).

[0076] Based on device status data and location information (distance d estimated via RSSI in beacon frames or coordinates (x, y) obtained from the device's built-in positioning module), a topology optimization objective function is constructed to minimize link loss and maximize network connectivity, specifically:

[0077] Define link quality factor (equipment i With equipment j between):

[0078]

[0079] Among them, is the path loss (derived from the RSSI in step S1, ), is the current bandwidth occupancy of the link, is the maximum bandwidth of the link, .

[0080] When there is a link (threshold ), trigger topology reconstruction, select the standby device k of as the relay, and update the network topology.

[0081] The dynamic selection mechanism of the networking protocol is specifically as follows: Select the appropriate protocol according to the service type parameters T (real-time) and D (delay tolerance):

[0082] If T≥0.8 and D≤50ms: Select a low-latency protocol (such as ZigBee, transmission rate 250kbps, latency <10ms);

[0083] If 0.3≤T<0.8 and 50ms<D≤500ms: Select a balanced protocol (such as Wi-Fi HaLow, transmission rate 150Mbps, latency 20 - 100ms);

[0084] If T<0.3 and D>500ms: Select a high-throughput protocol (such as LoRa, transmission rate 50kbps, latency >1s).

[0085] Protocol selection decision function: , where Prot is the selected protocol type.

[0086] The dynamic admission decision mathematical model is specifically as follows: First, define the admission index A , and comprehensively evaluate whether to admit a new device:

[0087]

[0088] L is the current network load (quantified as 0 - 1, L = used resources / total resources), (maximum allowable load);

[0089] P is the device priority (the result of the weighted model in step S2);

[0090] U is the service urgency (0 - 1);

[0091] Weight .

[0092] When If the new equipment is accepted, it will be accepted; otherwise, it will be rejected.

[0093] Step S2: Calculate device priorities using a weighted model and allocate resources based on those priorities;

[0094] In step S2, resource allocation based on priority is specifically as follows:

[0095] First, resources are divided into four categories and quantitative indicators are defined, including channel resources, computing resources, storage resources, and time resources.

[0096] Channel resources: Total number of channels Number of channels used Channel utilization ;

[0097] Computing resources: Total computing power (Unit: MIPS) Utilized computing power Calculate the utilization rate: ;

[0098] Storage resources: Total storage space (Unit: GB) Used space Storage utilization ;

[0099] Time resources: Total number of time slots (Unit: ms) Used time slots Time utilization rate .

[0100] Device i's requirement for resource type m (m=1, channel; 2, computing; 3, storage; 4, time)

[0101] in, For business type coefficients (k=1.2 for real-time business, k=0.8 for non-real-time business); Resource consumption rate per unit time (e.g., camera channel resources) ); For the duration of the business; Basic resource requirements (such as the minimum storage requirements for device startup).

[0102] Basic resource allocation: The allocation amount of device i for resource of type m. for:

[0103]

[0104] in For the total resources of the m-th class, Let i be the priority of device i, and N be the total number of devices.

[0105] The specific handling of insufficient resources is as follows: when At that time, for low-priority devices , Resource adjustments are made based on priority thresholds, specifically as follows:

[0106] Adjustment coefficient (The ratio of resource shortfall to the total allocation of low-priority equipment), adjusted resources for low-priority equipment: The released resources are used to supplement high-priority devices: To supplement the quantity, to meet the requirements .

[0107] Based on historical data (such as peak kitchen equipment load from 18:00 to 20:00 daily), use an LSTM model to predict resource demand at time t in the future. Pre-allocated resources: (to ensure minimum resources), in advance Reserve time to avoid peak-hour congestion.

[0108] Real-time monitoring and feedback of resource status, every interval Resource utilization rate ,when When the load exceeds the overload threshold, resource reallocation is triggered, and the adjustment strategy in step (iii) is repeated until... .

[0109] Step S3: Perform feature extraction and lightweight neural network processing on the input speech signal;

[0110] Step S4: Fine-tune the pre-trained model to achieve intent recognition and slot filling for kitchen instructions, and parse the kitchen instructions;

[0111] Step S5: Collect hardware performance indicators in real time, predict load trends, and adjust voltage frequency.

[0112] Step S6: Data encryption is performed at the wireless communication link layer, and a key distribution protocol is used to manage the session key.

[0113] In this invention, by judging the signal strength of beacon frames and using dynamic admission control based on load and priority, combined with weighted model calculation of device priority and priority queue algorithm for resource allocation, it can adapt to the characteristics of diverse equipment types and different functional urgency in kitchen scenarios, improve network access efficiency and resource utilization efficiency, reduce network congestion, and ensure priority access and resource supply for critical equipment.

[0114] The system performs pre-emphasis, frame-by-frame windowing, power spectrum calculation, Mel spectrum extraction, and PCA feature dimensionality reduction on the speech signal. Combined with a lightweight self-attention neural network architecture, sparse attention mechanism, and knowledge distillation technology, it reduces computational complexity while ensuring speech processing accuracy, enabling real-time speech processing on edge devices and improving the recognition accuracy and response speed of kitchen instructions.

[0115] By collecting hardware performance indicators in real time, using an LSTM neural network to predict load trends, and adjusting the voltage frequency through a proportional-integral controller, the system can make timely adjustments based on load changes during the operation of kitchen equipment, thereby reducing equipment energy consumption and improving equipment stability and performance.

[0116] Data encryption is performed using an AES-256 encryption module and CBC mode at the wireless communication link layer. Combined with a key distribution protocol based on a security association, ECDH algorithm for negotiating session keys, two-way authentication, and a regular key update mechanism, data security during kitchen equipment communication is ensured, and user privacy is protected.

[0117] Example 2

[0118] This embodiment is a further refinement of Embodiment 1. This embodiment provides a specific implementation method for a customized networking method for a kitchen scene:

[0119] Step 1: Configure the wireless communication module to broadcast at a preset period. The broadcast beacon frame carries a device identifier of "Smart Oven 001", a device type of "Kitchen Appliance", and initial priority information based on the initial set values ​​of device function urgency, data real-time performance, and fault impact.

[0120] The new device, "Smart Refrigerator 002," receives beacon frames according to the formula:

[0121]

[0122] Calculate the signal strength, where the path loss exponent n = 3.5, and the reference distance. At the reference distance, the signal strength A = -45 dBm. Assuming the distance between devices d = 2 m, the signal strength is: .

[0123] Due to preset access threshold Therefore, the new device "Smart Refrigerator 002" initiates an access request to the network. The network dynamically controls access based on the current load (such as the number of already connected devices, resource utilization of each device, etc.) and device priority (initial priority). If the network load is low and the device has a high priority, access is allowed.

[0124] Step 2, for the network-connected device "Smart Oven 001", the urgency of its function is F=8 (assuming a maximum score of 10), the real-time performance of its data is R=7, and the impact of its failure is I=9. Weighting coefficients. ), then priority .

[0125] Update cycle every time interval Priority is recalculated based on device status. (Assume total resource quantity...) Given a network with a total of 5 devices (N=5), and device priorities of P1=8.0, P2=7.5, P3=6.8, P4=7.2, and P5=6.5, the resource allocation for device "Smart Oven 001" is A1=(8.0 / (8.0+7.5+6.8+7.2+6.5))×100 ≈(8.0 / 36.0)×100 ≈ 22.22 units.

[0126] Step 3: Process the input kitchen voice command "Open the oven". First, pre-emphasize the command. The pre-emphasis coefficient is... The formula is: It enhances the high-frequency components, making the voice signal clearer.

[0127] Frame segmentation and windowing are performed, with a frame length L = 256 points and a frame shift M = 128 points, dividing the speech signal into multiple frames, each multiplied by a Hamming window function: Reduce spectrum leakage.

[0128] The power spectrum of each frame is calculated, and its frequency domain representation is obtained through Fast Fourier Transform (FFT). Then, the square of the modulus is calculated to obtain the power spectrum. .

[0129] The Mel spectrum is calculated using a Mel filter bank containing 40 filters, converting linear frequencies to Mel frequencies, which better matches the characteristics of human hearing. The calculation formula is as follows: ,in Let be the frequency response of the m-th Mel filter.

[0130] PCA is used for feature dimensionality reduction. The original feature matrix X is a Mel-spectral feature, and the covariance matrix is ​​calculated. Eigenvalue decomposition is performed on C, and the eigenvectors corresponding to the top 15 largest eigenvalues ​​are selected to form the transformation matrix U. After dimensionality reduction, the eigenvalues ​​are... This reduces feature dimensions and computational complexity.

[0131] Step 4 employs a lightweight self-attention neural network architecture, introducing a sparse attention mechanism to calculate attention only at key locations, thus reducing computational load. Simultaneously, knowledge distillation technology is combined, using a large pre-trained model as the teacher model and the lightweight model as the student model. The output of the teacher model guides the training of the student model, optimizing model performance.

[0132] Based on a BERT pre-trained model, fine-tuning was performed on a dataset containing 18,000 kitchen instructions, including different types of instructions such as "turn on the oven," "adjust the refrigerator temperature," and "start the range hood." A multi-task learning framework was employed to jointly optimize intent recognition and slot filling. The loss function is: Intended to identify cross-entropy loss: Where N=18000, and C is the number of intent categories (such as "device control", "temperature regulation", etc.). For real labels, The model predicts probabilities. Slot filling uses a CRF layer, and the loss function is... The Viterbi algorithm is used to decode the slot label sequence. For example, for the instruction "set oven temperature to 200 degrees", the slots include "device = oven" and "temperature = 200 degrees", and the specific intent and parameters of the kitchen instruction are obtained by parsing.

[0133] Step 5: Real-time data collection of hardware performance metrics for "Smart Oven 001," such as CPU utilization, memory usage, and temperature, is used as input to an LSTM neural network to predict the load trend for the next 5 minutes. The cell state update formula for the LSTM neural network is as follows: ,in Output for the forget gate. For input gate output, The candidate cell state is shown; the hidden state update formula is: , Output gate output.

[0134] The prediction results are mapped to load levels, such as low, medium, and high. Assuming the predicted load level is high, the error between the current load and the target load is e = 0.3, and the scaling factor... Integral coefficient The amount of change in the voltage frequency is adjusted. (Assuming the integration time is 1 minute) ≈0.03 + 0.01×0.3×60=0.03+0.18=0.21. Adjust the voltage frequency of the equipment according to this change to reduce the load.

[0135] Step 6: At the wireless communication link layer, the transmitted data is encrypted using an AES-256 encryption module in CBC mode. The encryption formula is as follows: , where is the plaintext data block, K represents the previous ciphertext block, and K is the session key. For example, the first plaintext block... The plaintext block P2 is XORed with the initial vector IV and then encrypted to obtain C1. The second plaintext block P2 is XORed with C1 and then encrypted to obtain C2, and so on, to ensure the confidentiality of the data.

[0136] A security association-based key distribution protocol is used, employing the ECDH algorithm to negotiate session keys. Upon device access, two-way authentication is performed to ensure the legitimacy of both communicating parties. Every [time period]... Automatic key updates enhance data security.

[0137] Example 2

[0138] This embodiment is a further refinement of Embodiment 1, mainly focusing on the specific implementation of the speech signal processing unit.

[0139] In the framing and windowing process in step 3, the frame length L = 300 points, the frame shift M = 150 points, and the Hanning window is used as the window function. The Mel filter bank contains 50 filters, covering a frequency range of 300Hz to 8000Hz, making it more suitable for the speech frequency characteristics in a kitchen environment. During PCA feature dimensionality reduction, the feature vectors corresponding to the top 20 largest eigenvalues ​​are selected to further preserve key features.

[0140] In step 4, the lightweight self-attention neural network has 4 layers, each containing 8 attention heads. The sparse attention mechanism uses sliding window attention with a window size of 16, calculating only the attention within the window, significantly reducing computational complexity. During knowledge distillation, the teacher model is a pre-trained RoBERTa model, and the student model is a lightweight self-attention model. The student model is optimized by minimizing the KL divergence between the outputs of the student and teacher models.

[0141] Example 3

[0142] This embodiment is a further refinement of Embodiment 2, mainly focusing on the specific implementation of the load monitoring and voltage / frequency regulation unit.

[0143] In step S5, the input layer of the LSTM neural network contains 10 neurons, corresponding to 10 collected hardware performance metrics (such as CPU utilization, memory usage, temperature, current, voltage, etc.), the hidden layer contains 128 neurons, and the output layer contains 3 neurons, corresponding to low, medium, and high load levels. When training the LSTM neural network, the Adam optimizer is used with a learning rate of 0.001, a batch size of 64, and a training cycle of 100 epochs to improve the accuracy of load prediction.

[0144] The parameters of the proportional-integral controller are adaptively adjusted according to the device type and load characteristics. For the "Smart Oven 001", the load changes rapidly during heating, so the proportional coefficient K_p is automatically adjusted to 0.15 and the integral coefficient Ki is adjusted to 0.015 to speed up the adjustment; during the heat preservation process, the load changes slowly, so the proportional coefficient K_p is adjusted to 0.05 and the integral coefficient Ki is adjusted to 0.005 to reduce adjustment fluctuations.

[0145] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0146] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A customized networking method for a kitchen scene, characterized in that: It includes the following steps: Step S1, control the wireless communication module to broadcast beacon frames, and the new device determines whether to initiate an access request according to the received beacon frame signal strength. Specifically: Control the wireless communication module to broadcast at a preset period. Broadcast beacon frames carrying device identifiers, device types, and initial priority information; new devices receive beacon frames and calculate signal strength using the following formula: Wherein, the path loss index n ranges from 3.0 to 4.0, and the distance between devices is d. A is the reference distance; A is the signal strength at the reference distance. The network performs dynamic admission control according to the load condition and device priority. Specifically: it is necessary to first collect multi-dimensional status data of the access device, specifically including device basic status parameters and service requirement parameters; Based on the device status data and location information, construct a topology optimization objective function to minimize link loss and maximize network connectivity. Specifically: Define link quality factor : in, For path loss, derived based on RSSI, , This represents the current bandwidth usage of the link. This is the maximum bandwidth of the link. ; =0.6, =0.4; When a link exists threshold When the value is 0.3, topology reconstruction is triggered, and the selection is made. The backup device k acts as a relay to update the network topology; Select an adaptation protocol according to the service type parameters T and D: If T≥0.8 and D≤50ms: Select a low-latency protocol; If 0.3≤T<0.8 and 50ms<D≤500ms: Select a balanced protocol; If T<0.3 and D>500ms: Select a high-throughput protocol; Where, T represents real-time performance, and D represents delay tolerance; Protocol selection decision function: , where Prot is the selected protocol type; Step S2, calculate the device priority using a weighted model and perform resource allocation based on the priority; calculate the device priority P using a weighted model, and the calculation formula is: In the formula, the weighting coefficient , The value range is 0.3-0.5; The value range is 0.2-0.4; The value range is 0.2-0.4, where F represents the urgency of the function, R represents the real-time nature of the data, and I represents the impact of the failure; each update cycle... Priorities are recalculated based on device status, and resources are allocated using a priority queue algorithm based on these priorities. The amount of resources allocated is... The calculation formula is: in, Let N be the priority of device i, and N be the total number of devices in the network. Total resources; Step S3, perform feature extraction and lightweight neural network processing on the input voice signal; Step S4, perform fine-tuning based on a pre-trained model to achieve intent recognition and slot filling of kitchen instructions, and parse kitchen instructions; Step S5, collect hardware performance indicators in real time, including CPU utilization rate, memory occupancy rate, temperature, current, and voltage, predict the load trend and adjust the voltage frequency. Specifically: map the prediction result to the load level; according to the load level, adjust the voltage frequency through a proportional-integral controller, and the adjustment formula is: in, The proportional coefficient represents the amount of voltage frequency adjustment. The value ranges from 0.05 to 0.15, and the integral coefficient is... The value range is 0.005-0.

015. This represents the error between the current load and the target load. Step S6, perform data encryption at the wireless communication link layer and manage session keys using a key distribution protocol.

2. The customized networking method for a kitchen scene according to claim 1, characterized in that: The wireless communication module broadcasts beacon frames. New devices determine whether to initiate an access request based on the signal strength of the received beacon frames. The signal strength A at the reference distance ranges from [value missing]. to When signal strength Greater than the preset access threshold At that time, the new device initiates an access request to the network; The range of values ​​is to .

3. The customized networking method for a kitchen scene according to claim 1, characterized in that: Perform feature extraction on the input voice signal, including: pre-emphasis processing, and the formula is: in, The preemphasis coefficient represents the signal value at time t. The value range is 0.9-1.0; The signal value at time t-1 is represented by the Mel spectrum calculated through frame-by-frame windowing; Principal component analysis (PCA) is used for feature dimensionality reduction.

4. The customized networking method for a kitchen scene according to claim 1, characterized in that: Framing and windowing processes divide the speech signal into frames of length L and frame shift M, and then multiply by a window function. The power spectrum is calculated using the following formula: In the formula, This represents the power spectral density value at a discrete frequency k; It is the processed discrete-time signal, where n is the discrete-time index; L represents the signal. The length of , i.e., the number of sample points in the discrete signal; The Mel filter bank contains M filters, and the formula for calculating the Mel spectrum is: In the formula, k represents a discrete frequency index value, used to traverse the frequency range and sum the relevant quantities from 0 to L / 2 - 1; This represents the frequency response value of the m-th filter in the Mel filter bank at frequency index k; Principal component analysis PCA, assuming the original feature matrix is X, calculate the covariance matrix. Specifically: Where C represents the covariance matrix; n represents the sample size, n-1 is used to obtain the covariance matrix for the unbiased estimate; X is the data matrix. Denotes the transpose of X; Perform eigenvalue decomposition on C, select the eigenvectors corresponding to the z largest eigenvalues ​​to form the transformation matrix U, and then perform dimensionality reduction on the eigenvalues. The value of z ranges from 10 to 20.

5. The customized networking method for a kitchen scene according to claim 1, characterized in that: Perform lightweight neural network processing on the voice signal, specifically using a lightweight self-attention neural network architecture, introduce a sparse attention mechanism to reduce the computational complexity, and optimize the model in combination with the knowledge distillation technology.

6. The customized networking method for a kitchen scene according to claim 1, characterized in that: Perform fine-tuning based on a pre-trained model, specifically based on a natural language processing pre-trained model, perform fine-tuning on a data set containing no less than 15,000 kitchen instructions, and use a multi-task learning framework to jointly optimize intent recognition and slot filling. The loss function is: Where L represents the total loss; Represents the weighting coefficients of the loss function; weighting coefficients The value range is 0.4-0.

6. The cross-entropy loss is calculated using the following formula: in, The CRF loss is filled to fit the slots, and the slot label sequence is decoded using the Viterbi algorithm to parse kitchen instructions. N is the number of samples, and C is the number of classes. For real labels, Predict probabilities for the model.

7. A customized networking method for a kitchen scene according to claim 1, characterized in that: Collect hardware performance indicators in real time, use an LSTM neural network to predict the future load trend, and the formula for updating the cell state of the LSTM neural network is: in, This represents the cell state at the current time t; This represents the output of the forget gate at the previous time t - 1; It represents the cell state at the previous time step t-1; This indicates element-wise multiplication; It is the input gate at the current time t; This represents the candidate cell state at the current time t.

8. A customized networking system for kitchen scenarios, characterized in that: The networking system is used to implement the networking method described in any one of claims 1-7. The networking system includes: A beacon broadcast and access control unit, used to control the wireless communication module to broadcast beacon frames, receive new device access requests, and perform dynamic admission control according to the load and priority; A priority management and resource allocation unit, used to calculate the device priority using a weighted model and perform resource allocation based on the priority; A voice signal processing unit, used to perform feature extraction and lightweight neural network processing on the input voice signal; The semantic understanding unit is used to fine-tune based on the pre-trained model and parse kitchen instructions; The load monitoring and voltage / frequency regulation unit is used to collect hardware performance indicators, predict load trends, and regulate voltage and frequency. The data encryption and key management unit is used to encrypt data at the wireless communication link layer and manage session keys using a key distribution protocol.

Images