A Deep Learning-Based Method and System for Predicting the Lifespan of Kitchen Air Conditioner Filter Devices

Abstract

This invention relates to the field of kitchen air conditioner filter life prediction technology, specifically to a deep learning-based method and system for predicting the life of kitchen air conditioner filters. The method includes: edge devices collecting real-time operating parameters of the kitchen air conditioner and extracting basic physical characteristic indicators of the filter; deploying a first local PINN-Transformer model on the edge gateway, performing data preprocessing and local life prediction inference, caching incremental gradient data, and uploading it to the cloud based on a secure fault-tolerant mechanism; deploying a second local PINN-Transformer model on the cloud, using a hybrid parallel architecture for distributed training, synchronously updating the global PINN-Transformer model, and distributing the model parameter package to each edge node after signature encryption. By constructing a filter life prediction system that integrates physical mechanisms and temporal learning, the prediction accuracy and generalization ability in high-fluctuation oil fume scenarios can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of kitchen air conditioner filter device life prediction, and particularly relates to a kitchen air conditioner filter device life prediction method and system based on deep learning. BACKGROUND

[0002] The popularity of kitchen air conditioners is steadily rising, and more and more families and catering businesses recognize the importance of installing air conditioners in the kitchen. This not only improves the working environment of chefs and staff, improves work efficiency, but also helps to prolong the service life of kitchen equipment. At the same time, kitchen air conditioners can also effectively control the spread of oil smoke, providing users with a more comfortable cooking and dining experience.

[0003] The service life of the kitchen air conditioner filter screen is usually affected by the degree of oil smoke attachment blockage and the frequency of cleaning, and its decay process has nonlinear characteristics. At present, the existing technical means mainly monitor through threshold alarm (such as using a differential pressure sensor) or a simple regression model, but these methods cannot effectively estimate the remaining service life of the kitchen air conditioner filter screen in advance. The existing kitchen air conditioner filter screen service life prediction method, for example: Chinese patent with publication number CN115875799A discloses an air conditioner filter screen clogging degree detection method, device and readable storage medium, which proposes that during the operation of the air conditioner, the noise value difference rate of the air conditioner fan blade, the temperature difference rate of the air outlet, and the air outlet speed difference rate are multiplied by the corresponding weight ratio to evaluate the clogging degree of the filter screen, and when the set value is reached, the user is reminded to clean the dust and particles accumulated on the air conditioner filter screen, but it cannot evaluate the performance of the kitchen air conditioner filter device. For example: Chinese patent with publication number CN109140692A discloses a detection method for a kitchen air conditioner filter device and a kitchen air conditioner, which proposes to use infrared rays to scan the kitchen air conditioner filter device to determine the ventilation performance of the filter device, but the uneven microstructure of the filter screen surface causes inherent speckle noise in infrared emission, resulting in image distortion. In addition, the temperature fluctuations caused by the airflow of the kitchen air conditioner during operation also affect the determination result. At the same time, the small contrast between the filter screen edge and the dried oil stains also causes edge recognition failure. For example: Chinese patent with publication number CN104568698A discloses a filter screen state detection device and method in an air conditioner, which proposes to use a wind pressure difference method to determine the clogging degree of the filter screen, but the cooking oil adhesion effect will cause a pressure difference response lag, and when the actual wind resistance increases by 30%, the sensor can only detect a 12% change (ASHRAE experimental data).

[0004] In summary, the existing kitchen air conditioner filter screen life prediction methods mainly have two defects: (1) the data-driven model (such as LSTM) has poor generalization in the oil smoke concentration mutation scene, and the reason is that the physical law is not considered; (2) the pure physical model is difficult to accurately describe the filter screen attenuation (such as microstructure damage caused by cleaning), resulting in large prediction deviation.

[0005] Therefore, the application provides a kitchen air conditioner filter device life prediction method and system based on deep learning. SUMMARY

[0006] In view of this, the application aims to provide a kitchen air conditioner filter device life prediction method and system based on deep learning.

[0007] According to a first aspect of the application, a kitchen air conditioner filter device life prediction method based on deep learning is provided, which is used in the cloud and a plurality of edge nodes interacting with the cloud, the edge nodes including edge devices and edge gateways interacting with the edge devices, the method comprising: the edge device collecting real-time kitchen air conditioner operating parameters, extracting filter device basic physical feature indicators, and sending the filter device basic physical feature indicators to the edge gateway; the edge gateway deploying a first local PINN-Transformer model, performing data preprocessing and local life prediction inference, caching incremental gradient data, and uploading to the cloud based on a safety fault-tolerant mechanism; the cloud deploying a second local PINN-Transformer model, receiving incremental gradient data from a plurality of edge nodes, using a hybrid parallel architecture for distributed training, synchronously updating a global PINN-Transformer model, and signing and encrypting the updated model parameters package and then distributing it to each edge node; wherein the first local PINN-Transformer model and the second local PINN-Transformer model constitute the global PINN-Transformer model.

[0008] Optionally, the real-time kitchen air conditioner operating parameters include the pressure difference on both sides of the kitchen air conditioner filter device, the air flow rate at the return air inlet, the air flow rate at the air outlet, the air temperature at the air outlet, the air humidity at the air outlet, the temperature at the return air inlet, the humidity at the return air inlet, PM2.5 in the kitchen, VOCs concentration in the kitchen, particulate matter concentration at the corresponding return air inlet in front of the kitchen air conditioner filter device, and particulate matter concentration in the return air inlet behind the kitchen air conditioner filter device.

[0009] Optionally, the edge device extracts the filter device basic physical feature indicators based on a sliding time window online, and the filter device basic physical feature indicators include the original filter efficiency, the original capacity attenuation rate, the cleaning frequency, and the cleaning time.

[0010] Optionally, the first local PINN-Transformer model deployed by the edge gateway is a Transformer encoder front-end layer and a local prediction head, the Transformer encoder front-end layer is used to perform data preprocessing, including linear interpolation of missing values, outlier detection, interpolation repair, processing of multi-dimensional time series feature input constructed by sliding window, generating filter device life label annotation, and the local prediction head is used to perform local life prediction inference according to the filter device life label annotation, and output the filter device remaining life prediction value.

[0011] Optionally, the data preprocessing includes: performing linear interpolation of missing values, isolated forest anomaly detection, spline interpolation repair and Gaussian noise injection on the filter device basic physical feature indicators; adopting a physical adaptive strategy to optimize the window length to obtain an adaptive sliding window; constructing multi-dimensional time series features based on the adaptive sliding window and performing normalization processing; inputting the normalized multi-dimensional time series features into the Transformer encoder front-end layer, extracting key time series dependencies through a multi-head self-attention mechanism, and outputting high-level hidden state sequences; performing time dimension average pooling on the high-level hidden state sequences, and compressing to filter device life label annotation through channel grouping and sparse regularization.

[0012] Optionally, the local life prediction inference includes: inputting the filter device life label annotation output by the Transformer into the local prediction head, and outputting the filter device remaining life prediction value.

[0013] Optionally, the second local PINN-Transformer model deployed in the cloud is a middle and rear segment layer of the Transformer encoder and a PINN constraint layer, the middle and rear segment layer of the Transformer encoder is used to receive incremental gradient data from multiple edge nodes, perform deep time series modeling on global feature representations aggregated by multiple edge nodes, capture complex nonlinear evolution patterns in long-term operation of the kitchen air conditioner, and the PINN constraint layer is used to embed a filter device attenuation dynamics equation as a soft constraint, calculate a physical residual term through automatic differentiation, and introduce a loss function to constrain the model output to comply with physical laws, and when the cumulative received incremental gradient data reaches a preset threshold, trigger an incremental learning process, combine the model with a dynamic model updating mechanism, use an AdamW optimizer during training, only update the adjustable physical coefficients of the PINN constraint layer and the weights of the local prediction head, freeze the original parameters of the Transformer encoder front-end layer, and pack as a model parameter package.

[0014] Optionally, the communication link between the edge gateway and the cloud further integrates a fault-tolerant mechanism, an adaptive width adjustment mechanism and an encrypted transmission mechanism.

[0015] Optionally, the model parameter package further embeds a device unique fingerprint.

[0016] According to a second aspect of the present application, a deep learning-based kitchen air conditioner filter device life prediction system is provided, which is used for a cloud and a plurality of edge nodes interacting with the cloud, the edge nodes including edge devices and edge gateways interacting with the edge devices, the edge devices including data acquisition modules and feature extraction modules, the edge gateways including local processing modules and cache uploading modules, the cloud including a cloud distributed model training module and a parameter issuing module; the data acquisition module is used for collecting real-time kitchen air conditioner operating parameters; the feature extraction module is used for extracting filter device basic physical feature indexes and sending the filter device basic physical feature indexes to the edge gateway; the local processing module is used for deploying a first local PINN-Transformer model, performing data preprocessing and local life prediction reasoning; the cache uploading module is used for caching incremental gradient data and uploading to the cloud based on a safety fault-tolerant mechanism; the cloud distributed model training module is used for deploying a second local PINN-Transformer model, receiving incremental gradient data from a plurality of edge nodes, performing distributed training using a hybrid parallel architecture, and synchronously updating a global PINN-Transformer model; the parameter issuing module is used for issuing the updated model parameter package to each edge node after signature encryption; wherein the first local PINN-Transformer model and the second local PINN-Transformer model constitute the global PINN-Transformer model.

[0017] The kitchen air conditioner filter device life prediction method and system based on deep learning provided by the application have the following beneficial effects: (1) By introducing a physical constraint mechanism, the prediction result is constrained to follow the filter device clogging dynamics principle, which can ensure the accuracy and reliability of life prediction; (2) By multi-dimensional time sequence feature learning, factors such as filter efficiency, capacity decay rate, and cleaning frequency of the kitchen air conditioner filter device are comprehensively considered, which can not only accurately quantify the damage effect of cleaning on the kitchen air conditioner filter device, but also fully grasp its working state; (3) By constructing a global PINN-Transformer model, the physical law and data-driven method are combined to realize collaborative optimization, thereby improving the life prediction effect of the kitchen air conditioner filter device; (4) By constructing the physical equation of the kitchen air conditioner filter device clogging dynamics, combined with the adaptive sliding window mechanism, the kitchen air conditioner filter device life prediction effect is optimized, the life prediction result misjudgment rate is reduced, the kitchen air conditioner life is prolonged, and the operation and maintenance cost is reduced; (5) The local edge node and cloud collaborative framework are adopted, and the hierarchical computing architecture can improve the model training efficiency and support dynamic model updating; (6) The communication link between the edge node and the cloud integrates a fault-tolerant mechanism, an adaptive width adjustment mechanism, and an encryption transmission mechanism, supports TCP-level breakpoint resume and transaction-type data rollback, dynamically adjusts the data compression rate and transmission rate, prevents model parameters from being stolen or illegally copied, and effectively reduces the data leakage risk. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The flowchart of the kitchen air conditioner filter device life prediction method based on deep learning of the embodiment of the application.

[0019] Figure 2 The structure block diagram of the kitchen air conditioner filter device life prediction system based on deep learning of the embodiment of the application.

[0020] Figure 3 The structure block diagram of the edge device of the embodiment of the application.

[0021] Figure 4 The structure block diagram of the edge gateway of the embodiment of the application.

[0022] Figure 5 The structure block diagram of the cloud of the embodiment of the application. DETAILED DESCRIPTION

[0023] Embodiments of the present application will be described in detail below, examples of which are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary only, for the purpose of explaining the present application, and should not be understood as limiting the present application, and the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0024] Embodiment one

[0025] The first aspect of the present application proposes a deep learning-based kitchen air conditioner filter device life prediction method, as shown in Figure 1 for the cloud and multiple edge nodes interacting with the cloud, the edge nodes include edge devices and edge gateways interacting with the edge devices, and the method comprises the following steps.

[0026] S1, the edge device collects real-time kitchen air conditioner operating parameters, extracts filter device basic physical characteristic indexes, and sends the filter device basic physical characteristic indexes to the edge gateway.

[0027] In some optional implementations of the present application, the real-time kitchen air conditioner operating parameters include the pressure difference on both sides of the kitchen air conditioner filter device, the air flow rate at the return air inlet, the air flow rate at the air outlet, the air temperature at the air outlet, the air humidity at the air outlet, the temperature at the return air inlet, the humidity at the return air inlet, PM2.5 in the kitchen, VOCs concentration in the kitchen, particulate matter concentration at the corresponding return air inlet in front of the kitchen air conditioner filter device, and particulate matter concentration in the corresponding return air inlet behind the kitchen air conditioner filter device.

[0028] In some optional implementations of the present application, the edge device extracts the filter device basic physical characteristic indexes based on a sliding time window online, and the filter device basic physical characteristic indexes include the original filtration efficiency, the original capacity attenuation rate, the cleaning frequency, and the cleaning time.

[0029] In some optional implementations of the present application, the original filtration efficiency of the kitchen air conditioner filter device at time t is calculated according to formula (1), formula (1) is , wherein, is the particulate matter concentration at the corresponding return air inlet in front of the kitchen air conditioner filter device; is the particulate matter concentration in the corresponding return air inlet behind the kitchen air conditioner filter device; is the original filtration efficiency; the original capacity attenuation rate of the kitchen air conditioner filter device at time t is calculated according to formula (2), formula (2) is: , wherein, is the original capacity attenuation rate.

[0030] In some optional implementations of this application, the extraction process of the basic physical characteristic indicators of the filtering device is completed in real time at the local edge device, retaining only low-dimensional, high-information-density derived features to significantly reduce the subsequent transmission load. All raw real-time kitchen air conditioner operating parameters are immediately cleared after local inference is completed, ensuring that sensitive information is not retained. Simultaneously, the edge device also integrates a lightweight anomaly detection mechanism, using a moving median filter to initially identify and mark abnormal readings. For example, a moving median filter (window size = 7) is used to... The sequence is preprocessed. If the current value deviates from the median by more than 3 times the MAD (Median Absolute Deviation), it is marked as an anomaly and temporarily stored at the edge gateway for further analysis.

[0031] In some optional implementations of this application, the cleaning method for the kitchen air conditioner filter includes the number of cleaning cycles and the cleaning time. Furthermore, all real-time kitchen air conditioner operating parameters are collected once per second during the user's cooking period and once per minute during other periods.

[0032] S2. The edge gateway deploys the first local PINN-Transformer model, performs data preprocessing and local lifetime prediction inference, caches incremental gradient data, and uploads it to the cloud based on a security and fault tolerance mechanism.

[0033] In some optional implementations of this application, the first local PINN-Transformer model deployed by the edge gateway is a Transformer encoder front-end layer and a local prediction head. The Transformer encoder front-end layer is used to perform data preprocessing, including linear interpolation of missing values, outlier detection, interpolation repair, processing of multi-dimensional temporal feature inputs constructed by sliding windows, and generating filter device lifetime label annotations. The local prediction head is used to perform local lifetime prediction inference based on the filter device lifetime label annotations and output the predicted value of the remaining lifetime of the filter device.

[0034] In some optional implementations of the present application, the first local PINN-Transformer model deployed by the edge gateway includes only the front-end layers (usually the first 4 ends) of the original Transformer encoder and a lightweight local prediction head, which mainly serves the following purposes: (1) preliminary feature extraction and context encoding: the edge gateway receives the filtered device basic physical feature indicators uploaded by the edge device, the front-end layers of the Transformer encoder perform preliminary feature extraction and context encoding on the multi-dimensional time series feature input constructed by the adaptive sliding window, and generate the filtered device life label label, effectively compressing the information dimension and retaining the key time series dependency; (2) local life prediction and real-time response: the local prediction head directly outputs the filtered device remaining life prediction value (i.e. the remaining effective operating hours) based on the filtered device life label label, achieving millisecond-level response; for example, when the filtered device life label label is detected to be less than 50% of the filtered device life label label, the edge gateway immediately triggers a cleaning reminder to ensure timely maintenance by the user; (3) resource adaptation and security guarantee: this deployment strategy can compress the model parameter quantity from 12M to about 1.2M, significantly reducing the computational load and memory occupancy, adapting to resource-constrained embedded environments (such as edge gateways based on ESP32 or Raspberry Pi). At the same time, only the minimum functional modules that can perform inference are retained, without complete training capabilities and physical equation structures, which can avoid model reverse engineering and sensitive parameter leakage, enhancing system security.

[0035] In some optional implementations of the present application, the data preprocessing includes: performing linear interpolation on missing values of the filtered device basic physical feature indicators, isolated forest anomaly detection, spline interpolation repair, and injecting Gaussian noise; adopting a physical adaptive strategy to optimize the window length to obtain an adaptive sliding window; constructing multi-dimensional time series features based on the adaptive sliding window and performing normalization processing; inputting the normalized multi-dimensional time series features into the front-end layers of the Transformer encoder, extracting key time series dependencies through the multi-head self-attention mechanism, and outputting high-level hidden state sequences; performing time dimension average pooling on the high-level hidden state sequences, and compressing them to filtered device life label labels through channel grouping and sparse regularization.

[0036] In some optional implementations of the present application, based on time series arrangement, linear interpolation is performed on missing values of the filtered device basic physical feature indicators, isolated forest anomaly detection, spline interpolation repair, and Gaussian noise injection, wherein the inserted missing values and / or the inserted values after spline interpolation repair are verified according to formula (3), and formula (3) is: , wherein, is the filtered device filtering efficiency decline rate; is a proportional coefficient; Gaussian noise is injected according to formula (4) to simulate sensor measurement error, and formula (4) is:​ In the formula, This is the data vector after noise injection; Let be the original data vector at time t; I is the noise vector; I is the identity matrix; The standard deviation of noise; The noise is a multivariate Gaussian distribution; Gaussian noise is injected into the processed filtration efficiency and capacity attenuation rate according to the above formula (4) to obtain a new... , .

[0037] In some optional implementations of this application, a physically adaptive strategy is used to optimize the window length to obtain an adaptive sliding window. Specific implementation methods include: traversing candidate window sizes... In the formula, Minimum window size; The maximum window size is determined by selecting the optimal window size that minimizes the prediction error according to formula (5). Formula (5) is: In the formula, It is a sequence of true values; The predicted value sequence is calculated based on the window size L; the first local Transformer-PINN model is trained, and the mean absolute percentage error (MAPE) of the validation set is calculated according to equation (6), which is: In the formula, n is the number of samples; MAPE is the mean absolute percentage error of the validation set; the sample with the smallest MAPE is selected. As the base window size; the window length is optimized through a physical adaptive mechanism according to formula (7), which is: In the formula, k is the adjustment range coefficient, which takes a value of 0.5; For oil fume compatibility factor, , The PM2.5 level in the kitchen at time t. This represents the maximum PM2.5 concentration in the kitchen. Specifically, it is the real-time normalized value of oil fume concentration (i.e., oil fume concentration factor). Factors affecting the cleaning of kitchen air conditioner filter devices. , The damage factor is affected by the cleaning method used to clean the filter of the kitchen air conditioner. This represents the cumulative number of cleaning sessions. This is the optimized window length.

[0038] In some optional implementations of this application, multi-dimensional temporal features are constructed based on an adaptive sliding window and then normalized. Specific implementations include: constructing multi-dimensional temporal features of a kitchen air conditioner filter using a sliding window, and outputting the multi-dimensional temporal features of the kitchen air conditioner filter according to formula (8), where formula (8) is: In the formula, T is the total length of the time series; L is the window length; and S is the window step size. The multi-dimensional temporal characteristics of the constructed kitchen air conditioner filter device; The maximum window index is used; key features of the kitchen air conditioner filter device are extracted; the key features include pretreatment filtration efficiency, pretreatment capacity decay rate and cumulative cleaning times, and the key features are normalized according to formula (9), which is: In the formula, X represents the standardized eigenvalues; X represents the original eigenvalues. The mean of the eigenvalues ​​X; denoted as the standard deviation of the eigenvalues.

[0039] In some optional implementations of this application, to solve the problem of the Transformer encoder lacking timing awareness, the position encoding PE of the injected data is determined according to formula (10), which is: In the formula, This refers to the time step position; i is the feature dimension; i is the dimension index; Transformer encoder calculation: inject Query, key, and Value multi-head self-attention according to formula (11) and formula (12), formula (11) is: Formula (12) is: In the formula, Q is the query matrix; K is the key matrix; V is the value matrix; and X is the input sequence matrix. The current transformation weight query matrix is ​​learnable; The current transformation weight key matrix is ​​learnable; It is a learnable linear transformation weight matrix; This is the attention score matrix; For each dimension of attention head, h is the number of attention heads; multi-head data splicing and linear transformation are performed according to formula (13), which is: In the formula, This represents the calculation result for the h-th attention head; To output the weight matrix, the input dimension is... The output dimension is Residual connectivity and layer normalization are performed according to formula (14), which is: In the formula, X is the output matrix of the first multi-head attention sublayer; X is the current input matrix; nonlinear transformations are introduced according to formulas (15) and (16) to enhance the model's fitting ability. Formula (15) is: Formula (16) is: In the formula, Z is the final output of the entire encoder; This is the weight matrix of the first layer of the feedforward neural network; This is the weight matrix of the second layer of the feedforward neural network; , All are bias vectors.

[0040] In some optional implementations of this application, the normalized multi-dimensional temporal features are deployed in the front-end layer (first 4 layers) of the Transformer encoder of the edge gateway. Specific implementations include: (1) Input embedding: mapping 4-dimensional features to a 128-dimensional space through learnable linear projection; (2) Position encoding: superimposing sine and cosine position encodings to preserve temporal order information; (3) Transformer encoding of layers 1-4: each layer contains a 3-head self-attention mechanism (h=3, ) and feedforward networks ( ); Layers 1-2 focus on local fluctuations (such as those caused by short-term oil fume disturbances). Oscillation); Layers 3-4 progressively model the medium- to long-term trend (e.g., (The continuous upward trend); each layer ensures stable gradient propagation through residual connections and layer normalization; finally, the high-level hidden state sequence is output: .

[0041] In some optional implementations of this application, time-dimensional average pooling is performed on the high-level hidden state sequence, and the sequence is compressed to filter device lifetime labeling through channel grouping and sparse regularization. The specific implementations include: (1) performing time-dimensional average pooling on the high-level hidden state sequence to obtain the global context vector; (2) using a channel grouping strategy to split it into two 64-dimensional subspaces, where subspace A: retains the 64 key feature channels with the highest attention integral scores during the training phase, mainly encoding dominant features such as "blockage trend intensity" and "oil fume exposure level"; subspace B: applies L1 sparse regularization to the remaining 64 dimensions to force the extraction of high-order statistical characteristics of residual dynamics (such as abnormal disturbance response and post-cleaning recovery mode); after splicing the two subspaces, they are compressed to a 64-dimensional low-dimensional representation (i.e., filter device lifetime labeling) through a linear projection layer. In the formula, This means concatenating 64-dimensional subspace A and 64-dimensional subspace B along the channel dimension to form a complete 128-dimensional input vector. Represents the learnable projection parameter matrix; Used to compensate for feature distribution shifts and improve representation robustness; This generates a compact, low-dimensional context code for use in subsequent prediction heads.

[0042] In some optional implementations of this application, the local lifetime prediction inference includes: inputting the filter device lifetime label output by the Transformer into the local prediction head, and outputting the predicted value of the remaining lifetime of the filter device.

[0043] In some optional implementations of this application, the filter device lifetime label is input into the local prediction head (single-layer fully connected network), and the remaining lifetime prediction value of the filter device is output. Specific implementations include: compressing the filter device lifetime label into a single vector according to formula (17), where formula (17) is: In the formula, It is a single vector; T is the total length of the input sequence; The state of the model at time step t; the predicted remaining lifespan of the filter device is output according to formula (18), which is: In the formula, The predicted value for the filter device output by the model represents the percentage of the filter device's remaining lifespan. 1 indicates a brand new filtration device; This is the weight matrix of the output layer; The bias vector is used; the loss function is constructed according to formula (19), which is: In the formula, This is a balancing parameter used to adjust the strength of the physical constraints, and its value is 0.3. Y represents the predicted lifespan loss; Y represents the actual percentage of the filter's lifespan. The physical information loss term is used to introduce the physical law constraint model of equipment degradation; the lifetime prediction loss is calculated according to formula (20), which is: In the formula, N is the number of samples in a training batch; Let i be the true label of the i-th sample; Let be the model's predicted value for the i samples.

[0044] In some alternative implementations of this application, the model can perform inference once every 15 minutes; for example, after a continuous stir-fry, the system detects... If the value increases by ≥15% three times consecutively (e.g., from 1.8→2.1→2.4→2.8), and the current predicted remaining lifespan of the filter is 23%, a cleaning reminder will be triggered if any of the following conditions are met: (1) the predicted remaining lifespan of the filter is ∈ [0, 30%], (2) If the value increases by ≥15% three times consecutively, the edge gateway immediately issues a cleaning reminder via the local user interface (e.g., an LED indicator turns red, or an app push notification), without relying on cloud-based decision-making, ensuring real-time performance and availability. The entire inference process, tested on an ARM Cortex-M7-based edge device, exhibits a latency of less than 15ms, meeting the real-time requirements of the edge.

[0045] In some optional implementations of this application, after inference is completed locally, incremental gradient data of model parameters is calculated based on the current batch data, and encrypted using AES-256-GCM mode before being cached in local non-volatile storage. The upload strategy is triggered by a composite mechanism, and the conditions may include: (1) the accumulated gradient is equivalent to 1000 hours of running data (approximately 1MB); (2) a sudden change in operating conditions is detected (such as...). (3) Network connection restored. It should be noted that the data uploaded to the cloud is only encrypted incremental gradient data, specifically the weight gradient of the local prediction head and the adjustable physical coefficient gradient of the PINN constraint layer. The weight gradient of the local prediction head is used to reflect the contribution direction of the low-dimensional context representation to the prediction of the remaining effective operating hours. The adjustable physical coefficient gradient (α, β, γ) of the PINN constraint layer is used to optimize the physical modeling accuracy of the coupling terms of inlet concentration, cleaning times and temperature and humidity in the clogging dynamics equation. It does not include the basic physical characteristic indicators of the filter device and the life label of the filter device, so as to balance model updates and privacy protection.

[0046] S3. Deploy the second local PINN-Transformer model in the cloud, receive incremental gradient data from multiple edge nodes, perform distributed training using a hybrid parallel architecture, synchronously update the global PINN-Transformer model, and send the updated model parameter package to each edge node after signing and encryption.

[0047] In some optional implementations of this application, the first local PINN-Transformer model and the second local PINN-Transformer model constitute a global PINN-Transformer model.

[0048] In some optional implementations of this application, the second local PINN-Transformer model deployed in the cloud is the mid-to-late stage layer of the Transformer encoder and the PINN constraint layer. The mid-to-late stage layer of the Transformer encoder is used to receive incremental gradient data from multiple edge nodes, perform deep temporal modeling on the global feature representation after aggregation of multiple edge nodes, and capture the complex nonlinear evolution pattern in the long-term operation of the kitchen air conditioner. The PINN constraint layer is used to embed the attenuation dynamic equation of the filter device as a soft constraint, calculate the physical residual term through automatic differentiation, and introduce a loss function to constrain the model output to conform to physical laws. When the accumulated received incremental gradient data reaches a preset threshold, the incremental learning process is triggered. Combined with the model having a dynamic model update mechanism, the AdamW optimizer is used during training to update only the adjustable physical coefficients of the PINN constraint layer and the weights of the local prediction head, freeze the original parameters of the front-end layer of the Transformer encoder, and package them into a model parameter package.

[0049] In some optional implementations of this application, a complete Transformer encoder's mid-to-late stages and a PINN (Physical Information Neural Network) constraint layer are deployed in the cloud to form the core computational module for high-precision joint training and physical law verification. The mid-to-late stages of the Transformer encoder are responsible for receiving incremental gradient data (encrypted incremental gradient data) uploaded by edge nodes, performing deep temporal modeling on the global feature representation aggregated from multiple nodes, and capturing the complex nonlinear evolution patterns during the long-term operation of the kitchen air conditioner. The PINN constraint layer is used to embed the attenuation dynamic equation of the filtering device as a soft constraint, automatically differentiates to calculate the physical residual term, and introduces it into the loss function to force the model output to conform to physical laws. This architecture supports dynamic adjustment of the physical loss weight during parameter updates, achieving adaptive fusion of data-driven and mechanism-driven approaches, significantly improving the model's generalization ability and prediction stability under extreme conditions.

[0050] In some optional implementations of this application, the second local PINN-Transformer model deployed in the cloud specifically includes the middle and later layers of the Transformer encoder (usually the 5th layer to the end layer) and the complete PINN constraint layer. Its core uses are as follows: (1) Deep temporal modeling and global feature fusion: The cloud receives encrypted incremental gradient data uploaded from multiple edge gateways, and after aggregation by federated averaging (FedAvg), it forms a global update signal; the middle and later layers of the Transformer encoder use its powerful self-attention mechanism to perform deep temporal modeling on the long-term running data after aggregation of multiple nodes, and capture the complex nonlinear evolution pattern of the filter device blockage process of the kitchen air conditioner under different usage habits, such as the rapid dust accumulation caused by periodic stir-frying, the influence of seasonal temperature and humidity changes on the adhesion rate, etc.; (2) Embedding of physical laws and calculation of residuals: According to formula (21), the PINN constraint layer embeds the filter device attenuation dynamic equation as a soft constraint, and formula (21) is: In the formula, This is the proportionality coefficient; Let be the rate of capacity decay of the filter device at time t; Let be the filtration efficiency of the filtration device at time t; This is the cleaning rate coefficient; The cleaning impact factor of the filter device is determined; the physical residual term of the PINN layer is calculated in real time using automatic differentiation technology, and it is introduced into the PINN loss function according to formula (22), which is: In the formula, The filtering efficiency of the filtering device at time t predicted by the neural network; used to constrain the model prediction results to conform to physical laws and improve the generalization of the model in extreme working conditions (such as a sudden increase in oil fumes during stir-frying); (3) Joint optimization and parameter update: The complete backpropagation algorithm is executed in the cloud, and the parameters of the middle and later layers of the Transformer encoder and the PINN constraint layer are jointly optimized. By dynamically adjusting the physical loss weight (based on the residual change rate feedback mechanism), the adaptive fusion of data-driven and mechanism-driven approaches is achieved to avoid overfitting or underfitting. In training, through The model is penalized for predictions that violate the equation, preventing it from operating in scenarios with high levels of cooking fumes (such as PM2.5 > 5 mg / m³). 3 It still maintains reasonable output, and the generalization error is reduced by more than 60%. According to the experimental comparison, the Transformer model without PINN constraint has an error of up to 31.5% in the hot pot scenario, which is reduced to 6.8% after adding PINN.

[0051] In some optional implementations of this application, during model updates, all parameters of the Transformer encoder front-end layer (first four layers) deployed locally on the edge gateway remain frozen and do not participate in any loading or replacement operations. These low-level parameters have been trained and solidified during the initial deployment phase and are mainly used to perform general feature extraction and context encoding, possessing good generalization capabilities. By freezing these parameters, redundant transmission and computational overhead are avoided, significantly reducing bandwidth consumption and flash memory write cycles, while also preventing embedded device storage wear and inference interruptions caused by frequent updates.

[0052] In some optional implementations of this application, a dynamic model update mechanism is used for the model update process. For example, when the accumulated received incremental gradient data (specifically, effective incremental gradient data) is equivalent to 1000 hours of equipment operation data (estimated at 1KB gradient generated per hour), the incremental learning process is triggered. During training, the AdamW optimizer is used, only updating the adjustable physical coefficients of the PINN constraint layer and the weights of the local prediction head, while freezing the original parameters of the Transformer encoder front-end layer. The average time taken is 47 minutes. After the update, the model parameters are encrypted and sent to each edge node.

[0053] In some optional implementations of this application, the communication link between the edge gateway and the cloud also integrates a fault tolerance mechanism, an adaptive width adjustment mechanism, and an encrypted transmission mechanism.

[0054] In some optional implementations of this application, the fault tolerance mechanism involved includes breakpoint resumption and data rollback. Specific implementation methods may include: uploading data in 64KB blocks, with each block appended with a SHA-256 hash value and a sequence number; in one transmission, a total of 17 data blocks need to be uploaded (the last block is 17KB), and the 12th block fails to transmit; after network recovery, the edge gateway resumes from the interruption point (the 12th block), retransmitting only the remaining 6 blocks, saving approximately 65% ​​bandwidth; the cloud maintains an "update transaction log" for each edge node, and when it detects that an incremental gradient data in a device's gradient is greater than... In the event of gradient explosion or device fingerprint mismatch, an automatic rollback operation is triggered to restore the system to the most recent consistent state (such as version v3.2), ensuring the atomicity and consistency of model updates.

[0055] In some optional implementations of this application, the communication process between the edge gateway and the cloud is implemented by a communication module built into the edge gateway and the cloud. This communication module has a built-in wide-sensing engine that monitors uplink throughput and latency in real time. For example, when throughput is detected to be <1Mbps during peak hours (20:00-22:00), the Brotli high-compression algorithm is enabled, achieving a compression ratio of 1:7.8 (compressing the original 1.088MB to 140KB), reducing transmission time from 12 seconds to 1.8 seconds. All uploaded data is encrypted using AES-256-GCM mode, providing confidentiality and integrity protection. The session key is dynamically negotiated and generated by the edge device and the cloud via the ECDH (Curve P-256) protocol, and is rotated every 24 hours to prevent long-term key leakage.

[0056] In some optional implementations of this application, the model parameter package also embeds a device-unique fingerprint.

[0057] In some optional implementations of this application, the updated model parameters are packaged into a lightweight update package, the structure of which consists of a data payload and security metadata: the payload encapsulates the adjustable physical coefficients (α, β, γ) triples of the PINN constraint layer and the weights of the local prediction header; the packet header embeds a unique device fingerprint generated by HMAC-SHA256 using the device hardware ID, public key certificate hash, and registration timestamp, achieving logical binding; the entire packet body is encrypted using the AES-256-GCM algorithm, and the session key is generated through negotiation between the edge and cloud ECDH (P-256 curve) to ensure the confidentiality and integrity of the transmission process and prevent man-in-the-middle attacks and parameter theft. After receiving the model parameter packet, the edge gateway first performs decryption and device fingerprint verification: it parses the packet header fingerprint and compares it with the local hardware fingerprint, and only allows the loading process to continue if there is a complete match. After successful verification, the system writes the received (α, β, γ) values ​​into the dedicated register address space corresponding to the PINN constraint layer, achieving dynamic calibration of the physical law modeling capability. Simultaneously, it replaces the weights of the original local prediction head with the weights of the newly received local prediction head, completing the accuracy optimization of the task output layer. If any verification fails, the model parameter package is discarded and a security audit log is triggered to ensure that the model asset is not illegally reused.

[0058] In some optional implementations of this application, a specific application scenario of the deep learning-based method for predicting the lifespan of a kitchen air conditioner filter is used as an example, such as: (1) Stable operating conditions (regular cooking), PM2.5 concentration is stable (≤1 mg / m3), cleaning frequency is normal (≤5 times / month); the mean absolute percentage error (MAPE) of the fixed window (L=168h) search is 6.2%, and the grid search (optimal) The MAPE of the previous method was 5.1%, while the MAPE of the combined strategy used in this application was 4.9%; the adaptive sliding window mechanism dynamically adjusts the window length according to formula (7), combined with the oil fume concentration factor. Factors affecting the cleaning of kitchen air conditioner filter devices In the stable scenario, the prediction accuracy was further optimized, and the error was reduced by 21% compared with the fixed window. (2) In the scenario of sudden increase in oil fume (stir-frying), PM2.5 increases instantaneously (>5 mg / m3), such as Chinese stir-frying; the prediction error of the Transformer model without PINN constraint is 31.5% (because physical laws are not considered). After adding PINN physical constraints, referring to formula (22), the error is reduced to 6.8%, and the generalization ability is improved by 78.4%. The introduction of physical constraints forces the model to follow the physical laws of blockage, referring to formula (21), to avoid prediction distortion caused by sudden changes in data; the combined strategy further optimizes MAPE to 7.3% in the stir-frying scenario (refer to Table 1 below), which is better than grid search (optimal). (3) In high-frequency cleaning scenarios, frequent cleaning of the filter (>10 times / month) leads to damage to the filter's microstructure; the MAPE of the fixed window (L=168h) search is 24.8%, and the grid search (optimal) The MAPE of the single-mode cleaning method was 8.6%, while that of the combined strategy was 5.4%. By quantifying the impact of cleaning cycles on the capacity decay rate, the model accurately captures the damage effect caused by cleaning. The combined strategy reduced the error by 78.2% compared to the fixed window method.

[0059] Table 1. Comparison of experimental data for fixed window, grid search, and combined strategies.

[0060]

[0061] As shown in Table 1, the combined strategy reduces the error by more than 50% in extreme scenarios, proving its effectiveness in balancing physical laws and data-driven approaches.

[0062] In some optional implementations of this application, the effectiveness of the edge-cloud collaborative framework is also verified. A dynamic model update mechanism is adopted. When the cumulative received incremental gradient data (specifically referring to effective incremental gradient data) is equivalent to 1000 hours of device operation data (estimated at 1KB gradient generated per hour), the incremental learning process is triggered. Only the adjustable physical coefficients of the PINN constraint layer and the weights of the local prediction head are updated, and the original parameters of the Transformer encoder front-end layer are frozen. The cloud training efficiency is improved by 40% (the hybrid parallel architecture supports multi-GPU synchronous gradient).

[0063] In some optional implementations of this application, the accuracy of predicting the lifespan of a kitchen air conditioner filter device is compared between Transformer models without PINN constraints and Transformer models with PINN constraints (test set MAE). Introducing PINN physical constraints into the Transformer model can significantly improve the generalization ability under extreme working conditions, and the error in the stir-frying scenario and the high-frequency cleaning scenario is reduced by more than 75%. The simulation comparison results are shown in Table 2.

[0064] Table 2 Simulation Comparison Results

[0065]

[0066] As shown in Table 2, the physical residual of the model decreased from 0.38 at the beginning of training to 0.02. The Transformer model without PINN constraint showed a problem of "anti-physical" rebound in prediction efficiency (efficiency exceeding 100% after cleaning) after t exceeded 120 hours. In contrast, the Transformer model with PINN constraint strictly followed the reasonable range of prediction results ∈ [65%, 98%].

[0067] Example 2

[0068] This embodiment, based on Embodiment 1 above, provides a deep learning-based system for predicting the lifespan of a kitchen air conditioner filter. Please refer to [link to previous document]. Figures 2-5This system is used to implement the deep learning-based method for predicting the lifespan of a kitchen air conditioner filter device as described in Embodiment 1 above. The system is mainly used for the cloud and multiple edge nodes that interact with the cloud. The edge nodes include edge devices and edge gateways that interact with the edge devices. The edge devices include a data acquisition module and a feature extraction module. The edge gateways include a local processing module and a cache upload module. The cloud includes a cloud-distributed model training module and a parameter distribution module. The data acquisition module is used to collect real-time operating parameters of the kitchen air conditioner; the feature extraction module is used to extract basic physical characteristic indicators of the filtration device and send these indicators to the edge gateway; the local processing module is used to deploy a first local PINN-Transformer model and perform data preprocessing and local lifetime prediction inference; the cache upload module is used to cache incremental gradient data and upload it to the cloud based on a security and fault tolerance mechanism; the cloud-based distributed model training module is used to deploy a second local PINN-Transformer model, receive incremental gradient data from multiple edge nodes, perform distributed training using a hybrid parallel architecture, and synchronously update the global PINN-Transformer model; the parameter distribution module is used to distribute the updated model parameter package to each edge node after signing and encryption; wherein, the first local PINN-Transformer model and the second local PINN-Transformer model constitute the global PINN-Transformer model.

[0069] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0070] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including a data acquisition unit, a rule determination unit, a weight calculation unit, and an anomaly determination unit. The names of these units do not necessarily limit the specific unit; for example, a data acquisition unit may also be described as a "unit for acquiring training data."

[0071] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0072] Obviously, those skilled in the art will understand that the various steps of the present invention described above can be performed in a manner different from that described above, and the simulation methods and experimental equipment include, but are not limited to, the above description. The steps of the present invention described above can be performed in a different order in certain circumstances, and the steps shown or described above can be performed separately. Therefore, the present invention is not limited to any particular combination of hardware and software.

[0073] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the life of a filter device of a kitchen air conditioner based on deep learning, characterized by, A plurality of edge nodes for and interacting with a cloud, the edge nodes including edge devices and edge gateways interacting with the edge devices, the method comprising: The edge device collects real-time kitchen air conditioner operating parameters, extracts filter device basic physical characteristic indicators, and sends the filter device basic physical characteristic indicators to the edge gateway; The edge gateway deploys a first local PINN-Transformer model, performs data preprocessing and local life prediction inference, caches incremental gradient data, and uploads to the cloud based on a secure fault-tolerant mechanism; The first local PINN-Transformer model deployed by the edge gateway includes a Transformer encoder front-end layer, which is used to perform data preprocessing, including: Performing linear interpolation on missing values, isolated forest anomaly detection, spline interpolation repair, and injecting Gaussian noise on the filter device basic physical characteristic indicators; The physical self-adaptive strategy is used to optimize the window length, and an adaptive sliding window is obtained, and a formula is as follows: , wherein k is an adjustment amplitude coefficient, and the value is 0.5; , is PM2.5 in the kitchen at time t, is the maximum value of PM2.5 in the kitchen, is a real-time oil fume concentration normalized value, that is, an oil fume concentration factor, is a kitchen air conditioner filter device cleaning influence factor, , is a cleaning damage coefficient, which is influenced by a kitchen air conditioner filter device cleaning method, is a cumulative cleaning frequency, is an optimized window length; Constructing multi-dimensional time series features based on an adaptive sliding window and performing normalization processing; Inputting the normalized multi-dimensional time series features into the Transformer encoder front-end layer, extracting key time series dependencies through a multi-head self-attention mechanism, and outputting high-level hidden state sequences; Performing time dimension average pooling on the high-level hidden state sequences, and compressing them to filter device life label annotations through channel grouping and sparse regularization; The cloud deploys a second local PINN-Transformer model, receives incremental gradient data from multiple edge nodes, adopts a hybrid parallel architecture for distributed training, synchronously updates a global PINN-Transformer model, and issues the updated model parameters to each edge node after signature encryption; The second local PINN-Transformer model deployed by the cloud includes a PINN constraint layer, which is used to embed filter device attenuation dynamics equations as soft constraints, calculate physical residual terms through automatic differentiation, and introduce a loss function to constrain the model output to comply with physical laws; Wherein, the first local PINN-Transformer model and the second local PINN-Transformer model constitute a global PINN-Transformer model. 2.The deep learning-based kitchen air conditioner filter device lifespan prediction method of claim 1, wherein The real-time kitchen air conditioner operating parameters include the pressure difference on both sides of the filter device of the kitchen air conditioner, the air flow rate at the return air inlet, the air flow rate at the air outlet, the air temperature at the air outlet, the air humidity at the air outlet, the temperature at the return air inlet, the humidity at the return air inlet, PM2.5 in the kitchen, VOCs concentration in the kitchen, particle concentration at the corresponding return air inlet before the filter device of the kitchen air conditioner, and particle concentration in the corresponding return air inlet after the filter device of the kitchen air conditioner. 3.The deep learning-based kitchen air conditioner filter device lifespan prediction method of claim 1, wherein The edge device extracts filter device basic physical characteristic indicators based on a sliding time window, including original filtration efficiency, original capacity attenuation rate, cleaning frequency, and cleaning time. 4.The deep learning-based kitchen air conditioner filter device lifespan prediction method of claim 1, wherein The Transformer encoder front-end layer includes missing value linear interpolation, outlier detection, interpolation repair, processing of multi-dimensional time series feature input constructed by sliding window, generation of filter device life label annotation, and the first local PINN-Transformer model deployed by the edge gateway further includes a local prediction head, which is used to perform local life prediction inference according to the filter device life label annotation and output a filter device remaining life prediction value. 5.The deep learning-based kitchen air conditioner filter device life prediction method of claim 1, wherein The local life prediction inference includes: The filter device life label annotation output by the Transformer is input into the local prediction head, and a filter device remaining life prediction value is output. 6.The deep learning-based kitchen air conditioner filter device life prediction method of claim 1, wherein The second local PINN-Transformer model deployed by the cloud further includes a middle and rear stage layer of the Transformer encoder, which is used to receive incremental gradient data from multiple edge nodes, perform deep time series modeling on global feature representations aggregated by the multiple edge nodes, capture complex nonlinear evolution patterns in long-term operation of the kitchen air conditioner, and trigger an incremental learning process when the cumulative received incremental gradient data reaches a preset threshold, combine the model with a dynamic model updating mechanism, use the AdamW optimizer during training, only update the adjustable physical coefficients of the PINN constraint layer and the weights of the local prediction head, freeze the original parameters of the Transformer encoder front-end layer, and pack them into a model parameter package. 7.The deep learning-based kitchen air-conditioner filter device life prediction method of claim 1, wherein The communication link between the edge gateway and the cloud further integrates a fault-tolerant mechanism, an adaptive width adjustment mechanism, and an encrypted transmission mechanism. 8.The deep learning-based kitchen air conditioner filter device life prediction method of claim 1, wherein The model parameter package further embeds a device unique fingerprint.

9. A deep learning-based system for predicting the life of a filter device of a kitchen air conditioner, characterized by, The cloud and multiple edge nodes interacting with the cloud, the edge nodes including edge devices and edge gateways interacting with the edge devices, the edge devices including data acquisition modules and feature extraction modules, the edge gateways including local processing modules and cache uploading modules, and the cloud including a cloud distributed model training module and a parameter issuing module; The data acquisition module is used to acquire real-time kitchen air conditioner operating parameters; The feature extraction module is used to extract filter device basic physical feature indicators and send the filter device basic physical feature indicators to the edge gateway; The local processing module is used to deploy a first local PINN-Transformer model and perform data preprocessing and local life prediction inference; The first local PINN-Transformer model includes a Transformer encoder front-end layer, which is used to perform data preprocessing, and the data preprocessing includes: Performing linear interpolation of missing values, isolated forest anomaly detection, spline interpolation repair, and injection of Gaussian noise on filter device basic physical feature indicators; The physical self-adaptive strategy is used to optimize the window length, and an adaptive sliding window is obtained, and a formula is as follows: , wherein k is an adjustment amplitude coefficient, and the value is 0.5; , is PM2.5 in the kitchen at time t, is the maximum value of PM2.5 in the kitchen, is a real-time oil fume concentration normalized value, that is, an oil fume concentration factor, is a kitchen air conditioner filter device cleaning influence factor, , is a cleaning damage coefficient, which is influenced by a kitchen air conditioner filter device cleaning method, is a cumulative cleaning frequency, is an optimized window length; Constructing multi-dimensional time series features based on adaptive sliding windows and performing normalization processing; The normalized multi-dimensional time sequence features are input into a front-end layer of a Transformer encoder, key time sequence dependencies are extracted through a multi-head self-attention mechanism, and a high-level hidden state sequence is output; The high-level hidden state sequence is subjected to time dimension average pooling, and is compressed to a filter device life label through channel grouping and sparse regularization; The cache upload module is configured to cache the incremental gradient data and upload the incremental gradient data to the cloud based on a safety fault-tolerant mechanism; The cloud distributed model training module is configured to deploy a second local PINN-Transformer model, receive the incremental gradient data from the multiple edge nodes, perform distributed training using a hybrid parallel architecture, and synchronously update a global PINN-Transformer model; The second local PINN-Transformer model deployed in the cloud includes a PINN constraint layer configured to embed a filter device attenuation dynamics equation as a soft constraint, calculate a physical residual term through automatic differentiation, and introduce a loss function to constrain the model output to comply with physical laws; The parameter delivery module is configured to deliver the updated model parameters to the edge nodes after signature encryption; The first local PINN-Transformer model and the second local PINN-Transformer model constitute the global PINN-Transformer model.

Images