Multi-unit pressure sensor self-calibration method and system based on federal learning

Abstract

The invention provides a multi-unit pressure sensor self-calibration method and system based on federated learning. The method comprises the steps that a multi-unit sensing node is located in a pre-built multi-unit sensing node-edge end-cloud layered star topology sensing network; the multi-unit sensing node carries out pressure compensation under environment calibration on actually measured pressure sensing data according to environment variable data through a pre-constructed pressure self-calibration model to generate self-calibration pressure data, the pressure self-calibration model is obtained by carrying out iterative training based on a federated learning framework, Byzantine fault tolerance and gradient cutting are realized, and the self-calibration pressure data is generated. The model convergence stability is ensured, and the system complexity and maintenance cost are reduced, so that the universality is improved; the problem of low cooperation efficiency of heterogeneous equipment is relieved, model convergence is accelerated, the resource utilization rate is improved, and for pressure sensor data heterogeneity, the calculation load is reduced, the communication efficiency is ensured, and the dynamic environment adaptability and calibration result accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, and in particular to a self-calibration method and system for multi-unit pressure sensors based on federated learning. Background Technology

[0002] Traditional calibration relies on offline calibration and external pressure sources, requiring manual intervention to establish input-output models. However, disassembly-based calibration cannot meet the demands of harsh operating conditions such as high temperatures and vibrations, exhibiting poor environmental adaptability. Furthermore, linear fitting struggles to suppress nonlinear errors and time-drift effects, resulting in insufficient long-term stability. Additionally, multiple sensors require individual manual calibration, leading to high costs. To overcome environmental limitations, researchers have proposed on-chip integrated self-calibration schemes, such as integrating temperature-controlled resistors into silicon carbide pressure-sensitive films. These schemes utilize the difference in thermal expansion between heterogeneous materials to simulate pressure input through electrothermal effects, achieving in-situ online calibration. However, calibration accuracy is constrained by the stability of the material's thermal properties, and rapid temperature changes easily introduce nonlinear errors, leading to inaccurate calibration results. Multimodal temperature compensation constructs temperature change models containing polynomial, periodic, and exponential terms, introducing dynamic weights and recursive optimization to achieve adaptive correction in complex temperature change scenarios. While these algorithms improve adaptability, they have stringent requirements for data quality and computing power, are limited in low-power scenarios, and suffer from high system complexity and maintenance costs, resulting in poor universality. Summary of the Invention

[0003] One objective of this invention is to provide a self-calibration method for multi-unit pressure sensors based on federated learning, achieving Byzantine fault tolerance and gradient pruning, ensuring model convergence stability, reducing system complexity and maintenance costs, thereby improving universality; alleviating the problem of low collaboration efficiency of heterogeneous devices, accelerating model convergence and improving resource utilization; reducing computational load for the heterogeneity of pressure sensor data, ensuring communication efficiency, and improving dynamic environment adaptability and calibration result accuracy. Another objective of this invention is to provide a self-calibration system for multi-unit pressure sensors based on federated learning. A further objective of this invention is to provide a computer-readable medium. A final objective of this invention is to provide a computer device.

[0004] To achieve the above objectives, this invention discloses a self-calibration method for multi-unit pressure sensors based on federated learning, comprising:

[0005] Multi-unit sensor nodes collect measured pressure sensing data and environmental variable data. The multi-unit sensor nodes are located in a pre-built multi-unit sensor node-edge-cloud hierarchical star topology sensor network.

[0006] Multi-unit sensing nodes use a pre-built pressure self-calibration model to perform pressure compensation under environmental calibration on measured pressure sensing data based on environmental variable data, generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework.

[0007] Preferably, the multi-unit sensing node includes a homogeneous redundant architecture and a heterogeneous fusion architecture;

[0008] The homogeneous redundancy architecture uses hardware cold backup and is configured with multiple pressure sensors of the same model.

[0009] The heterogeneous fusion architecture integrates multiple different types of pressure sensors and environmental sensors.

[0010] Preferably, the method further includes:

[0011] Based on the hierarchical star topology sensor network and multi-unit sensor node attribute data sent from the edge, the cloud generates a node scheduling list and determines the initial global model.

[0012] The edge device sends the initial global model sent by the cloud to the multi-unit sensor nodes according to the node scheduling list sent by the cloud.

[0013] The multi-unit sensing node trains the initial global model locally based on the local training dataset, generates an intermediate training model and encrypted model parameters, and sends the encrypted model parameters to the edge.

[0014] The edge and cloud perform hierarchical model aggregation based on the encrypted model parameters to build and update the training model;

[0015] The cloud calculates the contribution of multi-unit sensor nodes, generates an updated node scheduling list based on the contribution, and sends the updated node scheduling list to the edge. The edge sends iterative training messages to the multi-unit sensor nodes according to the updated node scheduling list, so that the multi-unit sensor nodes can perform iterative model training until a pressure self-calibration model is generated.

[0016] Preferably, the encryption model parameters include feature vectors and encryption model gradients;

[0017] The edge and cloud perform hierarchical model aggregation based on the encrypted model parameters to build and update the training model, including:

[0018] At the edge, a clustering algorithm is used to aggregate multi-unit sensing nodes based on feature vectors to generate initial grouping results;

[0019] At the edge, local sensitive hashing and security filtering algorithms are used to filter abnormal nodes in the initial grouping results based on the gradient of the encryption model, generate a set of candidate parameters, and upload the set of candidate parameters to the cloud.

[0020] The cloud uses a robust aggregation algorithm to aggregate the candidate parameter set coordinate by coordinate, filter out outliers in each dimension of the model parameters, and generate updated model parameters.

[0021] The cloud updates the initial global model according to the updated model parameters, generating an updated training model.

[0022] Preferably, before sending the updated node scheduling list to the edge, the method further includes:

[0023] The cloud uses a constructed validation dataset to validate the updated training model and generate model validation metrics.

[0024] The cloud platform generates a comprehensive model score based on model validation metrics.

[0025] The cloud determines whether the model's overall score is higher than the average score of the model in a specified historical consecutive rounds;

[0026] If not, continue with the step of sending the updated node scheduling list to the edge.

[0027] Preferably, the method further includes:

[0028] If so, the cloud will send the updated training model and the updated node scheduling list to the edge.

[0029] The edge device sends the updated training model to the multi-unit sensor node according to the updated node scheduling list, so that the multi-unit sensor node can iteratively train the updated training model until a pressure self-calibration model is generated.

[0030] Preferably, the method further includes:

[0031] Training log files are stored using blockchain technology.

[0032] This invention also discloses a self-calibration system for multi-unit pressure sensors based on federated learning, comprising:

[0033] The data acquisition unit is used to collect measured pressure sensing data and environmental variable data from the multi-unit sensing nodes. The multi-unit sensing nodes are located in a pre-built hierarchical star topology sensing network of multi-unit sensing nodes-edge-cloud.

[0034] The pressure self-calibration unit is used by multi-unit sensing nodes to perform pressure compensation under environmental calibration on measured pressure sensing data based on environmental variable data through a pre-built pressure self-calibration model, generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework.

[0035] The present invention also discloses a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0036] The present invention also discloses a computer device, including a memory and a processor, wherein the memory is used to store information including program instructions, and the processor is used to control the execution of the program instructions, wherein the processor executes the program to implement the method described above.

[0037] The present invention also discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the method described above.

[0038] This invention utilizes multi-unit sensor nodes to collect measured pressure sensing data and environmental variable data. These nodes are located in a pre-built hierarchical star topology sensor network consisting of multi-unit sensor nodes, edge devices, and the cloud. The multi-unit sensor nodes, through a pre-constructed pressure self-calibration model, perform pressure compensation under environmental calibration on the measured pressure sensing data based on the environmental variable data, generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework, achieving Byzantine fault tolerance and gradient pruning, ensuring model convergence stability, reducing system complexity and maintenance costs, thereby improving universality. It also alleviates the problem of low collaborative efficiency of heterogeneous devices, accelerates model convergence and improves resource utilization. Furthermore, it addresses the heterogeneity of pressure sensor data, reducing computational load, ensuring communication efficiency, and improving dynamic environmental adaptability and calibration result accuracy. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of a single-wafer etching and packaging technology for a multi-unit pressure sensor provided in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the structure of a multi-unit pressure sensor based on system-in-package technology, provided in an embodiment of the present invention.

[0042] Figure 3 This invention provides a schematic diagram of the hierarchical structure of a multi-unit pressure sensor self-calibration system based on federated learning, as shown in an embodiment of the invention.

[0043] Figure 4A schematic diagram of the topology network of a multi-unit pressure sensor self-calibration system based on federated learning is provided for an embodiment of the present invention.

[0044] Figure 5 A flowchart illustrating a self-calibration method for a multi-unit pressure sensor based on federated learning, provided in an embodiment of the present invention;

[0045] Figure 6 An interactive diagram illustrating another self-calibration method for multi-unit pressure sensors based on federated learning provided in this embodiment of the invention;

[0046] Figure 7 A schematic diagram of a multi-unit pressure sensor self-calibration system based on federated learning provided in an embodiment of the present invention;

[0047] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0048] 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.

[0049] To facilitate understanding of the technical solution provided in this application, the relevant content of the technical solution will be explained below. Federated learning provides a new path for distributed calibration. By training local models and aggregating encrypted parameters, it protects data privacy while integrating global features to optimize accuracy. It fully utilizes the characteristics and value of sensor data under actual working conditions, establishes a nonlinear mapping compensation model of environmental variables and stress response, and realizes a mature multi-physics coupling compensation mechanism, which has great potential in large-scale heterogeneous sensor networks. This application significantly improves data security, communication efficiency, model convergence speed, and system scalability by integrating technologies such as dynamic scheduling, secure screening, accurate contribution quantification, and efficient communication. It is suitable for large-scale heterogeneous sensor networks and has high engineering practical value.

[0050] Figure 1 This is a schematic diagram of a single-wafer etching and packaging technology for a multi-unit pressure sensor provided in an embodiment of the present invention, as shown below. Figure 1As shown, high integration is achieved by etching a multi-sensor structure array on a single wafer. Using MEMS bulk silicon processing and deep reactive ion etching (DRIE) technology, different sensing units such as piezoresistive, capacitive, and resonant sensors are integrated onto the same silicon chip with micron-level spacing, achieving physical isolation and signal independence between the sensing units. The sensitive structures are packaged using wafer-level bonding technology, combined with on-chip temperature and humidity compensation circuitry, enabling direct output of environmentally calibrated pressure data. Simultaneously, photolithography ensures the consistency of the sensing units, and wafer-level testing and screening improve yield. Its core advantages lie in ultra-high integration density and signal synchronization, but it relies on advanced photolithography equipment and a complex process chain, resulting in high mass production costs and process complexity.

[0051] Figure 2 This is a schematic diagram of a multi-unit pressure sensor based on system-level packaging technology, provided as an embodiment of the present invention. Figure 2 As shown, a multi-unit pressure sensor architecture is integrated through three-dimensional stacking and heterogeneous integration based on System-in-Package (SIP) technology. In the homogeneous redundancy architecture, multiple identical microelectromechanical system (MEMS) pressure sensor chips are vertically integrated onto a ceramic substrate using through-silicon via (TSV) technology, coupled with independent signal conditioning circuits and parallel power supply design to achieve single-point fault isolation and functional redundancy. In the heterogeneous fusion architecture, SIP technology integrates piezoresistive, resonant, and fiber optic sensor chips and temperature compensation chips on a stacked structure using gold-tin (Au-Sn) eutectic bonding, and multi-source data fusion is used to meet accuracy requirements. SIP packaging offers significant space efficiency advantages, and hermetic sealing and rewiring technologies improve environmental adaptability and mass production economics.

[0052] Circuit board-level multi-unit pressure sensor technology enables flexible deployment through modular design and distributed architecture. In the homogeneous redundancy architecture, two or more pressure sensor arrays are soldered onto a printed circuit board (PCB). Dynamic load balancing algorithms reduce the daily operating time of a single sensor, thus lowering the rate of sudden failures. The heterogeneous fusion architecture integrates two or more capacitive, piezoelectric, and resonant sensors on a multi-layer epoxy fiberglass board (FR4) substrate, combined with a multi-modal application-specific integrated circuit (ASIC) and a digital humidity sensor. A three-dimensional temperature field compensation model further reduces nonlinear errors. Modular replacement is supported, providing ample expansion margins. Costs are significantly lower than SIP (System-in-Package), and distributed power supply ensures minimal power consumption fluctuations during single-point failures, making it suitable for scenarios requiring high economic efficiency and maintainability.

[0053] From a process feasibility perspective, wafer-level technology, based on the integration of complementary metal-oxide-semiconductor (CMOS) and microelectromechanical systems (MEMS), achieves monolithic integration through wafer-level bonding, resulting in low mass production costs, but is limited by multi-sensor collaborative design. SiP technology relies on advanced packaging processes such as TSV 3D stacking and Au-Sn eutectic bonding, but requires equipment investment and is suitable for high-value-added scenarios. Circuit board-level technology uses standardized surface mount technology (SMT), compatible with FR4 substrates and surface mount bonding, but needs to address thermal stress and mechanical deformation issues under dense layouts. Employing different packaging levels, from wafer-level to system-level to circuit board-level, with progressively larger dimensions, forms a high-medium-low spectrum process coverage, allowing selection based on application requirements.

[0054] Figure 3 This invention provides a schematic diagram of the hierarchical structure of a multi-unit pressure sensor self-calibration system based on federated learning, as shown in the embodiment of the invention. Figure 3 As shown, the system hierarchy includes a hardware layer, a communication layer, and a software layer. The hardware layer employs a homogeneous redundancy and heterogeneous fusion architecture, integrating multi-principle sensors and environmental detection units. Through wafer-level etching, system-level packaging, and modular design, it achieves full-spectrum process coverage, improving data diversity and reliability. The communication layer combines encryption technology, adaptive compression, and data verification mechanisms to reduce transmission load and ensure security. The software layer constructs a multi-unit sensor node-edge-cloud hierarchical federated learning framework. Through dynamic scheduling, secure filtering, robust aggregation, and contribution metric algorithms, it achieves fault tolerance for abnormal nodes, gradient optimization, and collaborative resource allocation.

[0055] The hardware layer comprises clients, edge devices, and cloud services. Clients are pressure sensor nodes, edge devices are edge nodes, and cloud services are cloud servers. The hardware layer distributes the computing load through collaborative work among pressure sensor nodes, edge nodes, and cloud servers: low-power devices perform local training, edge nodes complete intermediate aggregation, and cloud servers perform global optimization, balancing resource efficiency and cost-effectiveness.

[0056] The communication layer runs through the data transmission process between multi-unit sensor nodes and the edge, and between the edge and the cloud, providing communication support. This layer includes privacy protection, secure transmission, and communication efficiency. It combines narrowband and high-speed links, and introduces data encryption and adaptive compression technologies to reduce communication overhead in low-bandwidth scenarios while ensuring transmission security, making large-scale device access feasible. Privacy protection uses encryption technology to distribute keys and update encryption gradients to ensure communication confidentiality; adaptive compression dynamically adjusts the compression ratio of model updates through gradient pruning and quantization to reduce communication volume; and the data verification mechanism uses an end-to-end data verification and confirmation mechanism with updates accompanied by hash checksums or digital signatures to ensure the integrity and correctness of transmission.

[0057] The software layer includes data cleaning, local training, dynamic grouping, security screening, robust aggregation, contribution quantification, parameter tuning, and subsequent training. The software layer integrates modular algorithms to achieve end-to-end closed-loop optimization: data cleaning and security screening ensure data quality, the robust aggregation (BULYAN) algorithm defends against malicious attacks, and contribution quantification and intelligent scheduling dynamically allocate resources to accelerate model convergence.

[0058] Figure 4 A schematic diagram of the topology network of a multi-unit pressure sensor self-calibration system based on federated learning, provided as an embodiment of the present invention, is shown below. Figure 4 As shown, this topology employs a star-shaped sensor network consisting of a cloud (100), edge (200), and multi-unit sensor nodes (300). High-efficiency collaboration is achieved through centralized global aggregation at the cloud (100) and regional management of the multi-unit sensor nodes (300) at the edge (200). The multi-unit sensor nodes (300) communicate with the nearest edge (200) via low-power links, reducing transmission latency and energy consumption. After local aggregation and security screening, the edge (200) interacts with the cloud (100) via high-speed links, reducing the direct load on the cloud (100). This structure supports large-scale device access and optimizes resource allocation through dynamic grouping and intelligent scheduling, balancing security and scalability.

[0059] The cloud (100) and edge (200) communication connections, and the edge (200) and multi-unit sensor nodes (300) communication connections, provide communication support during data transmission. To prevent information leakage or interference during transmission, the system employs quantum encryption technology for key distribution and encrypted gradient updates. Quantum key distribution generates unconditionally secure keys based on information theory between the communicating parties, allowing for the detection of any eavesdropping. Furthermore, for narrowband networks such as LoRa / NB-IoT, the compression ratio of model updates is dynamically adjusted based on network bandwidth. In federated learning, model gradients / parameters often have extremely high dimensionality, making direct transmission prohibitively expensive; gradient pruning and quantization compression can significantly reduce communication overhead. Under high-speed 5G links, the compression ratio can be appropriately reduced to ensure accuracy. Difference compensation before and after compression is used to minimize compression errors. Simultaneously, ensuring the complete and correct arrival of gradient update data packets is crucial in multi-hop transmission and unreliable network environments. Each uploaded update is accompanied by a hash checksum or digital signature. Upon receipt, the cloud (100) and edge (200) calculate and compare the hashes to verify data integrity, preventing undetected data tampering or packet loss during transmission. A fragmented retransmission protocol is established to split model updates into multiple fragments, which are sent with numbered segments. If a fragment is lost, only that fragment is retransmitted, thereby improving reliability.

[0060] In the technical solution provided by this invention, multi-unit sensing nodes collect measured pressure sensing data and environmental variable data. These nodes are located in a pre-built hierarchical star-topology sensing network consisting of multi-unit sensing nodes, edge devices, and the cloud. The multi-unit sensing nodes use a pre-constructed pressure self-calibration model to perform pressure compensation under environmental calibration on the measured pressure sensing data based on the environmental variable data, generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework, achieving Byzantine fault tolerance and gradient pruning, ensuring model convergence stability, reducing system complexity and maintenance costs, thereby improving universality. This alleviates the problem of low collaborative efficiency of heterogeneous devices, accelerates model convergence and improves resource utilization. Addressing the heterogeneity of pressure sensor data, it reduces computational load, ensures communication efficiency, and improves dynamic environmental adaptability and calibration result accuracy.

[0061] It is worth noting that, Figure 3 and 4 The hierarchical structure and topology of the federated learning-based multi-unit pressure sensor self-calibration system shown are also applicable to Figure 5 or Figure 6 The self-calibration method for multi-unit pressure sensors based on federated learning will not be elaborated here.

[0062] The following example uses a multi-unit pressure sensor self-calibration system based on federated learning as the execution subject to illustrate the implementation process of the multi-unit pressure sensor self-calibration method based on federated learning provided in this embodiment of the invention. It is understood that the execution subject of the multi-unit pressure sensor self-calibration method based on federated learning provided in this embodiment of the invention includes, but is not limited to, a multi-unit pressure sensor self-calibration system based on federated learning.

[0063] Figure 5 A flowchart of a self-calibration method for a multi-unit pressure sensor based on federated learning provided in an embodiment of the present invention is shown below. Figure 5 As shown, the method includes:

[0064] Step 101: Collect measured pressure sensing data and environmental variable data from multi-unit sensing nodes.

[0065] In this embodiment of the invention, the multi-unit sensing nodes are located in a pre-built hierarchical star topology sensing network of multi-unit sensing nodes-edge-cloud. The multi-unit sensing nodes include homogeneous redundancy architecture and heterogeneous fusion architecture. The homogeneous redundancy architecture uses hardware cold backup, configuring multiple pressure sensors of the same model to achieve repeated measurements of the same physical quantity, improving system reliability and fault tolerance. The heterogeneous fusion architecture integrates multiple pressure sensors and environmental sensors of different models, enabling multi-parameter acquisition, thereby improving measurement accuracy, adaptability, and environmental robustness.

[0066] In this embodiment of the invention, the environmental sensors include, but are not limited to, temperature sensors and humidity sensors. Specifically, pressure sensor data is obtained through actual measurement and acquisition; environmental sensor data is obtained through actual measurement and acquisition.

[0067] Furthermore, from the perspective of process feasibility, the goal is to achieve full-spectrum process coverage of integrated multi-unit sensing nodes. At the wafer level, high integration is achieved by etching multi-sensor structure arrays on a single wafer. At the die level, based on system-in-package technology, the architecture integration of multi-unit pressure sensors is achieved through three-dimensional stacking and heterogeneous integration. At the device level, integrated sensing units are constructed through modular design and distributed architecture.

[0068] Step 102: The multi-unit sensing node uses a pre-built pressure self-calibration model to perform pressure compensation under environmental calibration on the measured pressure sensing data based on environmental variable data, generating self-calibrated pressure data.

[0069] In this embodiment of the invention, the pressure self-calibration model is obtained through iterative training based on a federated learning framework.

[0070] Specifically, environmental variable data is input into the measured pressure sensing data to perform pressure compensation under environmental calibration in the pressure self-calibration model, and self-calibrated pressure data is output.

[0071] In the technical solution provided by this invention, multi-unit sensing nodes collect measured pressure sensing data and environmental variable data. These nodes are located in a pre-built hierarchical star-topology sensing network consisting of multi-unit sensing nodes, edge devices, and the cloud. The multi-unit sensing nodes use a pre-constructed pressure self-calibration model to perform pressure compensation under environmental calibration on the measured pressure sensing data based on the environmental variable data, generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework, achieving Byzantine fault tolerance and gradient pruning, ensuring model convergence stability, reducing system complexity and maintenance costs, thereby improving universality. This alleviates the problem of low collaborative efficiency of heterogeneous devices, accelerates model convergence and improves resource utilization. Addressing the heterogeneity of pressure sensor data, it reduces computational load, ensures communication efficiency, and improves dynamic environmental adaptability and calibration result accuracy.

[0072] Figure 6 An interaction diagram for another self-calibration method for multi-unit pressure sensors based on federated learning provided in this embodiment of the invention is shown below. Figure 6 As shown, the method includes:

[0073] Step 201: The cloud generates a node scheduling list and determines the initial global model based on the hierarchical star topology sensor network and multi-unit sensor node attribute data sent by the edge.

[0074] In this embodiment of the invention, each step is performed by a multi-unit pressure sensor self-calibration system based on federated learning.

[0075] Specifically, multi-unit sensor nodes report their own attribute data to the edge; the edge counts the number of its own multi-unit sensor nodes and, based on the generated hierarchical star topology sensor network, uploads the multi-unit sensor node attribute data, the number of its own multi-unit sensor nodes, and the hierarchical star topology sensor network to the cloud, providing a basis for the overall training plan; the cloud generates a node scheduling list based on the multi-unit sensor node attribute data, the number of its own multi-unit sensor nodes, and the hierarchical star topology sensor network according to preset screening conditions, and determines the initial global model.

[0076] The attribute data includes, but is not limited to, online status, resource metrics, and data status. Resource metrics include, but are not limited to, CPU / GPU utilization, memory usage, channel idle rate, battery-powered device ratio, and communication efficiency metrics. Data status includes, but is not limited to, data integrity and data volume.

[0077] In this embodiment of the invention, the screening criteria are based on multi-unit sensor nodes that are willing to participate and do not interfere with the daily tasks of the sensors. "Does not interfere with the daily tasks of the sensors" means that daily tasks have absolute priority, and federated training only utilizes idle sensor resources and communication redundancy periods. As an optional scheme, the screening criteria are: online status (online); resource indicators (CPU utilization less than 30% or memory usage less than 20%, battery-powered devices greater than 50%); communication efficiency indicators (stable network); priority upload of daily data; federated updates only uploaded on preset idle channels, such as during a certain period at night, or when bandwidth is redundant and the channel idle rate is greater than 80%; and data status (complete data). "Willing to participate" means that the multi-unit sensor nodes are willing to accept fairness principles, such as the implementation of reward and punishment mechanisms and resource usage. Simultaneously, data weights are initialized according to the sensitivity specified in the pressure sensor datasheet.

[0078] In this embodiment of the invention, the cloud selects multi-unit sensing nodes that do not interfere with the daily tasks of the sensors and are willing to participate based on screening criteria; it generates initialization priorities for the selected multi-unit sensing nodes based on the calibrated sensitivity initialization data weights, and selects multi-unit sensing nodes with specified priorities as multi-unit sensing nodes to participate in training according to the initialization priorities; it selects edge terminals to participate in training based on the number of multi-unit sensing nodes of the edge terminals themselves according to actual needs; and it obtains a node scheduling list based on the selected edge terminals to participate in training, the selected multi-unit sensing nodes, and their priorities.

[0079] In this embodiment of the invention, the cloud determines an initial global model based on preset user requirements, resource conditions, and attribute data of multi-unit sensor nodes. User requirements include data quality requirements, model performance requirements, and task objectives. Data quality requirements include, but are not limited to, sensor accuracy and real-time performance; model performance requirements include, but are not limited to, target accuracy and convergence speed; and task objectives include, but are not limited to, contribution metric methods and fairness guarantees. Resource conditions include, but are not limited to, communication resources, computing resources, and energy and network stability. Communication resources include, but are not limited to, bandwidth and latency metrics; computing resources include, but are not limited to, multi-unit sensor node computing power metrics, edge computing power metrics, and cloud computing power metrics; and energy and network stability include, but are not limited to, power supply metrics and network fluctuation metrics.

[0080] Specifically, by calculating the variance ratio of a specified indicator, if the variance ratio is less than a preset variance ratio threshold, it indicates that the data has a high degree of linearity, and a linear regression model is selected as the initial global model; if the variance ratio is greater than the preset variance ratio threshold, it indicates that the data has a high degree of nonlinearity, and the time-series autocorrelation coefficient is calculated for the specified indicator using the autocorrelation function to obtain the time-series autocorrelation coefficient; if the time-series autocorrelation coefficient is greater than a preset coefficient threshold, a Long Short-Term Memory Network (LSTM) is selected as the initial global model; if the time-series autocorrelation coefficient is less than or equal to the preset coefficient threshold, a Support Vector Machine (SVM) is selected as the initial global model.

[0081] It is worth noting that the specified indicators can be selected from user needs, resource conditions, and multi-unit sensor node attribute data according to actual needs, and this embodiment of the invention does not limit this.

[0082] Furthermore, the parameters of the global initial model are initialized, and each participating node in the node scheduling list is activated.

[0083] Furthermore, the cloud sends the node scheduling list and the initial global model to the edge devices participating in the training.

[0084] Step 202: The edge device sends the initial global model to the multi-unit sensor node according to the node scheduling list.

[0085] Specifically, the edge device forwards the initial global model to the multi-unit sensor node participating in the training as indicated by the node scheduling list.

[0086] Step 203: The multi-unit sensing node trains the initial global model locally based on the local training dataset, generates an intermediate training model and encrypted model parameters, and sends the encrypted model parameters to the edge.

[0087] In this embodiment of the invention, multi-unit sensor nodes acquire a local dataset, perform data cleaning on the local dataset to obtain a cleaned local dataset, ensuring data quality and thus guaranteeing model accuracy; the cleaned local dataset is then divided into a local training dataset and a local validation dataset. Data cleaning includes handling missing values ​​and outlier values.

[0088] Specifically, missing value handling is based on imputation using a noisy median, calculated as follows:

[0089]

[0090] in, The median with noise is given, median is used to calculate the median, and D is used to calculate the median. local [:,j] represents the local training data j, and Laplace(0,Δ / ∈) represents Laplace noise.

[0091] Specifically, outlier handling uses a robust Z-score for thresholding. Outliers above the upper threshold (3) or below the lower threshold (-3) require outlier handling, while the original sampled data remains unchanged within the range between the upper and lower thresholds.

[0092]

[0093] Among them, z ij This is an outlier; The median with added noise; is the calculated absolute deviation of the noise-added median; 0.6745 is the conversion factor between MAD and standard deviation under the standard normal distribution, which can be selected according to user needs; x ij This is the sampled data.

[0094] The local training dataset is input into the initial global model for local training, generating intermediate training models and model gradients. The model gradients are encrypted using data encryption technology to generate encrypted model gradients, ensuring that the gradients cannot be cracked by third parties or servers during transmission and aggregation, preventing threats, risks, and data leaks, while meeting privacy compliance requirements such as the General Data Protection Regulation (GDPR) and ensuring the security of user data.

[0095] It is worth noting that data encryption technologies include, but are not limited to, differential privacy or homomorphic encryption.

[0096] Calculate the feature vector of the local training dataset. The feature vector includes, but is not limited to, data homogeneity metrics, communication capability metrics, and computing power and energy consumption metrics. Communication metrics include, but are not limited to, upload speed, link type, bandwidth, and latency. Computing power and energy consumption metrics include, but are not limited to, processor speed, number of processors, and remaining battery power. Data homogeneity metrics specifically include the noisy median. Absolute deviation of median with added noise missing value ratio r miss Outlier ratio r outlier Added noise KL divergence KL divergence can effectively identify data distribution heterogeneity and provide a quantitative basis for model aggregation and contribution allocation. It is calculated using the following formula:

[0097]

[0098] in, KL divergence; C is the total number of categories; P local (c) represents the proportion of class c samples in the local training data of the multi-unit sensing node; Q global (c) is the global reference distribution, which is the proportion of samples of category c in the global distribution initialized and published by the cloud based on global data. It is used as a standard reference distribution to compare the deviation of local data. σ is Laplace noise. Setting Laplace noise can meet the differential privacy requirements and prevent the original data distribution from being derived from KL divergence.

[0099] Feature vectors are the basis for clustering and dividing multi-unit sensor nodes. By using feature vectors, multi-unit sensor nodes with similar data distribution, communication capabilities, computing power, and energy consumption can be grouped into the same group to achieve reasonable grouping.

[0100] The local training set data includes measured pressure sample values, environmental variable sample values, time variable sample values, and standard pressure sample values. The standard pressure sample values ​​are based on the measurements of the highest precision pressure sensor.

[0101] Furthermore, data encryption technology can be used to encrypt the feature vector, generating an encrypted feature vector to ensure the security of data transmission.

[0102] Step 204: The edge and cloud perform model aggregation in layers according to the encrypted model parameters to build and update the training model.

[0103] In this embodiment of the invention, the encryption model parameters include feature vectors and encryption model gradients.

[0104] In this embodiment of the invention, step 204 specifically includes:

[0105] Step 2041: At the edge, a clustering algorithm is used to aggregate multi-unit sensing nodes based on feature vectors to generate initial grouping results.

[0106] In this embodiment of the invention, the clustering algorithm is the Mini-Batch K-Means clustering algorithm. Specifically, the edge nodes execute the Mini-Batch K-Means clustering algorithm based on the feature vectors, with the initial number of groups being: Here, [·] represents the floor function. The initial centroid is randomly selected from high-resource multi-unit sensor nodes. In each round, 10% of the multi-unit sensor nodes are sampled to update the centroid until the centroid's movement distance is less than 10. -3 If the number of nodes in a group is less than a preset threshold, they are merged into the nearest group. Simultaneously, if the KL divergence within a group is too large, asynchronous reorganization is triggered. High-resource multi-unit sensor nodes are selected after quantifying multi-unit sensor nodes based on weighted computing power, communication, and energy consumption. As an optional approach, the communication capability index and computing power and energy consumption index of the feature vector of the multi-unit sensor node are weighted and quantized to obtain the quantization result. The quantization results are then sorted from largest to smallest, and the top 20% of multi-unit sensor nodes are identified as high-resource multi-unit sensor nodes.

[0107] It is worth noting that the quantity threshold can be set according to actual needs, and this embodiment of the invention does not limit it. As an optional solution, the quantity threshold is 5.

[0108] It is worth noting that, in order to more accurately match the actual data structure and system resource status, a dynamic resource awareness mechanism can be introduced to adaptively adjust the optimal number of groups k during operation.

[0109] Step 2042: At the edge, the initial grouping results are filtered for abnormal multi-unit sensor nodes based on the gradient of the encryption model using Local Sensitive Hash (LSH) and KRUM security filtering algorithm to generate a candidate parameter set, which is then uploaded to the cloud.

[0110] In this embodiment of the invention, in order to identify and exclude malicious nodes, Bayesian parameters are used to estimate the number of possible malicious nodes before edge aggregation.

[0111] During the cold start phase, the estimated parameters α0 = 1 and β0 = 1 are initialized based on a uniform prior, while retaining the data from the most recent W rounds, such as W = 50. After this round of training, it is assumed that in round t, the following condition is met: If there are 1 abnormal node, then:

[0112]

[0113] in, and Let be the estimated parameters for the t-th round of detection; and λ1 represents the estimated parameters for the (t-1)th round of detection; λ1 is the forgetting factor, typically taken as 0.9. Let be the number of abnormal nodes detected in the t-th round; Let g be the total number of nodes in group g during round t.

[0114] Calculate the posterior expectation as an estimate of the number of malicious nodes in the current round:

[0115]

[0116] in, n is an estimate of the number of malicious nodes in the current round. g The total number of participating multi-unit sensing nodes in the g-th group is used to describe the size of the group. and represents the estimated parameters for the t-th round of detection; [·] is the rounding function.

[0117] Perform LSH+KRUM filtering. LSH uses four independent hash functions to project the decrypted model gradients into a 64-dimensional space to calculate the hash buckets within each group. KRUM filtering is then performed on each group to exclude... Anomaly calculation. The MAD anomaly score is calculated as follows:

[0118]

[0119] Among them, Score i Anomaly score for the i-th multi-unit sensing node; θ ij Let j be the value of parameter j of the i-th multi-cell sensor node; Median(θ) j ) represents the median of parameter j for all multi-unit sensor nodes; ∈ represents the minimum value to prevent division by zero errors; MAD j d is the median absolute deviation of parameter j for all multi-unit sensing nodes; d is the maximum dimension, for example, 64.

[0120] In this embodiment of the invention, the higher the anomaly score, the greater the probability of an anomaly in the multi-unit sensor node; the lower the anomaly score, the lower the probability of an anomaly in the multi-unit sensor node. All multi-unit sensor nodes are sorted from highest to lowest anomaly score, and multi-unit sensor nodes with anomaly scores within a specified range are marked as anomaly nodes. A global list of anomaly nodes is compiled, and a specified number of cooldown cycles are applied to the anomaly nodes.

[0121] It is worth noting that both the specified range and the specified number of rounds can be set according to actual needs, and this embodiment of the invention does not limit them. As an optional solution, the specified range is the first 20%; the specified number of rounds is 3 rounds.

[0122] In this embodiment of the invention, the edge device filters out potentially abnormal multi-unit sensor nodes before cloud aggregation, performs Byzantine defense, uses LSH for approximate neighbor screening, calculates KRUM scores and filters out outliers, identifies and marks abnormal nodes, and obtains a set of filtered candidate parameters.

[0123] Step 2043: The cloud uses the robust aggregation (BULYAN) algorithm to aggregate the candidate parameter set coordinate by coordinate, filter out outliers in each dimension of the model parameters, and generate updated model parameters.

[0124] In this embodiment of the invention, the candidate parameter set includes filtered model parameters. The BULYAN algorithm is executed in the cloud to calculate the weighted median for the parameter dimensions of the candidate parameter set and perform coordinate median aggregation, removing outlier intermediate results, performing multiple rounds of median pruning, and generating updated model parameters.

[0125] Specifically, for each dimension j of the model parameters, the highest f and lowest f values ​​are removed from the m candidate updated values ​​of that dimension, retaining only the middle part; the average of the remaining m-2f values ​​is taken to obtain the aggregated result for that dimension; and the updated model parameters are generated based on the aggregated results for each dimension of the model parameters. This process is equivalent to applying a truncated average or median to each parameter coordinate. When abnormal updates exist, the parameter may exhibit extremely large or small values ​​at some coordinates, and these values ​​are removed during the pruning process without affecting the final average.

[0126] Step 2044: The cloud updates the initial global model according to the updated model parameters to generate an updated training model.

[0127] In this embodiment of the invention, the updated model parameters are applied to the initial global model to generate an updated training model.

[0128] This invention employs LSH+KRUM for rapid screening of abnormal nodes and BULYAN multi-layer aggregation algorithm to achieve Byzantine fault tolerance and gradient pruning, ensuring model convergence stability. Based on data homogeneity, computing power and communication indicators, dynamic clustering and grouping are combined with contribution quantification to prioritize scheduling of high-value nodes, alleviating the problem of low collaborative efficiency of heterogeneous devices, accelerating model convergence and improving resource utilization.

[0129] Step 205: The cloud performs contribution calculations on the multi-unit sensor nodes and generates an updated node scheduling list based on the contribution.

[0130] In this embodiment of the invention, the dynamic contribution formula is defined as follows:

[0131]

[0132] Where, φ iThe dynamic contribution of the i-th multi-unit sensing node; α(t) is the KL divergence correction weight for the i-th multi-unit sensing node; α(t) is the time decay function. For gradient consistency, and These are the local gradient and the local gradient, respectively; f i This is a data quality factor used to score the data quality of multi-unit sensor nodes. For example, if the proportion of missing or outlier values ​​is low and the data volume is large, this data quality factor increases; λ² is the outlier penalty coefficient; Score i The anomaly score is given for the i-th multi-unit sensing node.

[0133] Furthermore, the dynamic contribution is normalized, and the multi-unit sensor nodes are updated based on the normalized dynamic contribution. The node scheduling list is also updated, and the multi-unit sensor nodes with higher dynamic contribution are selected first in the next round, providing data support for the scheduling plan and dynamic grouping in subsequent rounds.

[0134] It is worth noting that this application generates dynamic contributions by using data quality, multi-unit sensor node participation, and gradient quality. If computing power is sufficient, Shapley statistics can also be used to calculate the contribution in a more refined manner.

[0135] This invention evaluates client contributions from multiple dimensions, including data quality and gradient quality, and combines historical record normalization priority to incentivize high-quality data participation, suppress interference from abnormal nodes, and optimize the fairness and efficiency of federated learning.

[0136] Step 206: The cloud uses the constructed validation dataset to validate the updated training model and generate model validation metrics.

[0137] In this embodiment of the invention, the verification dataset can be constructed according to actual needs, or the local verification dataset divided in step 203 can be used.

[0138] In this embodiment of the invention, model validation metrics include, but are not limited to, mean absolute error, root mean square error, goodness of fit, compensated error distribution, and maximum deviation of key pressure range.

[0139] Specifically, the updated training model is validated using a validation dataset, and the mean absolute error, root mean square error, goodness of fit, compensated error distribution, and maximum deviation in the key pressure range of the updated training model are output.

[0140] Step 207: The cloud generates a comprehensive model score based on the model validation metrics.

[0141] In this embodiment of the invention, the model verification indicators are weighted and fused according to the preset verification indicator weights to generate a model comprehensive score; the model comprehensive score is then normalized to obtain a normalized model comprehensive score.

[0142] Step 208: The cloud determines whether the model's overall score is higher than the average score of the model in the specified historical consecutive rounds. If yes, proceed to step 209; otherwise, proceed to step 211.

[0143] In this embodiment of the invention, if the model's overall score is higher than the average score of the model in a specified historical consecutive rounds, it indicates that the model needs to be updated, and step 209 is continued; if the model's overall score is lower than the average score of the model in a specified historical consecutive rounds, it indicates that the model does not need to be updated, and step 211 is continued.

[0144] It is worth noting that the number of consecutive historical rounds can be set according to actual needs, and this embodiment of the invention does not limit this. As an optional solution, the number of consecutive historical rounds is 5 consecutive times.

[0145] It is worth noting that in real-world scenarios, changes in model parameters and the dynamic nature of validation datasets cause variations in the overall model score in each round. If consecutive models achieve the same overall score, a fault-tolerance mechanism needs to be added. This mainly includes: parameter change detection, proactive injection of differential privacy noise, resampling of multi-unit sensor nodes, or constructing adversarial examples and adding stress pulse data.

[0146] Step 209: The cloud will send the updated training model and the updated node scheduling list to the edge.

[0147] In this embodiment of the invention, the cloud broadcasts the updated training model and the updated node scheduling list and implements incentive feedback, so that the edge device can receive the updated training model and the updated node scheduling list.

[0148] Step 210: The edge device sends the updated training model to the multi-unit sensor node according to the updated node scheduling list, so that the multi-unit sensor node can iteratively train the updated training model until a pressure self-calibration model is generated, and then continue to execute step 212.

[0149] In this embodiment of the invention, the edge device forwards the updated training model to the multi-unit sensor node with higher dynamic contribution indicated by the updated node scheduling list; this allows the multi-unit sensor node with higher dynamic contribution to iteratively train the updated training model until the model converges, thereby constructing a stress self-calibration model.

[0150] Furthermore, blockchain technology is used to store training log files. These training log files include all processes involved in each training and update, facilitating version backtracking and operational status recording.

[0151] Step 211: The cloud sends the updated node scheduling list to the edge, so that the edge can send iterative training messages to the multi-unit sensor nodes according to the updated node scheduling list, so that the multi-unit sensor nodes can iteratively train the intermediate training model until a pressure self-calibration model is generated, and then continue to execute step 212.

[0152] In this embodiment of the invention, the cloud broadcasts the updated node scheduling list and implements incentive feedback, enabling the edge device to receive the updated node scheduling list; the edge device forwards the iterative training message to the multi-unit sensor node with higher dynamic contribution indicated by the updated node scheduling list; the multi-unit sensor node with higher dynamic contribution iteratively trains the intermediate training model until the model converges, thus constructing a stress self-calibration model.

[0153] Furthermore, blockchain technology is used to store training log files. These training log files include all processes involved in each training and update, facilitating version backtracking and operational status recording.

[0154] Step 212: Collect measured pressure sensing data and environmental variable data from multi-unit sensing nodes.

[0155] In this embodiment of the invention, pressure sensor data is obtained through actual measurement and acquisition; environmental sensor data is obtained through actual measurement and acquisition.

[0156] Step 213: The multi-unit sensing node uses a pressure self-calibration model to perform pressure compensation under environmental calibration on the measured pressure sensing data based on environmental variable data, generating self-calibrated pressure data.

[0157] Specifically, environmental variable data is input into the measured pressure sensing data to perform pressure compensation under environmental calibration in the pressure self-calibration model, and self-calibrated pressure data is output.

[0158] This invention addresses the drift and heterogeneity characteristics of pressure sensors by integrating data cleaning, anomaly detection, and dynamic recovery mechanisms to achieve continuous self-calibration without manual intervention, thereby improving the consistency and reliability of the sensor network during long-term operation.

[0159] It is worth noting that the acquisition, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. The user information in the embodiments of this application was obtained through legal and compliant means, and the acquisition, storage, use, and processing of user information have been authorized and agreed upon by the client.

[0160] It is worth noting that the information collected in this application is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with the relevant laws, regulations and standards of the relevant countries and regions, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation portals are provided for users to choose to authorize or refuse.

[0161] It is worth noting that the technical solution provided in this application provides users with a corresponding operation entry point, allowing users to choose to agree to or reject the automated decision-making results; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0162] In the technical solution of the multi-unit pressure sensor self-calibration method based on federated learning provided in this invention embodiment, multi-unit sensing nodes collect measured pressure sensing data and environmental variable data. The multi-unit sensing nodes are located in a pre-built multi-unit sensing node-edge-cloud hierarchical star topology sensing network. The multi-unit sensing nodes use a pre-constructed pressure self-calibration model to perform pressure compensation under environmental calibration on the measured pressure sensing data based on the environmental variable data, generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework, realizing Byzantine fault tolerance and gradient pruning, ensuring model convergence stability, reducing system complexity and maintenance costs, thereby improving universality; alleviating the problem of low collaborative efficiency of heterogeneous devices, accelerating model convergence and improving resource utilization, reducing computational load for the heterogeneity of pressure sensor data, ensuring communication efficiency, and improving dynamic environmental adaptability and calibration result accuracy.

[0163] Figure 7 This is a schematic diagram of a self-calibration system for a multi-unit pressure sensor based on federated learning, provided in an embodiment of the present invention. This system is used to execute the aforementioned self-calibration method for a multi-unit pressure sensor based on federated learning, such as... Figure 7 As shown, the system includes: a data acquisition unit 11 and a pressure self-calibration unit 12.

[0164] The data acquisition unit 11 is used to collect measured pressure sensing data and environmental variable data from the multi-unit sensing nodes. The multi-unit sensing nodes are located in a pre-built multi-unit sensing node-edge-cloud hierarchical star topology sensing network.

[0165] The pressure self-calibration unit 12 is used by multi-unit sensing nodes to perform pressure compensation under environmental calibration on measured pressure sensing data based on environmental variable data through a pre-built pressure self-calibration model, thereby generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework.

[0166] In this embodiment of the invention, the system further includes: a node scheduling list generation unit 13, a sending unit 14, a local training unit 15, a model aggregation unit 16, and a list updating unit 17.

[0167] The node scheduling list generation unit 13 is used to generate a node scheduling list and determine the initial global model based on the hierarchical star topology sensor network and multi-unit sensor node attribute data sent by the edge end.

[0168] The sending unit 14 is used at the edge to send the initial global model sent by the cloud to the multi-unit sensing nodes according to the node scheduling list sent by the cloud.

[0169] The local training unit 15 is used by the multi-unit sensing node to train the initial global model locally based on the local training dataset, generate intermediate training models and encrypted model parameters, and send the encrypted model parameters to the edge.

[0170] The model aggregation unit 16 is used to aggregate models at the edge and cloud based on encrypted model parameters in a hierarchical manner to build and update the training model.

[0171] The list update unit 17 is used to calculate the contribution of multi-unit sensor nodes in the cloud, generate an updated node scheduling list based on the contribution, and send the updated node scheduling list to the edge terminal so that the edge terminal can send iterative training messages to the multi-unit sensor nodes according to the updated node scheduling list, so that the multi-unit sensor nodes can perform model iterative training until a pressure self-calibration model is generated.

[0172] In this embodiment of the invention, the encrypted model parameters include feature vectors and encrypted model gradients; the model aggregation unit 16 is specifically used at the edge to aggregate multi-unit sensing nodes according to feature vectors using a clustering algorithm to generate initial grouping results; at the edge, abnormal nodes are filtered in the initial grouping results according to the encrypted model gradients using local sensitive hashing and security filtering algorithms to generate a candidate parameter set, and the candidate parameter set is uploaded to the cloud; the cloud uses a robust aggregation algorithm to aggregate the candidate parameter set coordinate by coordinate, filtering outliers in each dimension of the model parameters to generate updated model parameters; the cloud updates the initial global model according to the updated model parameters to generate an updated training model.

[0173] In this embodiment of the invention, the system further includes: a model verification unit 18, a model comprehensive score generation unit 19, and a judgment unit 20.

[0174] The model validation unit 18 is used to validate the updated training model on the cloud using the constructed validation dataset and generate model validation metrics.

[0175] The model comprehensive score generation unit 19 is used to generate a model comprehensive score in the cloud based on the model validation indicators.

[0176] The judgment unit 20 is used to determine in the cloud whether the overall score of the model is higher than the average score of the model in the specified historical consecutive rounds; if not, the list update unit 17 is triggered to continue to execute the step of sending the updated node scheduling list to the edge.

[0177] In this embodiment of the invention, the sending unit 14 is further configured to, if the judgment unit 20 determines that the model's overall score is higher than the average score of the model in a specified historical consecutive rounds, send the updated training model and the updated node scheduling list to the edge end; the edge end sends the updated training model to the multi-unit sensor node according to the updated node scheduling list, so that the multi-unit sensor node iteratively trains the updated training model until a pressure self-calibration model is generated.

[0178] In this embodiment of the invention, the system further includes a blockchain storage unit 21.

[0179] The blockchain storage unit 21 is used to store training log files using blockchain technology.

[0180] In this embodiment of the invention, multi-unit sensor nodes collect measured pressure sensing data and environmental variable data. These multi-unit sensor nodes are located in a pre-built hierarchical star topology sensor network consisting of multi-unit sensor nodes, edge devices, and the cloud. The multi-unit sensor nodes use a pre-constructed pressure self-calibration model to perform pressure compensation under environmental calibration on the measured pressure sensing data based on the environmental variable data, generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework, achieving Byzantine fault tolerance and gradient pruning, ensuring model convergence stability, reducing system complexity and maintenance costs, thereby improving universality. It also alleviates the problem of low collaborative efficiency of heterogeneous devices, accelerates model convergence and improves resource utilization. Furthermore, it reduces computational load, ensures communication efficiency, and improves dynamic environmental adaptability and calibration result accuracy, addressing the heterogeneity of pressure sensor data.

[0181] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer device, specifically, a computer device can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0182] This invention provides a computer device, including a memory and a processor. The memory is used to store information including program instructions, and the processor is used to control the execution of the program instructions. When the program instructions are loaded and executed by the processor, they implement the steps of the above-described embodiment of the self-calibration method for multi-unit pressure sensors based on federated learning. For a detailed description, please refer to the above-described embodiment of the self-calibration method for multi-unit pressure sensors based on federated learning.

[0183] The following is for reference. Figure 8 It shows a schematic diagram of the structure of a computer device 600 suitable for implementing the embodiments of this application.

[0184] like Figure 8 As shown, the computer device 600 includes a central processing unit (CPU) 601, which can perform various appropriate tasks and processes based on programs stored in read-only memory (ROM) 602 or programs loaded from storage section 608 into random access memory (RAM) 603. The RAM 603 also stores various programs and data required for the operation of the computer device 600. The CPU 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0185] The following components are connected to I / O interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal feedback (LCD), etc., and speakers, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to I / O interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 610 as needed so that computer programs read from it can be installed in storage section 608 as needed.

[0186] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program tangibly embodied on a machine-readable medium, the computer program including program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611.

[0187] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0188] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0189] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0190] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0191] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0192] It should also be noted that 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 a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0193] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0194] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0195] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0196] This application can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0197] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0198] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A self-calibration method for multi-unit pressure sensors based on federated learning, characterized in that, The method includes: Multi-unit sensor nodes collect measured pressure sensing data and environmental variable data. The multi-unit sensor nodes are located in a pre-built multi-unit sensor node-edge-cloud hierarchical star topology sensor network. The multi-unit sensing node uses a pre-built pressure self-calibration model to perform pressure compensation under environmental calibration on the measured pressure sensing data based on the environmental variable data, generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework.

2. The self-calibration method for multi-unit pressure sensors based on federated learning according to claim 1, characterized in that, The multi-unit sensing node includes a homogeneous redundant architecture and a heterogeneous fusion architecture; The homogeneous redundancy architecture adopts a hardware cold backup method and is configured with multiple pressure sensors of the same model. The heterogeneous fusion architecture integrates multiple pressure sensors and environmental sensors of different models.

3. The self-calibration method for multi-unit pressure sensors based on federated learning according to claim 1, characterized in that, The method further includes: Based on the hierarchical star topology sensor network and multi-unit sensor node attribute data sent from the edge, the cloud generates a node scheduling list and determines the initial global model. The edge device sends the initial global model sent by the cloud to the multi-unit sensor nodes according to the node scheduling list sent by the cloud. The multi-unit sensing node trains the initial global model locally based on the local training dataset, generates an intermediate training model and encrypted model parameters, and sends the encrypted model parameters to the edge. The edge and cloud perform hierarchical model aggregation based on the encrypted model parameters to construct and update the training model; The cloud calculates the contribution of the multi-unit sensor nodes, generates an updated node scheduling list based on the contribution, and sends the updated node scheduling list to the edge terminal. The edge terminal then sends iterative training messages to the multi-unit sensor nodes according to the updated node scheduling list, enabling the multi-unit sensor nodes to perform iterative model training until the pressure self-calibration model is generated.

4. The self-calibration method for multi-unit pressure sensors based on federated learning according to claim 3, characterized in that, The encryption model parameters include feature vectors and encryption model gradients; The edge and cloud perform hierarchical model aggregation based on the encryption model parameters to construct and update the training model, including: At the edge, a clustering algorithm is used to aggregate multi-unit sensing nodes based on the feature vector to generate initial grouping results; At the edge, local sensitive hashing and security filtering algorithms are used to filter abnormal nodes in the initial grouping results according to the gradient of the encryption model, generate a set of candidate parameters, and upload the set of candidate parameters to the cloud. The cloud uses a robust aggregation algorithm to aggregate the candidate parameter set coordinate by coordinate, filter out outliers in each dimension of the model parameters, and generate updated model parameters. The cloud updates the initial global model according to the updated model parameters to generate an updated training model.

5. The self-calibration method for multi-unit pressure sensors based on federated learning according to claim 3, characterized in that, Before sending the updated node scheduling list to the edge, the method further includes: The cloud uses a constructed validation dataset to validate the updated training model and generate model validation metrics. The cloud platform generates a comprehensive model score based on the model validation metrics. The cloud determines whether the overall score of the model is higher than the average score of the model in a specified historical consecutive rounds; If not, continue with the step of sending the updated node scheduling list to the edge.

6. The self-calibration method for multi-unit pressure sensors based on federated learning according to claim 5, characterized in that, The method further includes: If so, the cloud will send the updated training model and the updated node scheduling list to the edge. The edge device sends the updated training model to the multi-unit sensor node according to the updated node scheduling list, so that the multi-unit sensor node iteratively trains the updated training model until the pressure self-calibration model is generated.

7. The self-calibration method for multi-unit pressure sensors based on federated learning according to claim 1, characterized in that, The method further includes: Training log files are stored using blockchain technology.

8. A self-calibration system for multi-unit pressure sensors based on federated learning, characterized in that, The system includes: The data acquisition unit is used to collect measured pressure sensing data and environmental variable data from multi-unit sensing nodes. The multi-unit sensing nodes are located in a pre-built multi-unit sensing node-edge-cloud hierarchical star topology sensing network. The pressure self-calibration unit is used by the multi-unit sensing nodes to perform pressure compensation under environmental calibration on the measured pressure sensing data based on the environmental variable data through a pre-built pressure self-calibration model, thereby generating self-calibrated pressure data. The pressure self-calibration model is obtained through iterative training based on a federated learning framework.

9. A computer-readable medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the self-calibration method for multi-unit pressure sensors based on federated learning as described in any one of claims 1 to 7.

10. A computer device comprising a memory and a processor, the memory for storing information including program instructions, and the processor for controlling the execution of the program instructions, characterized in that, When the program instructions are loaded and executed by the processor, they implement the self-calibration method for multi-unit pressure sensors based on federated learning as described in any one of claims 1 to 7.

11. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the self-calibration method for multi-unit pressure sensors based on federated learning as described in any one of claims 1 to 7.

Images