A self-cooling low-temperature bearing ring development system integrated with temperature control sensors

Abstract

The application integrates a passive wireless temperature control sensing unit and a pulse eddy current nondestructive testing module, synchronously collects temperature, frequency offset and lattice strain data of a low-temperature bearing ring under rotating working conditions, constructs a thermal-force coupling correlation model through a low-temperature tribology-thermal conduction coupling analysis module, and outputs a friction heat generation rate prediction value. Relying on a federal learning edge inference module, distributed model training is completed, combined with homomorphic encryption to guarantee data security, and the data island restriction is broken through. Based on an optimization strategy, the edge controller dynamically adjusts the sensing sampling interval through reinforcement learning, solves a multi-objective optimization problem by using a non-dominated sorting genetic algorithm, and drives the hierarchical start and stop of the self-cooling system. The application eliminates the physical interference of wired sensing, realizes adaptive temperature control and energy-saving cooling of the bearing ring under low-temperature working conditions, and improves the temperature control precision and system generalization ability.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology and discloses a self-cooling cryogenic bearing ring development system integrating a temperature control sensor. Background Technology

[0002] In conventional cryogenic bearing ring development systems, wired sensing structures are typically used for temperature measurement under rotating conditions. This involves attaching thermocouples or resistance temperature detectors (RTDs) to the ring surface and transmitting electrical signals to external data acquisition equipment via slip rings and leads. For strain detection of the ring material, a separate pulsed eddy current non-destructive testing (NDT) device is usually employed, scanning the ring surface at shutdown or low speed to obtain lattice strain data. Regarding temperature control strategies, existing systems rely on fixed temperature thresholds; when the detected temperature exceeds the preset threshold, the cooling system is activated at full power. For multi-condition model training, temperature and strain data from each condition node are uploaded to a central server for centralized model training to update the temperature control thresholds.

[0003] In the aforementioned existing technologies, the wired sensing structure's leads are subjected to centrifugal force and cooling fluid erosion under rotational conditions, altering the dynamic balance and assembly accuracy of the raceway. Furthermore, the independently operating pulsed eddy current detection and wired temperature measurement are disconnected in time and space, failing to reflect the real-time coupling relationship during dynamic rotation. Because multi-condition nodes use centralized data upload, in industrial scenarios, limitations on data privacy and transmission bandwidth prevent nodes from sharing raw data, leading to data silos in model training. Under these data silo limitations, the system cannot integrate real-time temperature data with material strain characteristics across domains to establish a dynamically coupled model, thus failing to output adaptive cooling control strategies with generalization capabilities under multi-condition environments. Summary of the Invention

[0004] The purpose of this invention is to provide a solution that can effectively address the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A self-cooling low-temperature bearing ring R&D system integrating a temperature control sensor includes: a surface acoustic wave passive wireless temperature control sensing unit, integrated into the non-contact working surface of the bearing ring, used to synchronously collect real-time temperature data and surface acoustic wave resonant frequency offset data under rotation conditions. The pulsed eddy current nondestructive testing module is used to synchronously acquire lattice strain data of the bearing ring surface material; The low-temperature tribology-heat conduction coupling analysis module receives real-time temperature data, resonant frequency shift data, and lattice strain data to establish a correlation model of frictional heat generation, heat conduction, and temperature deformation of the ring. The federated learning edge inference module uses real-time temperature data from multiple nodes as the core training samples to train a distributed model and outputs optimized cooling system trigger logic and temperature control sensor unit sampling strategy. The edge controller receives the optimized cooling system trigger logic and the temperature control sensor unit sampling strategy, dynamically adjusts the sampling interval of the surface acoustic wave passive wireless temperature control sensor unit, and drives the self-cooling system to perform graded start-up and shutdown.

[0006] Preferably, the method for constructing the correlation model using the low-temperature tribology-heat conduction coupling analysis module includes: converting the resonant frequency shift data into the equivalent elastic modulus attenuation characteristics on the surface acoustic wave propagation path, and converting the lattice strain data into the microscopic grain boundary slip barrier change characteristics; A heterogeneous feature fusion layer is constructed based on graph neural network, which maps the equivalent elastic modulus decay feature and the micro grain boundary slip barrier change feature to the same high-dimensional latent space, and extracts the thermal-mechanical coupling degradation characterization vector of bearing rings under low temperature conditions. Using the thermo-mechanical coupling degradation characterization vector as the state variable, the Hamiltonian dynamic equation system, which includes the partial differential equation of heat conduction and the differential equation of frictional power consumption, is solved, and the predicted value of dynamic frictional heat generation rate is output.

[0007] Preferably, the method for the federated learning edge inference module to perform distributed model training includes: constructing a local loss function using the predicted value of dynamic frictional heat generation rate at each edge node, and using an adaptive moment estimation optimizer to iterate the local model parameters. During the model parameter aggregation stage, a homomorphic encryption mechanism is introduced to encrypt the gradient update amount of the local model parameters, and the global model parameters are calculated on the parameter server side based on cosine similarity weighting. The global model parameters are distributed to each edge node for decryption and model update until the loss function converges. The output is an optimized strategy network that includes the cooling system trigger logic and the temperature control sensor unit sampling strategy.

[0008] Preferably, the method for dynamically adjusting the sampling interval of the passive wireless temperature control sensing unit of the surface acoustic wave by the edge controller includes: constructing a Markov decision process based on reinforcement learning, using real-time temperature data and predicted values ​​of dynamic frictional heat generation rate to form an environmental state space, using discrete adjustment step size of sampling interval to form an action space, and using energy consumption cost and temperature prediction error in the graded start-up and shutdown process of the self-cooling system to form a reward function. The edge controller obtains the optimal action strategy by solving the Bellman equation. When the real-time temperature data is detected to be in a stable range, the sampling interval is increased, and when a sudden increase in the real-time temperature data is detected, the sampling interval is decreased.

[0009] Preferably, the method for the edge controller to drive the self-cooling system to perform graded start-stop includes: transforming the optimized cooling system triggering logic into a multi-objective optimization control problem, taking the target temperature range of the bearing rings and the minimum cooling energy consumption as optimization objectives, and using a non-dominated sorting genetic algorithm to solve the Pareto front optimal solution set; The compromise solution is extracted from the Pareto front optimal solution set and mapped to the opening duty cycle of the multi-channel cooling valves in the self-cooling system and the flow distribution ratio of the cooling fluid. Based on the flow distribution ratio, a pulse width modulation signal is generated to drive the multi-channel cooling valves to perform staged start and stop.

[0010] Preferably, the method for constructing a heterogeneous feature fusion layer using a graph neural network includes: introducing a multi-head self-attention mechanism in the graph attention network layer of the graph neural network to calculate the spatial correlation weights between the equivalent elastic modulus decay feature nodes and the micrograin boundary slip barrier change feature nodes; To address the nonlinear noise in feature data under low-temperature conditions, a temporal filtering layer based on a long short-term memory network is connected in series after a multi-head self-attention mechanism to filter out high-frequency noise interference. Using the filtered spatial correlation weights as the initial boundary conditions for the generalized momentum and generalized coordinates in the Hamiltonian dynamic equations improves the convergence of the correlation model under low-temperature, strongly nonlinear conditions.

[0011] Preferably, the method for the federated learning edge inference module to calculate global model parameters on the parameter server includes: under the homomorphic encryption mechanism, performing outlier detection in the ciphertext state on the encrypted gradient update amount uploaded by multiple edge nodes, and eliminating malicious node gradients that deviate from the preset confidence interval by calculating the sum of squared Euclidean distances of the encrypted gradient update amount in the ciphertext domain; The preserved encrypted gradient update is subjected to a distributed secure aggregation based on secret sharing. The encrypted gradient update is split into multiple random shadows and distributed to each edge node using a Lagrange interpolation polynomial. The edge node locally calculates the shadow product and reconstructs the plaintext global model parameters.

[0012] Preferably, the method for edge controllers to solve the Bellman equation to obtain the optimal action policy includes: using a soft actor-commentator algorithm to iteratively solve the Bellman equation, and introducing reparameterization techniques in the commentator network to transform the stochastic policy into deterministic action output to reduce variance; For the continuous high-dimensional environmental state space composed of real-time temperature data, a variational autoencoder is embedded in the actor network of the soft actor-critic algorithm to compress the continuous high-dimensional environmental state space into low-dimensional normally distributed latent variables. The edge controller calculates the maximum entropy objective function based on the low-dimensional normally distributed latent variables and outputs a smoothly changing sampling interval with discrete adjustment step size.

[0013] Preferably, the method of using non-dominated sorting genetic algorithm to solve the optimal solution set of Pareto front includes: introducing Cauchy mutation operator to replace traditional Gaussian mutation operator in the crossover and mutation stage of non-dominated sorting genetic algorithm to enhance the algorithm's escape ability near local extrema of Pareto front. In the solution process, the constraints of the target temperature range of the bearing rings are transformed into an unconstrained optimization problem based on the penalty function. An adaptive penalty factor is constructed to dynamically adjust the penalty intensity according to the degree of violation of the constraints by individuals in the current population. By combining the differential vector strategy of the differential evolution algorithm to update the individual positions of the population, the convergence process of the optimal solution set of the Pareto front is accelerated.

[0014] Preferably, the method for calculating spatial correlation weights in the graph attention network layer includes: using a dynamic graph structure generation mechanism, dynamically adding or deleting topological connection edges in the graph attention network based on the cosine similarity change trend between the attenuation characteristics of the equivalent elastic modulus and the change characteristics of the micrograin boundary slip barrier at adjacent sampling times; In each attention branch of the multi-head self-attention mechanism, a forgetting gating mechanism based on gated recurrent units is introduced to apply an exponential decay mask to the spatial correlation weights of historical moments. The spatial correlation weights after applying an exponential decay mask are concatenated and linearly mapped to output a high-dimensional latent space feature vector with temporal forgetting properties.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention integrates a passive wireless surface acoustic wave temperature control sensing unit, eliminating the wired lead structure and removing the physical interference of the lead on the dynamic balance and assembly accuracy of the rings under rotating conditions. Addressing the model generalization and adaptive control problems under data silo constraints, this invention uses a graph neural network to map the equivalent elastic modulus decay characteristics and microscopic grain boundary slip barrier change characteristics to a high-dimensional latent space to extract thermo-mechanical coupling degradation characterization vectors. Based on this, the Hamiltonian dynamic equations are solved to establish a real-time dynamic coupling correlation model between temperature change and material strain, outputting a predicted value for the frictional heat generation rate. Combining federated learning and homomorphic encryption mechanisms, each edge node completes distributed model aggregation without leaking original data, breaking through the limitations of multi-condition data silos and enabling the trained cooling system trigger logic to have cross-condition generalization capabilities. Based on this trigger logic, the edge controller uses a non-dominated sorting genetic algorithm to solve the Pareto front optimal solution set, mapping the compromise solution to the opening duty cycle and flow distribution ratio of multiple cooling valves, realizing adaptive hierarchical cooling control based on minimum energy consumption and target temperature range.

[0016] 2. In the feature processing stage of the dynamically coupled model, the graph attention network layer introduces a dynamic graph structure generation mechanism and a forgetting gating mechanism for the gated recurrent unit. By dynamically adding and deleting topological connection edges and applying an exponential decay mask to historical weights, nonlinear high-frequency noise under low-temperature conditions is filtered out, improving the convergence of the Hamiltonian dynamic equations solution process. In the execution stage of adaptive regulation, the edge controller constructs a reinforcement learning Markov decision process, combining a soft actor-commentator algorithm and a variational autoencoder to compress the continuous high-dimensional environmental state space into low-dimensional normally distributed latent variables. The step size is discretely adjusted based on the sampling interval that smoothly changes according to the maximum entropy objective function, balancing the energy consumption cost and temperature prediction error in the staged start-up and shutdown process of the self-cooling system. In the model aggregation stage of federated learning, outlier detection and malicious node gradient removal are performed by calculating the sum of squared Euclidean distances in the encrypted domain. Lagrange interpolation polynomials are combined for secret sharing and distributed secure aggregation, ensuring data security during the model parameter reconstruction process. When solving for the optimal solution set of the Pareto front, the Cauchy mutation operator is introduced and combined with the differential vector strategy of the differential evolution algorithm to update the position of individuals in the population. This enhances the algorithm's ability to escape near local extrema and improves the efficiency of solving for the optimal solution set. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall system architecture of the present invention; Figure 2 This is a flowchart of the low-temperature tribology-heat conduction coupling modeling process of the present invention; Figure 3 This is a flowchart of the federated learning edge reasoning training process of the present invention; Figure 4 This is a flowchart illustrating the adaptive adjustment of the sampling interval of the present invention. Figure 5 This is a flowchart of the staged start-stop control process of the self-cooling system of the present invention; Figure 6 This is a flowchart of the heterogeneous feature fusion process of the graph neural network of the present invention. Detailed Implementation

[0018] The following provides a detailed description of the embodiments of the present invention. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0019] Example 1: Refer to Appendix Figure 1 The self-cooling cryogenic bearing ring R&D system integrating temperature control sensors includes a surface acoustic wave passive wireless temperature control sensing unit, a pulsed eddy current non-destructive testing module, a cryogenic tribology-thermal conduction coupling analysis module, a federated learning edge inference module, and an edge controller.

[0020] The surface acoustic wave (SAW) passive wireless temperature control sensing unit is integrated into the non-contact working surface of the bearing race. This non-contact working surface is the surface of the bearing race that does not contact the rolling elements, mating shaft, bearing housing, or sealing ring. It includes the non-matting area of ​​the outer cylindrical surface of the outer race, the non-matting area of ​​the inner bore surface of the inner race, and the non-raceway area of ​​the raceway end face, thus avoiding interference with the dynamic balance, assembly accuracy, and operational performance of the bearing race. The SAW passive wireless temperature control sensing unit includes a 128° YX-cut lithium niobate piezoelectric substrate, an interdigital transducer, a reflective grating array, and a passive radio frequency antenna. Passive excitation is achieved through radio frequency pulse signals emitted by an external reader / writer, simultaneously acquiring real-time temperature data and SAW resonant frequency offset data of the bearing race under rotating conditions. The temperature data acquisition of the SAW passive wireless temperature control sensing unit follows the following mapping relationship: in, To measure the resonant frequency, Reference temperature The reference resonant frequency is below. It is a first-order temperature coefficient. The second-order temperature coefficient was obtained in advance through a reference temperature calibration experiment; real-time temperature data was obtained by back-calculation using the measured resonant frequency. The surface acoustic wave resonant frequency offset data is the frequency offset after deducting the temperature effect, satisfying: in, To measure the resonant frequency, This is the theoretical resonant frequency corresponding to the current real-time temperature. These are resonant frequency offset data, used to characterize the change in the elastic modulus of the bearing race material.

[0021] The pulsed eddy current non-destructive testing module is fixedly installed on the non-rotating side of the bearing housing, maintaining a fixed lift-off distance of 0.5mm~2mm from the surface of the bearing race under test. It shares a synchronous trigger clock output by an edge controller with the surface acoustic wave passive wireless temperature control sensing unit, achieving strict alignment of sampling times and synchronously acquiring lattice strain data of the bearing race surface material. The pulsed eddy current non-destructive testing module includes a pulse excitation source, a differential detection coil, a signal conditioning circuit, and a high-speed AD acquisition unit. The pulse excitation source outputs a nanosecond-level rising pulse signal to the excitation coil, inducing an eddy current field on the bearing race surface. The differential detection coil acquires the induced voltage signal resulting from the superposition of the eddy current field and the excitation field. After signal conditioning and AD conversion, lattice strain data is output based on a preset strain-conductivity mapping model, where the mapping relationship satisfies: in, For lattice strain data, This represents the change in electrical conductivity of the surface material of the bearing race. , The strain coefficient is obtained in advance through calibration experiments.

[0022] The low-temperature tribology-heat conduction coupling analysis module is deployed in an embedded edge computing box. It is coupled to the reader / writer of the surface acoustic wave passive wireless temperature control sensing unit and the pulse eddy current non-destructive testing module via industrial Ethernet. It receives synchronously acquired real-time temperature data, resonant frequency offset data and lattice strain data, and establishes a correlation model of frictional heat generation-heat conduction-temperature deformation of the ring.

[0023] The federated learning edge inference module adopts a distributed architecture of multiple edge nodes and a parameter server. The multiple edge nodes are condition monitoring nodes for multiple bearing rings of the same model. Each edge node locally deploys the low-temperature tribology-heat conduction coupling analysis module, and the parameter server is deployed on a local central industrial control computer or a cloud server. The federated learning edge inference module uses the real-time temperature data collected by each edge node as the core training sample for distributed model training. Each edge node only uploads the encrypted model gradient update, without sharing the original collected data, thus overcoming the limitations of data silos. Finally, it outputs optimized cooling system trigger logic and temperature control sensor unit sampling strategy.

[0024] The edge controller employs a programmable logic controller or an embedded motion controller. It is coupled via an industrial bus to the reader / writer of the surface acoustic wave passive wireless temperature control sensor unit, the parameter server of the federated learning edge inference module, and the valve actuator of the self-cooling system. It receives the optimized cooling system trigger logic and the sampling strategy of the temperature control sensor unit, dynamically adjusts the sampling interval of the surface acoustic wave passive wireless temperature control sensor unit, and drives the self-cooling system to perform graded start-stop operations. The self-cooling system includes a cooling sleeve fitted around the outer ring of the bearing race, multiple independent cooling channels, a cryogenic fluid storage tank, a circulating pump, and multiple electrically controlled cooling valves. Each cooling channel corresponds to one electrically controlled cooling valve. By adjusting the valve opening duty cycle, the cooling fluid flow rate is regulated, achieving graded cooling control.

[0025] As a preferred embodiment, refer to Figure 2 The method for constructing the correlation model of frictional heat generation, heat conduction, and temperature deformation of the ring-shaped ring using the low-temperature tribology-heat conduction coupling analysis module includes the following steps: First, the resonant frequency shift data is converted into the equivalent elastic modulus attenuation characteristics along the surface acoustic wave propagation path. The propagation velocity of the surface acoustic wave and the material's equivalent elastic modulus satisfy the following relationship: in, The propagation speed of surface acoustic waves, The equivalent elastic modulus of the material. For material density, The material's Poisson's ratio. Combined with the relationship between the resonant frequency and propagation speed of surface acoustic waves. ( The wavelength of the surface acoustic wave (determined by the period of the interdigital transducer's fingers) is used to derive the equivalent elastic modulus attenuation characteristics: in, This represents the characteristics of equivalent elastic modulus attenuation. The reference elastic modulus of the material under strain-free conditions. This is the resonant frequency offset data. This is the reference resonant frequency.

[0026] Simultaneously, the lattice strain data is transformed into microscopic grain boundary slip barrier variation characteristics. Based on dislocation dynamics theory, the mapping relationship between grain boundary slip barrier and lattice strain satisfies: in, This describes the characteristics of grain boundary slip barrier changes. For lattice strain data, The Burgers vector of the material dislocation. This represents the activation volume for dislocation slip.

[0027] Secondly, a heterogeneous feature fusion layer is constructed based on a graph neural network. This layer maps the equivalent elastic modulus decay feature and the micrograin boundary slip barrier change feature to the same high-dimensional latent space, extracting the thermo-mechanical coupling degradation characterization vector of the bearing ring under low-temperature conditions. The heterogeneous feature fusion layer includes a graph attention network layer and a fully connected mapping layer. It treats the two types of heterogeneous features as node features of a graph structure, calculates the association weights between nodes through a graph attention mechanism, and completes the adaptive fusion of heterogeneous features. Then, it is mapped to a high-dimensional latent space through the fully connected layer, with an output dimension of [missing dimension]. Thermo-mechanical coupling degradation characterization vector .

[0028] Finally, using the thermo-mechanical coupling degradation characterization vector as the state variable, the Hamiltonian kinetic equations, which include the partial differential equations of heat conduction and the differential equations of frictional heat dissipation, are solved to output the predicted value of the dynamic frictional heat generation rate. The Hamiltonian function of the Hamiltonian kinetic equations is: in, For generalized coordinates, including the temperature field distribution of the bearing rings. With strain field distribution ; Generalized momentum, including heat flux density With stress tensor ; For the system's kinetic energy term, This represents the system's potential energy term. The Hamiltonian canonical equation is: The partial differential equations for heat conduction contained in the system of equations are: in, The specific heat capacity at constant pressure of the material. The thermal conductivity of the material. Let be the rate of frictional heat generation, i.e., the objective quantity to be solved. The differential equation for frictional heat dissipation contained in the system of equations is: in, Degradation characterization vector with thermo-mechanical coupling The relevant coefficient of friction, This refers to the contact stress between the bearing rings and the rolling elements. The relative sliding velocity of the contact surfaces. Let be the contact area. Given the initial boundary conditions, the Hamiltonian dynamic equations are solved using the fourth-order Runge-Kutta method to output the predicted dynamic frictional heat generation rate under different operating conditions.

[0029] As a preferred embodiment, refer to Figure 3 The method for the federated learning edge inference module to perform distributed model training includes the following steps: First, at each edge node, a local loss function is constructed using the predicted dynamic frictional heat generation rate, and an adaptive moment estimation optimizer is used to iterate the local model parameters. The local loss function is: in, The mean squared error loss term satisfies , This represents the number of local training samples. This is the predicted value of the frictional heat generation rate output by the model. The measured frictional heat generation rate is obtained by back-calculation using real-time temperature data; For L2 regularization terms, satisfying , The regularization coefficient is . These are the local model parameters. The parameter iteration process of the adaptive moment estimation optimizer satisfies: in, This is a first-moment estimate of the gradient. For the second moment estimation of the gradient, , The exponential decay rate is estimated by moments. for The gradient of the loss function at time step 1. For learning rate, To prevent the minimum value constant from being divided by zero.

[0030] Secondly, during the model parameter aggregation stage, the CKKS homomorphic encryption mechanism is introduced to update the gradient of the local model parameters. Encryption is performed to generate ciphertext of gradient update values. The data is then uploaded to the parameter server. Within the encrypted domain, the parameter server calculates global model parameters based on cosine similarity weighting, where the cosine similarity weights satisfy the following: in, The number of edge nodes. For the first Gradient update amount of each edge node This is the mean of the gradient update amounts for all edge nodes. This is the function for calculating cosine similarity.

[0031] Finally, the encrypted global model parameters obtained from the weighted calculation are sent to each edge node. Each edge node decrypts and updates the model using its local private key, repeating the local training and global aggregation process until the global loss function converges. The convergence condition is that the change in global loss is less than 1 / 2 for five consecutive rounds. The final output is an optimized strategy network that includes the cooling system trigger logic and the sampling strategy of the temperature control sensor unit. The input of the optimized strategy network is real-time temperature data and the predicted value of the frictional heat generation rate, and the output is the trigger threshold of the cooling system, the graded start-stop rules, and the sampling interval adjustment rules of the surface acoustic wave passive wireless temperature control sensor unit.

[0032] As a preferred embodiment, refer to Figure 4 The method for dynamically adjusting the sampling interval of the surface acoustic wave passive wireless temperature control sensing unit by the edge controller includes the following process: Construct a Markov decision process based on reinforcement learning, wherein the Markov decision process consists of quintuples. Composition, in which: Environmental state space It consists of real-time temperature data and predicted values ​​of dynamic frictional heat generation rate. The state vector at time t is , for Real-time temperature data at any given moment. for Predicted value of dynamic frictional heat generation rate at any given time; Action space The set of discrete adjustment step sizes for the sampling interval, the adjustment step size including The corresponding sampling interval adjustment range is 100ms~10s; This is the state transition probability matrix, which represents the pattern of environmental state changes after an action is performed. reward function It consists of the energy consumption cost and temperature prediction error during the staged start-up and shutdown process of the self-cooling system. The reward value at any given moment satisfies: in, for The energy consumption of the self-cooling system at all times. This is the current temperature value predicted based on data from the previous sampling time. This is the current measured temperature value. , These are preset weighting coefficients; This is the discount factor, and its value range is... This is used to characterize the weighting of future rewards.

[0033] The edge controller obtains the optimal action strategy by solving the Bellman optimal equation, which satisfies: in, The optimal action value function. To perform the action The state transitioned to in the next moment is This is an optional action for the next moment.

[0034] Based on the optimal action strategy obtained from the solution, when the edge controller detects that the real-time temperature data is in a stable range, that is, the temperature change rate is less than the preset stable threshold of 0.5K / min, it increases the sampling interval and reduces the sampling power consumption and data transmission volume; when it detects a sudden increase in the real-time temperature data, that is, the temperature change rate is greater than the preset sudden change threshold of 5K / min, it reduces the sampling interval and increases the sampling frequency to ensure the accuracy and response speed of temperature control monitoring.

[0035] As a preferred embodiment, refer to Figure 5 The method for the edge controller to drive the self-cooling system to perform graded start-up and shutdown includes the following process: The optimized cooling system triggering logic is transformed into a multi-objective optimization control problem, with the target temperature range of the bearing rings and the minimum cooling energy consumption as the optimization objectives. The mathematical expression of the multi-objective optimization problem is as follows: in, To optimize the temperature deviation target, The average temperature of the bearing raceway. Set the temperature for the target. , These are the lower and upper limits of the target temperature range; To optimize cooling energy consumption, This represents the total operating power of the self-cooling system. For the decision variable vector, For the first The duty cycle of the cooling valve opening. This represents the total number of cooling valves.

[0036] The Pareto front optimal solution set of the multi-objective optimization problem is solved by a non-dominated sorting genetic algorithm. The solution process includes population initialization, non-dominated sorting, crowding calculation, selection, crossover and mutation operations, and iteratively generating the Pareto front optimal solution set that satisfies the constraints.

[0037] The compromise solution is extracted from the Pareto front optimal solution set. The membership degree of each solution is calculated using the fuzzy membership function method, and the solution with the highest membership degree is selected as the compromise solution. The compromise solution is then mapped to the ratio of the opening duty cycle of the multi-way cooling valves in the self-cooling system to the flow distribution ratio of the cooling fluid. The cooling fluid flow rate corresponding to the cooling valves in the circuit meets the following requirements: in, For the first Flow distribution values ​​for the cooling valves. For the first The maximum rated flow rate of the cooling valve.

[0038] A pulse width modulation signal with a corresponding frequency is generated according to the flow distribution ratio. The duty cycle of the pulse width modulation signal is consistent with the opening duty cycle of the corresponding cooling valve. The pulse width modulation signal is output to the valve driver to drive multiple cooling valves to perform graded start and stop, thereby realizing adaptive cooling control based on the target temperature range and minimum energy consumption.

[0039] As a preferred embodiment, refer to Figure 6 The method for constructing a heterogeneous feature fusion layer in a graph neural network includes the following steps: In the graph attention network layer of the graph neural network, a multi-head self-attention mechanism is introduced to calculate the spatial correlation weights between the equivalent elastic modulus decay feature nodes and the micrograin boundary slip barrier change feature nodes. First, a heterogeneous graph structure is constructed. , where the node set Including elastic modulus nodes With grain boundary barrier nodes , The initial characteristic is the equivalent elastic modulus decay characteristic. The initial characteristics are the variations in the microscopic grain boundary slip barrier; edge set These are the topological connections between nodes. A linear transformation is applied to the initial features of each node to obtain the transformed feature vector: in, For nodes The initial feature vector, For a trainable linear transformation weight matrix, This is the transformed feature vector. (Compute node) With nodes Attention coefficient between: in, For attention weight vectors, For feature splicing operations, This is a non-linear activation function. The attention coefficients are normalized using the softmax function to obtain the spatial association weights. in, For nodes The set of adjacent nodes. Through Each independent attention head performs the above calculation, and the outputs of each attention head are concatenated to obtain the output features of the multi-head self-attention mechanism: in, For the first Spatial association weights calculated by each attention head For the first Linear transformation weight matrix for each attention head. It is a non-linear activation function.

[0040] To address the nonlinear noise in feature data under low-temperature conditions, a temporal filtering layer based on a long short-term memory network is connected in series after a multi-head self-attention mechanism to filter out high-frequency noise interference. The input of the long short-term memory network is the multi-head self-attention output feature sequence at consecutive sampling times, and its unit update process satisfies: in, for Input characteristics at time step This is the hidden state from the previous moment. This represents the cell state at the previous moment. Output for the forget gate. For input gate output, For output gate output, This represents the current state of the cell. The output features at the current time. and These are the trainable weight matrix and bias vector for the corresponding gating, respectively. It is the sigmoid activation function.

[0041] Using the filtered spatial correlation weights as the initial boundary conditions for the generalized momentum and generalized coordinates in the Hamiltonian dynamic equations, that is, mapping the filtered eigenvectors to the initial boundary values ​​of the temperature field, strain field, heat flux density and stress tensor, narrowing the solution range of the equations and improving the convergence of the correlation model under the low-temperature strong nonlinear state.

[0042] As a preferred embodiment, the method by which the federated learning edge inference module calculates global model parameters on the parameter server side includes the following process: Under homomorphic encryption, outlier detection is performed on the encrypted gradient updates uploaded by multiple edge nodes in the ciphertext state. Based on the ciphertext operation characteristics of the CKKS homomorphic encryption scheme, the sum of squared Euclidean distances of the encrypted gradient updates of each edge node in the ciphertext domain is calculated without decryption: in, For encrypted data, For the first Gradient update amount of each edge node This is the mean of the gradient update amounts for all edge nodes. Let be the dimension of the gradient vector. The index is the gradient dimension. A pre-defined confidence interval is set based on the chi-square distribution, with a confidence level of 95% and degrees of freedom equal to the gradient dimension. If the sum of squared Euclidean distances of a certain edge node exceeds the preset confidence interval, the node is determined to be a malicious node and its corresponding gradient update is removed.

[0043] A distributed secure aggregation based on secret sharing is performed on the retained encrypted gradient updates, using a Lagrange interpolation polynomial to split the encrypted gradient updates into multiple random shaded parts. The Shamir secret sharing scheme is adopted, and for each retaining node... Ciphertext gradient ,structure Second Lagrange interpolation polynomial: in, ~ The ciphertext coefficients are randomly generated. For the threshold value of secret sharing, The number of edge nodes to be retained is less than or equal to the number of edge nodes to be retained. The polynomial in The value at the location As a shadow, it is distributed to the corresponding Each edge node has 1 edge node. Receive shadows sent by all reserved nodes Calculate the sum of all shadows locally: The edge nodes will be calculated Uploaded to the parameter server, the parameter server receives at least indivual Then, the ciphertext of the global gradient is reconstructed using Lagrange interpolation: in, For the Lagrange basis function in The value at that position satisfies , This is the index value of the corresponding node.

[0044] The reconstructed global model parameters are sent in ciphertext to each edge node, which then decrypts them using its local private key to reconstruct the plaintext global model parameters, thus completing the update of the global model.

[0045] As a preferred embodiment, the method for the edge controller to solve the Bellman equation to obtain the optimal action strategy includes the following process: The Bellman equation is solved iteratively using the soft actor-critic algorithm, which is a maximum entropy reinforcement learning algorithm. Its optimization objective is to maximize the sum of the cumulative reward and the policy entropy. The objective function satisfies: in, The policy network to be optimized The trajectory generated by the strategy, This is a temperature coefficient used to balance cumulative rewards and strategy entropy. Let the entropy of the strategy satisfy... The soft actor-critic algorithm comprises two critic networks. , Two target commentator networks , and an actor network A reparameterization technique is introduced into the critic network to transform the stochastic policy into deterministic action outputs to reduce variance: the output of the actor network is set to the mean of the actions. with standard deviation Sampling from the standard normal distribution The action is obtained through reparameterization: Transform the action of random sampling into something related to The deterministic function enables the policy gradient to be backpropagated through the reparameterization path, reducing the variance of the gradient estimate.

[0046] For the continuous high-dimensional environmental state space composed of real-time temperature data, a variational autoencoder is embedded in the actor network of the soft actor-critic algorithm to compress the continuous high-dimensional environmental state space into low-dimensional normally distributed latent variables. The variational autoencoder includes an encoder and a decoder; the encoder receives the high-dimensional state vector. Output hidden variables mean with standard deviation The posterior distribution of the latent variables is obtained: Sample latent variables from the posterior distribution The decoder will contain hidden variables. Reconstruct to the original state The loss function of the variational autoencoder satisfies: in, Let KL divergence be the KL divergence. The prior is a standard normal distribution. The high-dimensional environment state vector is then processed by the trained variational autoencoder. Compressed into low-dimensional normally distributed latent variables As input for the actor network.

[0047] Edge controllers are based on low-dimensional normally distributed latent variables Calculate the maximum entropy objective function; the optimization objective of the actor network is to maximize: The optimization objective of the critic network is to minimize the mean squared error loss. in, For experience replay pool, , By iteratively optimizing the actor network and the critic network until the network converges, the optimal action strategy is obtained, and the output sampling interval is discretely adjusted with a smooth change.

[0048] As a preferred embodiment, the method for solving the Pareto front optimal solution set using a non-dominated sorting genetic algorithm includes the following steps: In the crossover and mutation stages of the non-dominated sorting genetic algorithm, the Cauchy mutation operator is introduced to replace the traditional Gaussian mutation operator, enhancing the algorithm's escape ability near local maxima in the Pareto front. The mutation process of the Cauchy mutation operator satisfies: in, For the individual before the mutation, For the mutated individual, The position parameter is 0 and the scale parameter is... Standard Cauchy distribution random numbers, To adapt to the variation scale, satisfy , This is the initial scale of variation. For the current generation, This represents the maximum number of generations.

[0049] In the solution process, the constraints of the target temperature range of the bearing rings are transformed into an unconstrained optimization problem based on a penalty function, and an adaptive penalty factor is constructed. First, the degree of individual constraint violation is calculated: in, To constrain the amount of violations, This indicates that the individual meets the constraints. The larger the value, the more severe the constraint violation. Construct a penalty function with an adaptive penalty factor: in, This is an adaptive penalty factor that dynamically adjusts based on the degree to which individuals in the current population violate the constraints: when the proportion of individuals violating the constraints in the population exceeds 50%, the penalty factor is increased. The value of is determined by the proportion of individuals in the population that violate the constraints; when the proportion is less than or equal to 50%, the value is reduced. The value of is used to achieve adaptive control that expands the search range in the early stage of evolution and ensures that constraints are satisfied in the later stage of evolution.

[0050] By combining the differential vector strategy of the differential evolution algorithm to update the position of individuals in the population, the convergence process of the Pareto front optimal solution set is accelerated. This is applied to each target individual in the population. Generate mutated individuals: in, , , For random selection, and Distinct individual indexes This is the scaling factor, and its value range is... For mutated individuals With target individuals Perform a binomial crossover operation to generate experimental individuals. .

Claims

1. A self-cooling cryogenic bearing ring development system integrating a temperature control sensor, characterized in that, include: A passive wireless surface acoustic wave temperature control sensor unit is integrated into the non-contact working surface of the bearing ring to synchronously collect real-time temperature data and surface acoustic wave resonant frequency offset data under rotation conditions. The pulsed eddy current nondestructive testing module is used to synchronously acquire lattice strain data of the bearing ring surface material; The low-temperature tribology-heat conduction coupling analysis module receives real-time temperature data, resonant frequency shift data, and lattice strain data to establish a correlation model of frictional heat generation, heat conduction, and temperature deformation of the ring. The federated learning edge inference module uses real-time temperature data from multiple nodes as the core training samples to train a distributed model and outputs optimized cooling system trigger logic and temperature control sensor unit sampling strategy. The edge controller receives the optimized cooling system trigger logic and the temperature control sensor unit sampling strategy, dynamically adjusts the sampling interval of the surface acoustic wave passive wireless temperature control sensor unit, and drives the self-cooling system to perform graded start-up and shutdown.

2. The self-cooling cryogenic bearing ring development system with integrated temperature control sensor according to claim 1, characterized in that, The method for constructing the correlation model in the low-temperature tribology-heat conduction coupling analysis module includes: converting the resonant frequency shift data into the equivalent elastic modulus attenuation characteristics on the surface acoustic wave propagation path, and converting the lattice strain data into the microscopic grain boundary slip barrier change characteristics. A heterogeneous feature fusion layer is constructed based on graph neural network, which maps the equivalent elastic modulus decay feature and the micro grain boundary slip barrier change feature to the same high-dimensional latent space, and extracts the thermal-mechanical coupling degradation characterization vector of bearing rings under low temperature conditions. Using the thermo-mechanical coupling degradation characterization vector as the state variable, the Hamiltonian dynamic equation system, which includes the partial differential equation of heat conduction and the differential equation of frictional power consumption, is solved, and the predicted value of dynamic frictional heat generation rate is output.

3. The self-cooling cryogenic bearing ring development system with integrated temperature control sensor according to claim 1, characterized in that, The method for the federated learning edge inference module to perform distributed model training includes: constructing a local loss function using the predicted value of dynamic frictional heat generation rate at each edge node, and using an adaptive moment estimator optimizer to iterate the local model parameters. During the model parameter aggregation stage, a homomorphic encryption mechanism is introduced to encrypt the gradient update amount of the local model parameters, and the global model parameters are calculated on the parameter server side based on cosine similarity weighting. The global model parameters are distributed to each edge node for decryption and model update until the loss function converges. The output is an optimized strategy network that includes the cooling system trigger logic and the temperature control sensor unit sampling strategy.

4. The self-cooling cryogenic bearing ring development system with integrated temperature control sensor according to claim 1, characterized in that, The method for dynamically adjusting the sampling interval of the passive wireless temperature control sensing unit of surface acoustic wave by the edge controller includes: constructing a Markov decision process based on reinforcement learning, using real-time temperature data and predicted values ​​of dynamic frictional heat generation rate to form the environmental state space, using the discrete adjustment step size of the sampling interval to form the action space, and using the energy consumption cost and temperature prediction error in the graded start-up and shutdown process of the self-cooling system to form the reward function. The edge controller obtains the optimal action strategy by solving the Bellman equation. When the real-time temperature data is detected to be in a stable range, the sampling interval is increased, and when a sudden increase in the real-time temperature data is detected, the sampling interval is decreased.

5. The self-cooling cryogenic bearing ring development system with integrated temperature control sensor according to claim 1, characterized in that, The method of driving the self-cooling system to perform graded start-stop by the edge controller includes: transforming the optimized cooling system triggering logic into a multi-objective optimization control problem, taking the target temperature range of the bearing ring and the minimum cooling energy consumption as optimization objectives, and using a non-dominated sorting genetic algorithm to solve the Pareto front optimal solution set; The compromise solution is extracted from the Pareto front optimal solution set and mapped to the opening duty cycle of the multi-channel cooling valves in the self-cooling system and the flow distribution ratio of the cooling fluid. Based on the flow distribution ratio, a pulse width modulation signal is generated to drive the multi-channel cooling valves to perform staged start and stop.

6. The self-cooling cryogenic bearing ring development system with integrated temperature control sensor according to claim 2, characterized in that, The method for constructing a heterogeneous feature fusion layer in a graph neural network includes: introducing a multi-head self-attention mechanism in the graph attention network layer of the graph neural network to calculate the spatial correlation weight between the equivalent elastic modulus decay feature node and the micrograin boundary slip barrier change feature node; To address the nonlinear noise in feature data under low-temperature conditions, a temporal filtering layer based on a long short-term memory network is connected in series after a multi-head self-attention mechanism to filter out high-frequency noise interference. Using the filtered spatial correlation weights as the initial boundary conditions for the generalized momentum and generalized coordinates in the Hamiltonian dynamic equations improves the convergence of the correlation model under low-temperature, strongly nonlinear conditions.

7. The self-cooling cryogenic bearing ring development system with integrated temperature control sensor according to claim 3, characterized in that, The method for the federated learning edge inference module to calculate global model parameters on the parameter server side includes: under the homomorphic encryption mechanism, performing outlier detection in the ciphertext state on the encrypted gradient update amount uploaded by multiple edge nodes, and eliminating malicious node gradients that deviate from the preset confidence interval by calculating the sum of squared Euclidean distances of the encrypted gradient update amount in the ciphertext domain. The preserved encrypted gradient update is subjected to a distributed secure aggregation based on secret sharing. The encrypted gradient update is split into multiple random shadows and distributed to each edge node using a Lagrange interpolation polynomial. The edge node locally calculates the shadow product and reconstructs the plaintext global model parameters.

8. The self-cooling cryogenic bearing ring development system with integrated temperature control sensor according to claim 4, characterized in that, Methods for edge controllers to solve the Bellman equation to obtain the optimal action policy include: using the soft actor-commentator algorithm to iteratively solve the Bellman equation, and introducing reparameterization techniques in the commentator network to transform the stochastic policy into deterministic action output to reduce variance; For the continuous high-dimensional environmental state space composed of real-time temperature data, a variational autoencoder is embedded in the actor network of the soft actor-critic algorithm to compress the continuous high-dimensional environmental state space into low-dimensional normally distributed latent variables. The edge controller calculates the maximum entropy objective function based on the low-dimensional normally distributed latent variables and outputs a smoothly changing sampling interval with discrete adjustment step size.

9. The self-cooling cryogenic bearing ring development system with integrated temperature control sensor according to claim 5, characterized in that, The method of using non-dominated sorting genetic algorithm to solve the optimal solution set of Pareto front includes: introducing Cauchy mutation operator to replace traditional Gaussian mutation operator in the crossover and mutation stage of non-dominated sorting genetic algorithm to enhance the algorithm's escape ability near local extrema of Pareto front. In the solution process, the constraints of the target temperature range of the bearing rings are transformed into an unconstrained optimization problem based on the penalty function. An adaptive penalty factor is constructed to dynamically adjust the penalty intensity according to the degree of violation of the constraints by individuals in the current population. By combining the differential vector strategy of the differential evolution algorithm to update the individual positions of the population, the convergence process of the optimal solution set of the Pareto front is accelerated.

10. The self-cooling cryogenic bearing ring development system with integrated temperature control sensor according to claim 6, characterized in that, The method for calculating spatial association weights in graph attention network layers includes: using a dynamic graph structure generation mechanism, dynamically adding or deleting topological connection edges in the graph attention network based on the cosine similarity change trend between the attenuation characteristics of the equivalent elastic modulus and the change characteristics of the micrograin boundary slip barrier at adjacent sampling times; In each attention branch of the multi-head self-attention mechanism, a forgetting gating mechanism based on gated recurrent units is introduced to apply an exponential decay mask to the spatial correlation weights of historical moments. The spatial correlation weights after applying an exponential decay mask are concatenated and linearly mapped to output a high-dimensional latent space feature vector with temporal forgetting properties.

Images