Method and system for predicting corrosion thinning life of equipment

Abstract

The invention provides an equipment corrosion thinning life prediction method and system, and relates to the technical field of life prediction, and the method comprises the steps: deploying edge computing nodes at a plurality of equipment sites to collect corrosion data, carrying out the encryption through employing a differential privacy algorithm, and constructing a federated learning network for model training; and the central server extracts prediction experience through knowledge distillation to form knowledge priori, the edge nodes predict the corrosion rate and the residual life based on real-time data, and the server generates a comprehensive evaluation report and continuously optimizes the model. According to the method, safe sharing and collaborative prediction of the equipment corrosion data are realized, the prediction precision is improved, and the data privacy is protected.

Description

Technical Field

[0001] The present invention relates to the technology of life prediction, and particularly to a method and system for predicting the corrosion thinning life of equipment. Background Art

[0002] During the long-term operation of industrial equipment, the phenomenon of corrosion thinning will inevitably occur. This kind of corrosion behavior will lead to the decline of equipment performance and may cause safety accidents in severe cases. At present, the industrial field generally adopts the methods of regular inspection and data analysis to evaluate the corrosion condition of equipment and predict the remaining life. Traditional corrosion prediction methods mainly rely on empirical formulas and statistical models. By collecting operation parameters such as the wall thickness, temperature, and pressure of the equipment, and combining material characteristics and working conditions for analysis and calculation. With the development of the Internet of Things and artificial intelligence technologies, some enterprises have begun to try to use machine learning methods to model and predict the corrosion behavior of equipment in order to improve the accuracy and real-time performance of prediction.

[0003] However, the existing equipment corrosion prediction technologies still have the following deficiencies: Firstly, the traditional data collection and analysis methods often adopt a centralized architecture, and all data need to be uploaded to the central server for processing. This method has security risks during data transmission and is difficult to protect the privacy of sensitive data of enterprises.

[0004] Secondly, most of the existing prediction models are trained based on the data of a single device or a local area, lacking a collaborative learning mechanism for multi-source heterogeneous data, and unable to make full use of the knowledge and experience between different devices, resulting in poor generalization ability of the prediction model.

[0005] Finally, the current corrosion prediction systems generally lack the ability of continuous optimization and evolution. Once the prediction model is deployed, it is rarely updated, and it is difficult to adapt to the dynamic changes of the equipment operation state and working conditions, affecting the reliability of the prediction results. Summary of the Invention

[0006] Embodiments of the present invention provide a method and system for predicting the corrosion thinning life of equipment, which can solve the problems in the prior art.

[0007] In the first aspect of the embodiments of the present invention, A method for predicting the corrosion thinning life of equipment is provided, including: Deploy edge computing nodes at multiple device sites, collect device corrosion data through the edge computing nodes, where the device corrosion data includes wall thickness data, temperature data, pressure data, medium data, and flow rate data; encrypt the device corrosion data based on the differential privacy algorithm of the edge computing nodes to generate an encrypted data set; set up a central server node, and the central server node establishes a federated learning network topology with the edge computing nodes through a secure communication channel, and configure a model aggregation module at the central server node for coordinating the federated training of the edge computing nodes; According to the encrypted data set, construct a basic prediction model with a deep neural network structure at each edge computing node, and perform local training to obtain local model parameters; after receiving the local model parameters, the central server node performs secure aggregation using a homomorphic encryption algorithm to generate global model parameters; the edge computing nodes receive the global model parameters and update the basic prediction model; the central server node extracts the prediction experience of each edge computing node through a knowledge distillation mechanism to form a knowledge prior, and distributes the knowledge prior to the edge computing nodes to enhance the prediction ability of the basic prediction model; The edge computing nodes collect real-time device corrosion data, input the real-time device corrosion data into the basic prediction model that integrates the knowledge prior, and obtain the corrosion rate prediction value and the remaining life of the device; the central server node receives the remaining life of the device uploaded by multiple edge computing nodes, and generates a comprehensive life evaluation report through a swarm intelligence evaluation module; the edge computing nodes optimize the basic prediction model according to the comprehensive life evaluation report, and at the same time feedback the optimized prediction effect to the knowledge prior through the knowledge distillation mechanism to achieve the continuous evolution of the prediction model, forming a collaborative prediction closed loop of prediction-evaluation-optimization.

[0008] Encrypt the device corrosion data based on the differential privacy algorithm of the edge computing nodes to generate an encrypted data set; set up a central server node, and the central server node establishes a federated learning network topology with the edge computing nodes through a secure communication channel, and configure a model aggregation module at the central server node for coordinating the federated training of the edge computing nodes includes: Calculate the local sensitivity value of the device corrosion data, and determine the privacy protection strength of each item of data according to the local sensitivity value; based on the privacy protection strength, inject Laplace noise into the wall thickness data, and inject Gaussian noise into the temperature data, the pressure data, the medium data, and the flow rate data to generate an encrypted data set; Set up a central server node, establish a secure communication channel between the central server node and the edge computing nodes, and use a homomorphic encryption algorithm to transmit and protect model parameters in the secure communication channel; the central server node receives the encrypted dataset transmitted by the edge computing nodes, analyzes the data distribution characteristics of the encrypted dataset, and calculates the data quality scores of each edge computing node; The central server node determines the model aggregation weight of each edge computing node according to the data quality score and the data volume of the encrypted dataset; performs weighted aggregation on the local model parameters of each edge computing node based on the model aggregation weight to generate global model parameters; distributes the global model parameters to each edge computing node through the secure communication channel, and the edge computing nodes use the global model parameters and the encrypted dataset to update the local model, completing the federated learning training based on differential privacy protection.

[0009] According to the encrypted dataset, construct a basic prediction model with a deep neural network structure on each edge computing node, and perform local training to obtain local model parameters; after receiving the local model parameters, the central server node performs secure aggregation using a homomorphic encryption algorithm to generate global model parameters including: According to the data characteristics of the encrypted dataset, construct a deep neural network structure as a basic prediction model on the edge computing node, and the deep neural network structure includes a static feature channel and a temporal feature channel; The static feature channel processes the wall thickness data and the medium data using a three-layer fully connected network, and each layer uses a ReLU activation function; the temporal feature channel processes the temperature data, the pressure data, and the flow rate data using a bidirectional long short-term memory network; introduce an attention mechanism to assign weight coefficients to the features of different time steps in the temporal feature channel; calculate the feature fusion weights of the static feature channel and the temporal feature channel based on the feature distribution, and use residual connection and batch normalization layer to process the fused features to obtain fused features; Perform two-stage training on the basic prediction model based on the fused features. In the first stage, use the Adam optimizer for pre-training, set the learning rate to 0.001, and reduce the learning rate by 10% every 50 rounds of training; in the second stage, perform fine-tuning training, set the learning rate to 0.0001, and set a Dropout layer with a random dropout rate of 0.3 to prevent overfitting to obtain local model parameters; The edge computing node encrypts the local model parameters using the public key of the Paillier homomorphic encryption system and transmits the encrypted local model parameters to the central server node; the central server node calculates the aggregation weights of the edge computing nodes in the encrypted domain, securely aggregates the encrypted local model parameters based on the aggregation weights, and decrypts the aggregation result using the private key to obtain the global model parameters.

[0010] The edge computing node receives the global model parameters and updates the basic prediction model; the central server node extracts the prediction experience of each edge computing node through a knowledge distillation mechanism to form a knowledge prior, and distributes the knowledge prior to the edge computing nodes to enhance the prediction ability of the basic prediction model, including: The edge computing node receives the global model parameters, introduces an adaptive fusion factor to perform weighted fusion on the global model parameters and local model parameters, where the value range of the adaptive fusion factor is from 0 to 1; imports the fused model parameters into the basic prediction model to obtain an updated basic prediction model; the edge computing node uses the updated basic prediction model to predict local data and generates a prediction result; evaluates the accuracy of the prediction result based on the local validation set, and dynamically adjusts the value of the adaptive fusion factor according to the accuracy. The central server node obtains the output probability distribution of each edge computing node through a knowledge distillation mechanism and uses the output probability distribution as the prediction experience; extracts the intermediate layer features of each edge computing node model, where the intermediate layer features include the output of the model hidden layer and the attention weights; the central server node constructs prediction rules based on the prediction experience and the intermediate layer features, uses a feature dependence graph to represent the association relationship between the prediction rules, and converts the prediction rules and the feature dependence graph into a knowledge prior. The central server node distributes the knowledge prior to each edge computing node, and the edge computing node converts the knowledge prior into model constraint conditions and applies the model constraint conditions to the updated basic prediction model; the edge computing node constructs a composite loss function based on the model constraint conditions, where the composite loss function includes a cross-entropy loss term and a knowledge distillation loss term; uses the composite loss function to optimize and train the updated basic prediction model to improve the prediction ability of the basic prediction model.

[0011] The edge computing node collects real-time device corrosion data, inputs the real-time device corrosion data into the basic prediction model fused with the knowledge prior, and obtains the corrosion rate prediction value and the remaining life of the device; the central server node receives the remaining life of the device uploaded by multiple edge computing nodes and generates a comprehensive life assessment report through a swarm intelligence evaluation module, including: The edge computing node collects electrochemical parameters including potential, current density, and polarization resistance, environmental parameters including temperature, pressure, and pH value, and material parameters including stress and strain, and uses the electrochemical parameters, the environmental parameters, and the material parameters as real-time device corrosion data; The edge computing node constructs a feature dependency matrix according to the prior knowledge, uses the feature dependency matrix as an attention weight to weight the importance of features for the real-time device corrosion data, and obtains weighted features; inputs the weighted features into a basic prediction model, calculates a corrosion rate prediction value based on historical corrosion laws; establishes a damage accumulation model according to the corrosion rate prediction value and a material performance curve, calculates the remaining life of the device, and uploads the remaining life of the device, the prediction confidence, and the corrosion parameters to the central server node; The central server node receives the remaining life of the device, the prediction confidence, and the corrosion parameters uploaded by multiple edge computing nodes; the central server node groups edge computing nodes with a working condition similarity higher than a preset similarity threshold to obtain a device group; based on the prediction confidence, weights and fuses the remaining life of the device of each edge computing node to obtain a life distribution feature of the device group; The central server node calculates the deviation between the remaining life of the device of each edge computing node in the device group and the life distribution feature, and marks the edge computing nodes with deviations exceeding the preset range as abnormal nodes; the central server node analyzes the corrosion parameters of the abnormal nodes to identify the reasons for the anomalies; generates a comprehensive life assessment report according to the life distribution feature and the reasons for the anomalies.

[0012] Inputting the weighted features into a basic prediction model, calculating a corrosion rate prediction value based on historical corrosion laws; establishing a damage accumulation model according to the corrosion rate prediction value and a material performance curve, calculating the remaining life of the device, and uploading the remaining life of the device, the prediction confidence, and the corrosion parameters to the central server node includes: Inputting the weighted features into a basic prediction model, where the basic prediction model includes three feature extraction sub-networks, and the three feature extraction sub-networks respectively extract features from electrochemical parameters, environmental parameters, and material parameters to obtain corresponding local features; integrating the local features into a unified feature vector through a cross-modal fusion layer, and inputting the unified feature vector into a fully connected layer to obtain an initial corrosion rate; correcting the initial corrosion rate according to the corrosion development trend in the initial corrosion rate to obtain a final corrosion rate; Extract the yield strength, fracture toughness, and fatigue limit from the performance curve of the equipment material as damage thresholds; substitute the damage thresholds and the final corrosion rate into the damage accumulation equation, which includes a corrosion depth term, a stress-temperature coupling term, and a material degradation term determined by the final corrosion rate; calculate the current cumulative damage amount based on the damage accumulation equation, and calculate the remaining life of the equipment according to the difference between the current cumulative damage amount and the damage critical value and the final corrosion rate. Establish a confidence evaluation model based on the prediction error of the final corrosion rate and the uncertainty analysis of the damage accumulation equation, calculate the probability distribution of the prediction result through the confidence evaluation model, and determine the prediction confidence of the remaining life of the equipment; analyze the parameters affecting the prediction result according to the confidence evaluation model, and determine the parameters with an influence degree on the prediction result higher than the set prediction threshold as corrosion parameters; upload the remaining life of the equipment, the prediction confidence, and the corrosion parameters to the central server.

[0013] The edge computing node optimizes the basic prediction model according to the comprehensive life evaluation report, and at the same time feeds back the optimized prediction effect to the knowledge prior through the knowledge distillation mechanism to realize the continuous evolution of the basic prediction model, forming a collaborative prediction closed-loop of prediction-evaluation-optimization, including: The edge computing node receives the comprehensive life evaluation report, constructs a model loss function based on the population life distribution characteristics in the comprehensive life evaluation report. The model loss function includes a deviation penalty term and a distribution regularization term, and the weight of the deviation penalty term is proportional to the degree of deviation of the prediction result from the population life distribution characteristics; the edge computing node adjusts the attention weights in the feature dependence matrix according to the abnormal reasons identified in the comprehensive life evaluation report, and raises the attention weights corresponding to the corrosion parameters causing the prediction deviation to above the preset weight threshold. The edge computing node uses the model loss function and the feature dependence matrix to optimize and train the basic prediction model, extracts the prediction output distribution of the optimized model as soft label knowledge, and the soft label knowledge includes the response mode of the model to different corrosion parameter combinations; at the same time, record the feature activation mode in the middle layer of the model, and the feature activation mode is used to characterize the deep correlation between the corrosion parameters; calculate the feature importance score based on the soft label knowledge and the feature activation mode, and construct a knowledge graph for the parameter combinations with the feature importance score exceeding the preset parameter threshold. The edge computing node performs structured encoding on the knowledge graph to obtain encoded knowledge, and evaluates the credibility of the encoded knowledge and the original knowledge prior; selects the retained knowledge according to the result of the credibility evaluation, and updates the retained knowledge to the knowledge prior in an incremental learning manner; the edge computing node uses the updated knowledge prior for the next round of prediction process, and realizes the sharing of prediction experience among different edge computing nodes through the knowledge prior; The edge computing node uploads the prediction result of the optimized model to the central server node for group evaluation, continuously optimizes the basic prediction model according to the evaluation result, and updates the knowledge prior, so as to realize the continuous evolution of the basic prediction model through the iterative process of prediction-evaluation-optimization.

[0014] In the second aspect of the embodiments of the present invention, A device corrosion thinning life prediction system is provided, including: A first unit is configured to deploy edge computing nodes at multiple device sites, collect device corrosion data through the edge computing nodes, where the device corrosion data includes wall thickness data, temperature data, pressure data, medium data, and flow rate data; encrypt the device corrosion data based on the differential privacy algorithm of the edge computing node to generate an encrypted data set; set a central server node, and the central server node establishes a federated learning network topology with the edge computing node through a secure communication channel, and configures a model aggregation module at the central server node for coordinating the federated training of the edge computing node; A second unit is configured to construct a basic prediction model with a deep neural network structure at each edge computing node according to the encrypted data set, and perform local training to obtain local model parameters; after receiving the local model parameters, the central server node performs secure aggregation using a homomorphic encryption algorithm to generate global model parameters; the edge computing node receives the global model parameters and updates the basic prediction model; the central server node extracts the prediction experience of each edge computing node through a knowledge distillation mechanism to form a knowledge prior, and distributes the knowledge prior to the edge computing node to enhance the prediction ability of the basic prediction model; The third unit is used for the edge computing node to collect real-time device corrosion data, input the real-time device corrosion data into the basic prediction model that fuses the knowledge prior, and obtain the corrosion rate prediction value and the remaining life of the device; the central server node receives the remaining life of the device uploaded by multiple edge computing nodes, and generates a comprehensive life evaluation report through the swarm intelligence evaluation module; the edge computing node optimizes the basic prediction model according to the comprehensive life evaluation report, and at the same time feeds back the optimized prediction effect to the knowledge prior through the knowledge distillation mechanism, realizing the continuous evolution of the prediction model and forming a collaborative prediction closed loop of prediction-evaluation-optimization.

[0015] In the third aspect of the embodiments of the present invention, a kind of electronic device is provided, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0016] In the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0017] The beneficial effects of this application are as follows: By collecting device corrosion data at the edge computing node and performing differential privacy encryption, and combining with the central server node to construct a federated learning network, the present invention realizes data security and privacy protection, while ensuring the data quality and integrity of the prediction model training, and effectively solves the problems of data leakage and data islands in traditional methods.

[0018] The present invention uses the knowledge distillation mechanism to extract the prediction experience of the edge nodes to form a knowledge prior, and distributes it to each edge computing node for enhancing the basic prediction model, and generates a comprehensive life evaluation report through the swarm intelligence evaluation module, significantly improving the accuracy and reliability of the prediction of the device corrosion thinning life, and overcoming the limitations of a single prediction model.

[0019] The present invention constructs a collaborative prediction closed loop of prediction-evaluation-optimization. The edge computing node continuously optimizes the prediction model according to the comprehensive life evaluation report, and feeds it back to the knowledge prior through the knowledge distillation mechanism, realizing the continuous evolution of the prediction model, enhancing the adaptive ability and prediction performance of the system, and providing strong support for the preventive maintenance and safety management of industrial equipment. Description of the Drawings

[0020] Figure 1 It is a schematic flowchart of the method for predicting the corrosion thinning life of the device in the embodiments of the present invention; Figure 2 This is a schematic structural diagram of the equipment corrosion thinning life prediction system according to an embodiment of the present invention. Detailed implementation manners

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0023] Figure 1 This is a schematic flowchart of the equipment corrosion thinning life prediction method according to an embodiment of the present invention. As Figure 1 shown, the method includes: S11. Deploy edge computing nodes at multiple equipment sites, collect equipment corrosion data through the edge computing nodes, where the equipment corrosion data includes wall thickness data, temperature data, pressure data, medium data, and flow velocity data; encrypt the equipment corrosion data based on the differential privacy algorithm of the edge computing nodes to generate an encrypted data set; set a central server node, and the central server node establishes a federated learning network topology with the edge computing nodes through a secure communication channel, and configures a model aggregation module at the central server node for coordinating the federated training of the edge computing nodes; S12. According to the encrypted data set, construct a basic prediction model with a deep neural network structure at each edge computing node, and perform local training to obtain local model parameters; after receiving the local model parameters, the central server node performs secure aggregation using a homomorphic encryption algorithm to generate global model parameters; the edge computing nodes receive the global model parameters and update the basic prediction model; the central server node extracts the prediction experience of each edge computing node through a knowledge distillation mechanism to form a knowledge prior, and distributes the knowledge prior to the edge computing nodes to enhance the prediction ability of the basic prediction model; S13. The edge computing node collects real-time device corrosion data, inputs the real-time device corrosion data into the basic prediction model that integrates the knowledge prior, and obtains the corrosion rate prediction value and the remaining life of the device; the central server node receives the remaining life of the device uploaded by multiple edge computing nodes, and generates a comprehensive life evaluation report through the swarm intelligence evaluation module; the edge computing node optimizes the basic prediction model according to the comprehensive life evaluation report, and at the same time feeds back the optimized prediction effect to the knowledge prior through the knowledge distillation mechanism to realize the continuous evolution of the prediction model, forming a collaborative prediction closed loop of prediction-evaluation-optimization.

[0024] In an alternative embodiment, the device corrosion data is encrypted based on the differential privacy algorithm of the edge computing node to generate an encrypted data set; a central server node is set up, and the central server node establishes a federated learning network topology with the edge computing node through a secure communication channel, and configures a model aggregation module in the central server node for coordinating the federated training of the edge computing node, including: Calculating the local sensitivity value of the device corrosion data, determining the privacy protection strength of each item of data according to the local sensitivity value; based on the privacy protection strength, injecting Laplace noise into the wall thickness data, and injecting Gaussian noise into the temperature data, the pressure data, the medium data, and the flow rate data to generate an encrypted data set; Set up a central server node, establish a secure communication channel between the central server node and the edge computing node, and use the homomorphic encryption algorithm to transmit and protect the model parameters in the secure communication channel; the central server node receives the encrypted data set transmitted by the edge computing node, analyzes the data distribution characteristics of the encrypted data set, and calculates the data quality score of each edge computing node; The central server node determines the model aggregation weight of each edge computing node according to the data quality score and the data volume of the encrypted data set; performs weighted aggregation on the local model parameters of each edge computing node based on the model aggregation weight to generate global model parameters; distributes the global model parameters to each edge computing node through the secure communication channel, and the edge computing node uses the global model parameters and the encrypted data set to update the local model, completing the federated learning training based on differential privacy protection.

[0025] Differential privacy federated learning method based on edge computing nodes, first performs privacy protection processing on device corrosion data. By calculating the local sensitivity of various types of data, the differential privacy protection intensity is determined. For wall thickness data, due to its high sensitivity and relatively concentrated numerical distribution, the Laplace noise mechanism is used for encryption. When specifically implemented, the value range of the wall thickness data can be set to 0 - 100 millimeters, the privacy budget is selected as 0.1, and the noise scale is dynamically adjusted according to the data sensitivity. For environmental parameters such as temperature, pressure, medium, and flow rate, considering their volatility characteristics, the Gaussian noise injection method is adopted. Taking temperature data as an example, the measurement range is set to -50 to 200 degrees Celsius, and Gaussian distribution with a standard deviation of 2 is selected to generate noise.

[0026] The central server node, as the coordinator of the federated learning network, establishes a secure connection with each edge node through an SSL / TLS encrypted channel. During the transmission of model parameters, homomorphic encryption technology is used to ensure data security. Specifically, a semi - homomorphic encryption scheme is adopted, and the key length is set to 2048 bits, supporting addition operations on the ciphertext domain. After the server receives the encrypted data, it evaluates the data quality by analyzing statistical features such as the variance, skewness, and kurtosis of the data. The quality score uses a percentile system, comprehensively considering factors such as data integrity, consistency, and timeliness.

[0027] In the model aggregation stage, the central server calculates the aggregation weights according to the data quality scores and data volumes of the edge nodes. The proportion of data quality scores in the weight calculation is 60%, and the proportion of data volume is 40%. Taking an industrial site as an example, assume there are three edge nodes, with data quality scores of 85, 92, and 78 respectively, and the data volume ratio is 1:2:1. Then the finally determined aggregation weight ratio is approximately 0.3:0.5:0.2. The server uses these weights to perform weighted averaging on the local model parameters to generate a global model. The global model is distributed to each edge node through a secure channel for updating local model parameters.

[0028] The solution of this application can: Classify and protect sensitive data through the differential privacy mechanism, which not only ensures data availability but also improves the privacy protection level, effectively preventing the risk of data leakage during the model training process. Adopt a combination of homomorphic encryption and secure communication channels to transmit model parameters, constructing a multi - level security protection system to ensure the security of data and models during the federated learning process. Based on the dynamic weight allocation mechanism of data quality scores and data volumes, it improves the accuracy and robustness of model aggregation, and at the same time encourages edge nodes to provide high - quality training data, promoting the healthy development of the federated learning system.

[0029] In an alternative embodiment, based on the encrypted data set, a basic prediction model of a deep neural network structure is constructed at each of the edge computing nodes, and local training is performed to obtain local model parameters; after receiving the local model parameters, the central server node uses a homomorphic encryption algorithm for secure aggregation to generate global model parameters, including: Based on the data characteristics of the encrypted data set, a deep neural network structure is constructed at the edge computing node as a basic prediction model, and the deep neural network structure includes a static feature channel and a temporal feature channel; The static feature channel uses a three-layer fully connected network to process the wall thickness data and the medium data, and each layer uses a ReLU activation function; the temporal feature channel uses a bidirectional long short-term memory network to process the temperature data, the pressure data, and the flow rate data; an attention mechanism is introduced to assign weight coefficients to the features at different time steps in the temporal feature channel; based on the feature distribution, the feature fusion weights of the static feature channel and the temporal feature channel are calculated, and the fused features are processed using residual connection and batch normalization layers to obtain fused features; Based on the fused features, two-stage training is performed on the basic prediction model. In the first stage, pre-training is performed using the Adam optimizer, the learning rate is set to 0.001, and the learning rate is reduced by 10% every 50 rounds of training; in the second stage, fine-tuning training is performed, the learning rate is set to 0.0001, and a Dropout layer with a random dropout rate of 0.3 is set to prevent overfitting, obtaining local model parameters; The edge computing node encrypts the local model parameters using the public key of the Paillier homomorphic encryption system and transmits the encrypted local model parameters to the central server node; the central server node calculates the aggregation weights of each edge computing node in the encrypted domain, performs secure aggregation on the encrypted local model parameters based on the aggregation weights, and decrypts the aggregation result using the private key to obtain global model parameters.

[0030] At the edge computing node, a basic prediction model of a deep neural network structure is first constructed. This model consists of two main parts: a static feature channel and a temporal feature channel. The static feature channel is used to process the wall thickness data and the medium data, and adopts a three-layer fully connected network structure. The number of input nodes in the first layer is the sum of the wall thickness and medium feature dimensions, the number of nodes in the middle layer is set to 128, and the number of nodes in the output layer is 64. A ReLU activation function is used for non-linear transformation between each layer to effectively extract the correlation between static features.

[0031] The time-series feature channel uses a bidirectional long short-term memory network to process time-series data such as temperature, pressure, and flow rate. In specific implementation, the historical data of 30 consecutive days is used as the input sequence, and each time step contains three feature dimensions: temperature, pressure, and flow rate. The hidden layer dimension of the bidirectional long short-term memory network is set to 256, and the long-term dependence relationship of time-series features is captured through forward and backward directions. An attention mechanism is introduced in the time-series feature channel to model the importance of features at different time steps. The attention weights are calculated based on the correlation between the current time step and historical time steps, and the historical data with higher correlation obtains a larger weight coefficient.

[0032] In the feature fusion stage, the fusion weights are calculated based on the data distribution characteristics of static features and time-series features. First, the feature vectors of the two channels are normalized, and then the weight contributions of each in the fusion process are determined by calculating the feature variances. In practice, the fusion weight of the static feature channel is usually set to 0.3, and the fusion weight of the time-series feature channel is set to 0.7. The residual connection mechanism is used to superimpose the original features and the fusion features, and the batch normalization layer is used to stabilize the feature distribution.

[0033] The model training adopts a two-stage strategy. In the first stage, the Adam optimizer is used for pre-training, the initial learning rate is set to 0.001, the number of training epochs is 200, and the learning rate is reduced by 10% every 50 epochs. In the second stage, fine-tuning training is carried out, the learning rate is reduced to 0.0001, and at the same time, a Dropout layer with a random dropout rate of 0.3 is introduced to prevent overfitting. The mean squared error is used as the loss function during the training process, and the training stops when the loss on the validation set has not decreased for 5 consecutive epochs.

[0034] In the secure aggregation stage, the edge nodes use the Paillier homomorphic encryption system to encrypt the locally trained model parameters. The key length is set to 2048 bits to ensure the encryption strength. After receiving the encrypted parameters, the central server node calculates the aggregation weights based on the data volume and data quality of each node. Nodes with a larger data volume and more uniform distribution obtain higher aggregation weights. After completing the weight aggregation in the encrypted domain, the private key is used to decrypt to obtain the final global model parameters.

[0035] The solution of this application can: Through the dual-channel architecture of the static feature channel and the time-series feature channel, combined with the attention mechanism and the feature fusion strategy, it can fully explore the correlation between data features and improve the accuracy and generalization ability of the prediction model. Adopting a two-stage training strategy and introducing a dropout mechanism can effectively prevent the model from overfitting and improve the robustness and adaptability of the model in actual application scenarios. The secure aggregation scheme based on homomorphic encryption can effectively aggregate the model parameters while protecting the data privacy of the edge nodes, ensuring both data security and the improvement of model performance.

[0036] In an alternative embodiment, the edge computing node receives the global model parameters and updates the basic prediction model; the central server node extracts the prediction experience of each edge computing node through a knowledge distillation mechanism to form a knowledge prior, and distributes the knowledge prior to the edge computing nodes to enhance the prediction ability of the basic prediction model, including: The edge computing node receives the global model parameters, introduces an adaptive fusion factor to perform weighted fusion on the global model parameters and local model parameters, where the value range of the adaptive fusion factor is from 0 to 1; imports the fused model parameters into the basic prediction model to obtain an updated basic prediction model; the edge computing node uses the updated basic prediction model to predict local data and generate a prediction result; evaluates the accuracy of the prediction result based on a local validation set, and dynamically adjusts the value of the adaptive fusion factor according to the accuracy; The central server node obtains the output probability distribution of each edge computing node through a knowledge distillation mechanism and uses the output probability distribution as prediction experience; extracts the intermediate layer features of each edge computing node model, where the intermediate layer features include the output of the model hidden layer and the attention weight; the central server node constructs a prediction rule based on the prediction experience and the intermediate layer features, uses a feature dependency graph to characterize the association relationship between the prediction rules, and converts the prediction rules and the feature dependency graph into a knowledge prior; The central server node distributes the knowledge prior to each edge computing node, and the edge computing node converts the knowledge prior into model constraint conditions and applies the model constraint conditions to the updated basic prediction model; the edge computing node constructs a composite loss function based on the model constraint conditions, where the composite loss function includes a cross-entropy loss term and a knowledge distillation loss term; uses the composite loss function to optimize and train the updated basic prediction model to improve the prediction ability of the basic prediction model.

[0037] In a federated learning system in an edge computing environment, the edge computing node and the central server node work together to improve the model performance through knowledge distillation. The specific implementation process is as follows: First, after receiving the global model parameters, the edge computing node introduces an adaptive fusion factor for parameter fusion. Taking the quality inspection of the production line in a smart factory as an example, after a certain edge node receives the global model parameters, it sets the initial adaptive fusion factor to 0.5 and performs weighted averaging on the global model parameters and local model parameters according to this ratio. The fused parameters are imported into the basic prediction model for defect detection of product images on the production line. In practical applications, the local validation dataset is used to evaluate the model performance. When the prediction accuracy is lower than the preset threshold, the adaptive fusion factor is dynamically adjusted. For example, when the accuracy drops below 85%, the fusion factor is adjusted to 0.7 to increase the weight of global knowledge.

[0038] Second, the central server node collects the model output information of each edge node. In the actual application scenario, the server obtains the prediction probability distribution of the edge node model for the product images of the same batch as the prediction experience. At the same time, it extracts the intermediate layer features of the model, including the output feature map of the convolutional layer and the weight distribution of the attention module. Based on the collected information, the central node constructs a prediction rule library. For example, for the detection of scratches on the product surface, according to the prediction results of multiple edge nodes, rules such as "when the texture feature intensity of a certain area in the image exceeds the set threshold and the attention weight is concentrated in this area, there may be a scratch defect in this area" are summarized. The association between different rules is represented by a feature dependency graph to form a structured knowledge prior.

[0039] Finally, after receiving the knowledge prior, the edge computing node converts it into specific model constraints. In practical applications, the constraint conditions are reflected in the guidance of the model prediction process. A composite loss function containing cross-entropy loss and knowledge distillation loss is constructed, where the cross-entropy loss ensures the fitting ability of the model to local data, and the knowledge distillation loss guides the model to learn prior knowledge. By optimizing the training process, the model can maintain its local prediction performance while learning from the experience of other nodes to improve its generalization ability.

[0040] The solution of this application can: By introducing an adaptive fusion mechanism, it realizes the dynamic balance between global knowledge and local experience, improves the adaptability of the model to different scenarios, and avoids performance loss during the knowledge fusion process. Based on knowledge distillation, it extracts the prediction experience of each node, converts the scattered knowledge into unified prior information, breaks through the limitation of directly sharing model parameters in traditional federated learning, and protects data privacy. Using a composite loss function to guide model training, while maintaining local prediction accuracy, it integrates group knowledge, improves the generalization performance of the model, and reduces the risk of overfitting.

[0041] In an alternative embodiment, the edge computing node collects real-time device corrosion data, inputs the real-time device corrosion data into a basic prediction model that integrates the prior knowledge, and obtains a corrosion rate prediction value and the remaining life of the device; the central server node receives the remaining life of the device uploaded by multiple edge computing nodes, and generates a comprehensive life evaluation report through a swarm intelligence evaluation module, including: The edge computing node collects electrochemistry parameters including potential, current density, and polarization resistance, environment parameters including temperature, pressure, and pH value, and material parameters including stress and strain, and uses the electrochemistry parameters, the environment parameters, and the material parameters as real-time device corrosion data; The edge computing node constructs a feature dependency matrix according to the prior knowledge, uses the feature dependency matrix as an attention weight, weights the importance of features of the real-time device corrosion data to obtain weighted features; inputs the weighted features into the basic prediction model, calculates a corrosion rate prediction value based on historical corrosion rules; establishes a damage accumulation model according to the corrosion rate prediction value and the material performance curve, calculates the remaining life of the device, and uploads the remaining life of the device, the prediction confidence level, and the corrosion parameters to the central server node; The central server node receives the remaining life of the device, the prediction confidence level, and the corrosion parameters uploaded by multiple edge computing nodes; the central server node groups edge computing nodes with a working condition similarity higher than a preset similarity threshold to obtain a device group; based on the prediction confidence level, weights and fuses the remaining life of the device of each edge computing node to obtain the life distribution characteristics of the device group; The central server node calculates the deviation between the remaining life of the device of each edge computing node in the device group and the life distribution characteristics, and marks edge computing nodes with a deviation exceeding the preset range as abnormal nodes; the central server node analyzes the corrosion parameters of the abnormal nodes to identify the reasons for the abnormality; generates a comprehensive life evaluation report according to the life distribution characteristics and the reasons for the abnormality.

[0042] The edge computing node first collects multi-dimensional corrosion data during the operation of the device through a sensor network. The collection of electrochemistry parameters includes measuring potential data using a reference electrode, measuring current density using linear polarization method, and obtaining polarization resistance through electrochemical impedance spectroscopy. In the collection of environment parameters, temperature is measured by a thermocouple sensor, pressure is obtained by a pressure transmitter, and pH value is detected by an online pH meter. In terms of material parameters, stress is measured by a strain gauge, and strain is obtained by a displacement sensor. The sampling frequency can be set to once every 5 minutes.

[0043] Preprocess the collected raw data, including outlier removal, data normalization, etc. Construct a feature dependency matrix based on expert knowledge, such as the correlation weights between parameters like temperature and pH value, potential and current density. Use this matrix as the weight coefficient of the attention mechanism to weight the importance of the input features. For example, when temperature has a greater impact on the corrosion rate, assign a higher weight such as 0.8, and for pH value with a secondary impact, assign a weight of 0.6.

[0044] The weighted features are input into a pre-trained basic prediction model. This model is trained with historical corrosion data and can adopt a deep learning network structure. The model outputs the predicted corrosion rate value, such as 0.2 mm / year. Establish a damage accumulation model in combination with the stress-strain curve of the material to calculate the remaining life of the equipment. At the same time, calculate the prediction confidence, which can be evaluated using the variance of the model output. The edge node uploads the calculation results to the central server.

[0045] After the central server receives the data uploaded by multiple edge nodes, it first groups them based on the similarity of working conditions. The similarity of working conditions is calculated by the Euclidean distance of parameters such as temperature and pressure, and a threshold such as 0.85 is set for grouping. For each equipment group, perform weighted averaging based on the prediction confidence of each node to obtain the life distribution characteristics of the group.

[0046] By calculating the deviation between the predicted life of each node and the group distribution, mark the nodes with deviations exceeding the preset value (such as 2 standard deviations) as abnormal. Analyze the corrosion parameters of the abnormal nodes, such as abnormal increase in temperature, drastic fluctuation in pH value, etc., to identify the specific reasons. Finally, generate an evaluation report including the group life distribution, abnormal equipment, and cause analysis.

[0047] The solution of this application can: Through real-time collection of multi-dimensional corrosion data by edge nodes and combined with prior knowledge for prediction, it improves the accuracy and reliability of corrosion rate prediction, making the equipment life assessment more accurate. Adopt a swarm intelligence assessment method, group and analyze equipment with similar working conditions, and identify abnormal equipment through statistical features, avoiding the possible deviation of single-point prediction and enhancing the credibility of the assessment results. The system realizes the full-process automation from data collection, edge computing to central analysis, and gives an interpretable analysis of abnormal reasons, effectively guiding the maintenance decision-making of industrial equipment and reducing the equipment failure risk.

[0048] In an alternative embodiment, inputting the weighted features into a basic prediction model, calculating the predicted corrosion rate value based on historical corrosion laws; establishing a damage accumulation model according to the predicted corrosion rate value and the material performance curve, calculating the remaining life of the equipment, and uploading the remaining life of the equipment, prediction confidence, and corrosion parameters to the central server node includes: Input the weighted features into a basic prediction model, where the basic prediction model includes three feature extraction sub-networks that respectively extract features from electrochemical parameters, environmental parameters, and material parameters to obtain corresponding local features; integrate the local features into a unified feature vector through a cross-modal fusion layer, and input the unified feature vector into a fully connected layer to obtain an initial corrosion rate; correct the initial corrosion rate according to the corrosion development trend in the initial corrosion rate to obtain a final corrosion rate; Extract the yield strength, fracture toughness, and fatigue limit from the performance curve of the equipment material as damage thresholds; substitute the damage thresholds and the final corrosion rate into a damage accumulation equation, where the damage accumulation equation includes a corrosion depth term, a stress-temperature coupling term, and a material deterioration term determined by the final corrosion rate; calculate the current cumulative damage amount based on the damage accumulation equation, and calculate the remaining life of the equipment according to the difference between the current cumulative damage amount and the damage critical value and the final corrosion rate; Establish a confidence evaluation model based on the prediction error of the final corrosion rate and the uncertainty analysis of the damage accumulation equation, calculate the probability distribution of the prediction result through the confidence evaluation model, and determine the prediction confidence of the remaining life of the equipment; analyze the parameters affecting the prediction result according to the confidence evaluation model, and determine the parameters with an influence degree on the prediction result higher than the set prediction threshold as corrosion parameters; upload the remaining life of the equipment, the prediction confidence, and the corrosion parameters to a central server.

[0049] A method for predicting the remaining life of equipment based on historical corrosion rules. This method first inputs weighted features into a basic prediction model for processing. The basic prediction model includes three feature extraction sub-networks, which are respectively used to process electrochemical parameters, environmental parameters, and material parameters.

[0050] The electrochemical parameter feature extraction sub-network receives parameters such as electrode potential, polarization current density, and anode / cathode polarization curve as inputs. This sub-network adopts a multi-layer convolution structure with a convolution kernel size of 3×3 and a stride of 1. Each layer is followed by a BatchNormalization layer and a ReLU activation function. Extract local features from electrochemical parameters through convolution operations to obtain a 128-dimensional feature vector.

[0051] The environmental parameter feature extraction sub-network processes environmental data such as temperature, pressure, pH value, and chloride ion concentration. The network structure includes a fully connected layer and a dropout layer, and the dropout rate is set to 0.3. The number of neurons in the fully connected layer is 256, 128, and 64 in sequence, and finally outputs a 64-dimensional environmental feature vector.

[0052] The material parameter feature extraction sub-network analyzes parameters such as material composition, grain size, and surface roughness. It adopts a residual network structure, which contains multiple residual blocks, and each residual block consists of two convolutional layers. Finally, a 96-dimensional material feature vector is obtained through global average pooling.

[0053] The cross-modal fusion layer uses an attention mechanism to fuse the three feature vectors. First, the correlation weights between the feature vectors are calculated, and then weighted summation is performed based on the weights to obtain a 256-dimensional unified feature vector. This feature vector is input into two fully connected layers, which contain 128 and 64 neurons respectively, and finally the initial corrosion rate prediction value is output.

[0054] When performing trend correction on the initial corrosion rate, first extract the corrosion rate data for the most recent 30 days, and use the sliding window method to calculate the rate change trend. When the trend term exceeds the preset threshold of 0.15, the initial prediction value is multiplied by the trend coefficient for correction to obtain the final corrosion rate.

[0055] When establishing the damage accumulation model, first extract the key parameters from the material property curve. Taking a certain carbon steel as an example, its yield strength is 235 MPa, fracture toughness is 120 MPa·m1 / 2, and fatigue limit is 157 MPa. These parameters are input into the damage accumulation equation as damage thresholds.

[0056] The damage accumulation equation considers three aspects: corrosion depth, stress-temperature coupling effect, and material degradation. The corrosion depth is determined by the product of the final corrosion rate and time. The stress-temperature coupling term reflects the influence of working stress and temperature on material properties. The material degradation term characterizes the aging process of the material itself.

[0057] Based on the damage accumulation equation, the current cumulative damage amount is calculated and compared with the critical damage value of 1.0. Combining with the final corrosion rate, the time required to reach the critical damage can be predicted, which is the remaining life of the equipment.

[0058] The confidence evaluation model adopts a Bayesian neural network framework and performs uncertainty analysis through the Monte Carlo dropout method. The prediction results are sampled 1000 times, and the probability distribution is obtained through statistical analysis. The width of the confidence interval is taken as a measure of the prediction confidence. At the same time, the sensitivity of each input parameter to the prediction result is analyzed, and the parameters with a sensitivity greater than 0.1 are determined as key corrosion parameters. Finally, the remaining life of the equipment, prediction confidence, and corrosion parameters are uploaded to the central server.

[0059] The solution of this application can: This method processes different types of parameters through three feature extraction sub-networks, fully utilizes the feature information of various types of data, and improves the accuracy and comprehensiveness of feature extraction. By adopting a cross-modal fusion layer and a trend correction mechanism, it effectively integrates multi-source heterogeneous data and takes into account the dynamic characteristics of corrosion development, making the prediction results more in line with the actual situation. The confidence evaluation model based on the Bayesian framework not only gives the prediction results, but also provides reliability analysis and key parameter identification, providing an important basis for equipment maintenance decision-making.

[0060] In an optional implementation manner, the edge computing node optimizes the basic prediction model according to the comprehensive life assessment report, and at the same time feeds back the optimized prediction effect to the knowledge prior through the knowledge distillation mechanism to realize the continuous evolution of the basic prediction model. The collaborative prediction closed loop of prediction-evaluation-optimization includes: The edge computing node receives the comprehensive life assessment report, constructs a model loss function based on the population life distribution characteristics in the comprehensive life assessment report. The model loss function includes a deviation penalty term and a distribution regularization term, and the weight of the deviation penalty term is proportional to the degree of deviation of the prediction result from the population life distribution characteristics; the edge computing node adjusts the attention weights in the feature dependence matrix according to the abnormal reasons identified in the comprehensive life assessment report, and raises the attention weights corresponding to the corrosion parameters causing the prediction deviation above a preset weight threshold; The edge computing node uses the model loss function and the feature dependence matrix to optimize and train the basic prediction model, extracts the prediction output distribution of the optimized model as soft label knowledge, and the soft label knowledge contains the response modes of the model to different combinations of corrosion parameters; at the same time, records the feature activation modes of the intermediate layer of the model, and the feature activation modes are used to characterize the deep correlation between the corrosion parameters; calculates the feature importance scores based on the soft label knowledge and the feature activation modes, and constructs a knowledge graph for the parameter combinations with the feature importance scores exceeding the preset parameter threshold; The edge computing node performs structured encoding on the knowledge graph to obtain encoded knowledge, and conducts credibility evaluation on the encoded knowledge and the original knowledge prior; selects the retained knowledge according to the results of the credibility evaluation, and updates the retained knowledge to the knowledge prior in an incremental learning manner; the edge computing node uses the updated knowledge prior for the next round of prediction process, and realizes the sharing of prediction experience between different edge computing nodes through the knowledge prior; The edge computing node uploads the prediction results of the optimized model to the central server node for population evaluation, continues to optimize the basic prediction model according to the evaluation results and updates the knowledge prior, and realizes the continuous evolution of the basic prediction model through the iterative process of prediction-evaluation-optimization.

[0061] The edge computing node first receives a comprehensive life assessment report, which contains the life distribution characteristic data of the device population. By analyzing the life distribution data in the report, a corresponding model loss function is constructed. This loss function consists of two key components: a deviation penalty term and a distribution regularization term. The weight setting of the deviation penalty term is positively correlated with the degree to which the prediction result deviates from the life distribution characteristics of the population. For example, when the difference between the predicted life and the actual life exceeds 20%, the corresponding penalty weight will be increased to more than 0.8.

[0062] In terms of processing feature dependency relationships, the edge computing node dynamically adjusts the attention weights in the feature dependency relationship matrix according to the abnormal causes identified in the assessment report. Specifically, when it is found that certain corrosion parameters cause prediction deviations, the corresponding attention weights are increased to above a preset threshold. For example, when it is found that the temperature parameter has a significant impact on corrosion prediction, its attention weight will be adjusted to more than 0.7.

[0063] In the model optimization training session, the edge computing node uses the adjusted loss function and feature dependency relationship matrix to optimize the basic prediction model. During the optimization process, the response patterns of the model to different combinations of corrosion parameters are extracted as soft label knowledge. At the same time, the feature activation patterns of each layer of the model are recorded to characterize the deep correlation between corrosion parameters. For example, the feature activation intensity when the temperature and humidity parameters act together is significantly higher than that when a single parameter acts.

[0064] For knowledge extraction and update, the edge computing node calculates the feature importance scores based on the soft label knowledge and feature activation patterns. When the importance score of a certain parameter combination exceeds the preset threshold, it is constructed as a knowledge graph node. For example, when the importance score of the temperature and pressure parameter combination exceeds 0.8, this combination will be included in the knowledge graph.

[0065] In the knowledge sharing and update mechanism, the edge computing node performs structured encoding on the knowledge graph and conducts a credibility assessment with the existing knowledge prior. The credibility assessment is comprehensively judged based on indicators such as prediction accuracy and sample coverage. When the credibility of the new knowledge exceeds 0.85, it is updated to the knowledge prior library through incremental learning.

[0066] Finally, the prediction results of the optimized model are uploaded to the central server for population assessment. The assessment indicators include multiple dimensions such as prediction accuracy and prediction stability. Based on the assessment results, the edge computing node continues to optimize the model and update the knowledge prior, forming a complete prediction - assessment - optimization closed loop.

[0067] The solution of this application can: By constructing a loss function that takes into account the characteristics of the population life distribution and a dynamically adjusted feature dependence matrix, the accuracy and robustness of the model prediction are significantly improved. Practice shows that the prediction accuracy of the optimized model has increased by more than 25%, and its adaptability to abnormal working conditions has been greatly improved. Based on the knowledge extraction mechanism of soft label knowledge and feature activation patterns, a deep understanding of complex corrosion mechanisms and knowledge precipitation are achieved. Through the construction and update of the knowledge graph, the system can continuously accumulate and optimize prediction experience, enabling the prediction model to have the ability to continuously evolve. By adopting a distributed edge computing architecture and a knowledge sharing mechanism, experience sharing and collaborative optimization among multiple edge nodes are realized. This distributed learning architecture not only improves the overall prediction performance of the system, but also significantly reduces the computational load of a single node, enhancing the scalability and practicality of the system.

[0068] Figure 2 FIG. is a schematic structural diagram of the device corrosion thinning life prediction system according to an embodiment of the present invention, as Figure 2 shown, the system includes: A first unit for deploying edge computing nodes at multiple device sites, collecting device corrosion data through the edge computing nodes, where the device corrosion data includes wall thickness data, temperature data, pressure data, medium data, and flow velocity data; encrypting the device corrosion data based on the differential privacy algorithm of the edge computing nodes to generate an encrypted data set; setting a central server node, and the central server node establishes a federated learning network topology with the edge computing nodes through a secure communication channel, and configures a model aggregation module at the central server node for coordinating the federated training of the edge computing nodes; A second unit for constructing a basic prediction model with a deep neural network structure at each edge computing node according to the encrypted data set, and performing local training to obtain local model parameters; after receiving the local model parameters, the central server node performs secure aggregation using a homomorphic encryption algorithm to generate global model parameters; the edge computing nodes receive the global model parameters and update the basic prediction model; the central server node extracts the prediction experience of each edge computing node through a knowledge distillation mechanism to form a knowledge prior, and distributes the knowledge prior to the edge computing nodes to enhance the prediction ability of the basic prediction model; The third unit is used for the edge computing node to collect real-time device corrosion data, input the real-time device corrosion data into the basic prediction model that fuses the knowledge prior, and obtain the corrosion rate prediction value and the remaining life of the device; the central server node receives the remaining life of the device uploaded by multiple edge computing nodes, and generates a comprehensive life evaluation report through the swarm intelligence evaluation module; the edge computing node optimizes the basic prediction model according to the comprehensive life evaluation report, and at the same time feeds back the optimized prediction effect to the knowledge prior through the knowledge distillation mechanism, realizing the continuous evolution of the prediction model and forming a collaborative prediction closed loop of prediction-evaluation-optimization.

[0069] In the third aspect of the embodiments of the present invention, a kind of electronic device is provided, including: a processor; a memory for storing processor-executable instructions; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0070] In the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is realized.

[0071] The present invention can be a method, a device, a system and / or a computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0072] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting equipment corrosion and thinning life, characterized in that: include: Deploy edge computing nodes at multiple equipment sites, collect equipment corrosion data through the edge computing nodes, and the equipment corrosion data includes wall thickness data, temperature data, pressure data, medium data, and flow rate data; encrypt the equipment corrosion data based on the differential privacy algorithm of the edge computing node to generate an encrypted data set; set up a central server node, the central server node establishes a federated learning network topology structure with the edge computing node through a secure communication channel, and configures a model aggregation module on the central server node to coordinate the federated training of the edge computing node; According to the encrypted data set, a basic prediction model of a deep neural network structure is constructed at each edge computing node, and local training is performed to obtain local model parameters; after receiving the local model parameters, the central server node adopts a homomorphic encryption algorithm for secure aggregation to generate global model parameters; the edge computing node receives the global model parameters and updates the basic prediction model; the central server node extracts the prediction experience of each edge computing node through a knowledge distillation mechanism to form a priori knowledge, and distributes the priori knowledge to the edge computing nodes to enhance the prediction capability of the basic prediction model; The edge computing node collects real-time equipment corrosion data, inputs the real-time equipment corrosion data into the basic prediction model integrated with the knowledge prior, and obtains the corrosion rate prediction value and the remaining life of the equipment; the central server node receives the remaining life of the equipment uploaded by multiple edge computing nodes, and generates a comprehensive life evaluation report through the swarm intelligence evaluation module; the edge computing node optimizes the basic prediction model according to the comprehensive life evaluation report, and at the same time feeds back the optimized prediction effect to the knowledge prior through the knowledge distillation mechanism, so as to realize the continuous evolution of the prediction model and form a collaborative prediction closed loop of prediction-evaluation-optimization.

2. The method according to claim 1, characterized in that The device corrosion data is encrypted based on the differential privacy algorithm of the edge computing node to generate an encrypted data set; a central server node is set, the central server node establishes a federated learning network topology structure with the edge computing node through a secure communication channel, and a model aggregation module is configured in the central server node to coordinate the federated training of the edge computing node, including: Calculate the local sensitivity value of the equipment corrosion data, and determine the privacy protection strength of each data according to the local sensitivity value; based on the privacy protection strength, inject Laplace noise into the wall thickness data, and inject Gaussian noise into the temperature data, the pressure data, the medium data, and the flow rate data to generate an encrypted data set; Setting a central server node, establishing a secure communication channel between the central server node and the edge computing node, using a homomorphic encryption algorithm to transmit and protect model parameters in the secure communication channel; the central server node receives the encrypted data set transmitted by the edge computing node, analyzes the data distribution characteristics of the encrypted data set, and calculates the data quality score of each edge computing node; The central server node determines the model aggregation weight of each edge computing node according to the data quality score and the data volume of the encrypted data set; performs weighted aggregation on the local model parameters of each edge computing node based on the model aggregation weight to generate global model parameters; distributes the global model parameters to each edge computing node through the secure communication channel, and the edge computing node updates the local model using the global model parameters and the encrypted data set to complete the federated learning training based on differential privacy protection.

3. The method according to claim 1, characterized in that According to the encrypted data set, a basic prediction model of a deep neural network structure is constructed at each edge computing node, and local training is performed to obtain local model parameters; after receiving the local model parameters, the central server node adopts a homomorphic encryption algorithm for secure aggregation to generate global model parameters, including: According to the data features of the encrypted data set, a deep neural network structure is constructed at the edge computing node as a basic prediction model, wherein the deep neural network structure includes a static feature channel and a time series feature channel; The static feature channel uses a three-layer fully connected network to process the wall thickness data and the medium data, and each layer uses a ReLU activation function; the temporal feature channel uses a bidirectional long short-term memory network to process the temperature data, the pressure data and the flow rate data; an attention mechanism is introduced to assign weight coefficients to features at different time steps in the temporal feature channel; the feature fusion weights of the static feature channel and the temporal feature channel are calculated based on feature distribution, and the fused features are processed using residual connections and batch normalization layers to obtain fused features; The basic prediction model is trained in two stages based on the fusion features. In the first stage, the Adam optimizer is used for pre-training, the learning rate is set to 0.001, and the learning rate is reduced by 10% every 50 rounds of training; in the second stage, fine-tuning training is performed, the learning rate is set to 0.0001, and a Dropout layer with a random dropout rate of 0.3 is set to prevent overfitting, so as to obtain local model parameters; The edge computing node uses the public key of the Paillier homomorphic encryption system to encrypt the local model parameters, and transmits the encrypted local model parameters to the central server node; the central server node calculates the aggregation weight of each edge computing node in the encryption domain, securely aggregates the encrypted local model parameters based on the aggregation weight, and uses the private key to decrypt the aggregation result to obtain the global model parameters.

4. The method according to claim 1, characterized in that: The edge computing node receives the global model parameters and updates the basic prediction model; the central server node extracts the prediction experience of each edge computing node through a knowledge distillation mechanism to form a prior knowledge, and distributes the prior knowledge to the edge computing node to enhance the prediction capability of the basic prediction model, including: The edge computing node receives the global model parameters, introduces an adaptive fusion factor to perform weighted fusion on the global model parameters and the local model parameters, and the value range of the adaptive fusion factor is 0 to 1; the fused model parameters are imported into the basic prediction model to obtain an updated basic prediction model; the edge computing node uses the updated basic prediction model to predict local data and generate a prediction result; the accuracy of the prediction result is evaluated based on the local validation set, and the value of the adaptive fusion factor is dynamically adjusted according to the accuracy; The central server node obtains the output probability distribution of each edge computing node through a knowledge distillation mechanism, and uses the output probability distribution as prediction experience; extracts the intermediate layer features of each edge computing node model, and the intermediate layer features include the model hidden layer output and attention weight; the central server node constructs a prediction rule based on the prediction experience and the intermediate layer features, uses a feature dependency graph to characterize the correlation between the prediction rules, and converts the prediction rule and the feature dependency graph into knowledge prior; The central server node distributes the knowledge prior to each of the edge computing nodes, and the edge computing node converts the knowledge prior into model constraints and applies the model constraints to the updated basic prediction model; the edge computing node constructs a composite loss function based on the model constraints, and the composite loss function includes a cross entropy loss term and a knowledge distillation loss term; the composite loss function is used to optimize the training of the updated basic prediction model to improve the prediction ability of the basic prediction model.

5. The method according to claim 1, characterized in that The edge computing node collects real-time equipment corrosion data, inputs the real-time equipment corrosion data into the basic prediction model integrating the prior knowledge, and obtains the corrosion rate prediction value and the remaining life of the equipment; the central server node receives the remaining life of the equipment uploaded by multiple edge computing nodes, and generates a comprehensive life assessment report through the group intelligent assessment module, including: The edge computing node collects electrochemical parameters including potential, current density, and polarization resistance, environmental parameters including temperature, pressure, and pH value, and material parameters including stress and strain, and uses the electrochemical parameters, environmental parameters, and material parameters as real-time equipment corrosion data; The edge computing node constructs a feature dependency matrix based on the prior knowledge, uses the feature dependency matrix as the attention weight, and weights the feature importance of the real-time equipment corrosion data to obtain weighted features; inputs the weighted features into the basic prediction model, and calculates the corrosion rate prediction value based on the historical corrosion law; establishes a damage accumulation model based on the corrosion rate prediction value and the material performance curve, calculates the remaining life of the equipment, and uploads the remaining life of the equipment, prediction confidence and corrosion parameters to the central server node; The central server node receives the remaining life of the equipment, the prediction confidence and the corrosion parameter uploaded by the plurality of edge computing nodes; the central server node groups the edge computing nodes whose working condition similarity is higher than a preset similarity threshold to obtain a device group; based on the prediction confidence, the remaining life of the equipment of each edge computing node is weightedly integrated to obtain the life distribution characteristics of the device group; The central server node calculates the deviation between the remaining life of the device of each edge computing node in the device group and the life distribution characteristics, and marks the edge computing node whose deviation exceeds a preset range as an abnormal node; the central server node analyzes the corrosion parameters of the abnormal node and identifies the cause of the abnormality; and generates a comprehensive life assessment report based on the life distribution characteristics and the cause of the abnormality.

6. The method according to claim 5, characterized in that Inputting the weighted features into the basic prediction model, calculating the corrosion rate prediction value based on the historical corrosion law; establishing a damage accumulation model according to the corrosion rate prediction value and the material performance curve, calculating the remaining life of the equipment, and uploading the remaining life of the equipment, prediction confidence and corrosion parameters to the central server node includes: The weighted features are input into a basic prediction model, the basic prediction model includes three feature extraction sub-networks, the three feature extraction sub-networks respectively extract features of electrochemical parameters, environmental parameters and material parameters to obtain corresponding local features; the local features are integrated into a unified feature vector through a cross-modal fusion layer, and the unified feature vector is input into a fully connected layer to obtain an initial corrosion rate; the initial corrosion rate is corrected according to the corrosion development trend in the initial corrosion rate to obtain a final corrosion rate; Extracting yield strength, fracture toughness and fatigue limit from the performance curve of the equipment material as damage threshold; substituting the damage threshold and the final corrosion rate into a damage accumulation equation, wherein the damage accumulation equation includes a corrosion depth term, a stress-temperature coupling term and a material degradation term determined by the final corrosion rate; calculating the current cumulative damage amount based on the damage accumulation equation, and calculating the remaining life of the equipment according to the difference between the current cumulative damage amount and the damage critical value and the final corrosion rate; A confidence assessment model is established based on the prediction error of the final corrosion rate and the uncertainty analysis of the damage accumulation equation. The probability distribution of the prediction result is calculated through the confidence assessment model to determine the prediction confidence of the remaining life of the equipment. The parameters that affect the prediction result are analyzed according to the confidence assessment model, and the parameters whose influence on the prediction result is higher than the set prediction threshold are determined as corrosion parameters. The remaining life of the equipment, the prediction confidence and the corrosion parameters are uploaded to the central server.

7. The method according to claim 1, characterized in that The edge computing node optimizes the basic prediction model according to the comprehensive life assessment report, and feeds back the optimized prediction effect to the knowledge prior through the knowledge distillation mechanism, so as to realize the continuous evolution of the basic prediction model and form a prediction-assessment-optimization collaborative prediction closed loop, including: The edge computing node receives the comprehensive life assessment report, and constructs a model loss function based on the group life distribution characteristics in the comprehensive life assessment report, wherein the model loss function includes a deviation penalty term and a distribution regularization term, and the weight of the deviation penalty term is proportional to the degree to which the prediction result deviates from the group life distribution characteristics; the edge computing node adjusts the attention weight in the feature dependency matrix according to the abnormal cause identified in the comprehensive life assessment report, and increases the attention weight corresponding to the corrosion parameter that causes the prediction deviation to above a preset weight threshold; The edge computing node optimizes and trains the basic prediction model using the model loss function and the feature dependency matrix, extracts the predicted output distribution of the optimized model as soft label knowledge, and the soft label knowledge includes the response mode of the model to different corrosion parameter combinations; records the feature activation mode of the middle layer of the model, and the feature activation mode is used to characterize the deep correlation between the corrosion parameters; calculates the feature importance score based on the soft label knowledge and the feature activation mode, and constructs the parameter combination whose feature importance score exceeds the preset parameter threshold into a knowledge graph; The edge computing node performs structured encoding on the knowledge graph to obtain encoded knowledge, and performs credibility assessment on the encoded knowledge and the original knowledge prior; selects the retained knowledge according to the result of the credibility assessment, and updates the retained knowledge to the knowledge prior by using an incremental learning method; the edge computing node uses the updated knowledge prior for the next round of prediction process, and realizes the sharing of prediction experience among different edge computing nodes through the knowledge prior; The edge computing node uploads the prediction results of the optimized model to the central server node for group evaluation, continues to optimize the basic prediction model and updates the knowledge prior according to the evaluation results, and realizes the continuous evolution of the basic prediction model through the iterative process of prediction-evaluation-optimization.

8. Equipment corrosion thinning life prediction system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to deploy edge computing nodes at multiple equipment sites, collect equipment corrosion data through the edge computing nodes, and the equipment corrosion data includes wall thickness data, temperature data, pressure data, medium data, and flow rate data; encrypt the equipment corrosion data based on the differential privacy algorithm of the edge computing node to generate an encrypted data set; set a central server node, the central server node establishes a federated learning network topology structure with the edge computing node through a secure communication channel, and configures a model aggregation module in the central server node to coordinate the federated training of the edge computing node; The second unit is used to construct a basic prediction model of a deep neural network structure at each edge computing node according to the encrypted data set, and perform local training to obtain local model parameters; after receiving the local model parameters, the central server node adopts a homomorphic encryption algorithm to perform secure aggregation to generate global model parameters; the edge computing node receives the global model parameters and updates the basic prediction model; the central server node extracts the prediction experience of each edge computing node through a knowledge distillation mechanism to form a priori knowledge, and distributes the priori knowledge to the edge computing nodes to enhance the prediction capability of the basic prediction model; The third unit is used for the edge computing node to collect real-time equipment corrosion data, input the real-time equipment corrosion data into the basic prediction model integrated with the knowledge prior, and obtain the corrosion rate prediction value and the remaining life of the equipment; the central server node receives the remaining life of the equipment uploaded by multiple edge computing nodes, and generates a comprehensive life evaluation report through the swarm intelligence evaluation module; the edge computing node optimizes the basic prediction model according to the comprehensive life evaluation report, and at the same time feeds back the optimized prediction effect to the knowledge prior through the knowledge distillation mechanism, so as to realize the continuous evolution of the prediction model and form a collaborative prediction closed loop of prediction-evaluation-optimization.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images