Part quality dynamic early warning system based on hierarchical federal learning framework

Abstract

The invention discloses a part quality dynamic early warning system based on a hierarchical federal learning framework, and the system comprises a part quality dynamic early warning platform which comprises a data processing unit, a part quality prediction unit and a dynamic early warning unit, and the data processing unit is used for preprocessing data; the part quality prediction unit is used for performing part life prediction by utilizing a CNN-Informer model on the basis of the data preprocessed by the data processing unit, and optimizing model hyper-parameters by taking a minimum symmetric average absolute percentage error SMAPE as a target function and adopting an improved big sugarcane-mouse algorithm; the dynamic early warning unit is used for deciding whether to give an early warning or not based on the prediction result; the system security protection platform is used for providing computing support through the cloud computing platform, verifying a user and guaranteeing data security; the enterprise user interaction platform is used for receiving the early warning result of the dynamic early warning unit and giving user feedback suggestions; the method can effectively improve the comprehensiveness and accuracy of part life prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine learning and intelligent industry, and particularly relates to a part quality dynamic early warning system based on a hierarchical federated learning framework. BACKGROUND

[0002] In modern industrial manufacturing, the stability of part quality is crucial for the overall performance of products and the economic benefits of enterprises. With the digital transformation of manufacturing, a large number of sensors are deployed on production equipment to collect real-time data such as temperature, vibration, pressure, and other types of data. However, how to efficiently utilize these massive amounts of data to accurately predict part quality and timely warn of potential problems has become a key challenge facing the industry.

[0003] Traditional part quality prediction methods often model and analyze based on single-enterprise internal data. This approach, on the one hand, due to the limitations of data volume and data diversity, results in insufficient generalization ability of the model, making it difficult to accurately cope with complex and variable production conditions. On the other hand, with the development of the industrial Internet, cross-enterprise collaboration is increasingly frequent, and there is a need for data sharing and cooperation between enterprises, but direct data sharing faces serious data security and privacy protection problems.

[0004] Federated learning, as an emerging distributed machine learning technology, allows multiple participants to collaboratively train models without sharing raw data, providing a new approach to solving cross-enterprise data cooperation problems. However, in practical applications, existing federated learning frameworks still have many shortcomings when dealing with complex tasks such as part quality dynamic early warning. For example, the dynamic optimization capability of the model is limited, making it difficult to adapt to changing production conditions and data characteristics; real-time response efficiency is not high, making it difficult to meet the demand for rapid part quality early warning; and the security and privacy protection mechanism is difficult to provide sufficient reliable protection against increasingly complex network attack methods. Therefore, developing a part quality dynamic early warning system with model dynamic optimization, real-time response efficiency, and strong privacy protection is of great significance for improving the intelligent level of industrial manufacturing and ensuring cross-enterprise data security cooperation. SUMMARY

[0005] The purpose of the present application is to provide a part quality dynamic early warning system based on a hierarchical federated learning framework that can improve prediction accuracy.

[0006] Technical solution: The part quality dynamic early warning system based on a hierarchical federated learning framework according to the present application comprises:

[0007] The component quality dynamic early warning platform includes a data processing unit, a component quality prediction unit, and a dynamic early warning unit. The data processing unit is used to preprocess the data; the component quality prediction unit is used to predict the component lifespan based on the preprocessed data; and the dynamic early warning unit is used to decide whether to issue an early warning based on the prediction results.

[0008] The system security protection platform is used to provide computing support through the cloud computing platform, verify users, and ensure data security;

[0009] The enterprise user interaction platform is used to receive the warning results from the dynamic warning unit and provide user feedback and suggestions.

[0010] Furthermore, the data processing unit is used to preprocess the data based on a sensor network consisting of temperature, vibration, and pressure sensors.

[0011] Furthermore, the part quality prediction unit is used to predict part life based on the data preprocessed by the data processing unit using the CNN-Informer model, and uses the minimum symmetric mean absolute percentage error (SMAPE) as the objective function, and adopts an improved big sugarcane rat algorithm to optimize the model hyperparameters.

[0012] Furthermore, the dynamic early warning unit is used to set an alarm threshold based on the prediction results and to determine whether to issue an early warning through a dynamic optimization mechanism.

[0013] Furthermore, the process of establishing the CNN-Informer prediction model in the part quality prediction unit includes the following steps:

[0014] (1.1) Extract the temporal and spatial features of temperature, vibration, and pressure data using the convolutional layers in a convolutional neural network (CNN);

[0015] (1.2) Input the data into the Informer model, assign higher weights to the temperature, vibration and pressure data features, reduce the dimension of the input features, thereby extracting historical moment information of the temperature, vibration and pressure data features of the Informer model and generating prediction results.

[0016] Further, the output of the CNN described in step (1.1) is represented as:

[0017]

[0018] In the formula, O I For output; W I and b I represents the weights and biases of the I-th convolutional layer, respectively; * represents the convolution operation; f is the activation function.

[0019] Further, the Informer model in step (1.2) adopts a probabilistic sparse self-attention mechanism to screen key queries to reduce computational complexity, and the expression of the probabilistic sparse self-attention mechanism is as follows:

[0020]

[0021] In the formula, Atten is a probabilistic sparse self-attention; Q is a query matrix; K is a key matrix; V is a value matrix; d is an input dimension; is only the first several Query sparse matrices; K T is the transpose of the K matrix;

[0022] The encoder of the Informer model utilizes an attention distillation mechanism to reduce features in the time dimension to process extended sequences, and the distillation process from the jth layer to the (j+1)th layer can be expressed as:

[0023]

[0024] In the formula, is the distillation process from the lth layer to the (l+1)th layer at time t; [·] AB is a multi-head sparse self-attention operation; Convld(·) is a one-dimensional convolution operation; ELU(·) is an activation function; MaxPool(·) is a maximum pooling operation;

[0025] The decoder of the Informer model performs cross-attention at each layer, utilizes the keys and values of the encoder to interact with the decoder queries, and generates predictions; the decoding expression is as follows:

[0026]

[0027] In the formula, Concat(·) is a concatenation operation; is a start token; is a placeholder for the prediction sequence using zero values.

[0028] Further, the part quality prediction unit takes the smallest symmetric average absolute percentage error (SMAPE) as the objective function, and adopts an improved giant hamster algorithm to optimize model hyperparameters, including:

[0029] (2.1) Initialize the population and parameters, and set the giant hamster population matrix as follows:

[0030]

[0031] Where X is the giant hamster population; x i,j is the value of the jth dimension proposed by the ith giant hamster; n is the total number of giant hamster population members; d is the number of decision variables;

[0032] (2.2) The randomly generated large mouse positions are brought into the objective function for evaluation, with each large mouse representing a candidate solution to the optimization problem, and the objective function is formulated as follows:

[0033]

[0034] where y i is the true value; is the predicted value;

[0035] (2.3) Update the population members, determine the new positions of the remaining mice in the search space according to the dominant male position, and update the optimal number of encoder layers, decoder layers, and attention heads, with the position formula as follows:

[0036]

[0037]

[0038] α = 2 × r × rand - r

[0039] β = 2 × r × μ - r

[0040] where X i is the new position of the i-th large mouse, is the value of its j-th dimension; x i,j is the value of the current position of the large mouse; x k,j is the dominant male position; x m,j is the randomly selected female mouse position; is the latest objective function value; F i is the current objective function value; F xk is the objective function value of the dominant male; C is a random number defined within the problem space boundary; r is the influence of simulating a rich food source; C iter is the current iteration number; MaxIter is the maximum number of iterations; α is the coefficient simulating the reduction of food sources; β is the coefficient that promotes the transfer of large mice to other available rich food sources within the breeding area; μ is a random number between [1, 4];

[0041] (2.4) Update the search space, develop areas with rich food sources according to the female mouse positions, and update the optimal number of encoder layers, decoder layers, and attention heads, with the position formula as follows:

[0042]

[0043] To improve this step, the Cauchy mutation strategy is introduced to expand the search scale of the large mouse algorithm, thereby improving the ability to jump out of local optima. The improved formula is as follows:

[0044]

[0045] wherein, is the improved new position; cauchy(0,1) is a standard cauchy distribution function; represents multiplication;

[0046] (2.5) determining whether the termination condition is met, if the termination condition is not met, returning to step (2.3); if the constraint condition is met, going to step (2.6);

[0047] (2.6) outputting the optimal number of encoder layers, the number of decoder layers and the number of attention heads.

[0048] Further, the dynamic early warning unit takes the basic threshold T d and the part life attenuation rate r d As an index to judge whether the system is early warning, give the historical life sequence [L t-n , L t-n+1 , …, L t ], the part life attenuation rate is expressed as follows:

[0049]

[0050] In the formula, ω k is the time attenuation weight; t is the sequence number; is the historical average life, δ is the smoothing constant;

[0051] When the current life L t is less than the basic threshold T d or the part life attenuation rate r d is greater than 0.1, the system issues an early warning; the early warning decision function is expressed as follows:

[0052]

[0053] Further, the dynamic optimization mechanism of the dynamic early warning unit optimizes the threshold by minimizing the loss function, which is expressed as follows:

[0054]

[0055] In the formula, l(·) is the cross-entropy loss function; Alert m is the m-node early warning result; λ is the true label, 1 = early warning is needed, 0 = not needed; λ is the regularization coefficient; Ω(·) is the regularization term to prevent overfitting.

[0056] Beneficial effects: Compared with the prior art, the application can effectively capture long-term dependencies when processing time series data by introducing Informer as an improved model of Transformer, and combining CNN which is good at extracting local features, both of which can take into account the local details and long-term trends of the data at the same time, improving the comprehensiveness and accuracy of part life prediction; the improved big mouse algorithm of the application is used to optimize the model hyperparameters, and through iterative optimization with the minimum symmetric mean absolute percentage error (SMAPE) as the objective function, the more optimal parameter configuration can be accurately found, and the prediction accuracy can be greatly improved; the dynamic optimization mechanism continuously optimizes the early warning threshold by minimizing the loss function, so that the dynamic early warning unit can flexibly adjust the alarm standard according to the actual prediction result, ensuring the timeliness and rationality of the early warning; the application is used in the field of industrial manufacturing, and in the aspects of part quality monitoring and cross-enterprise cooperation, the problems of one-sided prediction and insufficient accuracy of traditional models are solved, the response lag or false alarm caused by fixed early warning threshold is overcome, and data privacy protection is strengthened, providing strong support for cross-enterprise safe cooperation. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The system framework schematic diagram of the embodiment of the application is shown in the figure.

[0058] Figure 2 The CNN-Informer prediction model flowchart of the embodiment of the application is shown in the figure.

[0059] Figure 3 The big mouse algorithm flowchart of the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0060] In order to enable personnel in the technical field to better understand the application scheme, the technical solutions in the embodiments of the application will be described clearly and completely below in combination with the drawings in the embodiments of the application.

[0061] As shown in the figure, the part quality dynamic early warning system based on the hierarchical federated learning framework of the embodiment of the application comprises: Figure 1 The part quality dynamic early warning platform comprises a data processing unit, a part quality prediction unit and a dynamic early warning unit, the data processing unit is used for pre-processing data, the part quality prediction unit is used for part life prediction based on the pre-processed data of the data processing unit, and the dynamic early warning unit is used for determining whether to warn based on the prediction result.

[0062] The system security protection platform is used for providing computing support through the cloud computing platform, verifying the user and protecting data security.

[0063]

[0064] ​The enterprise user interaction platform is used for receiving the early warning result of the dynamic early warning unit and giving user feedback suggestions.

[0065] The data processing unit is configured to preprocess data based on a sensor network composed of temperature, vibration and pressure sensors.

[0066] The part quality prediction unit is configured to predict the life of the part based on the data preprocessed by the data processing unit using a CNN-Informer model, and to use an improved big mouse algorithm to optimize the model hyperparameters with the minimum symmetric mean absolute percentage error (SMAPE) as the objective function.

[0067] The dynamic early warning unit is configured to set an alarm threshold based on the prediction result and determine whether to issue an early warning through a dynamic optimization mechanism.

[0068] Specifically, as shown in Figure 2 The establishment process of the CNN-Informer prediction model includes the following steps:

[0069] (1.1) Extract the time and space features of temperature, vibration and pressure data using the convolution layer in the convolutional neural network (CNN);

[0070] (1.2) Input the data into the Informer model, assign higher weights to important features (temperature, vibration and pressure data), reduce the input feature dimension, extract more critical historical time information of the Informer model, and then generate the prediction result.

[0071] Further, the CNN output in step (1.1) is represented as:

[0072]

[0073] In the formula, O I is the output; W I and b I represent the weight and bias of the Ith convolutional layer, respectively; * represents the convolution operation; f is the activation function, usually the rectified linear unit.

[0074] Further, the Informer model in step (1.2) is represented as:

[0075] (1.2.1) The Informer model uses a probabilistic sparse self-attention mechanism to screen key queries and reduce computational complexity. The expression of the probabilistic sparse self-attention mechanism is as follows:

[0076]

[0077] where Atten is the probabilistic sparse self-attention; Q is the query matrix; K is the key matrix; V is the value matrix; and d is the input dimension; is only the first several Query sparse matrices; K T is the transpose of the K matrix.

[0078] (1.2.2) The encoder of the Informer model utilizes an attention distillation mechanism to reduce the features in the time dimension to process the extended sequence. The distillation process from the jth layer to the (j+1)th layer can be represented as:

[0079]

[0080] where, is the distillation process from the lth layer to the (l+1)th layer at time t; [·] AB is the multi-head sparse self-attention operation; Convld(·) is a one-dimensional convolution operation; ELU(·) is an activation function; and MaxPool(·) is a maximum pooling operation.

[0081] (1.2.3) The decoder of the Informer model performs cross-attention at each layer, utilizing the encoder keys and values to interact with the decoder queries to generate predictions. The decoding expression is as follows:

[0082]

[0083] where Concat(·) is a concatenation operation; is the start token; is a placeholder for the prediction sequence using zero values.

[0084] Further, the improved big mouse algorithm is used to optimize the model hyperparameters: the number of encoder layers, the number of decoder layers, and the number of attention heads, with the symmetric mean absolute percentage error (SMAPE) as the objective function.

[0085] As shown in Figure 3 , the steps of optimizing the model hyperparameters by the improved big mouse algorithm are as follows:

[0086] (2.1) Initialize the population and parameters, and the big mouse population matrix is as follows:

[0087]

[0088] where X is the big mouse population; x i,j is the value of the jth dimension proposed by the ith big mouse; n is the total number of big mouse population members; and d is the number of decision variables.

[0089] (2.2) Bring the randomly generated large mouse positions into the objective function for evaluation, and each large mouse represents a candidate solution to the optimization problem. The objective function formula is as follows:

[0090]

[0091] In the formula, y i is the true value; is the predicted value;

[0092] (2.3) Update the population members, and determine the new positions of the remaining mice in the search space according to the dominant male position, that is, update the optimal number of encoder layers, decoder layers, and attention heads. The position formula is as follows:

[0093]

[0094]

[0095] α = 2 × r × rand - r

[0096] β = 2 × r × μ - r

[0097] Where X i is the new position of the i-th large mouse, is the value of its j-th dimension; x i,j is the value of the current position of the large mouse; x k,j is the dominant male position; x m,j is the randomly selected female mouse position; is the latest objective function value; F i is the current objective function value; F xk is the objective function value of the dominant male; C is a random number defined within the problem space boundary; r is the influence of simulating a rich food source; C iter is the current iteration number; MaxIter is the maximum iteration number; α is the coefficient simulating the decrease of food source; β is the coefficient that promotes the large mouse to move to other available rich food sources in the breeding area; μ is a random number between [1, 4];

[0098] (2.4) Update the search space, and develop the area with rich food sources according to the female mouse position, that is, update the optimal number of encoder layers, decoder layers, and attention heads. The position formula is as follows:

[0099]

[0100] This step is improved by introducing the Cauchy mutation strategy to expand the search scale of the large mouse algorithm, and thus improve the ability to jump out of the local optimum. The improved formula is as follows:

[0101]

[0102] wherein, is the improved new position; cauchy(0,1) is a standard cauchy distribution function; represents multiplication.

[0103] (2.5) determining whether the termination condition is met, if the termination condition is not met, returning to step (2.3); if the constraint condition is met, going to step (2.6);

[0104] (2.6) outputting the optimal number of encoder layers, the number of decoder layers and the number of attention heads.

[0105] Further, the dynamic early warning unit takes the basic threshold T d and the part life attenuation rate r d as the index to judge whether the system needs to be warned, and gives a historical life sequence [L t-n , L t-n+1 , …, L t ], the part life attenuation rate is expressed as follows:

[0106]

[0107] In the formula, ω k is the time attenuation weight; t is the sequence number; is the historical average life, and δ is the smoothing constant.

[0108] When the current life L t is less than the basic threshold T d or the part life attenuation rate r d is greater than 0.1, the system issues a warning; the warning decision function is expressed as follows:

[0109]

[0110] Further, the dynamic optimization mechanism optimizes the threshold by minimizing the loss function, which is expressed as follows:

[0111]

[0112] In the formula, l(·) is the cross-entropy loss function; Alert m is the m-node warning result; λ is the true label, 1=need to warn, 0=don't need; λ is the regularization coefficient; Ω(·) is the regularization term to prevent overfitting.

Claims

1. A dynamic early warning system for part quality based on a hierarchical federated learning framework, characterized in that, include: The component quality dynamic early warning platform includes a data processing unit, a component quality prediction unit, and a dynamic early warning unit. The data processing unit is used to preprocess the data; the component quality prediction unit is used to predict the component lifespan based on the preprocessed data. The dynamic early warning unit is used to decide whether to issue an early warning based on the prediction results; The system security protection platform is used to provide computing support through the cloud computing platform, verify users, and ensure data security; The enterprise user interaction platform is used to receive the warning results from the dynamic warning unit and provide user feedback and suggestions.

2. The dynamic early warning system for part quality based on a hierarchical federated learning framework according to claim 1, characterized in that, The data processing unit is used to preprocess data based on a sensor network consisting of temperature, vibration, and pressure sensors.

3. The dynamic early warning system for part quality based on a hierarchical federated learning framework according to claim 1, characterized in that, The part quality prediction unit is used to predict part life based on the data preprocessed by the data processing unit using the CNN-Informer model, and uses the minimum symmetric mean absolute percentage error (SMAPE) as the objective function, and adopts the improved Big Cane Mouse algorithm to optimize the model hyperparameters.

4. The dynamic early warning system for part quality based on a hierarchical federated learning framework according to claim 1, characterized in that, The dynamic early warning unit is used to set alarm thresholds based on prediction results and to determine whether to issue an early warning through a dynamic optimization mechanism.

5. The dynamic early warning system for part quality based on a hierarchical federated learning framework according to claim 3, characterized in that, The process of establishing the CNN-Informer prediction model in the part quality prediction unit includes the following steps: (1.1) Extract the temporal and spatial features of temperature, vibration, and pressure data using the convolutional layers in a convolutional neural network (CNN); (1.2) Input the data into the Informer model, assign higher weights to the temperature, vibration and pressure data features, reduce the dimension of the input features, thereby extracting historical moment information of the temperature, vibration and pressure data features of the Informer model and generating prediction results.

6. The dynamic early warning system for part quality based on a hierarchical federated learning framework according to claim 5, characterized in that, The output of the CNN described in step (1.1) is represented as follows: In the formula, O I For output; W I and b I represents the weights and biases of the I-th convolutional layer, respectively; * represents the convolution operation; f is the activation function.

7. The dynamic early warning system for part quality based on a hierarchical federated learning framework according to claim 5, characterized in that, The Informer model described in step (1.2) uses a probabilistic sparse self-attention mechanism to filter key queries in order to reduce computational complexity. The expression for the probabilistic sparse self-attention mechanism is as follows: In the formula, Atten represents probabilistic sparse self-attention; Q is the query matrix; K is the key matrix; V is the value matrix; and d is the input dimension. A sparse matrix containing only the first few queries; K T This is the transpose of the K matrix; The encoder of the Informer model utilizes an attention distillation mechanism to reduce features in the time dimension in order to process extended sequences. The distillation process from layer j to layer j+1 can be represented as follows: In the formula, For time t, the distillation process proceeds from layer l to layer l+1; [·] AB For multi-head sparse self-attention operations; Convld(·) is a one-dimensional convolution operation; ELU(·) is an activation function; MaxPool(·) is a max pooling operation; The Informer model's decoder performs cross-attention at each layer, using the encoder's keys and values ​​to query and interact with the decoder to generate predictions; The decoding expression is as follows: In the formula, Concat(·) is the concatenation operation; It is the start marker; Zero values ​​are used as placeholders for the predicted sequence.

8. The dynamic early warning system for part quality based on a hierarchical federated learning framework according to claim 3, characterized in that, The part quality prediction unit uses the minimum symmetric mean absolute percentage error (SMAPE) as the objective function and employs an improved large cane rat algorithm to optimize the model hyperparameters, including: (2.1) Initialize the population and parameters. Assume the population matrix of the giant cane rat is as follows: Where X represents the large sugarcane rat population; x i,j d represents the value of the j-th dimension proposed for the i-th giant cane rat; n is the total number of giant cane rat population members; d is the number of decision variables; (2.2) Substitute the randomly generated positions of the large cane rats into the objective function for evaluation. Each large cane rat represents a candidate solution to the optimization problem. The objective function formula is as follows: In the formula, y i The actual value; This is a predicted value; (2.3) Update the population members. Determine the new positions of the remaining rats in the search space based on the position of the dominant male rat. Update the optimal number of encoder layers, decoder layers, and attention heads. The position formula is as follows: α=2×r×rand-r β=2×r×μ-r Among them, X i Let i be the new position of the i-th large cane rat. Its value in the j-th dimension; x i,j x is the value of the current position of the large cane rat; k,j Position of the dominant male mouse; x m,j The location of the randomly selected female mouse; The latest objective function value; F i This is the current objective function value; C represents the objective function value of the dominant male rat; C is a random number defined within the problem space boundary; r represents the effect of simulating abundant food sources; C iter α is the current iteration number; MaxIter is the maximum iteration number; α is the coefficient that simulates the reduction of food sources; β is the coefficient that prompts the large cane rat to move to other available abundant food sources in the breeding area; μ is a random number between [1,4]. (2.4) Update the search space. Develop food-rich areas based on the location of the female mouse, and update the optimal number of encoder layers, decoder layers, and attention heads. The location formula is as follows: This step is improved by introducing a Cauchy mutation strategy to expand the search scale of the Big Cane Rat algorithm, thereby enhancing its ability to escape local optima. The improved formula is as follows: in, The new position is the improved position; cauchy(0,1) is the standard Cauchy distribution function; Indicates multiplication; (2.5) Determine whether the termination condition is met. If the termination condition is not met, return to step (2.3); if the constraint condition is met, proceed to step (2.6). (2.6) The output yields the optimal number of encoder layers, decoder layers, and attention heads.

9. The dynamic early warning system for part quality based on a hierarchical federated learning framework according to claim 4, characterized in that, The dynamic early warning unit uses a basic threshold T d and component life decay rate r d As an indicator for determining whether a system should issue an early warning, given a historical lifetime sequence [L] t-n ,L t-n+1 ,…,L t The component life decay rate is expressed as follows: In the formula, ω k As a time decay weight; t is the sequence number; Historical average lifespan, δ is the smoothing constant; Current lifespan L t Less than the basic threshold T d Or component life decay rate r d When the value is greater than 0.1, the system issues an early warning; the early warning decision function is expressed as follows:

10. The dynamic early warning system for part quality based on a hierarchical federated learning framework according to claim 4, characterized in that, The dynamic optimization mechanism of the dynamic early warning unit optimizes the threshold by minimizing the loss function, as shown in the following expression: In the formula, l(·) is the cross-entropy loss function; Alert m This refers to the early warning results for node m. λ represents the true label, where 1 = warning needed, and 0 = no warning needed; λ is the regularization coefficient; Ω(·) is the regularization term to prevent over-adjustment.

Images