Cross-departmental device picture label error collaborative detection method and system

Abstract

The application discloses a cross-departmental equipment picture label error collaborative detection method and system, relates to the technical field of intelligent supervision and federated learning, and comprises a center Agent module, a customer Agent module and a database system module. In the collaborative training stage, the center Agent initializes a global model and distributes the global model to each customer Agent. Each customer Agent performs local training based on local equipment picture data, calculates model parameter update differences and uploads the model parameter update differences. The center Agent aggregates each difference to update the global model, and iterates to a preset round to obtain an optimized global model. In the label noise detection stage, each customer Agent inputs a local picture into the final global model, obtains a predicted label and a loss value, combines local inherent dimension characteristics, uses a Gaussian mixture model to perform clustering analysis on the loss value, and determines a label falling into a high cluster as a noise label. The application effectively protects data privacy in a sensitive scene, and improves the accuracy of label error detection and cross-departmental collaboration efficiency.

Description

Technical Field

[0001] This application relates to the fields of intelligent supervision and federated learning, and in particular to a collaborative detection method and system for cross-departmental device image labeling errors. Background Technology

[0002] Equipment cataloging is a fundamental and prerequisite step in the construction and management of equipment data resources. On the one hand, due to subjective judgment biases in the data annotation process or errors in data transmission, labeling noise is unavoidable in cross-departmental equipment image annotations. On the other hand, cross-departmental equipment image annotation information is inherently sensitive, making cross-departmental data sharing and collaborative work extremely difficult in terms of security protection. For these reasons, serious data resource quality issues exist during the cataloging process, severely impacting the overall effectiveness of equipment management.

[0003] Currently, training methods for label noise mainly fall into two categories: one is the use of non-explicit robust training methods, and the other is training methods based on label noise detection. However, non-explicit robust training methods have certain limitations when dealing with the ever-increasing volume of large-scale and cross-departmental multi-source device information. First, they lack the ability to handle sample specificity. Second, they have poor adaptability to large-scale heterogeneous data. Third, they are difficult to extract effective information from the noise.

[0004] Detection-based training methods can accurately detect noisy labels, which can, to some extent, prevent the model from being misled by incorrect labels and can also uncover potential useful information in noisy samples, better meeting the needs of federated learning for efficient modeling. Therefore, this application proposes a cross-departmental collaborative detection method and system for equipment image label errors to ensure the reliability of multi-source battlefield data and the efficiency of equipment cataloging. Summary of the Invention

[0005] The purpose of this application is to provide a cross-departmental collaborative detection method and system for equipment image labeling errors, which can ensure the reliability of multi-source battlefield data and the efficiency of equipment cataloging.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a cross-departmental collaborative detection method for image labeling errors in equipment, including: S1: Collaborative training phase; S1.1: Initialize the parameters of the global model through the central Agent to obtain the first global model, and then distribute the parameters of the first global model to n customer Agents, where n≥2; S1.2: After receiving the parameters of the first global model, each customer agent trains a local model based on the training dataset consisting of locally stored device images and corresponding real labels. The local model parameters are updated using the gradient descent algorithm, and the difference between the updated local model parameters and the received first global model parameters is calculated. ; S1.3: Each client agent determines whether the decrease in the loss value during this training exceeds a preset threshold. If so, it returns to step S1.2 to continue local training; otherwise, it calculates the difference. Uploaded to the central Agent; S1.4: The central agent receives parameter differences from all client agents. According to the preset aggregation algorithm, all differences are processed. Weighted fusion is performed to generate new global model parameters, which are then distributed to each customer agent as initial parameters for the next round. S1.5: Repeat steps S1.2 to S1.4 until the preset total number of training rounds is completed, and the final optimized global model is obtained, which is denoted as the second global model; S2: Tag Noise Detection Stage S2.1: Each customer agent inputs the local device image to be detected into the second global model to obtain the predicted label ŷ and the corresponding loss value LOSS for each image; S2.2: For each device image to be detected in the local dataset, each customer agent calculates the Local Intrinsic Dimension (LID) value of the device image in the local dataset; the LID value is used to measure the complexity of the data distribution between the image and its k nearest neighbor images. S2.3: Each customer agent uses the loss value (LOSS) and local intrinsic dimension (LID) values ​​of all images to be detected as feature vectors to form the detection dataset Rd; S2.4: Each customer agent uses a Gaussian mixture model to perform cluster analysis on the feature vectors in the detection dataset Rd to generate two probability distribution clusters; the two probability distribution clusters include: a low-mean cluster and a high-mean cluster; wherein, the low-mean cluster corresponds to clean labeled samples, and the high-mean cluster corresponds to noisy labeled samples; S2.5: Each client agent classifies the labels corresponding to images whose feature vectors fall into high-mean clusters as label noise and outputs the noise label detection results.

[0007] Optionally, the first global model is a FedNoiseCNN model, whose network structure includes convolutional layers, BatchNorm layers, and fully connected layers; wherein, the convolutional layer contains at least one 5×5 convolutional kernel for extracting fine-grained features from the device image; the BatchNorm layer introduces an illumination compensation factor to adapt to image data under different shooting environments; the output dimension of the fully connected layer is set according to the number of categories of device components.

[0008] Optionally, the local model training uses a weighted cross-entropy loss function, assigning a weight greater than 1 to key equipment components that affect combat performance, and assigning a weight less than 1 to non-key components.

[0009] Optionally, the expression for the local intrinsic dimension LID value is as follows: ; in, Image of the device to be tested; This represents the number of nearest neighbor samples; For nearest neighbor sample index; Image of the device to be tested With the The feature distance of the nearest neighbor samples; Image of the device to be tested of The largest feature distance among the nearest neighbor samples.

[0010] Optionally, the parameters of the global model are initialized through the central agent, specifically using the following formula: ; in, The first global model parameter matrix generated for initialization of the central Agent; The first global model parameter matrix contains the first... One parameter, Total number of model parameters 。

[0011] Optionally, the weights of the second global model are: ; in, For the first The second global model weight matrix generated after round-wise global aggregation; This is the serial number of the client agent; For client collection; For the first Local device image dataset for each client agent; For the first Local device image dataset for each client agent; This refers to the index number of the model parameter; For the first The client agent in the first... After one round of local training, the local model weight matrix is ​​updated.

[0012] Optionally, the expression for the Gaussian mixture model is: ; ; in, A collection of images of the device to be tested. The overall probability density; This represents the number of clusters in the Gaussian mixture model. For cluster indexing; For the first The weights of each cluster; It is a Gaussian distribution function; The loss value is the image of the device to be inspected; For the first The mean of each cluster; For data dimensions; is the matrix transpose symbol, and exp is a power of 10.

[0013] Optionally, the mean of the low-mean cluster The variance is between 0.05 and 0.07. Between 0.004 and 0.006; the mean of the high-mean cluster. The variance is between 0.7 and 0.8. Between 0.02 and 0.04.

[0014] Secondly, this application provides a cross-departmental collaborative detection system for image labeling errors in equipment, comprising: The central Agent module is used to initialize the parameters of the global model, obtain the first global model, and receive the parameter differences uploaded by each client Agent. The new global model is generated by aggregation and is referred to as the second global model. The parameters of the second global model are then distributed to each customer agent. There are n customer agent modules, each connected to a local data source of a department. These modules receive the initial global model parameters distributed by the central agent, train a local model based on local device image data, and calculate the parameter differences. And upload it, and after the collaborative training is completed, use the final optimized second global model to perform label noise detection; The database system module is used to store the first global model initialized by the central agent, the local model parameters of each customer agent, the second global model generated by the central agent, and the training logs of each stage. The customer agent modules and the central agent module are connected through an encrypted communication network, and during collaborative training and noise detection, they only exchange model parameters or their differences, without transmitting the original device images and label data.

[0015] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a collaborative detection method and system for cross-departmental device image labeling errors. Regarding strong privacy protection, this application employs a federated learning architecture, exchanging only model parameter differences throughout the process. The original device images and tag data are always stored locally, fundamentally eliminating the risk of sensitive data leakage and meeting the needs of high-security scenarios.

[0016] In terms of high detection accuracy, the model loss value LOSS is innovatively combined with the local intrinsic dimension LID of the sample as a joint feature for noise detection. LID can capture abnormal samples related to local data distribution that are difficult to find by LOSS alone, which significantly improves the detection accuracy and reliability of label noise, especially label noise of equipment images in complex fields. In terms of efficient collaboration, it has achieved automated and intelligent collaborative detection across multiple departments such as production, operation and maintenance, and warehousing, avoiding the high cost and low efficiency of traditional manual review, and greatly improving the processing speed of large-scale equipment images and the efficiency of equipment cataloging. In terms of domain adaptability, the model structure (such as specific-size convolutional kernels and illumination compensation factors) and training strategies (such as key component weighting) designed in the method are specifically optimized for the "tactical feature specificity" of images from specific devices, ensuring high availability and high reliability in real-world application scenarios. By pre-setting strict training rounds and model performance indicators, the method ensures sufficient training and high quality of the global model, providing a solid and reliable model foundation for subsequent noise detection and guaranteeing the credibility of the final detection results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1This is a flowchart illustrating a cross-departmental collaborative detection method for equipment image labeling errors in one embodiment of this application; Figure 2 This is a schematic diagram of a cross-departmental collaborative detection system for equipment image labeling errors, provided as an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] In one exemplary embodiment, such as Figure 1 As shown, a cross-departmental collaborative detection method for image labeling errors is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, it includes two stages: a collaborative training stage and a label noise monitoring stage. Wherein: S1: Collaborative training phase.

[0022] S1.1: Initialize the parameters of the global model through the central Agent to obtain the first global model, and then distribute the parameters of the first global model to n customer Agents, where n≥2.

[0023] In the initial training phase, the central agent randomly initializes the parameters of the convolutional layers, BatchNorm layers, and fully connected layers of the global model according to the FedNoiseCNN model structure.

[0024] To address the equipment image features in this embodiment, a four-layer convolutional layer design (3×3, 5×5, 3×3, 3×3) is used. The second 5×5 convolutional kernel is specifically designed to extract fine-grained features such as equipment barrel rifling and radar antenna array elements. The BatchNorm layer incorporates a specific scene illumination compensation factor to adapt to image data under different shooting environments, such as standard lighting in the production workshop, strong light / dust / low light at night, etc. The output dimension of the fully connected layer is set according to the equipment component category, such as setting the equipment scene to 3 (corresponding to "125mm barrel", "composite armor module", "track drive wheel").

[0025] set up θ For model parameter vectors, Given the total number of model parameters, the parameter matrix of the initial global model (i.e., the first global model) is represented as: ; in, The first global model parameter matrix generated for initialization of the central Agent; The first global model parameter matrix contains the first... One parameter.

[0026] In the scenario of this embodiment Values ​​range from 10240 to 15360 (adjusted based on the number of component categories), such as in a scenario with 3 types of components. =10240 (4 convolutional kernels with 6800 parameters, BatchNorm layers with 440 parameters, and fully connected layers with 3000 parameters).

[0027] S1.2: After receiving the parameters of the first global model, each customer agent trains a local model based on the training dataset consisting of locally stored device images and corresponding real labels. The local model parameters are updated using the gradient descent algorithm, and the difference between the updated local model parameters and the received first global model parameters is calculated. .

[0028] Customer Agent i ( After receiving the initial global model (i.e., the first global model) parameter matrix sent by the central agent, the local samples are input into the global model to obtain the prediction results. The difference between the prediction results and the true labels is calculated through the loss function (LOSS).

[0029] The local samples consist of images of specific equipment components. For example, the equipment production department sample consists of 50 images of core equipment components (512×512 resolution, including production serial number watermarks to prevent data leakage). The loss function uses weighted cross-entropy, assigning 1.5 times the weight to key components such as "gun barrel" and "radar antenna" and 0.8 times the weight to non-key components (such as "heat sink"), to ensure that the model prioritizes and accurately identifies component labels that affect performance.

[0030] Given a dataset Images of each device to be tested Each has a corresponding label y and a model, for any x Its predicted value is The real label is y The loss value (LOSS) is represented as: ; in, Image of the device to be tested; The total number of label categories for local device image samples; The nearest neighbor sample index; Image of the device to be tested The actual label encoding value; The first global model is the image of the device to be detected. The Class label prediction probability.

[0031] Update the local model parameters using gradient descent, and let the updated local model parameters be... Initial global parameters Calculate the difference between the updated parameters and the initial global parameters. Represented as: ; In the application scenario of this embodiment, the aforementioned "central agent" specifically refers to the collaborative scheduling server of the equipment support center, deployed on the core node of the intranet, and possessing anti-electromagnetic interference capabilities; the "customer agents" are the local training terminals of the equipment production department. Typically, the value is 2-5 (corresponding to 2-5 relevant departments such as production, operation and maintenance, warehousing, and repair), and each Agent establishes an encrypted connection through a Virtual Private Network (VPN).

[0032] S1.3: Each client agent determines whether the decrease in the loss value during this training exceeds a preset threshold. If so, it returns to step S1.2 to continue local training; otherwise, it calculates the difference. Uploaded to the central Agent.

[0033] When the loss value decreases by more than 1e-5, the amount of model parameter updates... Significant changes, customer agent i ( Continue updating the local model parameters using gradient descent, and calculate the difference between the updated parameters and the initial global parameters. Otherwise, the customer agent i ( The difference between the uploaded updated parameters and the initial global parameters will be calculated. To the central agent.

[0034] S1.4: The central agent receives parameter differences from all client agents. According to the preset aggregation algorithm, all differences are processed. Weighted fusion is performed to generate new global model parameters, which are then distributed to each customer agent as initial parameters for the next round.

[0035] S1.5: Repeat steps S1.2 to S1.4 until the preset total number of training rounds is completed, and the final optimized global model is obtained, which is denoted as the second global model.

[0036] The central agent receives the customer agent's information. i ( Uploaded Generate a new global model, namely the second global model: Let the system dataset be... Client m Local dataset Local model weights In the t At the end of the round, the second global model weight It will be updated as follows: ; in, S t Indicates the first t The set of clients participating in the model aggregation round. D i Indicates the client i Local dataset, Indicates the first t During the round, the local model weights of client m, This indicates the first element in the local dataset of client m. Feature data of each sample This indicates the first element in the local dataset of client m. The label (or target value) of each sample.

[0037] To ensure the controllability of the training process, the global iteration is terminated primarily by the "preset total number of rounds". If the global training rounds do not reach the preset upper limit, new global parameters will continue to be sent to the client.

[0038] In this embodiment, the total number of scenario rounds is preset to 25 (higher than the 20 rounds of a normal scenario), which ensures that the model fully learns the characteristics of military equipment; after each round, a "Military Model Log" is generated to record the round number and the departments involved. The global LOSS value and logs are stored in encrypted form using a key and can only be viewed by authorized personnel. After 25 rounds of training, the global model must have a prediction accuracy of ≥98.5% for the labels of core equipment components and ≥97% for the labels of radar equipment components before it can proceed to the subsequent noise detection stage.

[0039] S2: Tag noise detection stage.

[0040] S2.1: Each customer agent inputs the local image of the device to be detected into the second global model to obtain the predicted label for each image. And the corresponding loss value LOSS.

[0041] S2.2: For each image to be detected in the local dataset, each customer agent calculates the Local Intrinsic Dimension (LID) value of the image; the LID value is used to measure the complexity of the data distribution between the image and its k nearest neighbor images.

[0042] Given a set of device images X Each image in the set is a collection. X A "point" in the context of a reference image, where the reference image and its reference point are considered together. k The set of distances between the nearest neighbors, i.e., the locally intrinsic dimension (LID), is represented as:

[0043] in, Image of the device to be tested; This represents the number of nearest neighbor samples; For nearest neighbor sample index; Image of the device to be tested With the The feature distance of the nearest neighbor samples; Image of the device to be tested of The largest feature distance among the nearest neighbor samples.

[0044] Local Intrinsic Dimension (LID) is a metric that measures the complexity of the data distribution of a single device image among its nearest neighbor samples (other device images in the same department).

[0045] Let the threshold variable be LID_threshold The calculated LID value indicates that the label may be a noisy client if the LID value is greater than the variable; otherwise, it indicates that the label may be a noisy client.

[0046] That is, if the LID value is small (e.g., ≤3.0), it is determined to be a "clean sample area"; If the LID value is large (e.g., >3.0), it is determined to be "the area where the noise sample is located".

[0047] Assuming the equipment component combat support data scenario is set to 3.0, a labeling error could have serious consequences. Since LID(x) = 4.2 > 3.0, the local dataset of the field operations and maintenance department containing this sample is determined to have noise risk.

[0048] S2.3: Each client agent uses the loss value (LOSS) and local intrinsic dimension (LID) values ​​of all images to be detected as feature vectors to form the detection dataset Rd.

[0049] S2.4: Each customer agent uses a Gaussian mixture model to perform cluster analysis on the feature vectors in the detection dataset Rd to generate two probability distribution clusters; the two probability distribution clusters include a low-mean cluster and a high-mean cluster; wherein, the low-mean cluster corresponds to clean labeled samples, and the high-mean cluster corresponds to noisy labeled samples.

[0050] S2.5: Each client agent classifies the labels corresponding to images whose feature vectors fall into high-mean clusters as label noise and outputs the noise label detection results.

[0051] The expression for the Gaussian mixture model is: ; ; in, A collection of images of the device to be tested. The overall probability density; This represents the number of clusters in the Gaussian mixture model. For cluster indexing; For the first The weights of each cluster; It is a Gaussian distribution function; The loss value is the image of the device to be inspected; For the first The mean of each cluster; For data dimensions; is the matrix transpose symbol, and exp is a power of 10.

[0052] In the scenario, the dataset Rd is a set of LOSS values ​​for 800-1000 device images (e.g., 800 LOSS values ​​for the main equipment scenario). The Gaussian mixture model sets two clusters (lower clusters correspond to clean labels, and higher clusters correspond to noise labels), and the model training adopts usage-level data isolation technology to avoid LOSS values ​​leaking device features. Taking the equipment scenario as an example, after model training, the mean μ1=0.06 and variance Σ1=0.005 for the lower cluster (clean sample LOSS is concentrated in this range, such as "artillery barrel-089.jpg" LOSS=0.010498), and the mean μ2=0.75 and variance Σ2=0.03 for the higher cluster (noisy sample LOSS is concentrated in this range, such as "module-067.jpg" LOSS=0.9126).

[0053] In one exemplary embodiment, such as Figure 2As shown, a cross-departmental collaborative detection system for equipment image labeling errors is provided, including: The central Agent module is used to initialize the parameters of the global model, obtain the first global model, and receive the parameter differences uploaded by each client Agent. The new global model is generated by aggregation and is referred to as the second global model. The parameters of the second global model are then distributed to each customer agent. There are n customer agent modules, each connected to a local data source of a department. These modules receive the initial global model parameters distributed by the central agent, train a local model based on local device image data, and calculate the parameter differences. And upload it, and after the collaborative training is completed, use the final optimized second global model to perform label noise detection; The database system module is used to store the first global model initialized by the central agent, the local model parameters of each customer agent, the second global model generated by the central agent, and the training logs of each stage. The customer agent modules and the central agent module are connected through an encrypted communication network, and during collaborative training and noise detection, they only exchange model parameters or their differences, without transmitting the original device images and label data.

[0054] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0055] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A cross-departmental collaborative detection method for equipment image labeling errors, characterized in that, The cross-departmental collaborative detection method for equipment image labeling errors includes: S1: Collaborative training phase; S1.1: Initialize the parameters of the global model through the central Agent to obtain the first global model, and then distribute the parameters of the first global model to n customer Agents, where n≥2; S1.2: After receiving the parameters of the first global model, each customer agent trains a local model based on the training dataset consisting of locally stored device images and corresponding real labels, updates the local model parameters using the gradient descent algorithm, and calculates the difference between the updated local model parameters and the received first global model parameters. ; S1.3: Each client agent determines whether the decrease in the loss value during this training exceeds a preset threshold. If so, it returns to step S1.2 to continue local training; otherwise, it calculates the difference. Uploaded to the central Agent; S1.4: The central agent receives parameter differences from all client agents. According to the preset aggregation algorithm, all differences are processed. Weighted fusion is performed to generate new global model parameters, which are then distributed to each customer agent as initial parameters for the next round. S1.5: Repeat steps S1.2 to S1.4 until the preset total number of training rounds is completed, and the final optimized global model is obtained, which is denoted as the second global model; S2: Tag noise detection stage; S2.1: Each customer agent inputs the local image of the device to be detected into the second global model to obtain the predicted label for each image. and the corresponding loss value; S2.2: For each device image to be detected in the local dataset, each customer agent calculates the Local Intrinsic Dimension (LID) value of the device image in the local dataset; the LID value is used to measure the complexity of the data distribution between the image and its k nearest neighbor images. S2.3: Each customer agent uses the loss value (LOSS) and local intrinsic dimension (LID) values ​​of all images to be detected as feature vectors to form the detection dataset Rd; S2.4: Each customer agent uses a Gaussian mixture model to perform cluster analysis on the feature vectors in the detection dataset Rd to generate two probability distribution clusters; the two probability distribution clusters include: a low-mean cluster and a high-mean cluster; wherein, the low-mean cluster corresponds to clean labeled samples, and the high-mean cluster corresponds to noisy labeled samples; S2.5: Each client agent classifies the labels corresponding to images whose feature vectors fall into high-mean clusters as label noise and outputs the noise label detection results.

2. The cross-departmental collaborative detection method for equipment image labeling errors according to claim 1, characterized in that, The first global model is the FedNoiseCNN model, whose network structure includes convolutional layers, BatchNorm layers, and fully connected layers. The convolutional layers contain at least one 5×5 convolutional kernel to extract fine-grained features from device images. The BatchNorm layers introduce an illumination compensation factor to adapt to image data under different shooting environments. The output dimension of the fully connected layers is set according to the number of device component categories.

3. The cross-departmental collaborative detection method for equipment image labeling errors according to claim 1, characterized in that, The local model training uses a weighted cross-entropy loss function, assigning a weight greater than 1 to key equipment components that affect combat performance, and a weight less than 1 to non-key components.

4. The cross-departmental collaborative detection method for equipment image labeling errors according to claim 1, characterized in that, The expression for the local intrinsic dimension (LID) value is as follows: ； in, Image of the device to be tested; This represents the number of nearest neighbor samples; For nearest neighbor sample index; Image of the device to be tested With the The feature distance of the nearest neighbor samples; Image of the device to be tested of The largest feature distance among the nearest neighbor samples.

5. The cross-departmental collaborative detection method for equipment image labeling errors according to claim 1, characterized in that, The parameters of the global model are initialized through the central agent, specifically using the following formula: ； in, The first global model parameter matrix generated for initialization of the central Agent; The first global model parameter matrix contains the first... One parameter, This represents the total number of model parameters.

6. The cross-departmental collaborative detection method for equipment image labeling errors according to claim 1, characterized in that, The weights of the second global model are: ； in, For the first The second global model weight matrix generated after round-wise global aggregation; This is the serial number of the client agent; For client collection; For the first Local device image dataset for each client agent; For the first Local device image dataset for each client agent; This refers to the index number of the model parameter; For the first The client agent in the first... After one round of local training, the local model weight matrix is ​​updated.

7. The cross-departmental collaborative detection method for equipment image labeling errors according to claim 1, characterized in that, The expression for the Gaussian mixture model is: ； ； in, A collection of images of the device to be tested. The overall probability density; This represents the number of clusters in the Gaussian mixture model. For cluster indexing; For the first The weights of each cluster; It is a Gaussian distribution function; The loss value is the image of the device to be inspected; For the first The mean of each cluster; For data dimensions; is the matrix transpose symbol, and exp is a power of 10.

8. The cross-departmental collaborative detection method for equipment image labeling errors according to claim 1, characterized in that, Mean of low-mean clusters The variance is between 0.05 and 0.

07. Between 0.004 and 0.006; the mean of the high-mean cluster. The variance is between 0.7 and 0.

8. Between 0.02 and 0.

04.

9. A cross-departmental collaborative detection system for equipment image labeling errors, characterized in that, The cross-departmental equipment image labeling error collaborative detection system applies the cross-departmental equipment image labeling error collaborative detection method according to any one of claims 1 to 8, and the cross-departmental equipment image labeling error collaborative detection system comprises: The central Agent module is used to initialize the parameters of the global model, obtain the first global model, and receive the parameter differences uploaded by each client Agent. The new global model is generated by aggregation and is referred to as the second global model. The parameters of the second global model are then distributed to each customer agent. There are n customer agent modules, each connected to a local data source of a department. These modules receive the initial global model parameters distributed by the central agent, train a local model based on local device image data, and calculate the parameter differences. And upload it, and after the collaborative training is completed, use the final optimized second global model to perform label noise detection; The database system module is used to store the first global model initialized by the central agent, the local model parameters of each customer agent, the second global model generated by the central agent, and the training logs of each stage. The customer agent modules and the central agent module are connected through an encrypted communication network, and during collaborative training and noise detection, they only exchange model parameters or their differences, without transmitting the original device images and label data.

Images