Lubricating oil metal particle spectral quantitative analysis method based on federated learning

Abstract

This invention provides a method for quantitative spectral analysis of metal particles in lubricating oil based on federated learning, belonging to the field of lubricating oil condition monitoring technology based on federated learning. First, laser-induced breakdown spectra are acquired from lubricating oil samples from various industrial sites. After preprocessing, the peak intensity, integrated intensity, background correction intensity, and spectral line intensity ratio of seven metal elements are extracted and concatenated according to elemental order and spectral line center wavelength order to obtain a characteristic spectral line intensity sequence. This characteristic spectral line intensity sequence is input into a designed local concentration inversion model to obtain predicted concentration vectors for the seven metal elements. Simultaneously, a federated concentration inversion mechanism designed to address spectral differences across multiple industrial sites is used to determine the parameters of the officially invoked local model. This invention improves the accuracy of multi-metal concentration inversion under complex spectral conditions through a joint characterization mechanism and enhances the model's adaptability and stability in cross-site lubricating oil detection tasks based on the federated learning mechanism.

Description

Technical Field

[0001] This invention belongs to the field of lubricating oil condition monitoring technology based on federated learning, and particularly relates to a method for quantitative analysis of lubricating oil metal particles by spectrometry based on federated learning. Background Technology

[0002] Lubricating oil is a critical medium in gearboxes, engines, hydraulic systems, bearing assemblies, and large rotating machinery. It not only reduces friction, cools, and seals, but also carries metal particles generated by internal wear. Changes in the concentration of metal particles such as iron, copper, chromium, aluminum, nickel, tin, and lead in lubricating oil can reflect the degree of wear in components such as gears, bearings, bushings, pistons, and seals. Therefore, rapid, accurate, and continuous quantitative analysis of metal particles in lubricating oil is an important technical means to assess equipment wear conditions, formulate maintenance plans, and reduce the risk of sudden downtime.

[0003] Existing methods for detecting metal particles in lubricating oil mainly include the following categories.

[0004] First, there are detection methods based on laboratory chemical analysis. These methods typically employ inductively coupled plasma atomic emission spectrometry (ICP-AES), inductively coupled plasma mass spectrometry (ICP-MS), or atomic absorption spectrometry (AAS) to detect oil samples. While these methods offer high accuracy, samples require pretreatment and offline analysis in a laboratory, resulting in long processing times and failing to meet the demands of rapid detection and continuous monitoring in industrial settings. Furthermore, the sources of oil, additive systems, pollution backgrounds, and particle size distributions vary significantly across different industrial sites, making it difficult for laboratory models to directly adapt to the diverse data distributions from multiple locations.

[0005] Second, there are detection methods based on ferrography or particle counting. These methods can observe the morphology, quantity, and size distribution of wear particles in lubricating oil, making them suitable for wear type analysis. However, these methods are highly dependent on sample preparation, microscopic imaging, and manual interpretation, making it difficult to rapidly quantify the concentrations of multiple metal elements and to develop a unified automated analytical model.

[0006] Third, there is the lubricating oil detection method based on laser-induced breakdown spectroscopy. Laser-induced breakdown spectroscopy utilizes high-energy lasers to create plasma in oil samples and collects characteristic emission spectra, enabling rapid identification and quantitative analysis of metal elements. This method offers advantages such as high detection speed, low sample consumption, and suitability for field deployment. However, the complex matrix of lubricating oil means that oil viscosity, additive composition, particle size, laser energy fluctuations, and differences in the acquisition optical path all affect spectral intensity, resulting in different spectral responses for the same metal concentration at different sites. Training a concentration inversion model solely based on data from a single site results in insufficient model generalization ability; directly aggregating raw spectral data from various industrial sites poses a risk of leakage of equipment operation data, oil formulations, maintenance records, and production status information, making it difficult to promote across enterprises, workshops, and equipment.

[0007] Fourth, a centralized deep learning-based method for quantitative spectral analysis. This method gathers spectral samples from different sources onto a unified server to train the model, improving the model's ability to learn complex spectral features. However, centralized training requires uploading raw spectral data and concentration labels, leading to issues such as large data transmission volumes, privacy breaches of on-site data, and unclear data ownership boundaries. For industrial enterprises with data isolation requirements, centralized modeling is difficult to meet practical deployment needs.

[0008] In summary, existing methods for quantitative analysis of metal particles in lubricating oil still suffer from problems such as insufficient adaptability to field conditions, difficulty in sharing data across different fields, insufficient protection of the privacy of raw spectral data, low model update efficiency, and limited accuracy in multi-metal concentration inversion. Summary of the Invention

[0009] To address the above problems, this invention proposes a method for quantitative spectral analysis of lubricating oil metal particles based on federated learning, comprising the following steps: Laser-induced breakdown spectroscopy was performed on lubricating oil samples from various industrial sites to obtain raw spectral signals. After preprocessing, the peak intensity, integral intensity, background correction intensity, and spectral line intensity ratio of seven metal elements (iron, copper, chromium, aluminum, nickel, tin, and lead) were extracted and spliced ​​together according to element order and spectral line center wavelength order to obtain characteristic spectral line intensity sequences. The characteristic spectral line intensity sequence is input into the spectral coding sub-network, and the spectral line fusion feature is obtained through spectral line channel embedding, spectral line neighborhood convolution, element relationship attention and pooling compression. Then, it is input into the multi-element concentration prediction sub-network, and the predicted concentration vectors of seven metal elements are obtained through shared concentration characterization, element-specific prediction and element correlation correction. The concentration inversion error is obtained by comparing the predicted concentration vector with the true concentration vector. The local model update amount is generated, and the dynamic aggregation weight is calculated based on the number of samples, the update timestamp, and the local validation error. The global model update amount is then obtained by aggregating the weights. Global model parameters are generated based on the global model update volume and sent back to each industrial site to determine the local model parameters for formal use.

[0010] Preferably, the preprocessing includes wavelength calibration, background subtraction, smoothing and denoising, and intensity normalization, wherein the background subtraction is performed using a local baseline fitting method.

[0011] Preferably, the spectral coding sub-network is based on the first... Industrial site The Middle Characteristic spectral intensity sequence of a lubricating oil sample For input; First, the input feature spectral line intensity sequence is processed using a spectral line channel embedding layer. Channel mapping is performed; the spectral line channel embedding layer consists of a fully connected layer, a batch normalization layer, and a ReLU activation function, which is used to map the four types of spectral line features of each feature line to a unified feature dimension, thereby obtaining a primary spectral line feature map. ; Secondly, the primary spectral line characteristic map Spectral line neighborhood convolution processing is performed. This processing employs three parallel convolutional structures, using kernels of sizes 3, 5, and 7, respectively. The kernel of size 3 is used to extract local intensity changes between adjacent feature lines; the kernel of size 5 is used to extract the response relationship between multiple feature lines within the same metal element; and the kernel of size 7 is used to extract spectral interference relationships between different metal elements. The output results are then concatenated and linearly mapped to obtain a multi-scale spectral line feature map. ; Secondly, the multi-scale spectral feature map Elemental relationship attention processing is performed. An elemental relationship graph is constructed based on the spectral overlap, internal standard correlation, and common interference relationships of the oil matrix among seven metallic elements: iron, copper, chromium, aluminum, nickel, tin, and lead. Nodes in the elemental relationship graph represent the spectral characteristics of the corresponding metallic element, and edges represent the spectral interference relationships between different metallic elements. Subsequently, the information transfer weights between nodes of different elements are calculated through a graph attention layer, and the multi-scale spectral feature graph is then processed based on these information transfer weights. Weighted aggregation is performed to obtain cross-element spectral feature maps. ; Finally, the cross-element spectral line feature map Residual fusion and pooling compression are performed; residual fusion processing converts cross-elemental spectral feature maps. Compared with primary spectral line feature map Residual connections are performed; the pooling compression process consists of an attention pooling layer and a fully connected compression layer. The attention pooling layer is used to weight and summarize the contributions of different feature lines to the concentration inversion, and the fully connected compression layer is used to map the weighted and summarized features to spectral fusion features. .

[0012] Preferably, the multi-element concentration prediction sub-network is specifically: First, the spectral line fusion characteristics are analyzed using a shared concentration characterization layer. Global concentration feature extraction is performed. The shared concentration characterization layer consists of a first fully connected layer, a batch normalization layer, a ReLU activation function, a Dropout layer, and a second fully connected layer. It is used to extract the oil matrix response features, overall spectral intensity variation features, and laser energy fluctuation compensation features that are commonly dependent on the seven metal elements, thus obtaining the shared concentration characterization. ; Secondly, Seven element-specific prediction branches are input, each consisting of a fully connected layer, a ReLU activation function, and a single-element concentration output layer; the prediction branches start from the shared concentration characterization. The concentration-sensitive features of the corresponding metal elements are extracted, and the initial predicted concentrations of iron, copper, chromium, aluminum, nickel, tin, and lead are output respectively. The seven initial predicted concentrations are concatenated in a fixed order of iron, copper, chromium, aluminum, nickel, tin, and lead to obtain the initial concentration prediction vector. ; Secondly, an elemental correlation correction layer is used to... Elemental correlation correction is performed. The elemental correlation correction layer consists of a fully connected layer and a sigmoid activation function. It is used to generate correction weights based on the initial predicted concentrations of seven metal elements, and to compensate for inter-element prediction biases caused by spectral overlap, matrix effects, and laser energy fluctuations based on these correction weights, thus obtaining the corrected concentration prediction vector. ; Finally, the corrected concentration prediction vector will be... The input concentration is used to output a layer that employs the Softplus non-negative constraint activation function to correct the concentration prediction vector. Negative concentration values ​​are corrected to non-negative values, and predicted concentration vectors for seven metallic elements are output in a fixed order: iron, copper, chromium, aluminum, nickel, tin, and lead. .

[0013] Preferably, the specific process for obtaining the global model update amount is as follows: First, when the Industrial site Obtain the After obtaining the standard detection concentration results for each lubricating oil sample, the predicted concentration vectors of the seven metal elements are retrieved using the sample number, sampling time, equipment number, and lubricating oil circulation position as matching indices. The standard detection concentration results are recorded as the true concentration vector of the seven metal elements. Subsequently, the predicted concentration vector will be... With the true concentration vector Perform element-by-element comparison and calculate concentration inversion error. and the obtained characteristic spectral line intensity sequence With the true concentration vector Pairing is performed to create new local training samples, which are then written into the local spectral concentration dataset. Then, calculate the local validation error. Based on the updated local spectral concentration dataset For spectral coding subnetwork and multi-element concentration prediction subnetwork Corresponding local model parameters Perform training fine-tuning to generate local model update quantities. Next, the first Industrial site Update local model Local sample size Training round number Update timestamp and local verification error Federation upload data packets And upload it to the federated aggregation node; finally, the federated aggregation node processes the uploaded data packets according to the federated data packets. Calculate dynamic aggregate weights And update the local model uploaded by each industrial site currently participating in the aggregation. Perform quality-weighted aggregation to obtain the global model update amount. .

[0014] Preferably, the specific process for determining the local model parameters for the formal invocation is as follows: First, the federated aggregation node is the first Round global model parameters Based on this, combined with the obtained global model update amount , generate the first Round global model parameters Subsequently, the federated aggregation node will assign global model parameters. The data is then transmitted back to all participating industrial sites, along with the global model version number, training round number, and parameter generation time; then, the... Industrial site Receive global model parameters Next, the integrity of the parameters, the global model version number, and the training epoch number are verified. After the verification is passed, the global model parameters are... Write the parameters to the candidate model parameter storage area to form candidate local model parameters. Finally, the first Industrial site Based on local validation samples, the current local model parameters and candidate local model parameters Calculate the verification error separately, and determine the first verification error based on the local verification error criterion. The parameters of the local model that are officially used in the round.

[0015] Preferably, during training and fine-tuning, the characteristic spectral line intensity sequence is used. As input, spectral line fusion features are obtained through a spectral coding subnetwork, and then passed through a multi-element concentration prediction subnetwork. The predicted concentration vectors of seven metal elements were obtained. And compare the predicted concentration vector with the actual concentration vector. The error between them is used to construct the local training loss function. : ; In the formula, Indicates the first Industrial site The local training loss function, Indicates the first Industrial site The number of local samples, Represents the spectral coding subnetwork. This represents the multi-element concentration prediction subnetwork. Indicates the first Industrial site The Middle Characteristic spectral intensity sequences of a lubricating oil sample. Indicates the first Industrial site The Middle The true concentration vectors of seven metal elements in a lubricating oil sample. Indicates the first Industrial site In the The local model parameters used at the start of each training round. This represents the weight regularization coefficient.

[0016] Preferably, the step of determining the first error based on the local verification error criterion... The local model parameters used in the round are as follows: No. Industrial site Call the local validation samples that have already obtained the true concentration vector in the current round, and use the current local model parameters respectively. and candidate local model parameters Concentration inversion of seven types of metal elements was performed to obtain the current local model parameter validation error. And candidate local model parameter validation error If the candidate local model parameter validation error The validation error is no higher than the current local model parameters. With allowable error increment The sum of these values ​​determines whether the candidate local model parameters pass the local validation error check, and the candidate local model parameters are then set as follows: Determined as the number The local model parameters used in the round If the candidate local model parameter validation error The validation error is higher than the current local model parameters. With allowable error increment The sum of these values ​​determines that the candidate local model parameters have failed the local validation error check. Industrial site Instead of using candidate local model parameters, continue using the current local model parameters. Determined as the number The local model parameters used in the round .

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) Federated concentration inversion mechanism for spectral differences in multiple industrial sites: In view of the problem that the oil matrix, additive system, wear particle distribution and collection conditions vary greatly in different industrial sites, this invention incorporates the local model update results of multiple industrial sites into a unified federated aggregation process, so that the concentration inversion model can comprehensively learn the spectral response law of multiple sites and improve the adaptability and stability of the model in cross-site lubricating oil detection tasks.

[0018] (2) Joint characterization mechanism of characteristic spectral lines for seven types of wear metal elements: In view of the problem that the intensity of a single spectral line is easily affected by background noise, laser energy fluctuation and spectral interference, this invention focuses on seven target metal elements, namely iron, copper, chromium, aluminum, nickel, tin and lead, and uniformly extracts the peak intensity, integral intensity, background correction intensity and spectral line intensity ratio, forming a characteristic spectral line intensity sequence with consistent structure, thereby improving the accuracy of multi-metal concentration inversion under complex spectral conditions.

[0019] (3) Dynamic aggregation mechanism that integrates sample size and model quality: In response to the problems of uneven sample quantity, inconsistent model update quality and different update time in different industrial sites, this invention comprehensively considers the local sample quantity, update time and local validation error in the federated aggregation process, and performs differentiated weighting on the model update results of each industrial site, reducing the impact of low-quality updates on the global model and improving the training stability and continuous update capability of the global concentration inversion model. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall technical process of the present invention.

[0021] Figure 2 This is a flowchart of the local concentration inversion and detection result generation process of the present invention.

[0022] Figure 3 This is a flowchart illustrating the detection error feedback, federated aggregation, and local model parameter determination processes of this invention.

[0023] Figure 4 This is a comparison chart showing the dynamic evolution of global verification error in an embodiment of the present invention.

[0024] Figure 5 This is a distribution of multi-element concentration inversion values ​​under multiple field conditions in an embodiment of the present invention.

[0025] Figure 6 This invention provides an embodiment of the nonlinear response diagram of quality and timeliness to aggregation weights. Detailed Implementation

[0026] This invention proposes a method for quantitative analysis of lubricating oil metal particles based on federated learning. The overall technical approach is as follows: Figure 1 As shown: S1 Lubricating Oil Metal Particle Spectral Data Acquisition and Characteristic Spectral Line Intensity Sequence Construction. First, lubricating oil samples were collected from multiple industrial sites, and the sequence was obtained using a laser-induced breakdown spectral acquisition device. Industrial site The Middle The original spectral signal of each lubricating oil sample Subsequently, the original spectral signal was analyzed. Wavelength calibration, background subtraction, smoothing and denoising, and intensity normalization are performed to obtain a standardized spectral signal. Then, a target metal set was constructed for seven elements: iron, copper, chromium, aluminum, nickel, tin, and lead. The peak intensity, integrated intensity, background correction intensity, and spectral line intensity ratio of the characteristic spectral lines were extracted. Finally, the seven target metal elements were arranged in elemental order: iron, copper, chromium, aluminum, nickel, tin, and lead. Within each metal element group, the corresponding spectral line feature vectors were arranged in ascending order of the center wavelength of the characteristic spectral lines. These vectors were then concatenated to obtain the characteristic spectral line intensity sequence. .

[0027] S2 Local concentration inversion processing and metal particle detection results generation. First, the characteristic spectral line intensity sequence obtained in S1 is processed... Input spectral coding subnetwork By using spectral line channel embedding layers, spectral line neighborhood convolution processing, element-relation attention processing, and residual fusion and pooling compression, the characteristic spectral line intensity sequence is... Mapped to spectral line fusion features Subsequently, the spectral line fusion features were analyzed. Input multi-element concentration prediction subnetwork The predicted concentration vector is obtained through a shared concentration characterization layer, seven element-specific prediction branches, an element correlation correction layer, and a concentration output layer. Finally, the predicted concentration vectors of the seven metal elements were calculated. The metal particle concentration test results are entered into a table according to a fixed order of iron, copper, chromium, aluminum, nickel, tin and lead, and a quantitative analysis report of lubricating oil metal particles is generated.

[0028] S3 detection error feedback, local model update generation, and federated aggregation. First, when the... Industrial site Obtain the After obtaining the standard detection concentration results for each lubricating oil sample, the sample number, sampling time, equipment number, and lubricating oil circulation position are used as matching indices to call the predicted concentration vectors of the seven metal elements obtained from S2. The standard detection concentration results are recorded as the true concentration vector of the seven metal elements. Subsequently, the predicted concentration vector will be... With the true concentration vector Perform element-by-element comparison and calculate concentration inversion error. and the characteristic spectral line intensity sequence obtained from S1 With the true concentration vector Pairing is performed to create new local training samples, which are then written into the local spectral concentration dataset. Then, calculate the local validation error. Based on the updated local spectral concentration dataset For spectral coding subnetwork and multi-element concentration prediction subnetwork Corresponding local model parameters Perform training fine-tuning to generate local model update quantities. Next, the first Industrial site Update local model Local sample size Training round number Update timestamp and local verification error Federation upload data packets And upload it to the federated aggregation node; finally, the federated aggregation node processes the uploaded data packets according to the federated data packets. Calculate dynamic aggregate weights And update the local model uploaded by each industrial site currently participating in the aggregation. Perform quality-weighted aggregation to obtain the global model update amount. .

[0029] S4 global model parameter generation, candidate validation, and local model parameter determination. First, the federated aggregation node uses the... Round global model parameters Based on this, combined with the global model update amount obtained from S3 , generate the first Round global model parameters Subsequently, the federated aggregation node will assign global model parameters. The data is then transmitted back to all participating industrial sites, along with the global model version number, training round number, and parameter generation time; then, the... Industrial site Receive global model parameters Next, the integrity of the parameters, the global model version number, and the training epoch number are verified. After the verification is passed, the global model parameters are... Write the parameters to the candidate model parameter storage area to form candidate local model parameters. Finally, the first Industrial site Based on local validation samples, the current local model parameters and candidate local model parameters Calculate the verification error separately, and determine the first verification error based on the local verification error criterion. The local model parameters used in the formal process are then used to invert the concentration of subsequent lubricating oil samples.

[0030] The invention will be further described below with reference to specific embodiments.

[0031] I. Acquisition of Spectral Data of Lubricating Oil Metal Particles and Construction of Characteristic Spectral Line Intensity Sequence S1-1 Lubricating Oil Sample Collection and Spectroscopic Detection. In the industrial field. In this process, lubricating oil samples were collected according to the equipment number, sampling time, and lubricating oil circulation position. Each sample was then dropped onto the surface of a high-temperature resistant spectral sampling substrate to form an oil film sampling point of uniform thickness. The oil film sampling points were then excited using a laser-induced breakdown spectroscopy acquisition device to obtain the original spectral signal. .

[0032] ; In the formula, Industrial site The Middle The original spectral signal of each lubricating oil sample, Indicates the first Spectral intensity at each sampling wavelength point Indicates the first Each sampling wavelength point, This indicates the total number of spectral sampling wavelength points.

[0033] S1-2 Spectral Data Correction and Standardization Processing. This involves processing the original spectral signal... Wavelength calibration, background subtraction, smoothing and denoising, and intensity normalization are performed to obtain a standardized spectral signal. : ; In the formula, Indicates the first Industrial site The Middle Standardized spectral signals of a lubricating oil sample. Indicates the first Standardized spectral intensity at each sampling wavelength point Indicates the first Each sampling wavelength point, This indicates the total number of spectral sampling wavelength points.

[0034] Background subtraction is performed using a local baseline fitting method. Industrial site The Middle The lubricating oil sample was in the first The formula for calculating the background correction peak intensity at each characteristic spectral line is: ; In the formula, Indicates the first Industrial site The Middle The lubricating oil sample was in the first Background correction peak intensity at the characteristic spectral lines Indicates wavelength variable, Indicates the first The wavelength window where the characteristic spectral lines are located, Indicates the normalized spectral signal at wavelength Spectral intensity at that location Indicates the first The local baseline function within the wavelength window where the characteristic spectral lines are located.

[0035] S1-3 Selection of Characteristic Spectral Lines of Metallic Elements. Seven classes of metallic elements with clear indications of equipment wear in lubricating oil wear analysis were selected to form the target metal set. : ; In the formula, Represents the element iron. Represents the element copper. It represents the element chromium. Represents the element aluminum. Represents the element nickel. Represents the element tin. It represents the element lead.

[0036] For the target metal set For each metal element in the spectrum, the center wavelength of its characteristic spectral line is determined based on the standard spectral line library of the target metal element, and then analyzed in the standardized spectral signal. Using the center wavelength of this characteristic spectral line as the center, a wavelength window with a width of 0.40 nm is extracted by extending 0.20 nm in both the short-wavelength and long-wavelength directions. : ; In the formula, Indicates the first The wavelength window corresponding to each characteristic spectral line This indicates the first recorded line in the standard spectral library of the target metallic element. The center wavelength of each characteristic spectral line, 0.20 represents the wavelength width extending into both short and long wavelength directions, in nm.

[0037] Subsequently, in the wavelength window Internal calculation Industrial site The Middle The lubricating oil sample was in the first Peak intensity at characteristic spectral lines Integral strength and background correction intensity Among them, peak intensity Wavelength window Internally normalized spectral signal Maximum spectral intensity, integrated intensity Wavelength window Internally normalized spectral signal The spectral intensity integrals are calculated according to the following formulas: ; ; In the formula, Indicates the first Industrial site The Middle The lubricating oil sample was in the first Peak intensity at characteristic spectral lines, Indicates the first Industrial site The Middle The lubricating oil sample was in the first The integral intensity at each characteristic spectral line. Represents the normalized spectral signal At wavelength Spectral intensity at that location Indicates wavelength variable, Indicates the first The wavelength window corresponding to each characteristic spectral line.

[0038] The first Peak intensity of each characteristic spectral line Peak intensity of the internal standard spectral line corresponding to the same sample recorded in the standard spectral library of the target metal element The ratio is calculated to obtain the spectral line intensity ratio. : ; In the formula, Indicates the first Industrial site The Middle The lubricating oil sample was in the first The ratio of spectral line intensities at each characteristic spectral line. Indicates the first The characteristic spectral lines correspond to the internal standard spectral lines. peak intensity, This indicates the first recorded line in the standard spectral library of the target metallic element. The internal index spectral line number corresponding to each characteristic spectral line.

[0039] Therefore, we obtain the first... Industrial site The Middle The lubricating oil sample was in the first Spectral eigenvectors at each characteristic spectral line As shown below: ; In the formula, Indicates the first Industrial site The Middle The lubricating oil sample was in the first Spectral eigenvectors at characteristic spectral lines. Indicates the peak intensity of the characteristic spectral line. Indicates the integral intensity of the characteristic spectral line. Indicates the background correction peak intensity. This represents the intensity ratio of the characteristic spectral line to the corresponding internal standard spectral line.

[0040] S1-4 characteristic spectral line intensity sequence construction; target metal set The seven metal elements are arranged in the order of iron, copper, chromium, aluminum, nickel, tin, and lead. Within each metal element group, the corresponding spectral feature vectors are arranged in ascending order of the center wavelength of their characteristic spectral lines. These vectors are then concatenated to obtain the... Industrial site The Middle Characteristic spectral intensity sequence of a lubricating oil sample : ; In the formula, Indicates the first Industrial site The Middle Characteristic spectral intensity sequences of a lubricating oil sample. Indicates the first Industrial site The Middle The lubricating oil sample was in the first Spectral eigenvectors at characteristic spectral lines. This represents the total number of characteristic spectral lines corresponding to all target metallic elements.

[0041] S2. Deployment of the local concentration inversion model and generation of metal particle detection results. The overall process is as follows Figure 2 As shown: S2-1 Structure and feature extraction process of the spectral coding subnetwork. (The following is a partial translation of the original text, which is incomplete and requires further context.) Industrial site The Middle Characteristic spectral intensity sequence of a lubricating oil sample As input, the spectral coding subnetwork The main structure uses spectral line channel embedding, spectral line neighborhood convolution, element-relation attention, and pooling compression to progressively compress the feature spectral line intensity sequence. Mapped to.

[0042] Specifically: First, the input feature spectral line intensity sequence is processed using a spectral line channel embedding layer. Perform channel mapping. Characteristic spectral line intensity sequence. Depend on The spectral feature map is composed of spectral feature vectors corresponding to each feature line. Each feature vector contains four types of spectral features: peak intensity, integral intensity, background correction intensity, and internal standard intensity ratio. The spectral channel embedding layer consists of a fully connected layer, a batch normalization layer, and a ReLU activation function. It is used to map the four types of spectral features of each feature line to a unified feature dimension, resulting in a primary spectral feature map. The primary spectral line characteristic diagram Used to represent the intensity response, background correction response, and internal standard normalized response of each characteristic spectral line.

[0043] Secondly, the primary spectral line characteristic map Spectral line neighborhood convolution processing is performed. This processing employs three sets of parallel convolutional structures, using spectral line convolutional kernels of sizes 3, 5, and 7, respectively. Each set includes a spectral line convolutional layer, a batch normalization layer, and a ReLU activation function. The 3-size kernel extracts local intensity variations between adjacent feature lines, the 5-size kernel extracts response relationships between multiple feature lines within the same metal element, and the 7-size kernel extracts spectral interference relationships between different metal elements. The outputs of the three parallel convolutional structures are concatenated and linearly mapped to obtain a multi-scale spectral line feature map. .

[0044] Secondly, the multi-scale spectral feature map Elemental relation attention processing is performed. This process constructs an elemental relation graph based on the spectral overlap, internal standard correlation, and common interference relationships of the oil matrix among seven metallic elements: iron, copper, chromium, aluminum, nickel, tin, and lead. Nodes in the graph represent the spectral characteristics of the corresponding metallic element, and edges represent the spectral interference relationships between different metallic elements. Subsequently, the information transfer weights between nodes of different elements are calculated using a graph attention layer, and the multi-scale spectral feature map is then processed based on these weights. Weighted aggregation is performed to obtain cross-element spectral feature maps. This cross-elemental spectral feature map This is used to represent the interrelationships between spectral lines of the seven metallic elements. Finally, cross-elemental spectral line characteristic maps are analyzed. Residual fusion and pooling compression are performed. Residual fusion processing converts cross-elemental spectral feature maps... Compared with primary spectral line feature map Residual connections are performed to preserve weak response spectral lines and the original spectral line order information. Pooling compression processing consists of an attention pooling layer and a fully connected compression layer. The attention pooling layer is used to weight and summarize the contributions of different feature spectral lines to the concentration inversion, and the fully connected compression layer is used to map the weighted summaries to spectral line fusion features. : ; In the formula, Indicates the first Industrial site The Middle Spectral line fusion characteristics of individual lubricating oil samples Represents the spectral coding subnetwork. Indicates the first Industrial site The Middle Characteristic spectral intensity sequences of a lubricating oil sample.

[0045] S2-2 Structure and concentration inversion process of the multi-element concentration prediction subnetwork. Spectral line fusion features obtained from S2-1. The multi-element concentration prediction subnetwork is used as input. With shared concentration characterization, element-specific prediction, and elemental correlation correction as its main structure, spectral features are integrated layer by layer. Mapped to predicted concentration vectors of seven metal elements .

[0046] Specifically: First, the spectral line fusion characteristics are obtained by utilizing a shared concentration characterization layer. Global concentration feature extraction is performed. The shared concentration characterization layer consists of a first fully connected layer, a batch normalization layer, a ReLU activation function, a Dropout layer, and a second fully connected layer. It is used to extract the oil matrix response features, overall spectral intensity variation features, and laser energy fluctuation compensation features that are commonly dependent on the seven metal elements, thus obtaining the shared concentration characterization. This shared concentration characterization This is used to represent the global spectral information that is commonly relied upon for the concentration inversion of the seven metal elements. Secondly, the shared concentration characterization... Seven element-specific prediction branches are input. These branches correspond to iron, copper, chromium, aluminum, nickel, tin, and lead, respectively. Each element-specific prediction branch consists of a fully connected layer, a ReLU activation function, and a single-element concentration output layer. The seven element-specific prediction branches each start from a shared concentration characterization... The concentration-sensitive features of the corresponding metal elements are extracted, and the initial predicted concentrations of iron, copper, chromium, aluminum, nickel, tin, and lead are output respectively. The seven initial predicted concentrations are concatenated in a fixed order of iron, copper, chromium, aluminum, nickel, tin, and lead to obtain the initial concentration prediction vector. Next, the initial concentration prediction vector... Elemental correlation correction is performed. The elemental correlation correction layer consists of a fully connected layer and a sigmoid activation function. It is used to generate correction weights based on the initial predicted concentrations of seven metal elements, and to compensate for inter-element prediction biases caused by spectral overlap, matrix effects, and laser energy fluctuations, thereby obtaining a corrected concentration prediction vector. Finally, the corrected concentration prediction vector will be... Input concentration output layer. The concentration output layer uses the Softplus non-negative constraint activation function to correct the concentration prediction vector. Negative concentration values ​​are corrected to non-negative values, and predicted concentration vectors for seven metallic elements are output in a fixed order: iron, copper, chromium, aluminum, nickel, tin, and lead. : ; In the formula, Indicates the first Industrial site The Middle Predicted concentration vectors of seven metal elements in a lubricating oil sample. This represents the multi-element concentration prediction subnetwork. Indicates spectral line fusion characteristics, Indicates the first Industrial site In the The local model parameters used during round detection. The predicted concentration vectors of the seven metal elements. Expanding in the fixed order of iron, copper, chromium, aluminum, nickel, tin, and lead: ; In the formula, This indicates the predicted concentration of iron. This indicates the predicted concentration of copper. This indicates the predicted concentration of chromium. This indicates the predicted concentration of aluminum. This indicates the predicted concentration of nickel. This indicates the predicted concentration of tin. This indicates the predicted concentration of lead.

[0047] S2-3 Generation of metal particle concentration detection results and quantitative analysis report. This involves obtaining the predicted concentration vectors of seven metal elements. After that, the Industrial site Predicting concentration vectors for seven metal elements The metal particle concentration detection results are entered into the table according to a fixed order: iron, copper, chromium, aluminum, nickel, tin, and lead. The table includes sample number, sampling time, equipment number, lubricating oil circulation position, and predicted iron concentration. Predicted concentration of copper element Chromium element predicted concentration Predicted concentration of aluminum element Predicted concentration of nickel Predicted concentration of tin element Predicted concentration of lead Local model version number and detection time. Furthermore, the... Industrial site A quantitative analysis report of lubricating oil metal particles is generated based on the metal particle concentration detection results table. This report characterizes the concentration of metal wear particles in the tested equipment at the current sampling time. After obtaining the standard detection concentration results for the lubricating oil sample, error feedback, local model update generation, and federated aggregation are performed according to step S3.

[0048] S3, Detection Error Feedback, Local Model Update Generation, and Federated Aggregation S3-1 Standard detection concentration reception, prediction result matching, and concentration inversion error calculation. Industrial site Obtain the After obtaining the standard detection concentration results for each lubricating oil sample, the sample number, sampling time, equipment number, and lubricating oil circulation position are used as matching indices to retrieve the predicted concentration vector of seven metal elements generated in S2 for that lubricating oil sample. The standard detection concentration results are recorded as the true concentration vector of the seven metal elements. Subsequently, the predicted concentration vectors of the seven metal elements were generated. With the true concentration vector of seven types of metal elements Perform element-by-element comparison and calculate the concentration inversion error of the lubricating oil sample. : ; In the formula, Indicates the first Industrial site The Middle Concentration inversion error of each lubricating oil sample S2 represents the first... Industrial site The Middle Predicted concentration vectors of seven metal elements in a lubricating oil sample. Indicates the first Industrial site The Middle The true concentration vectors of seven metal elements were obtained from a lubricating oil sample through standard testing. This represents the summation of the absolute values ​​of the concentration errors of seven metallic elements: iron, copper, chromium, aluminum, nickel, tin, and lead.

[0049] S3-2 adds local training sample writing, local validation error calculation, and local model update generation. Industrial site The local computing node uses the sample number, sampling time, device number, and lubricating oil circulation position as matching indices to call the characteristic spectral intensity sequence generated in S1 for that lubricating oil sample. and the characteristic spectral line intensity sequence The true concentrations of the seven metal elements obtained from S3-1 are compared with those obtained from S3-1. Pairing is performed to create new local training samples, which are then written into the local spectral concentration dataset. : ; In the formula, Indicates the first Industrial site Local spectral concentration dataset, Indicates the first Industrial site The Middle Characteristic spectral intensity sequences of a lubricating oil sample. Indicates the first Industrial site The Middle The true concentration vectors of seven metal elements were obtained from a lubricating oil sample through standard testing. After adding new local training samples, the... Industrial site Based on the actual concentration vector already obtained in the current round Local validation error was calculated for each lubricating oil sample. : ; In the formula, Indicates the first Industrial site The local verification error, Indicates the first Industrial site The number of lubricating oil samples for which the true concentration vector has been obtained in the current round. Indicates the first The concentration inversion error of the first lubricating oil sample. Subsequently, the... Industrial site Based on the updated local spectral concentration dataset For the spectral coding subnetwork in S2 and multi-element concentration prediction subnetwork Corresponding local model parameters Perform training fine-tuning. During training fine-tuning, use the characteristic spectral line intensity sequence. As input, via the spectral coding subnetwork The spectral line fusion features are obtained, and then processed by a multi-element concentration prediction sub-network. The predicted concentration vectors of seven metal elements were obtained. And compare the predicted concentration vector with the actual concentration vector. The error between them is used to construct the local training loss function. : ; In the formula, Indicates the first Industrial site The local training loss function, Indicates the first Industrial site The number of local samples, This represents the spectral coding subnetwork in S2. This represents the multi-element concentration prediction subnetwork in S2. Indicates the first Industrial site The Middle Characteristic spectral intensity sequences of a lubricating oil sample. Indicates the first Industrial site The Middle The true concentration vectors of seven metal elements in a lubricating oil sample. Indicates the first Industrial site In the The local model parameters used at the start of each training round. This represents the weight regularization coefficient. Industrial site Based on the local training loss function Calculate local model parameters The gradient, combined with the local learning rate. Generate local model update volume : ; In the formula, Indicates the first Industrial site In the The amount of local model updates generated in each round. Indicates the first Industrial site The local learning rate, Represents the local training loss function For local model parameters The gradient. The local model update amount. Used to characterize the Industrial site The spectral coding subnetwork is adjusted based on newly added local training samples. and multi-element concentration prediction subnetwork The direction and magnitude of parameter changes obtained after training and fine-tuning.

[0050] S3-3 Federated Upload Data Packet Generation and Dynamic Aggregation Weight Calculation. Industrial site Update local model Local sample size Training round number Update timestamp and local verification error Federation upload data packets : ; In the formula, Indicates the first Industrial site Federated upload data packets uploaded to the federated aggregation node. Indicates the first Industrial site In the The amount of local model updates generated in each round. Indicates the first Industrial site The number of local samples, Indicates the training round number. Indicates updating the timestamp. Indicates the first Industrial site The local verification error.

[0051] No. Industrial site Upload federated data packets only to federated aggregation nodes The federated aggregation node receives federated upload data packets from each industrial site. Then, based on the number of local samples Update timestamp and local verification error Calculate the first Industrial site Unnormalized weights Specifically, the number of local samples. Used to characterize the degree to which local data from the industrial site supports model updates, update timestamp The local validation error is used to characterize the timeliness of the updated results of the industrial site model. Used to characterize the reliability of the local model update results for this industrial site. Unnormalized weights Calculate according to the following formula: ; In the formula, Indicates the first Industrial site Unnormalized weights Indicates the first Industrial site The number of local samples, Indicates the time decay coefficient. Indicates the current aggregation time and update timestamp The time difference between them Indicates the error suppression coefficient. Indicates the first Industrial site The local verification error. Subsequently, the federated aggregation node performs weight normalization on the industrial sites currently participating in the aggregation, obtaining the first... Industrial site Dynamic aggregate weights : ; In the formula, Indicates the first Industrial site Dynamic aggregation weights, This represents the collection of industrial sites currently participating in the aggregation. Indicates the first Industrial site Unnormalized weights Indicates the first Industrial site The unnormalized weights. Through the above dynamic aggregation weight calculation method, industrial sites with a larger sample size, more recent update time, and lower local validation error receive higher weights in this round of federated aggregation.

[0052] S3-4 Quality-weighted global model update aggregation and reliability improvement. Federated aggregation nodes use the dynamic aggregation weights obtained in S3-3. The amount of local model updates uploaded by each industrial site currently participating in the aggregation. Perform weighted aggregation to obtain the global model update amount. : ; In the formula, Indicates that the federated aggregation node is at the 1st The amount of global model update that is received in turn This represents the collection of industrial sites currently participating in the aggregation. Indicates the first Industrial site Dynamic aggregation weights, Indicates the first Industrial site The number of updates to the uploaded local model.

[0053] global model update amount Used to update the spectral coding subnetwork in S4 and multi-element concentration prediction subnetwork The corresponding global model parameters. Due to the global model update volume. Local model update volume generated from standard detection concentration results at various industrial sites The weighted result is obtained by dynamic aggregation of weights. This approach increases the proportion of updates from industrial sites with larger sample sizes, more recent updates, and lower local validation errors, while reducing the impact of updates from industrial sites with insufficient sample sizes, delayed updates, or large validation errors on the global model parameters. Therefore, after the global model parameters are fed back to each industrial site, each site uses the updated local model parameters to generate predicted concentration vectors for the seven metal elements in the next round of concentration inversion. This makes the next round of prediction results closer to the standard detection concentration results, thereby improving the reliability of subsequent lubricating oil metal particle concentration inversion results.

[0054] S4. Global model parameter generation, candidate validation, and local model parameter determination. The overall process is as follows Figure 3 As shown: S4-1 Global Model Parameter Generation. The federated aggregation node is in the... After the first round of federal aggregation is completed, the [number]th round will be [followed by the] [number]. Round global model parameters Based on this, combined with the global model update amount obtained from S3 , generate the first Round global model parameters : ; In the formula, Indicates the first Round global model parameters, Indicates the first Round global model parameters, Indicates that the federated aggregation node is at the 1st The amount of global model update received in turn. The global model parameters. It is obtained by aggregating the local model update volume uploaded from various industrial sites.

[0055] S4-2 Global model parameter feedback and candidate parameter establishment. The federated aggregation node will transmit the global model parameters generated in S4-1. The data is then transmitted back to all participating industrial sites, and the global model version number, training round number, and parameter generation time are simultaneously distributed. Industrial site Receive global model parameters Next, the integrity of parameters, the global model version number, and the training epoch number are verified; if the verification passes, the current local model parameters are not directly overwritten. Instead, first set the global model parameters Candidate model parameters are written to the candidate model parameter storage area of ​​the local compute node to form candidate local model parameters. : ; In the formula, Indicates the first Industrial site Receive the Candidate local model parameters are generated after rounding the global model parameters. This indicates the first node generated and returned by the federated aggregation node. Round global model parameters. Candidate local model parameters. It is only used for subsequent local verification error checking and will not be used in formal lubricating oil sample testing before the verification is passed.

[0056] S4-3 Local validation error verification and model parameter update determination. Industrial site Call the local validation samples that have already obtained the true concentration vector in the current round, and use the current local model parameters respectively. and candidate local model parameters Concentration inversion of seven types of metal elements was performed to obtain the current local model parameter validation error. And candidate local model parameter validation error If the candidate local model parameter validation error The validation error is no higher than the current local model parameters. With allowable error increment The sum of these values ​​determines whether the candidate local model parameters pass the local validation error check, and the candidate local model parameters are then set as follows: Determined as the number The local model parameters used in the round If the candidate local model parameter validation error The validation error is higher than the current local model parameters. With allowable error increment The sum of these values ​​determines that the candidate local model parameters have failed the local validation error check. Industrial site Instead of using candidate local model parameters, continue using the current local model parameters. Determined as the number The local model parameters used in the round .

[0057] S4-4 Local model parameter update results are determined. After completing local validation error verification and model parameter update determination, the... Industrial site Determine the first The local model parameters used in the round The local model parameters As the first Industrial site The model parameters called during subsequent lubricating oil sample concentration inversion are used to update the predicted concentration vectors of seven types of metal elements, the metal particle concentration detection result table, and the lubricating oil metal particle quantitative analysis report.

[0058] S5. Experimental Verification and Analysis: To quantitatively evaluate the technical performance of the proposed federated learning-based quantitative analysis method for metal particles in lubricating oil in terms of privacy and security, cross-site generalization, and multi-metal concentration inversion accuracy, this embodiment constructs a federated learning simulation experimental system consisting of five heterogeneous industrial sites. The experimental dataset originates from real-time extraction of lubricating oil samples from three typical rotating machinery types: wind power, mining, and shipbuilding. A federated process sample containing 20,000 sets of multi-channel standardized spectral features and corresponding concentration labels is processed through a feature spectral line encoding module. The experimental verification process focuses on three quantitative dimensions: the global convergence efficiency of the federated model, its adaptability to heterogeneous data across different sites, and the performance of the dynamic aggregation mechanism in suppressing inversion errors.

[0059] 1. Analysis of global convergence efficiency and training stability of the federated model Figure 4 The figure shows the global verification error convergence curve of the concentration inversion model proposed in this invention during the federated polymerization process. The horizontal axis represents the federated polymerization round, and the vertical axis represents the mean absolute error values ​​of the seven metal elements. Experiments compare the method of this invention with the traditional FedAvg centralized averaging algorithm. Analysis Figure 4This indicates that, affected by sensor aging and network jitter in industrial settings, the present invention enters a steady-state convergence region around the 120th iteration, and subtle oscillations consistent with engineering logic still exist in the later stages of convergence. In contrast, traditional algorithms converge extremely slowly and produce high residuals when processing heterogeneous spectral data. This demonstrates that the dynamic aggregation mechanism of the present invention can effectively suppress low-quality updates and significantly enhance the robustness of multi-site collaborative training.

[0060] 2. Robustness analysis of the accuracy of multi-metal concentration inversion across different sites Figure 5 The correlation distribution between predicted concentrations and standard detection values ​​of seven target metallic elements—iron, copper, chromium, aluminum, nickel, tin, and lead—at various sites is presented. Analysis Figure 5 It is known that models trained in a single field are subject to interference from laser energy fluctuations and matrix effects when applied across different fields, resulting in prediction errors typically exceeding 35%. This invention extracts cross-elemental response relationships through the feature spectral line encoding module constructed in step S2, and, in conjunction with the global model parameter distribution and update, suppresses the multi-element comprehensive inversion error across the entire range to within the industrial error band of approximately 20%. The scattering divergence characteristics exhibited by the high-concentration samples in the figure conform to the physical characteristics of LIBS laser-induced breakdown spectra in complex oil matrices, verifying the excellent generalization accuracy of this invention when processing non-stationary spectral data.

[0061] 3. Sensitivity analysis of dynamic aggregated weights on model update quality Figure 6 The distribution of unnormalized weights under different aggregation time difference constraints is shown. Analysis Figure 6 This demonstrates that the present invention successfully establishes a nonlinear mapping from update quality to weights by introducing a time decay coefficient and an error suppression coefficient. The non-stationary jitter characteristics exhibited by the curves in the figure simulate the real-world state of industrial field actuators under local computing power fluctuations and communication packet loss. When a local verification error at a certain site increases abnormally or updates are severely delayed, the aggregate weight of that site is automatically reduced, thereby protecting the purity of the global model parameters. This distribution characteristic proves that the present invention, while achieving multi-party collaborative modeling, constructs a federated parameter evolution system with self-purification capabilities, effectively solving the model drift risk caused by imbalanced samples in industrial fields.

[0062] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0063] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for quantitative analysis of lubricating oil metal particles based on federated learning, characterized in that, The process includes the following: Laser-induced breakdown spectroscopy was performed on lubricating oil samples from various industrial sites to obtain raw spectral signals. After preprocessing, the peak intensity, integral intensity, background correction intensity, and spectral line intensity ratio of seven metal elements (iron, copper, chromium, aluminum, nickel, tin, and lead) were extracted and spliced ​​together according to element order and spectral line center wavelength order to obtain characteristic spectral line intensity sequences. The characteristic spectral line intensity sequence is input into the spectral coding sub-network, and the spectral line fusion feature is obtained through spectral line channel embedding, spectral line neighborhood convolution, element relationship attention and pooling compression. Then, it is input into the multi-element concentration prediction sub-network, and the predicted concentration vectors of seven metal elements are obtained through shared concentration characterization, element-specific prediction and element correlation correction. The concentration inversion error is obtained by comparing the predicted concentration vector with the true concentration vector. The local model update amount is generated, and the dynamic aggregation weight is calculated based on the number of samples, the update timestamp, and the local validation error. The global model update amount is then obtained by aggregating the weights. Global model parameters are generated based on the global model update volume and sent back to each industrial site to determine the local model parameters for formal use.

2. The method for quantitative analysis of lubricating oil metal particles based on federated learning as described in claim 1, characterized in that: The preprocessing includes wavelength calibration, background subtraction, smoothing and denoising, and intensity normalization. The background subtraction is performed using a local baseline fitting method.

3. The method for quantitative analysis of lubricating oil metal particles based on federated learning as described in claim 1, characterized in that: The spectral coding subnetwork is based on the first... Industrial site The Middle Characteristic spectral intensity sequence of a lubricating oil sample For input; First, the input feature spectral line intensity sequence is processed using a spectral line channel embedding layer. Channel mapping is performed; the spectral line channel embedding layer consists of a fully connected layer, a batch normalization layer, and a ReLU activation function, which is used to map the four types of spectral line features of each feature line to a unified feature dimension, thereby obtaining a primary spectral line feature map. ; Secondly, the primary spectral line characteristic map Spectral line neighborhood convolution processing is performed. This processing employs three parallel convolutional structures, using kernels of sizes 3, 5, and 7, respectively. The kernel of size 3 is used to extract local intensity changes between adjacent feature lines; the kernel of size 5 is used to extract the response relationship between multiple feature lines within the same metal element; and the kernel of size 7 is used to extract spectral interference relationships between different metal elements. The output results are then concatenated and linearly mapped to obtain a multi-scale spectral line feature map. ; Secondly, the multi-scale spectral feature map Elemental relationship attention processing is performed. An elemental relationship graph is constructed based on the spectral overlap, internal standard correlation, and common interference relationships of the oil matrix among seven metallic elements: iron, copper, chromium, aluminum, nickel, tin, and lead. Nodes in the elemental relationship graph represent the spectral characteristics of the corresponding metallic element, and edges represent the spectral interference relationships between different metallic elements. Subsequently, the information transfer weights between nodes of different elements are calculated through a graph attention layer, and the multi-scale spectral feature graph is then processed based on these information transfer weights. Weighted aggregation is performed to obtain cross-element spectral feature maps. ; Finally, the cross-element spectral line feature map Perform residual fusion and pooling compression processing; Residual fusion processing will transform cross-elemental spectral feature maps Compared with primary spectral line feature map Residual connections are performed; the pooling compression process consists of an attention pooling layer and a fully connected compression layer. The attention pooling layer is used to weight and summarize the contributions of different feature lines to the concentration inversion, and the fully connected compression layer is used to map the weighted and summarized features to spectral fusion features. .

4. A method for quantitative analysis of lubricating oil metal particles based on federated learning as described in claim 1 or 3, characterized in that: The multi-element concentration prediction sub-network is specifically as follows: First, the spectral line fusion characteristics are analyzed using a shared concentration characterization layer. Perform global concentration feature extraction; The shared concentration characterization layer consists of a first fully connected layer, a batch normalization layer, a ReLU activation function, a Dropout layer, and a second fully connected layer. It is used to extract the oil matrix response characteristics, overall spectral intensity variation characteristics, and laser energy fluctuation compensation characteristics that are commonly dependent on seven metal elements, thus obtaining the shared concentration characterization. ; Secondly, Seven element-specific prediction branches are input, each consisting of a fully connected layer, a ReLU activation function, and a single-element concentration output layer; the prediction branches start from the shared concentration characterization. The concentration-sensitive features of the corresponding metal elements are extracted, and the initial predicted concentrations of iron, copper, chromium, aluminum, nickel, tin, and lead are output respectively. The seven initial predicted concentrations are concatenated in a fixed order of iron, copper, chromium, aluminum, nickel, tin, and lead to obtain the initial concentration prediction vector. ; Secondly, an elemental correlation correction layer is used to... Elemental correlation correction is performed. The elemental correlation correction layer consists of a fully connected layer and a sigmoid activation function. It is used to generate correction weights based on the initial predicted concentrations of seven metal elements, and to compensate for inter-element prediction biases caused by spectral overlap, matrix effects, and laser energy fluctuations based on these correction weights, thus obtaining the corrected concentration prediction vector. ; Finally, the corrected concentration prediction vector will be... The input concentration is used to output a layer that employs the Softplus non-negative constraint activation function to correct the concentration prediction vector. Negative concentration values ​​are corrected to non-negative values, and predicted concentration vectors for seven metallic elements are output in a fixed order: iron, copper, chromium, aluminum, nickel, tin, and lead. .

5. The method for quantitative analysis of lubricating oil metal particles based on federated learning as described in claim 1, characterized in that: The specific process for obtaining the global model update amount is as follows: First, when the Industrial site Obtain the After obtaining the standard detection concentration results for each lubricating oil sample, the predicted concentration vectors of the seven metal elements are retrieved using the sample number, sampling time, equipment number, and lubricating oil circulation position as matching indices. The standard detection concentration results are recorded as the true concentration vector of the seven metal elements. Subsequently, the predicted concentration vector will be... With the true concentration vector Perform element-by-element comparison and calculate concentration inversion error. and the obtained characteristic spectral line intensity sequence With the true concentration vector Pairing is performed to create new local training samples, which are then written into the local spectral concentration dataset. Then, calculate the local validation error. Based on the updated local spectral concentration dataset For spectral coding subnetwork and multi-element concentration prediction subnetwork Corresponding local model parameters Perform training fine-tuning to generate local model update quantities. Next, the first Industrial site Update local model Local sample size Training round number Update timestamp and local verification error Federation upload data packets And upload it to the federated aggregation node; finally, the federated aggregation node processes the uploaded data packets according to the federated data packets. Calculate dynamic aggregate weights And update the local model uploaded by each industrial site currently participating in the aggregation. Perform quality-weighted aggregation to obtain the global model update amount. .

6. The method for quantitative analysis of lubricating oil metal particles based on federated learning as described in claim 5, characterized in that: The specific process for determining the local model parameters for the formal invocation is as follows: First, the federated aggregation node is the first Round global model parameters Based on this, combined with the obtained global model update amount , generate the first Round global model parameters ; Subsequently, the federated aggregation node will transfer the global model parameters. The data is then transmitted back to all participating industrial sites, along with the global model version number, training round number, and parameter generation time; then, the... Industrial site Receive global model parameters Next, the integrity of the parameters, the global model version number, and the training epoch number are verified. After the verification is passed, the global model parameters are... Write the parameters to the candidate model parameter storage area to form candidate local model parameters. Finally, the first Industrial site Based on local validation samples, the current local model parameters and candidate local model parameters Calculate the verification error separately, and determine the first verification error based on the local verification error criterion. The parameters of the local model that are officially used in the round.

7. The method for quantitative analysis of lubricating oil metal particles based on federated learning as described in claim 5, characterized in that: During training and fine-tuning, the characteristic spectral line intensity sequence is used. As input, spectral line fusion features are obtained through a spectral coding subnetwork, and then passed through a multi-element concentration prediction subnetwork. The predicted concentration vectors of seven metal elements were obtained. And compare the predicted concentration vector with the actual concentration vector. The error between them is used to construct the local training loss function. : ； In the formula, Indicates the first Industrial site The local training loss function, Indicates the first Industrial site The number of local samples, Represents the spectral coding subnetwork. This represents the multi-element concentration prediction subnetwork. Indicates the first Industrial site The Middle Characteristic spectral intensity sequences of a lubricating oil sample. Indicates the first Industrial site The Middle The true concentration vectors of seven metal elements in a lubricating oil sample. Indicates the first Industrial site In the The local model parameters used at the start of each training round. This represents the weight regularization coefficient.

8. The method for quantitative analysis of lubricating oil metal particles based on federated learning as described in claim 6, characterized in that: The first is determined according to the local verification error criterion. The local model parameters used in the round are as follows: No. Industrial site Call the local validation samples that have already obtained the true concentration vector in the current round, and use the current local model parameters respectively. and candidate local model parameters Concentration inversion of seven types of metal elements was performed to obtain the current local model parameter validation error. And candidate local model parameter validation error If the candidate local model parameter validation error The validation error is no higher than the current local model parameters. With allowable error increment The sum of these values ​​determines whether the candidate local model parameters pass the local validation error check, and the candidate local model parameters are then set as follows: Determined as the number The local model parameters used in the round If the candidate local model parameter validation error The validation error is higher than the current local model parameters. With allowable error increment The sum of these values ​​determines that the candidate local model parameters have failed the local validation error check. Industrial site Instead of using candidate local model parameters, continue using the current local model parameters. Determined as the number The local model parameters used in the round .

Images