A Method and System for Monitoring Conductive Deposition Based on Gas-Liquid Phase Separation

Abstract

This invention discloses a method and system for monitoring conductive deposition based on gas-liquid phase separation, relating to the fields of artificial intelligence and machine learning. The method includes: acquiring and preprocessing a multi-dimensional time-series dataset using a monitoring system; constructing a MsDL-AimTS fusion model, inputting the preprocessed multi-dimensional time-series dataset into the fusion model, predicting resistance change trends based on a multi-scale time-series dynamic learning framework, and identifying early subtle anomalies based on an enhanced sequence-image contrast learning framework; calculating the resistance change rate, temperature influence factor, and monitoring duration based on the preprocessed multi-dimensional time-series dataset, and calculating the conductive deposition coefficient; and outputting the deposition risk level and determining compliance based on the conductive deposition coefficient, resistance change trend prediction results, and early anomaly identification results according to preset fusion judgment rules. This invention provides more efficient and reliable technical support for the performance evaluation of electric vehicle lubricants and the prevention of transmission system failures.

Description

Technical Field

[0001] The embodiments of the present invention relate to the fields of artificial intelligence and machine learning technology, and in particular to a conductive deposition monitoring method and system based on gas-liquid phase separation. Background Technology

[0002] Exposed components of electric vehicle drive systems, such as windings and electrical contacts, are prone to forming conductive deposits due to lubricants, which can lead to catastrophic failures. Although existing conductive deposit detection equipment can perform basic monitoring in both the gas and liquid phases, it relies solely on manual analysis of resistance change data, resulting in low data processing efficiency, inability to accurately predict deposition trends, and insensitivity to early, subtle resistance changes.

[0003] Traditional detection techniques lack the ability to deeply mine historical data, making it difficult to establish a correlation model between multiple parameters (temperature, time, resistance) and the conductive deposition state. Specifically, they have the following shortcomings:

[0004] First, existing methods mostly employ single-scale modeling, failing to comprehensively capture the dynamic features of different time scales in time-series data, resulting in limited accuracy in predicting depositional trends. Second, the lack of adaptive model parameter optimization mechanisms makes it difficult to adjust model configurations according to the dynamic characteristics of different monitoring scenarios, leading to poor versatility. Third, relying solely on single time-series modal data for analysis makes it difficult to capture structural change features corresponding to weak resistance fluctuations, resulting in insufficient sensitivity in identifying early anomalous signals. Fourth, semantic distortion is prone to occur during data augmentation, affecting the model's robustness to real monitoring data. Fifth, the lack of effective loss function guidance for cross-modal feature fusion makes it difficult to achieve deep alignment between numerical and structural information, further limiting the accuracy of anomaly identification.

[0005] These shortcomings prevent traditional detection technologies from providing early warnings of potential faults, and the detection results are easily affected by human analytical errors, limiting their reliability and making it difficult to meet the high precision and high sensitivity requirements of electric vehicle transmission systems for conductive deposition monitoring. Summary of the Invention

[0006] This invention provides a conductive deposition monitoring method and system based on gas-liquid phase separation, which provides more efficient and reliable technical support for the performance evaluation of electric vehicle lubricants and the prevention of transmission system failures.

[0007] In a first aspect, embodiments of the present invention provide a method for monitoring conductive deposition based on gas-liquid phase separation, comprising:

[0008] A multi-dimensional time-series dataset is collected based on a pre-built multi-module collaborative monitoring system, and the multi-dimensional time-series dataset is preprocessed.

[0009] Construct a MsDL-AimTS fusion model, wherein the MsDL-AimTS fusion model includes a multi-scale temporal dynamic learning framework and an enhanced sequence-image contrast learning framework;

[0010] The preprocessed multidimensional time series dataset is input into the MsDL-AimTS fusion model, and resistance change trend is predicted according to the multi-scale time series dynamic learning framework, and early weak anomaly identification is performed according to the enhanced sequence-image contrast learning framework.

[0011] The resistance change rate, temperature influence factor and monitoring duration are calculated based on the preprocessed multidimensional time series dataset, and the conductivity deposition coefficient is calculated based on the resistance change rate, temperature influence factor and monitoring duration.

[0012] Based on the predicted results of the conductivity deposition coefficient, resistance change trend, and early anomaly identification, the deposition risk level is output and the qualification is determined according to the preset fusion judgment rules.

[0013] In a preferred embodiment, the data preprocessing of the multidimensional time-series dataset includes:

[0014] Different interpolation strategies are used to fill in missing values ​​based on the degree of continuity of missing data in the multidimensional time series dataset;

[0015] An anomaly detection method based on statistical distribution identifies and removes outliers from the multidimensional time-series dataset.

[0016] Time alignment of data from different monitoring sources in the multidimensional dataset is performed based on timestamps;

[0017] The time-series data, after missing value imputation, outlier removal, and data alignment, are standardized to unify the numerical scale of each monitoring parameter.

[0018] As a preferred embodiment, predicting the resistance change trend based on the multi-scale time-series dynamic learning framework includes:

[0019] Configure multiple reservoir computing units with different time scales;

[0020] The reservoir computing unit performs dynamic representation learning on the preprocessed time series data to obtain an initial multi-scale dynamic representation.

[0021] An iterative interval narrowing search algorithm is adopted, with the combined loss function as the optimization objective, to adaptively adjust the radius of the connection weight spectrum of each reservoir calculation unit in order to obtain the optimized model parameters.

[0022] Based on the optimized model parameters, dynamic representation learning is performed again to obtain the final multi-scale dynamic representation.

[0023] The final multi-scale dynamic representation is fused into a unified high-dimensional feature vector and input into a machine learning model to predict future resistance change trends and deposition formation risks.

[0024] As a preferred implementation, early weak anomaly identification is performed based on an enhanced sequence-image contrast learning framework, including:

[0025] The preprocessed multidimensional time-series dataset is input into the time-series encoder and the image encoder respectively to extract time-series numerical features and image structural features.

[0026] Cross-modal alignment and fusion of the temporal numerical features and image structural features are performed through a contrastive learning mechanism to obtain cross-modal fused features;

[0027] Anomaly detection is performed based on the cross-modal fusion features. When the detection result meets the preset conditions, it is determined to be an early anomaly and an early warning is triggered.

[0028] In a preferred embodiment, the step of performing cross-modal alignment and fusion of the temporal numerical features and image structural features through a contrastive learning mechanism to obtain cross-modal fused features includes:

[0029] Data augmentation is performed on data samples in the multidimensional time-series dataset to generate an augmented view, and prototype contrastive loss is calculated based on the augmented view;

[0030] The time-series numerical features and image structural features are subjected to bidirectional comparative learning to calculate the sequence-image contrast loss.

[0031] The prototype contrast loss and the calculated sequence-image contrast loss are weighted and fused together, and the temporal encoder and image encoder are jointly optimized through backpropagation.

[0032] Based on the jointly optimized temporal encoder and image encoder, temporal numerical features and image structural features are extracted again, respectively.

[0033] The cross-modal fusion feature is obtained by fusing the re-extracted temporal numerical features with the image structural features.

[0034] In a preferred embodiment, calculating the prototype contrast loss based on the enhanced view includes:

[0035] The enhanced view is input into a temporal encoder to extract enhanced features, and sample prototype representations are generated by mean aggregation.

[0036] Calculate intra-prototype loss and inter-prototype loss based on the sample prototype representation;

[0037] The intra-prototype loss and inter-prototype loss are weighted and fused to obtain the prototype contrast loss.

[0038] In a preferred embodiment, the temporal numerical features and image structural features are subjected to bidirectional contrast learning, and the sequence-image contrast loss is calculated, including:

[0039] The time-series numerical features and image structural features are used as positive sample pairs to construct a cross-modal contrastive learning task;

[0040] Based on the cross-modal contrastive learning task, calculate the bidirectional naive contrastive loss and geodesic mixing loss respectively;

[0041] The naive contrast loss and the geodesic mixing loss are weighted and fused to obtain the sequence-image contrast loss.

[0042] In a preferred embodiment, the calculation of the resistance change rate, temperature influence factor, and monitoring duration based on the preprocessed multidimensional time-series dataset, and the calculation of the conductivity deposition coefficient based on the resistance change rate, temperature influence factor, and monitoring duration, includes:

[0043] Based on the current resistance value and the initial resistance value in the multidimensional time series dataset, the resistance change rate is calculated;

[0044] Based on the actual temperature values ​​and preset standard temperature values ​​in the multidimensional time-series dataset, the temperature influence factor is calculated.

[0045] Based on the timestamp information of the multidimensional time series dataset, the cumulative monitoring duration from the start of monitoring to the current moment is calculated;

[0046] The conductivity deposition coefficient is obtained by multiplying the resistance change rate, temperature influence factor and monitoring duration by their respective weighting coefficients and then summing them by weight.

[0047] In a preferred embodiment, based on the conductivity deposition coefficient, the predicted resistance change trend, and the early anomaly identification results, and according to a preset fusion judgment rule, the deposition risk level is output and a pass / fail determination is performed, including:

[0048] Based on a comprehensive evaluation of the conductivity deposition coefficient, the predicted resistance change trend, and the early anomaly identification results, the deposition risk is divided into multiple levels and corresponding qualification judgment results are output.

[0049] Secondly, embodiments of the present invention also provide a conductive deposition monitoring system based on gas-liquid phase separation, comprising:

[0050] The data acquisition and preprocessing module is used to acquire multidimensional time-series datasets based on a pre-built multi-module collaborative monitoring system, and to preprocess the multidimensional datasets.

[0051] The MsDL-AimTS fusion model construction module is used to construct the MsDL-AimTS fusion model, wherein the MsDL-AimTS fusion model includes a multi-scale temporal dynamic learning framework and an enhanced sequence-image contrast learning framework.

[0052] The resistance change trend prediction module is used to input the preprocessed multi-dimensional time series dataset into the MsDL-AimTS fusion model, predict the resistance change trend according to the multi-scale time series dynamic learning framework, and identify early weak anomalies according to the enhanced sequence-image contrast learning framework.

[0053] The conductive deposition coefficient calculation module is used to calculate the resistance change rate, temperature influence factor and monitoring duration based on the preprocessed multi-dimensional time series dataset, and to calculate the conductive deposition coefficient based on the resistance change rate, temperature influence factor and monitoring duration.

[0054] The deposition risk level determination module is used to output the deposition risk level and determine its compliance based on the conductivity deposition coefficient, the resistance change trend prediction result, and the early anomaly identification result, according to the preset fusion determination rules.

[0055] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising:

[0056] One or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the conductive deposition monitoring method based on gas-liquid phase separation as described in any embodiment of the present invention.

[0057] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the conductive deposition monitoring method based on gas-liquid phase separation as described in any embodiment of the present invention.

[0058] The present invention achieves the following beneficial effects:

[0059] (1) By constructing the MsDL-AimTS fusion model, this invention combines the dual advantages of multi-scale time-series dynamic learning and cross-modal comparative learning to achieve accurate prediction of resistance change trends and efficient identification of early weak anomalies, while improving monitoring efficiency and result reliability.

[0060] (2) This invention proposes a multi-scale time-series dynamic learning method, which comprehensively captures the dynamic characteristics of time-series data at different time scales through multiple reservoir computing units with different recursive connection steps, providing more comprehensive feature support for sedimentation trend prediction.

[0061] (3) This invention proposes an iterative interval narrowing search algorithm to optimize the spectrum radius of each reservoir and use the combined loss function as the optimization target to achieve adaptive adjustment of model parameters and improve the adaptability of the model to different monitoring scenarios.

[0062] (4) This invention proposes an enhanced sequence-image contrast learning framework, which enhances the ability to identify early abnormal signals such as small resistance fluctuations by fusing and contrasting time-series and image dual-modal features, and solves the problem of insufficient sensitivity of single-modal monitoring.

[0063] (5) The prototype contrast loss of this invention effectively avoids semantic distortion caused by data augmentation, improves the robustness of the model to monitoring data disturbances, and ensures the stability and discriminability of feature representation.

[0064] (6) The present invention designs a sequence-image contrast loss, including naive contrast loss and geodesic hybrid loss, to achieve deep alignment and fusion of dual-modal features, fully explore the complementary value of numerical information and structural information, further improve the accuracy of early anomaly identification, and ultimately provide more efficient and reliable technical support for the performance evaluation of electric vehicle lubricants and the prevention of transmission system failures. Attached Figure Description

[0065] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0066] Figure 1 This is a flowchart of a conductive deposition monitoring method based on gas-liquid phase separation provided in an embodiment of the present invention;

[0067] Figure 2 This is a schematic diagram of the overall framework of the conductive deposition monitoring method based on gas-liquid phase separation provided in the embodiments of the present invention;

[0068] Figure 3 This is a flowchart of resistance change trend prediction based on a multi-scale time-series dynamic learning framework provided in an embodiment of the present invention;

[0069] Figure 4 This is a flowchart of the enhanced sequence-image contrast learning framework for early weak anomaly identification provided in this embodiment of the invention;

[0070] Figure 5 This is a schematic diagram of the structure of a conductive deposition monitoring system based on gas-liquid phase separation provided in an embodiment of the present invention;

[0071] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0073] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0074] Example 1: As Figure 1 The diagram shows a flowchart of a conductive deposition monitoring method 100 based on gas-liquid phase separation provided in Embodiment 1 of the present invention. The method 100 specifically includes the following steps:

[0075] S110. Collect multi-dimensional time-series datasets based on a pre-built multi-module collaborative monitoring system, and perform data preprocessing on the multi-dimensional time-series datasets.

[0076] In some embodiments, the monitoring system that enables multi-module collaborative operation specifically includes:

[0077] The sample containment module, temperature monitoring module, high-precision resistance acquisition unit, testing module, circuit module, intelligent data processing module and auxiliary components, etc. Among them, the temperature monitoring module is equipped with a gas phase temperature sensor (monitoring the temperature of the gas phase environment above the sample) and a liquid phase temperature sensor (monitoring the temperature of the liquid phase inside the sample), with a sampling frequency of 1Hz.

[0078] This monitoring system synchronously monitors the gas phase loop resistance (Rg) and liquid phase loop resistance (Rl) using a high-precision resistance acquisition unit, and synchronously records the gas phase temperature (Tg) and liquid phase temperature (Tl) using a temperature monitoring module. The data is timestamped with the resistance data (error ≤ 1ms), and a segmented acquisition strategy is employed. For example, data is acquired every 10 seconds for the first 100 hours, every 30 seconds from 100 to 300 hours, and every 60 seconds from 300 to 500 hours, balancing the data volume with the need to capture key information. After 500 hours of continuous monitoring, a time-series dataset containing five dimensions of parameters—timestamp, Rg, Rl, Tg, and Tl—is generated.

[0079] In addition, the monitoring system has a built-in outlier detection mechanism. When a parameter exceeds the physical reasonable range (such as resistance < 1kΩ or temperature > 250℃), it automatically marks the parameter and triggers repeated acquisition. If a device failure occurs during the acquisition process, the system automatically records the interruption time and starts the compensation acquisition mode after recovery to ensure data integrity. Through systematic hardware configuration, strict environmental control and multi-dimensional synchronous acquisition strategy, the monitoring system provides high-quality basic data for subsequent deep learning analysis, effectively avoiding the phase interference, data loss and time misalignment problems existing in traditional acquisition methods.

[0080] In some embodiments, to provide high-quality input, the multidimensional time-series dataset undergoes data preprocessing, including:

[0081] (1) Missing value handling: For the missing data marked during the collection process, an improved interpolation algorithm is adopted: For segments with ≤5 consecutive missing data points, local weighted regression (LOESS) interpolation is used; for segments with >5 consecutive missing data points, a generative adversarial network is constructed based on historical sample data under the same temperature conditions to generate imputation data and label the confidence level (≥0.85 is acceptable).

[0082] (2) Outlier removal: Extreme values ​​in resistance and temperature data are initially screened using the 3σ criterion. Local standard deviations are calculated using a sliding window (window size = 30 data points) to identify sudden abnormal fluctuations (such as instantaneous voltage interference). After removing outliers, linear interpolation is used to fill in the gaps.

[0083] (3) Data alignment: The gas / liquid phase data are synchronized twice based on the timestamp to ensure that the time deviation between Rg and Tg, and between Rl and Tl is ≤0.5ms, thus eliminating the phase difference caused by hardware delay.

[0084] (4) Data standardization: Zero-mean standardization is performed on the preprocessed time series data to unify the data scale and provide a stable input distribution for model training.

[0085] S120. Construct the MsDL-AimTS fusion model, wherein the MsDL-AimTS fusion model includes a multi-scale temporal dynamic learning framework and an enhanced sequence-image contrast learning framework.

[0086] In some embodiments, a MsDL-AimTS fusion model is constructed, which combines the advantages of multi-scale temporal dynamic learning and cross-modal comparative learning to achieve resistance change trend prediction and early anomaly identification.

[0087] S130. Input the preprocessed multidimensional time series dataset into the MsDL-AimTS fusion model, predict the resistance change trend according to the multi-scale time series dynamic learning framework, and identify early weak anomalies according to the enhanced sequence-image contrast learning framework.

[0088] In some embodiments, predicting resistance change trends based on a multi-scale temporal dynamic learning framework includes:

[0089] Multiple reservoir computing units with different time scales are configured; the preprocessed time-series data are dynamically represented and learned through the reservoir computing units to obtain an initial multi-scale dynamic representation; an iterative interval narrowing search algorithm is adopted, with the combined loss function as the optimization objective, to adaptively adjust the connection weight spectrum radius of each reservoir computing unit to obtain optimized model parameters; dynamic representation learning is performed again based on the optimized model parameters to obtain the final multi-scale dynamic representation; the final multi-scale dynamic representation is fused into a unified high-dimensional feature vector and input into a machine learning model to predict future resistivity change trends and deposition formation risks.

[0090] Specifically, configure K reservoir computing units (default). Each reservoir corresponds to The connection step length is determined by establishing a mapping relationship between historical data and future states through a k-step prediction task. The reservoir size is fixed at 20, the input weight scaling factor is set to 0.5, and the regularization parameter is... The reservoir connectivity is 0.5.

[0091] (a) Dynamic representation learning

[0092] For the preprocessed five-dimensional time series data (T is the time step), the hidden state of each k-hop reservoir is calculated using the following formula:

[0093]

[0094] in, The preprocessed five-dimensional time series dataset contains "timestamp-Rg-RI-Tg-TI"; Time series data in the first The five-dimensional parameter vector of the time step; : The total time step of the time series data; : Five-dimensional real vector space; : No. The first skip reservoir in the first step The hidden state vector at each time step; : Recursive connection step length of the reservoir; : Activation functions are used to perform nonlinear transformations on the reservoir hidden state, thereby enhancing the feature representation capability. : No. The connection weight spectrum radius of each reservoir is optimized through iterative interval narrowing search (IINS) to control the stability of the hidden state; : Normalized weight matrix of reservoir internal connectivity, dimension is The spectral radius is set to 1 to ensure numerical stability. Input weight matrix, used to map five-dimensional time series data to the reservoir hidden state space; : The initial hidden state vector of the reservoir, with a dimension of 20, and the initial values ​​are set to all zeros; Time step index (from 1 to 1) ).

[0095] The hidden state is mapped to future sequence elements by output weights. :

[0096]

[0097] in, : No. The first skip reservoir in the first step The hidden state vector of the time step stores the dynamic feature monitoring parameters of historical monitoring data; : No. The output weight matrix of each skip-step reservoir is used to capture the temporal dependency of the monitoring parameters, with a dimension of [missing information]. The solution formula is:

[0098]

[0099] in, : Target matrix, composed of arrive It consists of all the real five-dimensional monitoring parameter vectors, with dimensions of ; : Hidden state matrix, composed of the first The reservoir is in arrive The hidden state vectors at each time step are concatenated column by column; Hidden State Matrix The transpose of the matrix, with dimension 0; : Regularization parameters for ridge regression; Identity matrix, dimensions and Consistency (i.e.) This ensures the validity of matrix operations; The inverse of the regularized hidden state covariance matrix, used to solve for the optimal output weights. .

[0100] (b) Adaptive optimization

[0101] This invention proposes an Iterative Interval Narrowing Search (IINS) algorithm to optimize the spectral radius of each reservoir. The optimization objective is the combined loss function. :

[0102]

[0103] in, : Spectral radius vectors of all reservoirs, each Corresponding to the The radius of the connection weight spectrum of a skip-step reservoir; Loss balance coefficient, used to adjust the contribution weights of fitting loss and separation loss; : No. The fitting loss for each reservoir is calculated using the following formula:

[0104]

[0105] in, The Frobenius norm is used to calculate the magnitude of a matrix or vector, measuring the difference between the predicted and actual values ​​of a monitoring parameter. Total number of training samples; : Regularization hyperparameters of the fitting loss; : No. The sample at the th The output weight matrix of each reservoir has a dimension of . ; : No. The sample at the th The hidden state matrix of each reservoir has dimensions of . ; : No. The target matrix of each sample is composed of arrive The true five-dimensional monitoring parameters consist of the following dimensions: ; : No. The separation loss of each reservoir enhances the discriminative power of dynamic characterization of different types of samples, aiding in anomaly identification. The formula is as follows:

[0106]

[0107] in, : with the The total number of samples in the same category is used to normalize the logarithmic term in the separation loss. Sample index; : No. The category label for each sample, with values ​​of normal or abnormal; : No. The first reservoir The sample and the first The RBF kernel similarity of the dynamic representations of two samples measures the degree of similarity in the temporal dynamic features of the two samples. The formula is as follows:

[0108]

[0109] in, RBF kernel parameters control the sensitivity of similarity calculation and affect the class differentiation effect; : No. The sample at the th The output weight matrix of each reservoir has a dimension of . .

[0110] (c) Prediction of resistance change trend

[0111] Based on learned multi-scale dynamic representation The future resistivity trend and potential deposition formation time are predicted using a random forest classifier (100 decision trees).

[0112] To adapt to a random forest classifier (100 decision trees), multi-scale dynamic representations are needed. The features are concatenated into a unified high-dimensional feature vector. Let each... The dimension is (5 is the input parameter dimension, 20 is the reservoir size), then the concatenated feature vector Dimensions The mathematical expression is:

[0113]

[0114] in, : No. A multi-scale dynamic representation set of samples, containing The output weight matrix of each reservoir, Dimensions ; : No. The high-dimensional feature vector of each sample is composed of multi-scale dynamic representations concatenated; Concatenate Dimensional splicing operation, along the number of reservoirs Dimensional integration of all This enables multi-scale feature fusion.

[0115] Furthermore, for the input feature vector , No. Decision Tree The prediction result is (0 = no deposition risk, 1) Potential deposits, (Present sedimentation), final ensemble prediction results of random forest for:

[0116]

[0117] in, : Decision tree index, corresponding to a single decision tree in a random forest; : No. decision trees for the first The prediction result for each sample takes a value of ,in No sedimentation risk Potential sediments, Deposits are about to form; Predicting candidate values ​​for the categories, i.e. Any of the deposition risk levels; Random forest for the first The ensemble prediction results for each sample represent the depositional risk level that received the most votes from the 100 decision trees. : Indicator function, when the first Prediction results of decision trees equals category The value is 1 if the condition is met, and 0 otherwise. :category The total number of votes received across 100 decision trees is used to determine the final integrated prediction result; The operation is to select the category corresponding to the maximum value, that is, to select the sedimentation risk level with the most votes as the final prediction result.

[0118] In some embodiments, early weak anomaly identification based on an enhanced sequence-image contrast learning framework includes:

[0119] The preprocessed multidimensional time-series dataset is input into the time-series encoder and the image encoder respectively to extract time-series numerical features and image structural features. The time-series numerical features and image structural features are aligned and fused across modally through a contrastive learning mechanism to obtain cross-modal fusion features. Anomaly detection is performed based on the cross-modal fusion features. When the detection results meet the preset conditions, they are judged as early anomalies and an early warning is triggered.

[0120] Specifically, cross-modal alignment and fusion of the temporal numerical features and image structural features are performed through a contrastive learning mechanism to obtain cross-modal fused features, including:

[0121] Data augmentation is performed on data samples in a multidimensional time-series dataset to generate an augmented view, and a prototype contrast loss is calculated based on the augmented view. Temporal numerical features and image structural features are subjected to bidirectional contrast learning to calculate a sequence-image contrast loss. The prototype contrast loss and the calculated sequence-image contrast loss are weighted and fused, and the temporal encoder and image encoder are jointly optimized through backpropagation. Based on the jointly optimized temporal encoder and image encoder, temporal numerical features and image structural features are re-extracted. The re-extracted temporal numerical features and image structural features are fused to obtain the cross-modal fusion feature.

[0122] Further, the prototype contrast loss is calculated based on the enhanced view, including:

[0123] The enhanced view is input into the temporal encoder to extract enhanced features, and sample prototype representations are generated by mean aggregation. Then, the intra-prototype loss and inter-prototype loss are calculated based on the sample prototype representations. Finally, the intra-prototype loss and inter-prototype loss are weighted and fused to obtain the prototype contrast loss.

[0124] Furthermore, a bidirectional contrastive learning process is performed between temporal numerical features and image structural features to calculate the sequence-image contrastive loss, including:

[0125] A cross-modal contrastive learning task is constructed by using temporal numerical features and image structural features as positive sample pairs. Then, based on the cross-modal contrastive learning task, bidirectional naive contrastive loss and geodesic mixture loss are calculated respectively. The naive contrastive loss and geodesic mixture loss are weighted and fused to obtain the sequence-image contrastive loss.

[0126] Specifically, this invention extracts early, weak anomalous features using an enhanced sequence-image contrastive learning (AimTS) framework. The core of this framework lies in enhancing the ability to identify early anomalous signals such as minute resistance fluctuations through dual-modal feature fusion and contrastive learning. This is achieved through two main components: prototype contrastive learning and sequence-image contrastive learning.

[0127] First, prototype comparative learning includes the following steps:

[0128] (a) Image conversion

[0129] Multidimensional time-series data is plotted as line charts by color-coding variables (time axis as x-axis, parameter values ​​as y-axis), with different variables identified by distinct colors (e.g., Rg as red, Rl as blue, Tg as green, Tl as yellow). The subplots corresponding to each variable are then stitched together into a uniform-sized RGB image. This transformation process preserves the trend and structural characteristics of the time-series data, making previously imperceptible subtle fluctuations appear as distinguishable morphological differences in the image, providing a structural basis for cross-modal comparisons.

[0130] (b) Dual-modal feature extraction

[0131] Numerical features of the data are extracted using a temporal encoder (TSEncoder, which employs a 3-layer convolutional neural network structure). It captures the quantitative patterns of resistance and temperature changes over time; and extracts structural features using an image encoder (based on a lightweight ResNet18 model). This captures morphological patterns (such as abrupt slope changes and fluctuation frequency variations) in the line graph. To achieve comparability of cross-modal features, a nonlinear projection head consisting of two fully connected layers is used to map the features of both modalities to a low-dimensional space of the same dimension, resulting in:

[0132] ;

[0133] in, : No. The temporal numerical feature vector of each sample is extracted by the temporal encoder and has a high dimension. Feature extraction function of a temporal encoder; : No. One original time series sample; : No. Image structural feature vectors of each sample; The feature extraction function of the image encoder is built based on the ResNet18 lightweight model, and the input is the RGB image after temporal sample transformation; Image : No. RGB images of time-series samples after conversion; A nonlinear projection head for time-series features, consisting of two fully connected layers, projects high-dimensional time-series features. Mapped to a 256-dimensional low-dimensional space; A nonlinear projection head for image features, consisting of two fully connected layers, projects high-dimensional image features. Mapped to a 256-dimensional low-dimensional space; : No. The low-dimensional temporal feature vectors of each sample after projection, with dimensions of _ ... Dimension, used for cross-modal contrastive learning; : No. The low-dimensional image feature vectors of each sample after projection have a dimension of 256 and are used for cross-modal contrastive learning.

[0134] (c) Prototype Comparison Learning

[0135] To address the semantic distortion that may result from single data augmentation (such as excessive jitter potentially altering the true physical meaning of resistance data), robust time-series data representations are constructed through multi-augmented view aggregation and contrastive learning.

[0136] (c1) Prototype generation

[0137] For each time series sample Five augmentation methods (jitter, scaling, time warp, slicing, and window warp) are randomly selected from the data augmentation library. Two sets of augmented views with different parameters are generated for each method, resulting in 10 augmented samples (i.e.,...). and These enhanced views are then input into the TS encoder to obtain the corresponding high-dimensional representations. and By eliminating semantic biases from individual enhancements through mean aggregation, a prototype representation of the sample is generated:

[0138]

[0139] in, : No. The first original sample was passed through the... The first set of enhanced views generated by this enhancement method;

[0140] : No. The first original sample was passed through the... The second set of enhanced views generated by this enhancement method; Enhanced View High-dimensional temporal feature vectors extracted by the TS encoder; Enhanced View High-dimensional temporal feature vectors extracted by the TS encoder; : No. The first set of prototype representations for each sample is aggregated from the mean of high-dimensional temporal features of five enhanced views, and then... The projection yields a dimension of 256. : No. The second set of prototype representations for each sample consists of high-dimensional temporal features from five enhanced views. After mean aggregation, The projection yields a dimension of 256.

[0141] This prototype integrates features from various enhanced scenarios, preserving the core semantics of the samples while improving robustness to data perturbations.

[0142] (c2) Intra-prototype loss

[0143] To avoid feature dominance of prototype representation by some augmented views, the distance constraints between different augmented views are adjusted by adaptive temperature parameters. First, the Euclidean distance between different augmented views of the same sample is calculated. The greater the distance, the more significant the feature differences brought about by the enhancement; then, the softmax function is used to map the distance to a temperature parameter. The greater the distance, the higher the temperature, and the weaker the separation penalty for that group of views; conversely, the smaller the distance, the stronger the penalty, ensuring that the features of each enhanced view are evenly distributed.

[0144]

[0145] in, : No. In the nth sample species and first The Euclidean distance between views generated by the enhancement method Euclidean distance calculation function : No. In the nth sample species and first An adaptive temperature parameter for the enhanced view is used to adjust the penalty intensity for contrast loss; : Base temperature parameters, ensuring the temperature is positive and stable; Euclidean distance The exponential transformation is used to map distance to a temperature adjustment factor; : No. In the nth sample The distance exponent of the enhanced view and all 5 enhanced views is used to normalize the temperature parameter.

[0146] Furthermore, based on adaptive temperature calculation of intra-prototype contrast loss, two sets of views with the same enhancement method (such as two views with different parameters of jitter enhancement) are forced to become positive sample pairs, and views with different enhancement methods are forced to become negative sample pairs:

[0147]

[0148] in, : No. In the nth sample Two sets of views for different enhancement methods and The similarity score, i.e., the similarity between positive sample pairs, is calculated as follows: ( for Features after projection); : No. In the nth sample species and first The similarity score of the enhanced view is calculated as follows: ( for Features after projection); Indicator function, when It takes a value of 1 when the condition is met and 0 otherwise, and is used to filter negative sample pairs with different enhancement methods. : No. The first sample Enhanced view The low-dimensional feature vector after processing by the time encoder and projection head has 256 dimensions. : No. The first sample Enhanced view The low-dimensional feature vector after processing by the time encoder and projection head has a dimension of 256.

[0149] (c3) inter-prototype loss

[0150] To enhance the discriminatory power between different samples, two prototypes of the same sample ( and The original samples are considered positive samples, and all prototypes of other samples are considered negative samples. The contrastive loss is used to bring the prototypes of the same type of samples closer together and push the prototypes of different types of samples further apart.

[0151]

[0152] in, : No. Inter-prototype loss for each sample, : No. The first set of prototype representations for each sample; : No. The second set of prototype representations for each sample; : Global temperature parameter, default value is 0.07. Training batch size, which is the number of samples input to the model for each training iteration; : Sample index within the batch; Indicator function, when It takes a value of 1 when the condition is met and 0 otherwise, and is used to filter negative sample pairs from different samples. : No. The first set of prototype representations for each sample; : No. The second set of prototype representations for each sample; : No. The similarity index of a sample with all negative sample pairs in the batch, where the negative sample includes two sets of prototype representations of other samples.

[0153] (c4) Total prototype loss

[0154] Through weighting coefficients Balancing the contributions of intra-prototype and inter-prototype losses ensures the model possesses both robustness and discriminative power.

[0155]

[0156] in, Total prototype loss; Loss balance coefficient, default value is 0.7; The loss of all samples within a batch is averaged, and 2B is the normalization coefficient. : No. The intra-prototype loss for each sample is used to constrain the feature distribution of different augmented views of the same sample.

[0157] In some embodiments, sequence-image contrast learning specifically includes the following steps:

[0158] To integrate the numerical information of time-series data with the structural information of images, a unified feature space is constructed through cross-modal contrastive learning to compensate for the information deficiencies of a single modality: (d1) Naive contrastive loss

[0159] Time series samples and their corresponding images are considered positive sample pairs (e.g., Rg-RI-Tg-TI time series data and the converted RGB image at a certain time point), while time series data and images of other samples in the same batch are considered negative sample pairs. A bidirectional contrastive loss is used to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs.

[0160]

[0161] in, : No. The image-temporal bidirectional contrast loss for each sample is used to maximize the similarity between image features and corresponding temporal features, and minimize the similarity with other temporal features. : No. The time-to-image bidirectional contrast loss for each sample, its effect on... Complementarity ensures bidirectional alignment of dual-modal features; The cosine similarity function is used to calculate the similarity between two low-dimensional feature vectors, and its value ranges from 1 to 2. A larger value indicates that the features are more similar; : No. The low-dimensional image feature vector of each sample after projection has 256 dimensions, capturing the structural information of the line graph. : No. The low-dimensional time-series feature vector of each sample after projection has 256 dimensions, capturing the numerical variation pattern of the monitoring parameters. : Global temperature parameter, default value is 0.07. Training batch size, default value is 16. : No. The low-dimensional temporal feature vector of each sample, with a dimension of 256. : No. The low-dimensional image feature vectors of each sample are 256-dimensional. The naive contrast loss is calculated by... and On average, Loss calculation from image modality to temporal modality Loss is calculated from temporal modality to image modality to ensure bidirectional alignment of bimodal features:

[0162]

[0163] in, Naive cross-modal contrastive loss; The two-way contrast loss of all samples in the batch is averaged, and 2B is the normalization coefficient (considering the two-way loss).

[0164] (d2) Geodesic mixed loss

[0165] Considering that bimodal features may be distributed in different subspaces, direct comparison may lead to bias. Therefore, a geodesic mixing strategy is designed to generate cross-modal hybrid representations to fill the subspace gaps. First, temporal features are calculated. Image features Geodesic angle on a unit hypersphere Then, based on the Beta distribution, the mixing coefficients are sampled. Generate hybrid features along the geodesic direction This ensures that the hybrid features contain both numerical and structural information.

[0166]

[0167] in, Cross-modal hybrid feature vector, 256 dimensions, along temporal features Image features The geodesic direction is generated, and it also includes monitoring parameter numerical information and polyline plot structure information; : Mixture coefficient, from the Beta distribution obtained from sampling, The default value is 0.1, which controls the mixing ratio of temporal features and image features; Low-dimensional image feature vectors, 256 dimensions; Low-dimensional temporal feature vectors, with 256 dimensions; Temporal characteristics Image features Geodesic angles on a unit hypersphere, Temporal characteristics In image features Unit vectors in orthogonal directions ensure that mixed features are generated along geodesic directions.

[0168] The mixed features are used as negative samples and compared with the original positive sample pairs to further expand the feature distance of the out-of-class samples:

[0169]

[0170] in, : No. Image-mixed contrast loss for each sample : No. The time-mixed contrast loss for each sample, and its effect on... Complementary features enhance the alignment between temporal features and corresponding image features; : No. Cross-modal hybrid feature vectors of samples.

[0171] Geodesic mixed loss The calculation formula is:

[0172]

[0173] (d3) Total contrast loss

[0174] Through weighting coefficients By balancing naive contrastive loss and hybrid contrastive loss, we can ensure the basic alignment of bimodal features while enhancing the discriminative power of the feature space through hybrid features.

[0175]

[0176] in, Sequence-image total contrast loss, balancing naive contrast loss and mixed contrast loss. Loss balance coefficient.

[0177] In some embodiments, the prototype contrastive loss described above is combined with the sequence-image contrastive loss to jointly optimize the parameters of the TS encoder and the image encoder, enabling the model to learn features that combine robustness, discriminativeness, and cross-modal consistency. The total model loss is [not specified]. The formula is:

[0178]

[0179] The MsDL-AimTS fusion model is trained based on the total loss function of the aforementioned model. After training, the time-series data to be detected and the transformed image are input, and the model extracts dual-modal fusion features. When the similarity between the features and the features of normal samples is lower than a preset threshold, and this low similarity state persists for more than 3 monitoring periods, it is determined as an early abnormal signal, triggering an early warning prompt, thus achieving accurate identification of the initial stage of conductive deposition.

[0180] S140. Calculate the resistance change rate, temperature influence factor and monitoring duration based on the preprocessed multidimensional time series dataset, and calculate the conductivity deposition coefficient based on the resistance change rate, temperature influence factor and monitoring duration.

[0181] In some embodiments, the resistance change rate is calculated based on the current resistance value and the initial resistance value in the multidimensional time-series dataset;

[0182] Based on the actual temperature values ​​and preset standard temperature values ​​in the multidimensional time series dataset, the temperature influence factor is calculated; based on the timestamp information of the multidimensional time series dataset, the cumulative monitoring time from the start of monitoring to the current moment is calculated; the resistance change rate, temperature influence factor and monitoring time are multiplied by their corresponding weighting coefficients and then weighted and summed to obtain the conductivity deposition coefficient.

[0183] Specifically, conductivity deposition coefficient The calculation formula is:

[0184]

[0185] in, Weighting coefficient for the rate of change of resistance. , Weighting coefficients of temperature influence factors. , Weighting coefficient for monitoring duration And satisfy ; The change in resistance is the difference between the current monitored resistance and the initial resistance. For real-time monitoring of gas phase or liquid phase circuits; Initial resistance: The resistance value of the gas or liquid phase circuit at the start of monitoring (0 hours) (unit: ); : Rate of change of resistance; Actual temperature; Standard temperature; Temperature influence factor Monitoring duration is the cumulative time from the start of monitoring to the current moment.

[0186] S150. Based on the conductivity deposition coefficient, the resistance change trend prediction results, and the early anomaly identification results, the deposition risk level is output and the qualification is determined according to the preset fusion judgment rules.

[0187] In some embodiments, based on a comprehensive evaluation of the conductivity deposition coefficient, the predicted resistance change trend, and the early anomaly identification results, the deposition risk is divided into multiple levels and corresponding qualification judgment results are output.

[0188] Based on the above resistance trend prediction results, early anomaly identification results, and CDF values, the corresponding phase (gas phase / liquid phase) is deemed unqualified if any of the following conditions are met:

[0189] (a) A preset threshold (the judgment threshold is determined based on historical sedimentary sample data, and its value is generally between 2.0 and 3.0).

[0190] (b) Resistance trend forecast results show that resistance changes in the next 72 hours will lead to the risk of copper circuit continuity.

[0191] (c) Early anomaly identification results detected early anomalies and the degree of anomaly continued to worsen within three consecutive monitoring cycles (corresponding acquisition intervals).

[0192] Based on the above CDF values, prediction results, and anomaly conditions, the deposition risk level is output as follows: Low: no anomaly and CDF < 2; Medium: Or a single abnormality; High: meets the non-compliance criteria or ).

[0193] According to embodiments of the present invention, the following beneficial effects are achieved:

[0194] (1) This invention proposes to construct the MsDL-AimTS fusion model, which combines the advantages of multi-scale time-series dynamic learning and cross-modal comparative learning, to achieve accurate prediction of resistance change trends and efficient identification of early weak anomalies, while improving monitoring efficiency and result reliability.

[0195] (2) This invention proposes a multi-scale time-series dynamic learning (MsDL) method, which comprehensively captures the dynamic characteristics of time-series data at different time scales through multiple reservoir computing units with different recursive connection steps, providing more comprehensive feature support for sedimentation trend prediction.

[0196] (3) The present invention proposes an iterative interval narrowing search (IINS) algorithm to optimize the spectral radius of each reservoir. With the combined loss function as the optimization objective, the model parameters are adaptively adjusted to improve the model's adaptability to different monitoring scenarios.

[0197] (4) This invention proposes an enhanced sequence-image contrast learning (AimTS) framework, which enhances the ability to identify early abnormal signals such as small resistance fluctuations by fusing and contrasting time-series and image dual-modal features, and solves the problem of insufficient sensitivity of single-modal monitoring.

[0198] (5) The prototype contrast loss of this invention effectively avoids semantic distortion caused by data augmentation, improves the robustness of the model to monitoring data disturbances, and ensures the stability and discriminability of feature representation.

[0199] (6) The present invention designs a sequence-image contrast loss, including naive contrast loss and geodesic hybrid loss, to achieve deep alignment and fusion of dual-modal features, fully explore the complementary value of numerical information and structural information, further improve the accuracy of early anomaly identification, and ultimately provide more efficient and reliable technical support for the performance evaluation of electric vehicle lubricants and the prevention of transmission system failures.

[0200] Example 2: Figure 5 This is a schematic diagram of a conductive deposition monitoring system based on gas-liquid phase separation provided in Embodiment 2 of the present invention, as shown below. Figure 5 As shown, the system includes:

[0201] The data acquisition and preprocessing module 510 is used to acquire multidimensional time-series datasets based on a pre-built multi-module collaborative monitoring system, and to perform data preprocessing on the multidimensional datasets.

[0202] MsDL-AimTS fusion model construction module 520 is used to construct the MsDL-AimTS fusion model, wherein the MsDL-AimTS fusion model includes a multi-scale temporal dynamic learning framework and an enhanced sequence-image contrast learning framework.

[0203] The resistance change trend prediction module 530 is used to input the preprocessed multi-dimensional time series dataset into the MsDL-AimTS fusion model, predict the resistance change trend according to the multi-scale time series dynamic learning framework, and identify early weak anomalies according to the enhanced sequence-image contrast learning framework.

[0204] The conductive deposition coefficient calculation module 540 is used to calculate the resistance change rate, temperature influence factor and monitoring duration based on the preprocessed multi-dimensional time series dataset, and to calculate the conductive deposition coefficient based on the resistance change rate, temperature influence factor and monitoring duration.

[0205] The deposition risk level determination module 550 is used to output the deposition risk level and make a qualification determination based on the conductivity deposition coefficient, the resistance change trend prediction result and the early anomaly identification result, according to the preset fusion determination rules.

[0206] The conductive deposition monitoring system based on gas-liquid phase separation provided in the embodiments of the present invention can execute the conductive deposition monitoring method based on gas-liquid phase separation provided in any of the embodiments of the present invention. It has the corresponding functions and beneficial effects of executing the conductive deposition monitoring method based on gas-liquid phase separation. For detailed process, please refer to the relevant operations of the conductive deposition monitoring method based on gas-liquid phase separation in the foregoing embodiments.

[0207] Example 3: Figure 6 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0208] like Figure 6 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0209] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0210] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes the dynamic prediction method for disaster-prone processes based on a multiphysics coupling model described above.

[0211] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0212] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for monitoring conductive deposition based on gas-liquid phase separation, characterized in that, include: A multi-dimensional time-series dataset is collected based on a pre-built multi-module collaborative monitoring system, and the multi-dimensional time-series dataset is preprocessed. Construct a MsDL-AimTS fusion model, wherein the MsDL-AimTS fusion model includes a multi-scale temporal dynamic learning framework and an enhanced sequence-image contrast learning framework; The preprocessed multidimensional time series dataset is input into the MsDL-AimTS fusion model, and resistance change trend is predicted according to the multi-scale time series dynamic learning framework, and early weak anomaly identification is performed according to the enhanced sequence-image contrast learning framework. The resistance change rate, temperature influence factor and monitoring duration are calculated based on the preprocessed multidimensional time series dataset, and the conductivity deposition coefficient is calculated based on the resistance change rate, temperature influence factor and monitoring duration. Based on the predicted results of the conductivity deposition coefficient, resistance change trend, and early anomaly identification, the deposition risk level is output and the qualification is determined according to the preset fusion judgment rules.

2. The method according to claim 1, characterized in that, The data preprocessing of the multidimensional time-series dataset includes: Different interpolation strategies are used to fill in missing values ​​based on the degree of continuity of missing data in the multidimensional time series dataset; An anomaly detection method based on statistical distribution identifies and removes outliers from the multidimensional time-series dataset. Time alignment is performed on data from different monitoring sources in the multidimensional time-series dataset based on timestamps; The time-series data, after missing value imputation, outlier removal, and data alignment, are standardized to unify the numerical scale of each monitoring parameter.

3. The method according to claim 2, characterized in that, Based on the aforementioned multi-scale temporal dynamic learning framework, resistance change trend prediction is performed, including: Configure multiple reservoir computing units with different time scales; The reservoir computing unit performs dynamic representation learning on the preprocessed time series data to obtain an initial multi-scale dynamic representation. An iterative interval narrowing search algorithm is adopted, with the combined loss function as the optimization objective, to adaptively adjust the radius of the connection weight spectrum of each reservoir calculation unit in order to obtain the optimized model parameters. Based on the optimized model parameters, dynamic representation learning is performed again to obtain the final multi-scale dynamic representation. The final multi-scale dynamic representation is fused into a unified high-dimensional feature vector and input into a machine learning model to predict future resistance change trends and deposition formation risks.

4. The method according to claim 3, characterized in that, Early weak anomaly detection based on an enhanced sequence-image contrast learning framework, including: The preprocessed multidimensional time-series dataset is input into the time-series encoder and the image encoder respectively to extract time-series numerical features and image structural features. Cross-modal alignment and fusion of the temporal numerical features and image structural features are performed through a contrastive learning mechanism to obtain cross-modal fused features; Anomaly detection is performed based on the cross-modal fusion features. When the detection result meets the preset conditions, it is determined to be an early anomaly and an early warning is triggered.

5. The method according to claim 4, characterized in that, The step of performing cross-modal alignment and fusion of the temporal numerical features and image structural features through a contrastive learning mechanism to obtain cross-modal fused features includes: Data augmentation is performed on data samples in the multidimensional time-series dataset to generate an augmented view, and prototype contrastive loss is calculated based on the augmented view; The time-series numerical features and image structural features are subjected to bidirectional comparative learning to calculate the sequence-image contrast loss. The prototype contrast loss and the calculated sequence-image contrast loss are weighted and fused together, and the temporal encoder and image encoder are jointly optimized through backpropagation. Based on the jointly optimized temporal encoder and image encoder, temporal numerical features and image structural features are extracted again, respectively. The cross-modal fusion feature is obtained by fusing the re-extracted temporal numerical features with the image structural features.

6. The method according to claim 5, characterized in that, The calculation of prototype contrast loss based on the enhanced view includes: The enhanced view is input into a temporal encoder to extract enhanced features, and sample prototype representations are generated by mean aggregation. Calculate intra-prototype loss and inter-prototype loss based on the sample prototype representation; The intra-prototype loss and inter-prototype loss are weighted and fused to obtain the prototype contrast loss.

7. The method according to claim 5, characterized in that, The temporal numerical features and image structural features are subjected to bidirectional comparative learning, and the sequence-image contrast loss is calculated, including: The time-series numerical features and image structural features are used as positive sample pairs to construct a cross-modal contrastive learning task; Based on the cross-modal contrastive learning task, calculate the bidirectional naive contrastive loss and geodesic mixing loss respectively; The naive contrast loss and the geodesic mixing loss are weighted and fused to obtain the sequence-image contrast loss.

8. The method according to claim 7, characterized in that, The process involves calculating the resistance change rate, temperature influence factor, and monitoring duration based on the preprocessed multidimensional time-series dataset, and then calculating the conductivity deposition coefficient based on these parameters. Based on the current resistance value and the initial resistance value in the multidimensional time series dataset, the resistance change rate is calculated; Based on the actual temperature values ​​and preset standard temperature values ​​in the multidimensional time-series dataset, the temperature influence factor is calculated. Based on the timestamp information of the multidimensional time series dataset, the cumulative monitoring duration from the start of monitoring to the current moment is calculated; The conductivity deposition coefficient is obtained by multiplying the resistance change rate, temperature influence factor and monitoring duration by their respective weighting coefficients and then summing them by weight.

9. The method according to claim 4, characterized in that, Based on the predicted results of the conductivity deposition coefficient, resistance change trend, and early anomaly identification, and according to the preset fusion judgment rules, the deposition risk level is output and a qualification judgment is performed, including: Based on a comprehensive evaluation of the conductivity deposition coefficient, the predicted resistance change trend, and the early anomaly identification results, the deposition risk is divided into multiple levels and corresponding qualification judgment results are output.

10. A conductive deposition monitoring system based on gas-liquid phase separation, characterized in that, include: The data acquisition and preprocessing module is used to acquire multidimensional time-series datasets based on a pre-built multi-module collaborative monitoring system, and to perform data preprocessing on the multidimensional time-series datasets. The MsDL-AimTS fusion model construction module is used to construct the MsDL-AimTS fusion model, wherein the MsDL-AimTS fusion model includes a multi-scale temporal dynamic learning framework and an enhanced sequence-image contrast learning framework. The resistance change trend prediction module is used to input the preprocessed multi-dimensional time series dataset into the MsDL-AimTS fusion model, predict the resistance change trend according to the multi-scale time series dynamic learning framework, and identify early weak anomalies according to the enhanced sequence-image contrast learning framework. The conductive deposition coefficient calculation module is used to calculate the resistance change rate, temperature influence factor and monitoring duration based on the preprocessed multi-dimensional time series dataset, and to calculate the conductive deposition coefficient based on the resistance change rate, temperature influence factor and monitoring duration. The deposition risk level determination module is used to output the deposition risk level and determine its compliance based on the conductivity deposition coefficient, the resistance change trend prediction result, and the early anomaly identification result, according to the preset fusion determination rules.

Images