Equipment state prediction method based on weak supervised learning

Abstract

The invention discloses an equipment state prediction method based on weak supervised learning. According to the method, through innovative means such as a hierarchical labeling strategy, an active learning module and an anomaly detection and automatic labeling module, the requirement for manual data labeling is greatly reduced. Meanwhile, various semi-supervised learning models are fused, data distribution is comprehensively learned, a multi-modal fusion attention mechanism is innovatively introduced, heterogeneous data are comprehensively utilized, and the prediction precision is improved. According to the method, a data enhancement and expansion technology is further designed, the generalization ability of the model is enhanced, online incremental learning is supported, and continuous optimization of the model is achieved. According to the method, the future state of the equipment can be efficiently and accurately predicted under the condition that the annotation data is poor, the prediction cost is reduced, and the method has important theoretical value and wide application prospects.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning and pattern recognition, and in particular to a device state prediction method based on weakly supervised learning. Background Art

[0002] Equipment status prediction is one of the key links to ensure the normal and efficient operation of the production system. It can timely and accurately predict possible failures or abnormal conditions of the equipment, providing an important basis for formulating maintenance strategies and taking preventive measures. Traditional equipment status prediction methods mainly rely on manual experience and calculations based on physical models. With the continuous improvement of the level of modern industrial automation, sensor networks and information systems are widely used, accumulating a large amount of equipment operation data, which provides the possibility for data-driven intelligent prediction methods.

[0003] In recent years, machine learning and deep learning technologies have been widely used and studied in the field of equipment status prediction. However, most prediction models based on supervised learning require a large amount of labeled training data. Acquiring large-scale labeled data usually requires a lot of manpower and cost, which has become the main bottleneck restricting the application of such methods. On the other hand, although unsupervised or semi-supervised learning methods do not require labeled data, they often fail to accurately capture the complex patterns of equipment status because they do not use any prior knowledge.

[0004] Therefore, there is an urgent need for a new device status prediction method that can use limited labeled data and a large amount of unlabeled data to efficiently learn the inherent laws of the device's operating status through innovative technical means in the case of scarce labeled data, thereby achieving accurate prediction of the device's future status. This will greatly reduce the annotation cost of the prediction model and expand its scope of application in practical application scenarios. Summary of the invention

[0005] In order to solve the problems of the prior art, an embodiment of the present invention provides a device state prediction method based on weakly supervised learning. The technical solution is as follows:

[0006] On the one hand, a device status prediction method based on weakly supervised learning is provided. This method combines a variety of innovative technologies, including hierarchical labeling strategy, active learning module, anomaly detection automatic labeling module, data enhancement and expansion, integrated semi-supervised learning model and multimodal fusion attention mechanism, so as to efficiently and accurately predict the device status in scenarios with scarce labeled data. The specific steps are as follows:

[0007] a) Collect multimodal equipment operation data including time series data, image data and text data, and pre-process the raw data;

[0008] b) Based on domain knowledge and expert experience, a hierarchical annotation strategy is implemented for equipment status, with fine annotation of the most important abnormal status and coarse-grained annotation of other status, so as to effectively allocate manual annotation resources;

[0009] c) Set up an active learning module to select unlabeled samples with the highest uncertainty based on the current model state, submit them to the manual labeling queue, and continuously accumulate high-quality labeled samples;

[0010] d) Set up an anomaly detection automatic labeling module to automatically label the normal / abnormal status of a large amount of unlabeled data using algorithms such as isolation forest and one-class SVM;

[0011] e) Enhance and expand limited labeled data through transformation and other means, and use the temporal continuity of time series data to generate a large number of supervised sequence samples;

[0012] f) Construct a semi-supervised learning integrated framework including models such as generative adversarial networks, variational autoencoders, and smoothing kernels, use labeled data as a supervised learning source, and perform semi-supervised learning on unlabeled data to comprehensively learn the overall distribution of data;

[0013] g) Design corresponding encoders to extract features for different modal data, use cross-modal attention mechanism to fuse multi-modal features, and obtain a comprehensive representation of device status;

[0014] h) Input the fused multimodal features into the decoder network to predict the probability distribution of the device state in the future time step;

[0015] i) New data is collected regularly, and incorporated into the training set after review and automatic annotation, and the model is incrementally learned online to continuously improve performance.

[0016] Furthermore, in step a), the raw data is cleaned, standardized and synchronized to ensure the quality and consistency of the multimodal data.

[0017] Furthermore, in step b), the abnormal state is divided into multiple levels according to factors such as the severity and frequency of occurrence of the abnormal state.

[0018] Furthermore, for the highest-level critical abnormal status, the annotation details include rich information such as the cause of occurrence, scope of impact, duration, and expected losses.

[0019] Furthermore, in step c), the active learning module calculates the uncertainty score of each unlabeled sample based on the prediction result and confidence of the current model, and selects some samples with the highest uncertainty for labeling.

[0020] Further, in step d), the anomaly detection automatic labeling module is an isolation forest model or a one-class support vector machine model trained on a large amount of unlabeled data.

[0021] Furthermore, in step e), the data enhancement method includes injecting Gaussian noise, interpolation and other transformation operations into the original data.

[0022] Furthermore, in step f), the integrated semi-supervised learning model includes a generative adversarial network, a variational autoencoder, a smoothing kernel model, a semi-supervised random forest, and the like.

[0023] Furthermore, in step g), the cross-modal attention mechanism performs weighted fusion of different modal features by calculating the correlation weights between multimodal features.

[0024] Furthermore, in step h), the decoder network adopts a sequence-to-sequence structure to output a time series prediction of the device state.

[0025] Furthermore, the method is applied to the fields of industrial intelligent manufacturing, predictive maintenance, etc.

[0026] The technical solution provided by the embodiment of the present invention has the following beneficial effects:

[0027] 1. Reduced annotation costs

[0028] The present invention significantly reduces the demand for a large amount of manually labeled data and saves a lot of manpower and financial costs through innovative means such as hierarchical labeling strategies, active learning modules, and automatic labeling for anomaly detection.

[0029] 2. Improved prediction accuracy

[0030] The present invention integrates a variety of advanced semi-supervised learning models to comprehensively learn the overall distribution of data and capture the complex patterns of device status. At the same time, it innovatively introduces a multimodal data fusion mechanism, comprehensively utilizes the complementary information of heterogeneous data, and further improves the accuracy of status prediction.

[0031] 3. Enhanced generalization ability of prediction

[0032] Through data enhancement and expansion techniques, the present invention greatly expands the scale and diversity of training data, improves the generalization ability of the model to unseen data, and ensures the effectiveness of prediction performance in practical application scenarios.

[0033] 4. Support online model optimization

[0034] The present invention designs a mechanism to support online incremental learning, which can continuously utilize newly collected operating data, continuously optimize and improve model performance, and realize sustainable evolution of the model.

[0035] 5. Has broad application prospects

[0036] The present invention can be widely used in industrial manufacturing, predictive maintenance, remote health monitoring and other fields, providing important technical support for the intelligent operation and maintenance management of equipment, improving production efficiency, and reducing operation and maintenance costs, which has important practical significance.

[0037] In general, the present invention innovatively integrates multiple advanced technologies to form a complete weakly supervised learning framework, which can efficiently and accurately predict the status of equipment under the condition of scarce labeled data, and has significant theoretical value and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0039] Figure 1 It is a flow chart of a device state prediction method based on weakly supervised learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0041] The present invention aims to provide a device state prediction method based on weakly supervised learning, which can effectively capture the patterns and laws of device operation states by using limited labeled data and a large amount of unlabeled data in the case of scarce labeled data, thereby achieving efficient and accurate prediction of the future state of the device.

[0042] This method comprehensively applies a variety of innovative technologies, including hierarchical labeling strategies, active learning modules, anomaly detection automatic labeling modules, data enhancement and expansion, integrated semi-supervised learning models, and multimodal fusion attention mechanisms, to build a complete prediction framework.

[0043] See also Figure 1 The specific implementation steps of this method are as follows:

[0044] a) Data collection and preprocessing

[0045] Collect multimodal equipment operation data including time series data (such as equipment sensor readings, control instructions, etc.), image data (equipment monitoring images), and text data (operation and maintenance logs, fault descriptions, etc.). Perform pre-processing such as cleaning, standardization, and synchronization on the raw data to ensure data quality.

[0046] Specifically, time series data may include sensor readings such as temperature, pressure, vibration, and control parameters such as equipment speed and power. These data are usually collected at fixed time intervals (such as every second or every minute). Image data may come from surveillance cameras on the surface or inside of the equipment, capturing changes in the appearance of the equipment or internal motion. Text data may include unstructured information such as operators' daily records and maintenance reports.

[0047] In the preprocessing stage, data cleaning is first performed to remove outliers and missing values. For example, for time series data, moving average or interpolation methods can be used to fill missing values; for outliers, they can be identified and processed by setting thresholds or using statistical methods (such as the 3σ principle). Secondly, data standardization is performed to convert data of different dimensions to the same scale. Commonly used methods include Min-Max normalization and Z-score normalization. Finally, multimodal data is synchronized in time to ensure that data from different sources are aligned in the time dimension, which is crucial for subsequent multimodal fusion.

[0048] b) Hierarchical annotation strategy

[0049] Based on domain knowledge and expert experience, the equipment status is divided into multiple levels with different severity. The highest level of critical abnormal status (such as failures that cause downtime) is finely labeled, including rich information such as the cause, scope of impact, duration, and expected loss. For other levels of status, only coarse-grained normal / abnormal binary labels are performed. This hierarchical strategy can effectively allocate limited manual labeling resources and focus on key anomalies.

[0050] For example, the device status can be divided into four levels:

[0051] 1. Normal operation status

[0052] 2. Slight abnormality (does not affect production)

[0053] 3. Moderate abnormal state (affecting production efficiency)

[0054] 4. Serious abnormal state (causing downtime)

[0055] For level 4 severe abnormal conditions, the annotation information may include:

[0056] - Cause: such as "spindle bearing overheating"

[0057] - Scope of impact: such as "the entire production line stops running"

[0058] -Duration: such as "4 hours"

[0059] -Expected loss: such as "production capacity loss of approximately RMB 2 million"

[0060] For level 2 and 3 abnormal states, they are simply marked as "abnormal" without further subdivision. This strategy ensures a deep understanding of key abnormalities while reducing the overall burden of labeling work.

[0061] c) Active Learning Module

[0062] Set up an active learning module to calculate the uncertainty score of each unlabeled sample based on the prediction results and confidence of the current model. Select some samples with the highest uncertainty and submit them to the manual labeling queue. After manual labeling, they will be incorporated into the training set to continuously accumulate high-quality labeled data and actively optimize model performance.

[0063] When implementing it, you can take the following steps:

[0064] 1. Use the current model to predict unlabeled samples

[0065] 2. Calculate the prediction uncertainty of each sample, you can use indicators such as information entropy or prediction variance

[0066] 3. Select the top-k samples with the highest uncertainty (such as top-100)

[0067] 4. Submit these samples to human experts for annotation

[0068] 5. Add the newly annotated samples to the training set and update the model

[0069] This process can be done periodically, such as once a week or every time a certain number of new samples are collected. In this way, the model can continuously learn the most valuable and challenging samples, thereby quickly improving performance.

[0070] d) Automatic labeling of anomaly detection

[0071] Set up an anomaly detection automatic labeling module. Use algorithms such as isolation forest and one-class support vector machine to train unsupervised anomaly detection models on a large amount of unlabeled data, and give preliminary normal / abnormal labels to the unlabeled data. After manual review, these automatically labeled data are incorporated into the training set to further expand the amount of labeled data.

[0072] Taking Isolation Forest as an example, its working principle is:

[0073] 1. Randomly select a feature

[0074] 2. Randomly select a split point

[0075] 3. Recursively split the data until all points are isolated

[0076] 4. Outliers are usually easier to isolate, so fewer segmentations are required

[0077] By setting a suitable threshold, the data can be divided into two categories: normal and abnormal. This method is particularly suitable for processing high-dimensional data and has high computational efficiency.

[0078] In practical applications, the parameters of the anomaly detection model can be adjusted and verified using a small amount of labeled data before being applied to a large amount of unlabeled data. The results of automatic labeling should be sampled and reviewed by human experts to ensure accuracy.

[0079] e) Data enhancement and expansion

[0080] For limited manual and automatically annotated data, data enhancement is performed by injecting Gaussian noise, interpolation and other transformations to generate more annotated samples. At the same time, the temporal continuity of time series data is utilized to generate a large number of supervised sequence samples through sliding windows to further expand the training data.

[0081] Specific data enhancement techniques can include:

[0082] 1. Time Series Data:

[0083] -Add Gaussian noise: Y′=Y+N(0,σ 2 ), where Y is the original data, Y′ is the enhanced data, N(0,σ 2 ) means the mean is 0 and the variance is σ 2 Gaussian noise.

[0084] -Time Warp: Randomly stretch the timeline

[0085] - Amplitude scaling: Y′=aY, where a is a scaling factor, usually randomly selected within a certain range (eg, between 0.8 and 1.2).

[0086] 2. Image data:

[0087] - Random rotation, translation, scaling

[0088] -Change brightness, contrast

[0089] - Added Gaussian blur

[0090] 3. Text Data:

[0091] - Synonym replacement

[0092] - Randomly insert, delete or swap word order

[0093] For time series data, a sliding window method can be used to generate a large number of supervised sequence samples. For example, for a time series of length T, the window size can be set to w and the step size to s, then (Tw) / s+1 samples can be generated. Each sample contains w time steps of input data and the target output of the subsequent h time steps.

[0094] f) Integrated semi-supervised learning model

[0095] Build an integrated framework that includes multiple semi-supervised learning models such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Smoothing Kernel Models, and Semi-supervised Random Forests. Use manually / automatically labeled samples as the data source for supervised learning, and use a large amount of unlabeled data for semi-supervised learning to comprehensively learn the potential distribution of the entire data. Integrate the prediction results of a single model to give full play to the different advantages of each model.

[0096] The characteristics and functions of each model are as follows:

[0097] 1. GAN: Captures the true distribution of data through adversarial learning between the generator and the discriminator. It can be used to generate virtual samples that conform to the actual distribution and further expand the training set.

[0098] 2.VAE: Learns low-dimensional latent representations of data, which helps discover the intrinsic structure and patterns of the data.

[0099] 3. Smooth kernel model: Exploit the local similarity of data to propagate label information between labeled and unlabeled samples.

[0100] 4. Semi-supervised random forest: It combines the powerful predictive ability of random forest and the advantages of semi-supervised learning, and can effectively utilize unlabeled data.

[0101] These models can be trained in parallel, and then their predictions can be integrated by voting, weighted averaging, or stacking. The ensemble method can dynamically adjust the weights of each model based on its performance on the validation set to achieve the best overall performance.

[0102] g) Multimodal fusion attention mechanism

[0103] For different modal data such as time series, images and text, the corresponding encoder network is designed for feature extraction. Using the cross-modal attention mechanism, the modal features are weightedly fused by calculating the correlation weights between different modal features to obtain a comprehensive representation of the device status.

[0104] The specific implementation steps are as follows:

[0105] 1. Feature extraction:

[0106] -Time series data: Use LSTM or TCN (Temporal Convolutional Network) to extract time series features

[0107] - Image data: Use CNN (such as ResNet) to extract spatial features

[0108] -Text data: Use BERT or other pre-trained language models to extract semantic features

[0109] 2. Attention calculation:

[0110] -For each pair of modalities (i, j), calculate the attention weight: a ij =softmax(f(h i ,h j )), where h i and h j are the feature representations of the i-th and j-th modalities respectively, and f is the similarity function (such as dot product).

[0111] -where f can be a simple dot product or a more complex neural network.

[0112] 3. Feature Fusion:

[0113] -Weighted sum of features for each modality: Among them, h i ′ is the fused feature representation.

[0114] 4. Final characterization:

[0115] -Concatenate or sum all fused features to get the final multimodal representation

[0116] This mechanism allows the model to dynamically focus on the most relevant information in different modalities, improving the efficiency of feature utilization and the accuracy of prediction.

[0117] h) State prediction and online optimization

[0118] The fused multimodal features are input into the sequence-to-sequence decoder network to output the time series probability prediction of the device state.

[0119] The decoder network can use LSTM or Transformer structure to capture long-term dependencies. The prediction output can be the state probability distribution of multiple time steps in the future, for example:

[0120] P(state_t+1|feature_t),P(state_t+2|feature_t),...,P(state_t+n|feature_t)

[0121] Each state can be one of the multiple levels defined previously.

[0122] New operating data is collected regularly, and after manual review and automatic annotation, it is incorporated into the training set. The new data is used to perform incremental online learning on the model to continuously improve performance.

[0123] Incremental learning can be achieved by:

[0124] 1. For neural network models, you can use mini-batch gradient descent and only fine-tune using new data

[0125] 2. For tree-based models (such as random forests), you can dynamically add new decision trees or update the structure of existing trees

[0126] 3. Evaluate model performance regularly; if performance degrades, more comprehensive retraining may be required

[0127] i) Application areas

[0128] This invention can be widely used in industrial intelligent manufacturing, predictive maintenance and other fields, providing strong support for equipment health status monitoring and fault prevention.

[0129] Specific application scenarios include but are not limited to:

[0130] 1. Equipment status monitoring and predictive maintenance of industrial production lines

[0131] 2. Health management of turbine units in power plants

[0132] 3. Life prediction of aircraft components in the aerospace field

[0133] 4. Optimal control of HVAC systems in smart buildings

[0134] 5. Pipeline leak warning in the petrochemical industry

[0135] 6. Wind turbine performance prediction and maintenance scheduling in wind farms

[0136] By deploying this system, enterprises can significantly reduce unplanned downtime, optimize maintenance plans, and improve equipment utilization, thereby significantly reducing operating costs and improving production efficiency.

[0137] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A device state prediction method based on weakly supervised learning, characterized in that: This method combines a variety of innovative technologies, including hierarchical labeling strategy, active learning module, anomaly detection automatic labeling module, data enhancement and expansion, integrated semi-supervised learning model and multimodal fusion attention mechanism, so as to efficiently and accurately predict the device status in scenarios with scarce labeled data. The specific steps are as follows: a) Collect multimodal equipment operation data including time series data, image data and text data, and pre-process the raw data; b) Based on domain knowledge and expert experience, a hierarchical annotation strategy is implemented for equipment status, with fine annotation of the most important abnormal status and coarse-grained annotation of other status, so as to effectively allocate manual annotation resources; c) Set up an active learning module to select unlabeled samples with the highest uncertainty based on the current model state, submit them to the manual labeling queue, and continuously accumulate high-quality labeled samples; d) Setting up an anomaly detection automatic labeling module, using the isolation forest and one-class SVM algorithms to automatically label the normal / abnormal status of a large amount of unlabeled data; e) Enhance and expand the limited labeled data through transformation, and use the temporal continuity of time series data to generate a large number of supervised sequence samples; f) Construct a semi-supervised learning integrated framework including generative adversarial networks, variational autoencoders and smooth kernel models, use labeled data as a supervised learning source, and perform semi-supervised learning on unlabeled data to comprehensively learn the overall distribution of data; g) Design corresponding encoders to extract features for different modal data, use cross-modal attention mechanism to fuse multi-modal features, and obtain a comprehensive representation of device status; h) Input the fused multimodal features into the decoder network to predict the probability distribution of the device state in the future time step; i) New data is collected regularly, and incorporated into the training set after review and automatic annotation, and the model is incrementally learned online to continuously improve performance.

2. The device status prediction method according to claim 1, characterized in that: In step a), the raw data are cleaned, standardized, and synchronized to ensure the quality and consistency of the multimodal data.

3. The device status prediction method according to claim 1, characterized in that: In step b), the abnormal state is divided into multiple levels according to the severity and frequency of occurrence of the abnormal state.

4. The device state prediction method according to claim 3, characterized in that: For the highest-level critical abnormal conditions, the annotation details include rich information such as the cause of occurrence, scope of impact, duration, and expected loss.

5. The device status prediction method according to claim 1, characterized in that: In step c), the active learning module calculates the uncertainty score of each unlabeled sample based on the prediction results and confidence of the current model, and selects some samples with the highest uncertainty for labeling.

6. The device state prediction method according to claim 1, characterized in that: In step d), the anomaly detection automatic labeling module is an isolation forest model or a one-class support vector machine model trained on a large amount of unlabeled data.

7. The device status prediction method according to claim 1, characterized in that: In step e), the data enhancement method includes injecting Gaussian noise and interpolation transformation operations into the original data.

8. The device status prediction method according to claim 1, characterized in that: In step f), the integrated semi-supervised learning model includes a generative adversarial network, a variational autoencoder, a smoothing kernel model and a semi-supervised random forest.

9. The device state prediction method according to claim 1, characterized in that: In step g), the cross-modal attention mechanism performs weighted fusion of different modal features by calculating the correlation weights between multimodal features.

10. The device status prediction method according to claim 1, characterized in that: In step h), the decoder network adopts a sequence-to-sequence structure and outputs a time series prediction of the device state.

Images