Equipment health state intelligent identification-oriented low-quality data robust learning method

Abstract

The invention discloses an equipment health state intelligent identification-oriented low-quality data robust learning method. The method comprises the following steps of: obtaining multi-source sensor time sequence data of an electromechanical composite transmission system and preprocessing the multi-source sensor time sequence data; constructing a multi-sensor fusion identification network composed of a feature encoder, a projection head and a classification head; generating neighbor consensus tags and consistency measurement for samples according to the neighbor consistency of the embedded space, dynamically screening high-confidence samples and similar sample pairs, performing characterization pre-training by using a joint target of supervised contrast learning, classified cross entropy and sample pair similarity, and introducing a hybrid enhanced smooth discrimination boundary to obtain a robust encoder; freezing a robust encoder, training a new classification head only by using a high-confidence set, and obtaining a state recognition model; a state recognition model is used for directly reasoning a complete channel sample, a feature channel adaptability calibration module is introduced into a sensor channel data missing scene to dynamically compensate feature distribution offset, and probability distribution of each state category is output.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for mechanical equipment, and more specifically to a robust learning method for low-quality data for intelligent identification of equipment health status. Background Technology

[0002] With the rapid development of modern industrial technology, complex mechanical equipment, represented by high-speed trains, wind turbines, and large machine tools, plays an increasingly important role in the national economy. Among them, the electromechanical composite transmission system, as the core power unit, undertakes the key task of efficiently transmitting power. Maintenance personnel hope to accurately identify the type and cause of failures in the electromechanical composite transmission system in the early stages, take timely maintenance measures, reduce risks at the source, and improve equipment availability and integrity.

[0003] With the advancement of artificial intelligence and Internet of Things technologies, intelligent state recognition methods based on deep learning have gradually overcome the reliance on expert experience and manual features and have been widely applied. However, traditional intelligent recognition models usually require complete and high-quality training data, which is often difficult to meet in actual engineering. The data is often noisy, mainly in two aspects: First, accurate labeling of large-scale monitoring data often requires downtime for disassembly and inspection, which is costly and time-consuming. Moreover, it is affected by factors such as personnel experience, recording errors, and operational oversights, resulting in label noise in the training data, including incorrect and missing labels, which misleads model learning, causing feature representation distortion and a decline in generalization performance. Second, due to hardware aging, line faults, signal interruptions, or interference from harsh working conditions, recognition applications often encounter missing sensor channel data, leading to a mismatch between the input distribution and the training distribution, which significantly reduces the recognition performance of multi-sensor fusion models and affects the reliability of decision-making.

[0004] To address label noise, existing research has proposed methods such as noise-robust learning, sample selection, and consistency regularization. To address channel missing issues, strategies such as data imputation, robust network structures, and efficient parameter fine-tuning have been explored. However, the former may mistakenly delete valuable samples, while the latter may introduce inaccurate imputation or require repeated design and training for different missing patterns, resulting in high engineering costs. In addition, existing research usually addresses the two types of problems separately, while collaborative modeling research for label noise and channel missing issues is relatively limited, and cannot balance accuracy, stability, and parameter efficiency in resource-constrained scenarios.

[0005] Therefore, how to provide a novel intelligent state recognition method that can learn robust knowledge from noisy data that is not precisely labeled, and can flexibly and efficiently adapt to the dynamic changes of incomplete sensor data during diagnosis, so as to meet the needs of engineering applications, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] Therefore, the application provides a low-quality data robust learning method for equipment health state intelligent identification to solve the problems in the background art.

[0007] To achieve the above object, the application adopts the following technical solutions.

[0008] A low-quality data robust learning method for equipment health state intelligent identification, comprising the following steps:

[0009] S1. Obtaining time series data of a plurality of sensors of an electromechanical composite transmission system and performing preprocessing;

[0010] S2. Constructing a multi-sensor fusion identification network composed of a feature encoder, a projection head and a classification head;

[0011] S3. Generating a near neighbor consensus label and a consistency measure for the sample according to the embedding space near neighbor consistency, dynamically screening high confidence samples and same class sample pairs, performing pre-training of a supervised contrast learning, a classification cross entropy and a sample pair similarity joint target, and introducing a mixed enhanced data enhancement strategy to smooth the discrimination boundary to obtain a robust encoder;

[0012] S4. Freezing the robust encoder, training a new classification head only with the high confidence set, and obtaining a high-precision state recognition model;

[0013] S5. Using the trained state recognition model to directly infer the complete channel samples to output the probability distribution of each state category; for the sensor channel data missing scene, a feature channel adaptive calibration module is introduced in the state recognition model to dynamically compensate the feature distribution deviation, and the probability distribution of each state category is output.

[0014] Preferably, the specific content of step S1 comprises:

[0015] The original long sequence signal is cut into a sample set according to a fixed window size; each channel of each sample is independently normalized to zero mean and unit variance; and the processed data set is divided into a training set, a validation set and a test set.

[0016] Preferably, in step S2, the specific content of the multi-sensor fusion identification network is:

[0017] For the input sample , the encoder maps the input sample to a high-dimensional feature representation ;

[0018] The projection head further maps the high-dimensional feature to a low-dimensional embedding for measuring similarity ;

[0019] Classification Head Map high-dimensional features to unnormalized logits vectors. ;

[0020] in, These are the learnable parameters for the encoder, projector head, and classifier head, respectively.

[0021] Normalized to the class posterior probability distribution using the Softmax function:

[0022]

[0023] in, This indicates that the sample belongs to the first... The probability of a class.

[0024] Preferably, step S3 includes the following:

[0025] S31. Perform several rounds of unsupervised contrastive learning warm-up to obtain a preliminary separable feature space;

[0026] S32. Measure the similarity between samples in the embedding space. Based on the similarity, perform a similarity clustering algorithm to retrieve each sample in the embedding space, count the label distribution of its nearest neighbor samples, and select the category with the highest proportion in the distribution as the nearest neighbor consensus label of the sample.

[0027] S33. Quantify the consistency between the original label and the neighborhood consensus by calculating the consistency loss;

[0028] S34. Filter confidence samples according to the consistency loss threshold, construct a confidence set, and construct a set of confidence sample pairs in the confidence set based on the label consistency principle and feature similarity threshold;

[0029] S35. The total loss function is the weighted sum of supervised contrast loss, classification cross-entropy loss and sample pair similarity loss. Mixed augmentation data augmentation is introduced to linearly combine two samples to generate virtual samples. The mixed label of the predicted virtual sample is directly used as a supervision signal for the calculation of classification cross-entropy loss.

[0030] S36. In each training round, a small batch of data is randomly sampled, the confidence set and the confidence sample pair set are dynamically updated, the Adam optimizer is used to backpropagate the total loss, and the learnable parameters of the encoder, projector head and classifier head are updated until the learning rate decay or early stopping mechanism is triggered, and finally a converged and stable robust encoder is obtained.

[0031] Preferably, the consistency loss calculation method in step S33 is as follows:

[0032]

[0033] in, For the sample Consistency loss, For the sample The original tag, For the sample In its embedded space The label distribution of the nearest neighbor samples, To find the sample in the nearest neighbor samples Original tag The proportion of the same sample A numerical stability constant is set to prevent logarithmic overflow. The smaller the value of the consistency loss, the better the original label. The higher the consistency with the consensus of the neighborhood, the higher its reliability.

[0034] The method for generating nearest neighbor consensus labels for each sample based on neighborhood consensus is as follows:

[0035]

[0036] in, For the sample The label of neighborly consensus Indicates in the sample Nearest neighbor label distribution Choose the category with the highest percentage As its consensus label.

[0037] Preferably, step S34, the specific content of constructing the confidence set and the confidence sample pair set, is as follows:

[0038] S341. Based on the consistency loss threshold Filter confidence samples and construct confidence sets for:

[0039]

[0040] S342. Based on the principle of label consistency, a preliminary set is obtained. :

[0041]

[0042] S343. Utilizing feature similarity thresholds Further purification yields a set. :

[0043]

[0044] in, Indicates the sample The sample-label pairs consisting of the original labels and their original labels. and Indicates the sample The sample-label pairs consisting of the original labels and their original labels. Indicates sample and samples Feature similarity in the embedding space and These are the preset consistency loss threshold and feature similarity threshold, respectively;

[0045] S344. Taking the union of the sets yields the final set of confidence sample pairs. :

[0046] .

[0047] Preferably, the total loss function is:

[0048]

[0049] Among them, supervised comparison loss for:

[0050]

[0051] in, For a small batch of sample index set, For anchor samples The set of positive samples, This refers to temperature hyperparameters.

[0052] Classification cross-entropy loss for:

[0053]

[0054] in, For the target distribution of the th Each component, if using blend enhancement, is a soft label;

[0055] Sample Pair Similarity Loss for:

[0056]

[0057] Where, q ij Let these be the probabilities of two samples being of the same type. These are similar pointers.

[0058] Preferably, step S4 includes the following:

[0059] S41. Fix the encoder parameters obtained in step S3, remove the projection head and the old classification head, connect the new classification head to the encoder output feature and initialize it randomly;

[0060] S42. Using a high-confidence set as training data, fine-tune the new classification head with the goal of minimizing the standard cross-entropy loss;

[0061] S43. Early stopping verification is used to determine convergence, and finally a high-precision state recognition model is obtained.

[0062] Preferably, in step S5, for scenarios where sensor channel data is missing, the state recognition model introduces a feature channel adaptive calibration module to dynamically compensate for feature distribution offset, and outputs the specific content of the probability distribution of each state category as follows:

[0063] Without altering the existing encoder and classification head, a lightweight, learnable feature channel adaptive calibration module is introduced to perform independent affine transformations on the intermediate feature maps of each effective channel within the encoder.

[0064] For specific missing patterns, only the affine parameters are fine-tuned to align the feature distribution;

[0065] The features calibrated by the feature channel adaptive calibration module are sent to the classification head to obtain the final category probability distribution. The category with the highest probability is used as the final fault diagnosis result to complete the state classification of the sample with missing sensor channels.

[0066] The preferred method for affine transformation is as follows:

[0067]

[0068] in, For the first Intermediate characteristics of an effective channel, The transformed features, and They are respectively with Learnable scaling and translation vectors for shape matching. This indicates element-wise multiplication;

[0069] The method for fine-tuning only the affine parameters to align the feature distribution for a specific missing pattern is as follows:

[0070] Use a small-scale, label-reliable calibration dataset Training is performed using cross-entropy loss as the target:

[0071]

[0072] in, This indicates an encoder with an embedded feature channel adaptive calibration module;

[0073] During training, only the affine parameters are considered. Gradient updates are performed while the parameters of the backbone network and are kept frozen. .

[0074] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a robust learning method for low-quality data for intelligent identification of equipment health status. It adopts neighborhood consistency-driven high-confidence sample / sample pair screening and supervised comparative learning to learn a high-fidelity representation of the feature space on noisy training data; and uses feature channel adaptive calibration to perform small-scale parameter adaptation to cope with multiple sensor missing combinations in the deployment stage, reduce the need for full retraining and reduce the instability caused by interpolation. It can achieve robust and efficient intelligent status identification for industrial data with both label uncertainty and channel missing.

[0075] Specifically, this invention reduces the interference of label noise on model training and improves the robustness and generalization ability of the recognition model by adopting a phased robust learning strategy based on neighborhood consistency screening; by introducing lightweight feature channel adaptive calibration, it reduces the retraining cost of the model to adapt to different missing modes and enhances the efficiency and robustness of engineering deployment. Attached Figure Description

[0076] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0077] Figure 1 A schematic diagram illustrating a robust learning method for low-quality data in intelligent identification of equipment health status provided by this invention;

[0078] Figure 2 A schematic diagram of phased robust training based on neighborhood consistency provided by the present invention;

[0079] Figure 3 This is a schematic diagram of the adaptive calibration of the feature channel provided by the present invention;

[0080] Figure 4 A schematic diagram showing the performance comparison of the method of the present invention and the comparative method under different tag noise conditions;

[0081] Figure 5 This diagram illustrates the performance comparison between the method of the present invention and the comparative method in the case of channel absence. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] This invention discloses a robust learning method for low-quality data in intelligent identification of equipment health status, comprising the following steps:

[0084] S1. Acquire and preprocess the multi-source sensor timing data of the electromechanical composite transmission system;

[0085] S2. Construct a multi-sensor fusion recognition network consisting of a feature encoder, a projection head, and a classification head;

[0086] S3. Generate nearest neighbor consensus labels and consistency measures for samples based on the nearest neighbor consistency in the embedding space, dynamically select high-confidence samples and similar sample pairs, and perform pre-training of the representation using the joint objectives of supervised contrastive learning, classification cross-entropy and sample pair similarity. Introduce a hybrid augmentation data enhancement strategy to smooth the discrimination boundary and obtain a robust encoder.

[0087] S4. Freeze the robust encoder and train the new classification head only with the high-confidence set to obtain a high-precision state recognition model;

[0088] S5. For complete channel samples, the trained state recognition model is used to directly infer and output the probability distribution of each state category; for scenarios where sensor channel data is missing, a feature channel adaptive calibration module is introduced into the state recognition model to dynamically compensate for feature distribution offset and output the probability distribution of each state category.

[0089] To further implement the above technical solution, the specific content of step S1 includes:

[0090] Multi-source sensor time series data ,in, For sampling length, The number of sensor channels; the corresponding (potentially noisy) category label is denoted as... Missing tags are indicated by an empty tag;

[0091] The original long sequence signal is divided into a sample set by a fixed window size; each channel of each sample is independently normalized to zero mean and unit variance; the processed dataset is divided into a training set, a validation set and a test set.

[0092] To further implement the above technical solution, the specific content of the multi-sensor fusion identification network in step S2 is as follows:

[0093] For input samples encoder Map the input samples to high-dimensional feature representations ;

[0094] Projector head High-dimensional features Further mapping to a low-dimensional embedding for measuring similarity ;

[0095] Classification Head Map high-dimensional features to unnormalized logits vectors. ;

[0096] in, These are the learnable parameters for the encoder, projector head, and classifier head, respectively.

[0097] Normalized to the class posterior probability distribution using the Softmax function:

[0098]

[0099] in, This indicates that the sample belongs to the first... The probability of a class.

[0100] To further implement the above technical solutions, such as Figure 2 The specific content of step S3 includes:

[0101] S31. Perform several rounds of unsupervised contrastive learning warm-up to obtain a preliminary separable feature space;

[0102] S32. Measure the similarity between samples in the embedding space, and based on the similarity, for each sample... The nearest clustering algorithm is used to retrieve the labels of the nearest samples in the embedding space. The category with the highest proportion in the distribution is selected as the nearest neighbor consensus label for this sample. ;

[0103] S33. Quantify the consistency between the original label and the neighborhood consensus by calculating the consistency loss;

[0104] S34. Filter confidence samples according to the consistency loss threshold, construct a confidence set, and construct a set of confidence sample pairs in the confidence set based on the label consistency principle and feature similarity threshold;

[0105] S35. The total loss function is the weighted sum of supervised contrast loss, classification cross-entropy loss and sample pair similarity loss. Mixed augmentation data augmentation is introduced to linearly combine two samples to generate virtual samples. The mixed label of the predicted virtual sample is directly used as a supervision signal for the calculation of classification cross-entropy loss.

[0106] S36. In each training epoch, a mini-batch of data is randomly sampled, and the confidence set and confidence sample pair set are dynamically updated. The Adam optimizer is used to backpropagate the total loss, and the learnable parameters of the encoder, projector head, and classifier head are updated. , ,in, The learning rate is used until the learning rate decays or an early stopping mechanism is triggered, eventually resulting in a convergent and stable robust encoder.

[0107] In the embedding space, the similarity between samples is based on the cosine similarity measure:

[0108]

[0109] in, This is the embedding vector.

[0110] To further implement the above technical solution, step S33, the method for quantifying the consistency between the original label and the neighborhood consensus through consistency loss calculation, is as follows:

[0111]

[0112] in, For the sample Consistency loss, For the sample The original tag, For the sample In its embedded space The label distribution of the nearest neighbor samples, To find the sample in the nearest neighbor samples Original tag The proportion of the same sample A numerical stability constant is set to prevent logarithmic overflow. The smaller the value of the consistency loss, the better the original label. The higher the consistency with the consensus of the neighborhood, the higher its reliability.

[0113] The method for generating nearest neighbor consensus labels for each sample based on neighborhood consensus is as follows:

[0114]

[0115] in, For the sample The label of neighborly consensus Indicates in the sample Nearest neighbor label distribution Choose the category with the highest percentage As its consensus label.

[0116] To further implement the above technical solution, step S34, which involves constructing the confidence set and the confidence sample pair set, includes the following:

[0117] S341. Based on the consistency loss threshold Filter confidence samples and construct confidence sets for:

[0118]

[0119] S342. Based on the principle of label consistency, a preliminary set is obtained. :

[0120]

[0121] S343. Utilizing feature similarity thresholds Further purification yields a set. :

[0122]

[0123] in, Indicates the sample The sample-label pairs consisting of the original labels and their original labels. and Indicates the sample The sample-label pairs consisting of the original labels and their original labels. Indicates sample and samples Feature similarity in the embedding space and These are the preset consistency loss threshold and feature similarity threshold, respectively;

[0124] S344. Taking the union of the sets yields the final set of confidence sample pairs. :

[0125] .

[0126] To further implement the above technical solution, the total loss function is:

[0127]

[0128] Among them, supervised comparison loss for:

[0129]

[0130] in, For a small batch of sample index set, For anchor samples The set of positive samples, This refers to temperature hyperparameters.

[0131] To achieve smooth discrimination boundaries and enhance encoder robustness, a mixed-up data augmentation strategy is introduced, which augments two random samples from the training set. and The labels are linearly combined to generate a virtual sample located in the feature space of the original sample. By training the model to predict the mixed label of virtual samples, the model can learn a smoother and more linear decision transition between different categories, thus avoiding overfitting of the model to the training data.

[0132] The specific generation process is as follows:

[0133]

[0134] in, Sampling from the Beta distribution The tag vector is a one-hot encoded tag vector;

[0135] The resulting label is a soft label in the form of a probability distribution, which will be directly used as a supervision signal for the subsequent calculation of classification cross-entropy loss;

[0136] Classification cross-entropy loss for:

[0137]

[0138] in, For the target distribution of the th Each component, if using blend enhancement, is a soft label;

[0139] Sample Pair Similarity Loss The aim is to constrain the output of the classification head by utilizing the similarity and dissimilarity relationships between sample pairs, specifically:

[0140] Two samples Perform an inner product and map it to the same probability q using the Sigmoid function. ij :

[0141]

[0142] With similar pointers For supervision, assign 1 to similar classes and 0 to dissimilar classes, and calculate the binary cross-entropy loss:

[0143]

[0144] Where, q ij Let these be the probabilities of two samples being of the same type. These are similar pointers.

[0145] To further implement the above technical solution, step S4 includes the following:

[0146] S41. The encoder obtained in step S3 With parameters fixed, project the head Compared with the old classification head Remove, and connect a new classification head to the encoder output features. And initialize randomly;

[0147] S42. Using high-confidence sets As training data, with the goal of minimizing the standard cross-entropy loss, a new classification head is used. Make fine adjustments;

[0148]

[0149] in, Indicates the first Class probability;

[0150] S43. Early stopping verification is used to determine convergence, ultimately obtaining a high-precision state recognition model. .

[0151] In this embodiment, when the sensor channels of the sample to be tested are complete, the trained state recognition model is used directly. Perform forward reasoning, output the probability distribution of each state category, and use the category corresponding to the highest probability as the recognition result: .

[0152] When some sensor channel data is missing, the distribution of input data will change, leading to a decrease in the performance of the pre-trained model.

[0153] To further implement the above technical solutions, such as Figure 3 In step S5, for scenarios with missing sensor channel data, the state recognition model introduces a feature channel adaptive calibration module to dynamically compensate for feature distribution offset, and outputs the specific content of the probability distribution of each state category as follows:

[0154] Without modifying the already fixed encoder With classification head Under this premise, a lightweight, learnable feature channel adaptive calibration (AC) module is introduced to perform intermediate feature maps for each effective channel within the encoder. Perform independent affine transformations;

[0155] For specific missing patterns, only the affine parameters are fine-tuned to align the feature distribution;

[0156] The features, after being calibrated by the feature channel adaptive calibration module, are sent to the classification head. The final category probability distribution is obtained, and the category with the highest probability is used as the final fault diagnosis result to complete the state classification of the sample with missing sensor channels.

[0157] To further implement the above technical solution, the affine transformation method is as follows:

[0158]

[0159] in, For the first Intermediate characteristics of an effective channel, The transformed features, and They are respectively with Learnable scaling and translation vectors for shape matching. This indicates element-wise multiplication;

[0160] The method for fine-tuning only the affine parameters to align the feature distribution for a specific missing pattern is as follows:

[0161] Use a small-scale, label-reliable calibration dataset Training is performed using cross-entropy loss as the target:

[0162]

[0163] in, This indicates an encoder with an embedded feature channel adaptive calibration module;

[0164] During training, only the affine parameters are considered. Gradient updates are performed while the parameters of the backbone network and are kept frozen. ;

[0165] In this way, the model only needs to be trained with less than 1% of the total number of parameters to dynamically compensate for the feature distribution shift caused by missing channels, thus achieving robust inference for any combination of missing sensor data.

[0166] In another embodiment, taking the health status identification of an electromechanical composite transmission system as an example, the method of the present invention is verified using an urban rail train bogie test bench.

[0167] The experimental setup was built at a 1:2 scale, using a three-phase asynchronous motor applied at different speeds, and a hydraulic loading device to simulate different load conditions. Data acquisition covered 24 channels with a sampling frequency of 64kHz. The channel types included three-dimensional vibration, three-phase current, speed, and sound, totaling 24 channels. Specific health status settings included 5 categories for motors (normal state, bearing failure, short circuit, rotor bending, and rotor bar breakage), 9 categories for gearboxes (normal state, bearing inner ring failure, bearing outer ring failure, bearing rolling element failure, bearing cage failure, gear crack, gear wear, gear tooth breakage, and gear tooth loss), and 5 categories for axle boxes (normal state, bearing inner ring failure, bearing outer ring failure, bearing rolling element failure, and bearing cage failure). There were a total of 9 operating conditions. 100 long sequence samples were collected for each category and each operating condition, with a sampling length of 640,000.

[0168] The original sequence was divided into samples of length 1024 using a sliding window, and zero-mean unit variance standardization was performed on each channel. The training set, validation set, and test set were then divided in a 6:2:2 ratio.

[0169] To verify the robustness during the training phase, three training sets with label uncertainty were constructed: no noise, 10% incorrect labels + 5% missing labels, and 15% incorrect labels + 10% missing labels. To evaluate the channel missing adaptation capability during the deployment phase, 1, 3, and 5 sensor channels were randomly removed from the test set, corresponding to mild, moderate, and severe incomplete input scenarios, respectively.

[0170] The experiment used the method proposed in this invention to identify the state of the motor, gearbox, and axle box. Based on this, for ease of comparison, two robust training baselines were set during the training phase: ordinary supervised learning and MOIT+; two parameter adaptation baselines were set during the identification phase: BitFit and LoRA. The basic network structure is shown in Table 1, the relevant training parameters are shown in Table 2, and the robust training comparison results are shown in Tables 3 and 4. Figure 4 The results of the comparison of adaptation methods are shown in Table 4 and Figure 5 ;

[0171] Table 1. Summary of Basic Network Structure:

[0172]

[0173] Table 2 Summary of Training-Related Parameters:

[0174]

[0175] Table 3. Summary of accuracy at different label noise levels:

[0176]

[0177] Table 4. Summary of accuracy at different channel missing levels:

[0178]

[0179] The experimental results show that the recognition performance of the method of the present invention is significantly better than that of the comparison method under different tag noise and sensor channel missing conditions, and achieves robust and efficient state recognition more effectively.

[0180] A computer-readable storage medium having a computer program stored thereon, characterized in that, when executed by a processor, the computer program implements a robust learning method for low-quality data for intelligent identification of equipment health status.

[0181] A processing terminal includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, characterized in that the processor, when executing the computer program, implements a robust learning method for low-quality data for intelligent identification of equipment health status.

[0182] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0183] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A low-quality data robust learning method for equipment health state intelligent identification, characterized in that, The method comprises the following steps: S1. Obtain multi-source sensor time series data of the electromechanical composite transmission system and perform preprocessing; S2. Construct a multi-sensor fusion recognition network composed of a feature encoder, a projection head, and a classification head; S3. Generate a near neighbor consensus label and a consistency measure for the sample according to the embedding space near neighbor consistency, dynamically screen high confidence samples and same class sample pairs, perform representation pre-training on a joint target of supervised contrastive loss, classification cross-entropy, and sample pair similarity, introduce a mixed augmented data augmentation strategy to smooth the discrimination boundary, and obtain a robust encoder; S4. Freeze the robust encoder, train a new classification head only with the high confidence set, and obtain a high-precision state recognition model; S5. Use the trained state recognition model to directly infer the probability distribution of each state category for complete channel samples; for sensor channel data missing scenarios, a feature channel adaptive calibration module is introduced in the state recognition model to dynamically compensate for feature distribution deviation, and the probability distribution of each state category is output.

2. The low-quality data robust learning method for equipment health state intelligent identification according to claim 1, wherein, The specific content of step S1 includes: Divide the original long sequence signal into samples according to a fixed window size; independently perform zero-mean unit variance normalization processing on each channel of each sample; and divide the processed data set into a training set, a validation set, and a test set.

3. The low-quality data robust learning method for equipment health state intelligent identification according to claim 1, wherein, In step S2, the specific content of the multi-sensor fusion recognition network is: mapping input samples , encoder mapping input samples to high-dimensional feature representations ; Projection head high-dimensional features are further mapped to low-dimensional embeddings for measuring similarity ; Classification head Mapping high-dimensional features to unnormalized logits vectors ; wherein, are the learnable parameters of the encoder, the projection head, and the classification head, respectively; After normalization by the Softmax function, the category posterior probability distribution is obtained: ； wherein, represents the probability that the sample belongs to the class.

4. The low-quality data robust learning method for equipment health state intelligent identification according to claim 1, wherein, The specific content of step S3 includes: S31. Perform several rounds of unsupervised contrastive learning preheating to obtain a preliminarily separable feature space; S32. Measure the similarity between samples in the embedding space, perform a similar clustering algorithm retrieval in the embedding space for each sample based on the similarity, count the label distribution of the near neighbor samples, and select the category with the highest proportion in the distribution as the near neighbor consensus label of the sample; S33. Quantify the consistency between the original label and the neighborhood consensus through consistency loss calculation; S34. Screen confidence samples according to the consistency loss threshold, construct a confidence set, and construct a confidence sample pair set in the confidence set based on the label consistency principle and the feature similarity threshold; S35. Use the weighted sum of the supervised contrastive loss, the classification cross-entropy loss, and the sample pair similarity loss as the total loss function, introduce mixed augmented data augmentation, linearly combine the two samples to generate a virtual sample, and use the mixed label of the predicted virtual sample as the supervision signal for classification cross-entropy loss calculation; S36. In each training round, randomly sample a small batch of data, dynamically update the confidence set and the confidence sample pair set, use the optimizer Adam to perform back propagation on the total loss, and update the learnable parameters of the encoder, the projection head, and the classification head until the learning rate decay or early stopping mechanism is triggered, and finally obtain a converged and stable robust encoder.

5. The low-quality data robust learning method for equipment health state intelligent identification according to claim 4, characterized in that, The consistency loss calculation method in step S33 is: ； wherein, consistency loss of the sample, consistency loss of the sample, original label of the sample, original label of the sample, label distribution of the k neighbor samples in the embedding space of the sample, label distribution of the k neighbor samples in the embedding space of the sample, proportion of the samples identical to the sample in the neighbor samples, proportion of the samples identical to the sample in the neighbor samples, original label original label numerical stability constant set to prevent overflow of logarithmic operation, the smaller the value of the consistency loss, the higher the consistency of the original label with the neighborhood consensus, and the higher the reliability thereof; The method for generating a near neighbor consensus label for each sample based on the neighborhood consensus is: ； wherein, is a consensus label of the sample , represents selecting the class with the highest proportion in the neighbor label distribution of the sample , as its consensus label. ​ 6. The low-quality data robust learning method for equipment health state intelligent identification according to claim 5, characterized in that, In step S34, the specific content of constructing the confidence set and the confidence sample pair set is: S341. In accordance with the consistency loss threshold Screening the confident samples, constructing the confident set For: ； S342. Screening based on the label consistency principle, to obtain a preliminary set : ； S343. Utilize feature similarity threshold Further purify to obtain collection : ； wherein, represents a sample-label pair consisting of a sample and its original label, and represents a sample-label pair consisting of a sample and its original label, represents a sample and a sample a feature similarity in the embedding space, and are respectively a preset consistency loss threshold and a feature similarity threshold. S344. Take the union to get the final set of pairs of confidence samples : 。 7. The low-quality data robust learning method for equipment health state intelligent identification according to claim 6, characterized in that, The total loss function is: ； wherein the supervised contrastive loss is: ； wherein, is a set of indices of a small batch of samples, is an anchor sample is a set of positive samples of the anchor sample, is a temperature hyperparameter; Classification cross-entropy loss is: ； in, For the target distribution of the th Each component, if using blend enhancement, is a soft label; Sample pair similarity loss is: ； where q ij is the homogeneity probability of two samples, is the homogeneity indicator.

8. The low-quality data robust learning method for equipment health state intelligent identification according to claim 1, wherein, The specific content of step S4 includes: S41. Fix the encoder parameters obtained in step S3, remove the projection head and the old classification head, connect a new classification head on the encoder output features and randomly initialize it; S42. Fine-tune the new classification head using the high-confidence set as training data, aiming to minimize the standard cross-entropy loss; S43. Determine convergence by early stopping and finally obtain a high-precision state recognition model.

9. The low-quality data robust learning method for equipment health state intelligent identification according to claim 1, wherein, In step S5, for the sensor channel data missing scenario, a feature channel adaptive calibration module is introduced in the state recognition model to dynamically compensate for feature distribution deviation, and the specific content of the output probability distribution of each state category is: Without changing the solidified encoder and classification head, a lightweight and learnable feature channel adaptive calibration module is introduced to perform independent affine transformation on the intermediate feature maps of each effective channel in the encoder; For a specific missing pattern, only the affine parameters are fine-tuned to align the feature distribution; The calibrated features are sent to the classification head to obtain the final category probability distribution, and the category with the highest probability is taken as the final fault diagnosis result, completing the state classification of the current sample with sensor channel missing.

10. The low-quality data robust learning method for equipment health state intelligent identification according to claim 9, wherein, The method of affine transformation is: ； wherein, is the intermediate feature for the th active channel, is the transformed feature, and are a learnable scaling vector and a translation vector, respectively, that are shape-matched to the , and denotes element-wise multiplication; The method of fine-tuning only the affine parameters for a specific missing pattern to align the feature distribution is: Using a smaller, labeled calibration dataset Training with cross-entropy loss as the objective: ； wherein, represents an encoder in which a feature channel adaptive calibration module is embedded; During training, only the affine parameters are updated with gradients, while the parameters of the backbone network and are kept frozen: .

Images