A fault-sensitive self-supervised equipment anomaly detection method and system

Abstract

The application discloses a kind of self-supervision equipment anomaly detection method and system sensitive to fault, comprising: obtaining equipment state monitoring data;Based on equipment state monitoring data, construct contrast learning sample pair;Contrast feature learning module is constructed and trained, and the input contrast learning sample pair is mapped into feature space;L2 regularization processing is carried out to feature space;Based on principal component analysis, the feature space after regularization processing is reduced in dimension, and the feature representation satisfying preset requirement is extracted;Health state threshold determination is carried out based on the feature representation satisfying preset requirement, and equipment anomaly detection result is obtained.The application turns from the traditional detection model susceptible to non-key information interference to the anomaly detection model only sensitive to fault, significantly improves the fault detection and diagnosis performance under the constraint of real world, alleviates the previous limitations, and extends the applicability of deep learning to equipment health state monitoring in real environment.

Description

Technical Field

[0001] This invention belongs to the fields of deep learning and artificial intelligence technology, and specifically relates to a fault-sensitive self-supervised equipment anomaly detection method and system. Background Technology

[0002] Anomaly detection of equipment plays a crucial role in preventing malfunctions, improving safety, reducing maintenance costs, optimizing maintenance plans, and enhancing production efficiency. Timely detection and analysis of equipment anomalies ensure the normal operation of equipment, thereby achieving more sustainable production and operation.

[0003] With increasing investment in condition monitoring equipment, methods driven by large-scale training data have shown outstanding performance in anomaly detection. However, in real-world safe production environments, equipment failures are extremely rare, and failure samples are scarce. This makes the available training datasets typically neither representative nor large-scale. Furthermore, existing anomaly detection algorithms often overlook the fact that anomalies in the data are not only caused by failures but also by changes in the operating environment. These non-critical factors can lead to discrete or continuous shifts in the underlying data distribution. If failures are detected every time the operating environment changes, anomaly detection will generate many false alarms, thus lacking deployability in actual operation.

[0004] Therefore, in order to simultaneously achieve invariance to non-critical factors and sensitivity to faults, this invention proposes a fault-sensitive self-monitoring equipment anomaly detection method and system. Summary of the Invention

[0005] To address the constraints of scarce fault data and labels, and to improve the robustness of anomaly detection methods to changes in data unrelated to health status and their sensitivity to potential faults, this invention proposes a fault-sensitive self-supervised equipment anomaly detection method and system. To solve the adaptability problem of contrastive learning in the absence of erroneous data, a time index is chosen to quantify the similarity of different health statuses, constructing contrastive learning sample pairs. To address the gradient vanishing and degradation problems in deep network training, residual connections are introduced, and a ResNet1D network is selected as the backbone architecture of the contrastive learning feature encoder model. By constructing a contrastive feature learning module, the aim is to learn a feature representation that is invariant to operational or environmental factors but sensitive to changes in equipment health status. For unsupervised health status monitoring without faults, to accurately and quickly detect equipment anomalies, a health index is constructed based on the characteristic that the distance between the feature space and the health class can represent the severity of the defect. This index effectively reflects the basic state information of the equipment over time. Health detection of the equipment is completed according to a threshold, and the fault time is calibrated.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A fault-sensitive self-monitoring equipment anomaly detection method includes the following steps:

[0008] Acquire equipment status monitoring data;

[0009] Based on the equipment status monitoring data, construct comparative learning sample pairs;

[0010] Construct and train a contrastive feature learning module, and based on the trained contrastive feature learning module, map the input contrastive learning sample pairs into the feature space;

[0011] The feature space is then subjected to L2 regularization.

[0012] Based on principal component analysis, the dimensionality of the regularized feature space is reduced, and feature representations that meet preset requirements are extracted.

[0013] Health status thresholds are determined based on feature representations that meet preset requirements, and equipment anomaly detection results are obtained.

[0014] Preferably, the comparative learning sample pairs include: positive pairs and negative pairs; wherein, the positive pairs are data samples recorded within a fixed time period after the equipment is put into operation, and the positive pairs include: anchor samples and positive samples; the negative pairs are other samples besides the data samples recorded within the fixed time period after the equipment is put into operation, i.e., negative samples.

[0015] Preferably, the contrastive feature learning module selects the ResNet1D network as the backbone architecture and introduces residual connections.

[0016] Preferably, the contrastive loss function is used to optimize the parameters of the contrastive feature learning module through the backpropagation algorithm to complete the training of the contrastive feature learning module.

[0017] Preferably, the method for determining the health status threshold based on feature representations that meet preset requirements and obtaining equipment anomaly detection results includes:

[0018] Based on the characteristic that the distance of health classes in the feature space represents the severity of defects, a health index that can effectively reflect the basic state information of equipment over time is constructed.

[0019] Based on benchmark data, the threshold for detecting and judging abnormalities in health indicators is determined.

[0020] Based on the anomaly detection threshold, complete the health check of the equipment and calibrate the fault time.

[0021] The present invention also provides a fault-sensitive self-supervised equipment anomaly detection system, comprising: an acquisition module, a construction module, a training module, a regularization module, an extraction module, and a detection module;

[0022] The acquisition module is used to acquire equipment status monitoring data;

[0023] The construction module is used to construct comparative learning sample pairs based on the equipment status monitoring data;

[0024] The training module is used to construct and train a contrastive feature learning module, and based on the trained contrastive feature learning module, to map the input contrastive learning sample pairs into the feature space.

[0025] The regularization module is used to perform L2 regularization on the feature space;

[0026] The extraction module is used to reduce the dimensionality of the regularized feature space based on principal component analysis and extract feature representations that meet preset requirements.

[0027] The detection module is used to determine the health status threshold based on the feature representation that meets the preset requirements, and to obtain the equipment abnormality detection result.

[0028] Preferably, the comparative learning sample pairs include: positive pairs and negative pairs; wherein, the positive pairs are data samples recorded within a fixed time period after the equipment is put into operation, and the positive pairs include: anchor samples and positive samples; the negative pairs are other samples besides the data samples recorded within the fixed time period after the equipment is put into operation, i.e., negative samples.

[0029] Preferably, the contrastive feature learning module selects the ResNet1D network as the backbone architecture and introduces residual connections.

[0030] Preferably, the contrastive loss function is used to optimize the parameters of the contrastive feature learning module through the backpropagation algorithm to complete the training of the contrastive feature learning module.

[0031] Preferably, the detection module includes: an index construction unit, a threshold determination unit, and a fault calibration unit;

[0032] The indicator construction unit is used to construct health indicators that can effectively reflect the basic state information of equipment over time, based on the characteristic that the distance of health classes in the feature space represents the severity of defects.

[0033] The threshold determination unit is used to determine the abnormal detection threshold of health indicators based on the benchmark data.

[0034] The fault calibration unit is used to complete the health check of the equipment based on the anomaly detection judgment threshold and to calibrate the fault time.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] This invention provides a fault-sensitive, self-supervised equipment anomaly detection method and system. It solves the problem of distinguishing between data changes caused by changes in health status and data changes caused by non-critical information factors (domain transformation) when there is little or no fault data. Unlike previous anomaly detection methods, this invention shifts from traditional detection models susceptible to interference from non-critical information to an anomaly detection model sensitive only to faults. This significantly improves fault detection and diagnosis performance under real-world constraints, alleviates previous limitations, and extends the applicability of deep learning to equipment health status monitoring in real-world environments. Attached Figure Description

[0037] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the overall process of anomaly detection using the self-monitoring equipment of the present invention;

[0039] Figure 2 This is the network structure of the contrastive feature learning module of the present invention;

[0040] Figure 3 This is a flowchart of the equipment anomaly detection module of the present invention. Detailed Implementation

[0041] 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.

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Example 1

[0044] The overall process of this invention is as follows: Figure 1As shown, firstly, to address the adaptability issue of contrastive learning in the absence of error data, a time-based metric is chosen to quantify the similarity of different health conditions, constructing contrastive learning sample pairs. Secondly, to address the gradient vanishing and degradation problems in deep network training, residual connections are introduced, and a ResNet1D network is selected as the backbone architecture of the contrastive learning feature encoder model. Then, the contrastive feature learning encoder is trained using contrastive learning loss, aiming to learn a feature representation that is invariant to operational or environmental factors on the one hand, but sensitive to changes in equipment health on the other. Finally, for unsupervised health status monitoring in fault-free situations, to accurately and quickly detect equipment anomalies, based on the characteristic that the distance between health classes in the feature space can represent the severity of defects, a health index is constructed that effectively reflects the basic state information of the equipment over time. Health detection of the equipment is completed based on a threshold, and the fault time is calibrated.

[0045] Step 1: Construct an equipment anomaly detection dataset

[0046] The system receives equipment condition monitoring data and splits the initial dataset, with all data from defective equipment and randomly selected healthy equipment forming the test dataset. Since accurate data on the actual condition of the equipment is difficult to obtain, it relies on labels provided by domain experts. The collected equipment monitoring data undergoes preliminary splitting; monitoring data from equipment that has been reported to have obvious defects is included in the "test" dataset. However, defect information may also exist in equipment data that has not yet been reported. The remaining data is then randomly divided into "training" and "test" datasets. All "test" data requires evaluation by domain experts; only data samples marked as healthy by the domain experts are added to the test dataset.

[0047] To ensure the optimal solution is obtained within a limited number of iterations, the data is normalized. This experiment uses the MinMaxScaler function in PyTorch. Min-max scaling is applied to the data, and through translation and scaling operations, the data eventually converges to the range [0,1]. The normalization calculation formula is as follows:

[0048]

[0049] Step 2: Construct contrastive learning sample pairs:

[0050] Considering the varying availability of data and labels for different tasks in complex systems such as equipment, different criteria are needed to select data triples so that the contrastive learning loss function can adapt. In supervised learning, labels are used to form data triples, where samples with the same label form positive pairs with the anchor sample, and samples with different labels form negative pairs. However, in reality, only healthy equipment state samples may be available, and the available training dataset is unlikely to contain data from fault conditions. Therefore, unsupervised feature learning is required. Under unsupervised conditions, it is impossible to directly impose sensitivity to faults or select data triples based on labels. To address the adaptability problem of contrastive learning in the absence of erroneous data, two facts are utilized: 1) the condition of equipment is typically monitored over discrete time periods; 2) due to normal degradation, the health status of each piece of equipment usually changes over time.

[0051] Based on the above facts, even if there is no available fault data in the dataset, it contains data from various non-critical information changes and data from various degradation states. Different degradation states can be considered as different health states within the health hierarchy.

[0052] Because samples separated by time exhibit different degradation states, a time-based index is chosen to quantify the similarity of different health conditions: data samples that are close in time are considered similar (positive samples), while data samples that are far apart in time are considered dissimilar (negative samples). In this invention, data samples recorded within a fixed time period after equipment operation are selected as the positive (x) pairs. a As an anchor sample, x p As positive samples, to ensure variability under different operational and environmental conditions, all other samples are considered dissimilar (x). n ).

[0053] Step 3: Construct a contrastive feature learning encoder

[0054] The ResNet1D network was chosen as the backbone architecture of the contrastive feature learning encoder model, such as Figure 2 As shown. ResNet1D is a one-dimensional convolutional neural network architecture based on residual networks. It introduces residual connections to solve the gradient vanishing and degradation problems in deep network training, and is used for feature extraction tasks of time-series data. The feature encoder network architecture is as follows:

[0055] 1) Input layer: The time sequence of equipment status monitoring in the input layer.

[0056] 2) Convolutional layer 1: Uses one-dimensional convolution kernels to perform convolution operations on the input signal, extracts the local temporal features of the signal, and sets appropriate convolution kernel size and number to capture features at different scales.

[0057] 3) Batch normalization layer 1: Batch normalize the output of convolutional layer 1 to accelerate network convergence and improve the robustness of the model.

[0058] 4) Activation function layer 1: Use an activation function (such as ReLU) to perform a nonlinear transformation on the output of convolutional layer 1, introducing nonlinear features.

[0059] 5) Residual Block 1: Consists of two convolutional layers, a batch normalization layer, and an activation function layer. First, the output of convolutional layer 1 is used as input, passing through convolutional layer 2, batch normalization layer 2, and activation function layer 2 to obtain intermediate features. Then, the output of convolutional layer 1 is added to the intermediate features element-wise to obtain the output of residual block 1.

[0060] 6) Residual Block Stacking: Residual Block 2 to Residual Block N: Similar to Residual Block 1, stacking multiple residual blocks increases the network depth and allows for more complex feature extraction of the signal. The appropriate number of residual blocks needs to be determined based on the timing, signal complexity, and task requirements.

[0061] 7) Global Average Pooling Layer: Performs global average pooling on the output of the last residual block, converting the feature sequence into a fixed-length feature vector. Global average pooling aggregates the temporal features of the signal into global features, preserving important information.

[0062] 8) Fully connected layer: Connects the output of the global average pooling layer to one or more fully connected layers, maps the features to the final feature vector space, and uses Dropout to prevent overfitting.

[0063] 9) Feature Encoding Layer: The output of the fully connected layer is used as the output of the feature encoding layer to obtain a discriminative feature representation. The feature encoding layer is trained using a contrastive loss function to minimize the distance between positive sample pairs while maximizing the distance between negative sample pairs.

[0064] 10) Output layer: Outputs feature representations for subsequent anomaly detection.

[0065] Step 4: Train the contrastive feature learning module

[0066] Contrastive learning aims to learn a feature representation that groups semantically similar data closer together and pushes semantically dissimilar data further apart. For equipment anomaly detection, it is assumed that data samples recorded under similar health conditions are considered semantically similar, while samples recorded under dissimilar health conditions are considered semantically dissimilar. This module chooses a contrastive learning method to learn feature representations, learning the differences between normal and abnormal samples. It does not require different domains to achieve domain-invariant embeddings and can improve generalization ability to unpredictable domain shifts. The core of this module is the contrastive loss function, the semi-hard triple loss, defined as follows:

[0067]

[0068] Where, x a For anchor samples, x p For positive samples (sharing semantic meaning with the anchor), x n Let f(.) represent semantically distinct negative samples and health states, where ||.|| is the coded sample, ||.|| is the distance metric, and ò is the marginal parameter. Minimizing the distance between the anchor point and positive sample pairs increases invariance to non-informational changes; increasing the distance between the anchor point and negative sample pairs increases sensitivity to faults. Furthermore, to emphasize invariance within the same health state, the sum of all positive samples within the batch is used.

[0069] The encoder model parameters are optimized using a contrastive loss function and backpropagation. During training, positive and negative sample pairs are randomly selected to ensure the encoder obtains discriminative feature representations. The Adam algorithm is used for parameter optimization, and hyperparameters such as the learning rate are adjusted based on the convergence of the loss function. The trained encoder can map input samples into a feature space and obtain discriminative feature representations. These features can be used for various machine learning tasks, such as classification, clustering, and detection.

[0070] Step 5: Regularize the feature space

[0071] To reduce the complexity of the anomaly detection model and avoid overfitting, L2 regularization is applied to the feature space to improve generalization ability. L2 regularization uses the sum of the squares of the parameter weights as a penalty term, making the parameter weights smoother. The expression for the L2 regularization term is as follows: it calculates the square root of the sum of the squares of each element in the weight vector w.

[0072]

[0073] Where α||w||2 2This is the L2 regularization term. Due to the multiplication operation between the input and weights, the network tends to use all input features rather than heavily relying on a small subset of them, thereby improving the model's generalization ability and reducing the risk of overfitting.

[0074] Step 6: Feature Space Dimensionality Reduction

[0075] Principal component analysis (PCA) is used to reduce the dimensionality of the feature space. PCA transforms the original data into principal components, allowing for better extraction of important information from the original data while reducing its dimensionality. This compresses the data, improves the efficiency of data analysis, and makes the results easier to understand and interpret. The specific implementation process of the algorithm is as follows:

[0076] (1) Let the characteristic space be X:

[0077]

[0078] (2) Calculate the covariance matrix of the feature space: that is, the matrix formed by the covariance between features.

[0079]

[0080] in,

[0081]

[0082] (3) Calculate the eigenvalues ​​and eigenvectors of R:

[0083] Eigenvalues:

[0084] λ1≥λ2≥L≥λ p ≥0

[0085] R is a positive semi-definite matrix, and Feature vector:

[0086]

[0087] Among them, the eigenvalues ​​represent the importance of the corresponding eigenvectors, and the eigenvectors represent the main direction of change. (4) Calculate the principal component contribution rate and the cumulative contribution rate:

[0088]

[0089]

[0090] (5) Give the principal components

[0091] Generally, the first, second, ..., m-th (m < p) principal components corresponding to the feature values ​​with a cumulative contribution rate greater than 80% are selected to ensure that the obtained principal components can cover most of the effective information. The i-th principal component is: Z i =a 1i X1+a 2i X2+L+a pi X p (i = 1, 2, L, m).

[0092] Step 7: Equipment Health Status Anomaly Detection

[0093] like Figure 3 As shown, for unsupervised health status monitoring without faults, a health index is constructed that can effectively reflect the basic status information of equipment over time for equipment anomaly detection. Based on the characteristic that the distance between health classes in the feature space can represent the severity of defects: similar health statuses are close to each other; slightly different health statuses are slightly farther apart; substantially different health statuses are far apart. Based on the feature space processed by principal component analysis for dimensionality reduction, the Mahalanobis distance metric is chosen to construct the health index. Mahalanobis distance is a measure of similarity between two samples, reflecting the degree of difference between the sample and the normal state. For data points x and y with a covariance matrix of ∑, the Mahalanobis distance can be expressed as:

[0094]

[0095] Assuming the sample data follows a multivariate normal distribution, with the mean as the center of the model and the covariance matrix describing the model's dispersion, the anomaly detection threshold is determined first. The feature space of the normal equipment baseline data is obtained, and its mean and covariance matrix are calculated. The mean is used to calculate the Mahalanobis distance from the training set to this mean. The threshold for anomaly detection is then determined according to the rule of three standard deviations from the distribution center.

[0096] Next, the sample mean and covariance matrix of the contrastive feature space after dimensionality reduction by principal component analysis are calculated for the training and test sets. The mean of the training set is used as the centroid, and the Mahalanobis distance from the test set to this centroid is calculated. Then, the equipment status is calibrated based on the results of the threshold judgment. If the Mahalanobis distance of the data point is less than the threshold, it is calibrated as "False"; otherwise, it is calibrated as "True". If it is calibrated as "True" three times consecutively, the point is judged to be the moment of failure.

[0097] Example 2

[0098] The present invention also provides a fault-sensitive self-supervised equipment anomaly detection system, comprising: an acquisition module, a construction module, a training module, a regularization module, an extraction module, and a detection module;

[0099] The acquisition module is used to acquire equipment status monitoring data;

[0100] The construction module is used to construct comparative learning sample pairs based on the equipment status monitoring data;

[0101] The training module is used to construct and train the contrastive feature learning module, and based on the trained contrastive feature learning module, the input contrastive learning sample pairs are mapped into the feature space;

[0102] The regularization module is used to perform L2 regularization on the feature space;

[0103] The extraction module is used to reduce the dimensionality of the regularized feature space based on principal component analysis and extract feature representations that meet preset requirements.

[0104] The detection module is used to determine the health status threshold based on the feature representation that meets the preset requirements, and to obtain the equipment anomaly detection results.

[0105] In this embodiment, the comparative learning sample pairs include: positive pairs and negative pairs; wherein, the positive pairs are data samples recorded within a fixed time period after the equipment is running, and the positive pairs include: anchor samples and positive samples; the negative pairs are other samples besides the data samples recorded within the fixed time period after the equipment is running, i.e., negative samples.

[0106] In this embodiment, the contrast feature learning module selects the ResNet1D network as the backbone architecture and introduces residual connections.

[0107] In this embodiment, the contrastive loss function is used to optimize the parameters of the contrastive feature learning module through the backpropagation algorithm, thereby completing the training of the contrastive feature learning module.

[0108] In this embodiment, the detection module includes: an index construction unit, a threshold determination unit, and a fault calibration unit;

[0109] The indicator construction unit is used to construct health indicators that can effectively reflect the basic state information of equipment over time, based on the characteristic that the distance of health classes in the feature space represents the severity of defects.

[0110] The threshold determination unit is used to determine the threshold for detecting and judging abnormalities in health indicators based on benchmark data;

[0111] The fault calibration unit is used to complete the health check of the equipment based on the abnormality detection judgment threshold and to calibrate the fault time.

[0112] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A fault-sensitive self-monitoring equipment anomaly detection method, characterized in that, Includes the following steps: Acquire equipment status monitoring data; Based on the equipment status monitoring data, construct comparative learning sample pairs; Construct and train a contrastive feature learning module, and based on the trained contrastive feature learning module, map the input contrastive learning sample pairs into the feature space; The feature space is then subjected to L2 regularization. Based on principal component analysis, the dimensionality of the regularized feature space is reduced, and feature representations that meet preset requirements are extracted. Health status thresholds are determined based on feature representations that meet preset requirements to obtain equipment anomaly detection results. The contrastive learning sample pairs include positive pairs and negative pairs; wherein, the positive pairs are data samples recorded within a fixed time period after the equipment is put into operation, and the positive pairs include anchor samples and positive samples; the negative pairs are other samples besides the data samples recorded within the fixed time period after the equipment is put into operation, i.e., negative samples. The contrastive feature learning module selects ResNet1D network as the backbone architecture and introduces residual connections; the feature encoder network architecture is as follows: Input layer: The time sequence of equipment status monitoring at the input layer; Convolutional layer 1: Uses a one-dimensional convolutional kernel to perform convolution operations on the input signal, extracts the local temporal features of the signal, and sets the size and number of convolutional kernels to capture features at different scales; Batch normalization layer 1: Performs batch normalization on the output of convolutional layer 1; Activation function layer 1: The output of convolutional layer 1 is nonlinearly transformed using an activation function to introduce nonlinear features; Residual Block 1: It consists of two convolutional layers, a batch normalization layer, and an activation function layer. First, the output of convolutional layer 1 is used as input, and the intermediate features are obtained by passing through convolutional layer 2, batch normalization layer 2, and activation function layer 2. Then, the output of convolutional layer 1 is added to the intermediate features element-wise to obtain the output of residual block 1. Residual block stacking: Residual block 2 to residual block N: The depth of the network is increased by stacking multiple residual blocks, which allows for feature extraction of the signal; the number of residual blocks is set according to the timing, signal complexity and task requirements. Global average pooling layer: Performs global average pooling on the output of the last residual block, converting the feature sequence into a fixed-length feature vector; global average pooling aggregates the temporal features of the signal into global features; Fully connected layer: Connects the output of the global average pooling layer to one or more fully connected layers, maps features to the final feature vector space, and uses Dropout to prevent overfitting; Feature encoding layer: The output of the fully connected layer is used as the output of the feature encoding layer to obtain a discriminative feature representation; the feature encoding layer is trained by using a contrastive loss function to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs. Output layer: Outputs feature representations for subsequent anomaly detection; The contrastive loss function is used to optimize the parameters of the contrastive feature learning module through backpropagation algorithm, thus completing the training of the contrastive feature learning module; the contrastive loss function, semi-hard triple loss, is defined as: ； in, For anchor samples, As a positive sample, Let f(.) represent semantically distinct negative samples and health statuses, where ||.|| is the encoded sample and ||.|| is the distance metric. The marginal parameters are used; by minimizing the distance between the anchor point and the positive sample pair, invariance to non-information changes is increased; by increasing the distance between the anchor point and the negative sample pair, sensitivity to faults is increased; and the sum of all positive samples within the batch is used. The encoder model parameters are optimized using a contrastive loss function via backpropagation; the trained encoder maps the input samples into the feature space and obtains discriminative feature representations.

2. The fault-sensitive self-monitoring equipment anomaly detection method according to claim 1, characterized in that, Methods for determining health status thresholds and obtaining equipment anomaly detection results based on feature representations that meet preset requirements include: Based on the characteristic that the distance between a feature space and a health class represents the severity of a defect, a health index that can effectively reflect the basic state information of equipment over time is constructed. Based on benchmark data, the threshold for detecting and judging abnormalities in health indicators is determined. Based on the anomaly detection threshold, complete the health check of the equipment and calibrate the fault time.

3. A fault-sensitive self-monitoring equipment anomaly detection system, said system being used to implement the method described in any one of claims 1-2, characterized in that, include: The module includes an acquisition module, a construction module, a training module, a regularization module, an extraction module, and a detection module. The acquisition module is used to acquire equipment status monitoring data; The construction module is used to construct comparative learning sample pairs based on the equipment status monitoring data; The training module is used to construct and train a contrastive feature learning module, and based on the trained contrastive feature learning module, to map the input contrastive learning sample pairs into the feature space. The regularization module is used to perform L2 regularization on the feature space; The extraction module is used to reduce the dimensionality of the regularized feature space based on principal component analysis and extract feature representations that meet preset requirements. The detection module is used to determine the health status threshold based on the feature representation that meets the preset requirements, and to obtain the equipment abnormality detection result.

4. The fault-sensitive self-monitoring equipment anomaly detection system according to claim 3, characterized in that, The comparative learning sample pairs include positive pairs and negative pairs; wherein, the positive pairs are data samples recorded within a fixed time period after the equipment is put into operation, and the positive pairs include anchor samples and positive samples; the negative pairs are other samples besides the data samples recorded within the fixed time period after the equipment is put into operation, i.e., negative samples.

5. The fault-sensitive self-monitoring equipment anomaly detection system according to claim 3, characterized in that, The contrastive feature learning module selects ResNet1D network as the backbone architecture and introduces residual connections.

6. The fault-sensitive self-monitoring equipment anomaly detection system according to claim 3, characterized in that, The contrastive loss function is used to optimize the parameters of the contrastive feature learning module through backpropagation algorithm, thereby completing the training of the contrastive feature learning module.

7. The fault-sensitive self-monitoring equipment anomaly detection system according to claim 3, characterized in that, The detection module includes: an index construction unit, a threshold determination unit, and a fault calibration unit; The indicator construction unit is used to construct health indicators that can effectively reflect the basic state information of equipment over time, based on the characteristic that the distance of health classes in the feature space represents the severity of defects. The threshold determination unit is used to determine the abnormal detection threshold of health indicators based on the benchmark data. The fault calibration unit is used to complete the health check of the equipment based on the anomaly detection judgment threshold and to calibrate the fault time.

Images