A key equipment intelligent diagnosis method and system for rotating machinery

Abstract

The application discloses a kind of key equipment intelligent diagnosis method and system for rotating machinery, it is related to the technical field of fault diagnosis and prediction, including, with physical inspiration as guidance, in feature learning stage, introduce bounded metric encapsulation mechanism, traditional unbounded feature space is mapped into bounded similarity manifold, to suppress high-dimensional space distance concentration effect and improve inter-class separability and intra-class compactness.In bounded metric space, realize multi-fault identification by combining near neighbor or support vector classifier.The application has significant improvement in weak fault (germination period) detection, composite / multiclass fault distinction and robustness under low signal-to-noise ratio condition, has good explainability and engineering deployment value, is suitable for the predictive maintenance and online health monitoring of wind power, rail transit, aerospace and process manufacturing and other scenes.

Description

Technical Field

[0001] This invention relates to the technical field of fault diagnosis and prediction, and in particular to an intelligent diagnostic method and system for key equipment of rotating machinery. Background Technology

[0002] With the increasing complexity of mechanical systems and the diversification of operating conditions, achieving health monitoring and fault diagnosis of critical components has become an important research direction in the field of intelligent operation and maintenance. Traditional fault detection methods typically rely on the time-domain or frequency-domain characteristics of vibration signals, such as root mean square value, kurtosis, and envelope spectrum. These methods are computationally simple but sensitive to noise, and often struggle to guarantee diagnostic accuracy under complex operating conditions.

[0003] In recent years, machine learning and deep learning-based methods have been increasingly applied to fault diagnosis of mechanical systems. For example, models such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Temporal Convolutional Networks (TCN) have demonstrated good feature extraction and classification capabilities under large-scale data-driven conditions. However, these methods have two prominent problems: first, they are highly dependent on the quantity and quality of training samples, making it difficult to obtain large-scale, high-quality data in real-world industrial scenarios; second, their internal mapping process lacks physical constraints, leading to insufficient interpretability of diagnostic results.

[0004] On the other hand, metric learning methods exhibit certain advantages under small sample conditions because they can directly model the similarity between samples. While existing metrics such as Euclidean distance and Mahalanobis distance are commonly used, they are all unbounded functions, which can easily lead to excessive diffusion of similarity distribution, resulting in unstable thresholds when constructing classification discrimination thresholds. Furthermore, traditional metric learning often only optimizes inter-class or intra-class distances, lacking a joint balance between the two, leading to insufficient discriminative power in feature embeddings.

[0005] Furthermore, existing fault detection methods often perform poorly when handling multiple fault scenarios (such as the simultaneous presence of inner race and rolling element faults), easily leading to feature aliasing and misclassification. Meanwhile, threshold setting methods typically rely on experience or fixed rules, making it difficult to adaptively adjust under different operating conditions, resulting in high false alarm and false negative rates.

[0006] Therefore, existing technologies still have shortcomings in terms of feature extraction stability, boundedness of metric functions, discriminativeness of embedding space, and adaptability of threshold construction. How to improve fault detection and multi-fault classification performance under strong noise conditions while ensuring the physical rationality of the method has become a pressing technical problem to be solved in this field. Summary of the Invention

[0007] In view of the aforementioned existing problems, the present invention is proposed.

[0008] Therefore, this invention provides an intelligent diagnostic method for key equipment of rotating machinery to solve the problems of insufficient stability of existing mechanical system fault detection and classification methods, unstable threshold due to unbounded metric functions, weak discriminative power of feature embedding, and poor performance in identifying multiple faults.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0010] In a first aspect, the present invention provides an intelligent diagnostic method for key equipment of rotating machinery, which includes collecting the state signals of rotating machinery during the operation of the mechanical system and extracting features in the time domain, frequency domain and time-frequency domain.

[0011] The differences in features are mapped by a monotonically bounded encapsulation function; using the processing results of the monotonically bounded encapsulation function in each dimension, the unbounded distance in the original feature space is transformed into a bounded similarity metric space.

[0012] Metric learning is performed in the bounded similarity metric space to jointly optimize the minimum inter-class margin and the maximum intra-class dispersion to obtain a discriminative embedding representation;

[0013] An adaptive threshold is generated based on the distance distribution obtained from the metric learning; the adaptive threshold is set by minimizing the sum of the false alarm rate and the false negative rate obtained from the distance distribution, and is used for fault detection.

[0014] Within the bounded similarity metric space, a classifier is invoked to identify multiple faults and output the corresponding fault categories.

[0015] The monotonically bounded encapsulation function includes, based on the feature set, using a nonlinear feature mapping operator to extend the original d-dimensional features to a q-dimensional high-dimensional feature space, amplifying the complex nonlinear differences of the mechanical system under different working conditions, thereby enhancing the feature representation capability; in the high-dimensional space, arbitrary real-valued differences are compressed and normalized to ensure that feature differences are mapped to a bounded interval.

[0016] As a preferred embodiment of the intelligent diagnostic method for key equipment of rotating machinery described in this invention, the status signal uses vibration acceleration signal as the core diagnostic basis; it is acquired by acceleration sensors deployed in key parts of the equipment, and sampled under different rotational speeds, loads, and environmental interference conditions;

[0017] The vibration signals cover a variety of typical operating conditions, including but not limited to healthy conditions, outer ring faults, inner ring faults, rolling element faults, and combined faults.

[0018] The state signal is converted into a discrete time series by an analog-to-digital converter module: ;

[0019] in, This represents the vibration amplitude at the Lth sampling point, with a sampling frequency of... The total number of sampling points is L.

[0020] As a preferred embodiment of the intelligent diagnostic method for key equipment of rotating machinery described in this invention, the original vibration sequence is... According to fixed window length Divide the data by the sliding step size h to obtain the following results. A window segment, This represents a segment of the i-th window;

[0021] After standardizing each window, in the standardized signal segment The above uses Fast Fourier Transform to map signal segments to the frequency domain;

[0022] Time-frequency feature calculation is performed using wavelet packet decomposition: the signal is decomposed into different frequency bands, and sub-band coefficients are obtained. This represents the subband coefficient of the m-th frequency band within the i-th window; and utilizes... Calculate the energy of each sub-band; under the i-th window, the sub-band energy of the m-th frequency band is expressed as: Simultaneously, in each subband, indicators such as energy entropy, envelope spectrum peak value, and high-frequency energy ratio are extracted.

[0023] For the i-th window, all feature combinations extracted in the time domain, frequency domain, and time-frequency domain are used to form a feature vector. The feature set consists of all window fragments: ;

[0024] in, The labels correspond to different fault types.

[0025] As a preferred embodiment of the intelligent diagnostic method for key equipment of rotating machinery described in this invention, wherein: the monotonically bounded encapsulated function The domain is The range is , satisfy:

[0026]

[0027] in, Show encapsulated functions The maximum output value is the upper bound of the metric mapping, which can restrict the original unbounded feature differences to a finite interval; Let the absolute difference between any two samples be along a certain feature dimension. ; This represents the feature obtained by mapping sample u along the k-th dimension; This represents the feature obtained by mapping sample v along the k-th dimension.

[0028] Define any two samples Bounded distance measurement:

[0029]

[0030] in, This represents the adjustable norm parameter, used to control the aggregation method of feature differences across different dimensions; This is a normalization factor to ensure that distances are comparable across different dimensions.

[0031] As a preferred embodiment of the intelligent diagnostic method for key equipment of rotating machinery described in this invention, the minimum inter-class interval includes, denoted as follows: Let the sets of healthy state and fault state samples be respectively denoted as... and Construct sample pairs from the set: samples of any two health states. Pair of fault state samples ;

[0032] To enhance the separability between fault categories, a minimum interval constraint between categories is defined:

[0033]

[0034] in, The margin threshold in the feature space is used to adjust the separability between healthy sample pairs and faulty sample pairs; the constraint ensures that the embedding distance between different state categories is as large as possible, thereby expanding the inter-class margin.

[0035] The maximum intra-class dispersion includes, for healthy sample pairs The goal is to minimize the embedding distance. For faulty sample pairs Then, by adjusting the intermediate parameters This controls the distribution range, limiting intraclass differences to a physically interpretable bounded interval;

[0036] The comprehensive optimization objective can be expressed as:

[0037]

[0038] The first item guarantees that the intra-class distance does not exceed the parameter. The second term ensures that the distance between classes is not less than the parameter. This allows for the combined optimization of intra-class compactness and inter-class separation. This indicates that an optimization search is performed within the set of all possible feature mapping functions.

[0039] As a preferred embodiment of the intelligent diagnostic method for key equipment of rotating machinery described in this invention, the adaptive threshold includes, within the feature space, a pair of health status samples... Pair of fault state samples Statistical analysis was performed on the distance distribution to obtain the distribution of the health category. Distribution of fault classes ;

[0040] Introducing detection thresholds As an adjustable parameter, it is adaptively set by jointly minimizing the sum of the false alarm rate and the false negative rate. The optimization objective is defined as: ;

[0041] in, This indicates that when the detection threshold is The false alarm rate is the probability that a healthy sample is mistakenly identified as faulty. This represents the false negative rate at that threshold, which is the probability that a faulty sample is mistakenly identified as healthy. This represents the optimal detection threshold obtained by minimizing the sum of the false positive probability of healthy samples and the false negative probability of faulty samples.

[0042] As a preferred embodiment of the intelligent diagnostic method for key equipment of rotating machinery described in this invention, the detection threshold includes, for any test sample pair (u, v), if If the sample pairs belong to different categories, it means that healthy samples and faulty samples can be distinguished; if If the sample pair is determined to belong to the same category, it indicates either the healthy category or the same type of fault category.

[0043] As a preferred embodiment of the intelligent diagnostic method for key equipment of rotating machinery described in this invention, the classifier includes: inputting the sample to be tested into a trained bounded metric embedding space, calculating the distance relationship between the sample to be tested and each known category sample, and combining the classification decision rules to output the corresponding fault category.

[0044] The classifier can be either a nearest neighbor classifier or a support vector machine classifier.

[0045] When using a nearest neighbor classifier, the fault category to which the test sample belongs is determined based on the nearest neighbor distance relationship between the test sample and each category of samples in the training set in the embedding space.

[0046] When a support vector machine classifier is used, different fault category regions are divided in the embedding space using hyperplanes. By determining the relative position of the test sample to each hyperplane, the fault category to which the test sample belongs is determined.

[0047] As a preferred embodiment of the intelligent diagnostic method for key equipment of rotating machinery described in this invention, wherein: in the metric learning, a strategy is adopted to adaptively select the hyperparameters involved in the feature space;

[0048] The hyperparameters include wrapper functions. Feature space dimension q and norm parameter Specifically:

[0049] The selection strategy must satisfy the following properties: monotonicity, meaning the output value does not decrease as the input value increases; and boundedness, meaning the output value is limited by a finite constant as the input approaches infinity. Differentiability and smoothness; continuous differentiability within the domain.

[0050] The selection strategy for the feature space dimension q: The feature space dimension q is adaptively determined based on the complexity of the fault diagnosis task; it is adaptively set by minimizing the weighted sum of classification error and model complexity on the validation set.

[0051]

[0052] in, The optimal feature space dimension; Let q be the classification error rate of the model on the validation set given a dimension q; q is the regularization parameter; Complexity(q) is the model complexity, representing the complexity of the model under a given dimension q.

[0053] norm parameter Selection strategy: The value range is limited to the interval [1,3], and is adaptively determined through grid search to match the actual fault mode distribution characteristics of the target mechanical system.

[0054] Secondly, this invention provides an intelligent diagnostic system for key equipment of rotating machinery, comprising: a data acquisition unit for acquiring state signals of the rotating machinery during operation and extracting features in the time domain, frequency domain, and time-frequency domain; a transformation unit for mapping feature differences using a monotonically bounded encapsulation function; and using the processing results of the monotonically bounded encapsulation function in each dimension to transform the unbounded distance in the original feature space into a bounded similarity metric space; an optimization unit for performing metric learning in the bounded similarity metric space, jointly optimizing the minimum inter-class margin and the maximum intra-class dispersion to obtain a discriminative embedding representation; generating an adaptive threshold based on the distance distribution obtained by the metric learning; and setting the adaptive threshold by minimizing the sum of the false positive rate and the false negative rate obtained by judging the distance distribution for fault detection; and an identification unit for calling a classifier within the bounded similarity metric space to achieve multi-fault identification and output the corresponding fault category.

[0055] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent diagnostic method for key equipment of rotating machinery as described in the first aspect of the present invention.

[0056] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent diagnostic method for key equipment of rotating machinery as described in the first aspect of the present invention.

[0057] The beneficial effects of this invention are as follows: This invention transforms unbounded metrics into bounded similarity by using bounded encapsulation functions, solving the problems of strong diffusion and difficulty in stably setting thresholds in traditional metric functions; by jointly optimizing inter-class margins and intra-class dispersion, it improves the discriminative power of the embedding space and enhances the model's generalization ability under small sample conditions; through a distribution-based adaptive threshold construction mechanism, it effectively reduces false alarm and false negative rates, achieving more robust fault detection; and through multiple fault identification modules and an adaptive hyperparameter selection mechanism, it improves the classification accuracy and applicability of the method under complex working conditions. In summary, the mechanical system fault detection and classification method based on physically constrained bounded metric learning proposed in this invention not only improves the diagnostic accuracy of mechanical systems under diverse working conditions but also possesses both physical rationality and engineering applicability, demonstrating high promotional value and application prospects. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0059] Figure 1 A flowchart for an intelligent diagnostic method for key equipment in rotating machinery.

[0060] Figure 2 A comparison chart showing the classification accuracy of different methods for each fault type.

[0061] Figure 3 This is the confusion matrix diagram of the present invention.

[0062] Figure 4 The figure shows the ablation experiment results of this invention. Detailed Implementation

[0063] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0064] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0065] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0066] Reference Figure 1 As one embodiment of the present invention, this embodiment provides an intelligent diagnostic method for key equipment of rotating machinery, comprising the following steps:

[0067] S1: Collect the state signals of rotating machinery during the operation of the mechanical system, and extract features in the time domain, frequency domain, and time-frequency domain.

[0068] Firstly, for the operating status of mechanical systems (such as rotating components like bearings, gears, and rotors), vibration acceleration signals are collected as the core diagnostic basis. These vibration signals cover various typical operating conditions, including but not limited to healthy conditions, outer race faults, inner race faults, rolling element faults, and combined faults. The signals are acquired by acceleration sensors deployed at key parts of the equipment and sampled under different speeds, loads, and environmental disturbances. The samples are then converted into discrete-time series signals via an analog-to-digital converter.

[0069]

[0070] in, This represents the vibration amplitude at the Lth sampling point, with a sampling frequency of... The total number of sampling points is L.

[0071] To facilitate modeling and feature extraction, the original vibration sequence was... According to fixed window length Divide the data by the sliding step size h to obtain the following results. A window segment:

[0072]

[0073] To eliminate the influence of differences in signal dimensions under different operating conditions, for each window Standardization process:

[0074]

[0075] in, and These represent the mean and standard deviation of the window, respectively. It is a stability constant used to avoid division by zero.

[0076] In standardized signal segments The time-domain statistical features that reflect the overall volatility and impact of the signal are extracted, including but not limited to: mean, variance, kurtosis, skewness and peak factor. These time-domain features can reflect the variation of the impact pulse under different fault modes.

[0077] The signal segment is further mapped to the frequency domain using a Fast Fourier Transform (FFT):

[0078]

[0079] in, Indicates the first A window segment in the frequency index The complex spectral values ​​are obtained through the Fast Fourier Transform (FFT) in the frequency domain. Their squared modulus is... This corresponds to an approximate value for the power spectral density; The original vibration signal is divided into multiple segments after sliding slicing, with each segment indexed by a window number. Indicates the range as ; For time indexing, in the 1st In each segment, Used to mark the location of sampling points within the segment, with a value range of 1 to ; Indicates the first The first segment The standardized amplitude of each sampling point. This amplitude is the result of the original vibration signal after mean removal and normalization processing, ensuring the comparability of different segments in terms of dimensions and amplitude; Window length, i.e., the number of sampling points contained in each slice. It determines the spectral resolution and temporal resolution, and is usually selected based on the characteristic frequencies of the mechanical system being monitored; For frequency index, the range is In the discrete frequency domain representation of FFT, Corresponding to the actual frequency:

[0080]

[0081] For complex exponential basis functions, we define orthogonal bases used to extract different frequency components in the Discrete Fourier Transform. Based on these, frequency domain features are extracted, including but not limited to: total spectral energy, spectral centroid, and spectral bandwidth. These frequency domain indicators can reflect the migration and abnormal peak characteristics of the spectral energy distribution of rotating machinery under fault conditions.

[0082] For vibration signals operating under non-stationary conditions, wavelet packet decomposition (WPD) is introduced to calculate time-frequency characteristics. Through wavelet packet decomposition, the signal is decomposed into different frequency bands, yielding sub-band coefficients: Let represent the subband coefficient of the m-th frequency band under the i-th window; and calculate the energy of each subband. The subband energy of the m-th frequency band under the i-th window is expressed as: Furthermore, indicators such as energy entropy, envelope spectrum peak value, and high-frequency energy ratio can be extracted to effectively capture the non-stationary characteristics of early, weak faults. Time-frequency features can maintain the characterization of fault energy accumulation under non-stationary operation, avoiding mismatch in the single frequency domain under speed drift. Finally, all features extracted from the time domain, frequency domain, and time-frequency domain are combined to form a feature vector. ,in , The total number of selected features, i.e., the feature dimension; Represents a d-dimensional space. The feature set consists of all window fragments:

[0083]

[0084] in, The labels correspond to different fault types. This step achieves a systematic mapping from the original vibration signal to the feature space, which not only ensures the distinguishability of features under different fault modes, but also provides high-quality input data for subsequent bounded metric learning and intelligent classification.

[0085] S2: Map the differences in features using a monotonically bounded encapsulation function; using the processing results of the monotonically bounded encapsulation function in each dimension, convert the unbounded distance in the original feature space into a bounded similarity metric space.

[0086] Based on the feature set obtained in S1, a nonlinear feature mapping operator is introduced. , used to convert the original 3D features extended to A high-dimensional feature space is used to enhance feature representation capabilities. For two different sample vectors... Its representation in the extended space is:

[0087]

[0088] in, For the first Nonlinear projection of each feature dimension. This indicates transpose.

[0089] This feature extension mechanism amplifies and highlights the complex nonlinear differences of mechanical systems under different operating conditions, thus facilitating subsequent bounded metric calculations. This type of feature extension not only enhances the separability between classes but also ensures that the subtle differences between different fault modes (such as outer ring faults, inner ring faults, rolling element faults, and compound faults) are more clearly distinguished in the new feature space, improving detection sensitivity and robustness.

[0090] To avoid the unboundedness and concentration effects of traditional Euclidean or Mahalanobis distances in high-dimensional spaces, a monotone bounded encapsulation function is introduced. Its domain is The range is ,satisfy:

[0091]

[0092] in, Let be the absolute difference between any two samples along a certain feature dimension, i.e. . Show encapsulated functions The maximum output value is the upper bound of the metric mapping. Its function is to restrict the original unbounded feature differences to a finite interval, avoiding excessive diffusion or unbounded growth. The determination method can be based on the physical limits of the mechanical system (such as vibration amplitude or energy upper limit), the statistical distribution of sample differences (such as high quantiles), or engineering experience. It should be noted that... It is not obtained by directly solving a formula, but rather as design parameters for constraint functions, determined through a combination of physical interpretation and data-driven approaches. Mathematically speaking, This means that no matter how large the feature differences z of the input are, the output will be constrained to Within a finite interval. From an engineering perspective, This reflects the "reasonable limit" in physics, which corresponds to the upper limit of the vibration amplitude, energy, or damage degree of a mechanical system, and cannot be increased indefinitely.

[0093] This function can compress and normalize arbitrary real-valued differences, ensuring that feature differences are mapped to bounded intervals. This avoids excessively large values ​​dominating the learning process and improves the comparability of features at different scales. Furthermore, by introducing upper bound constraints, this encapsulation function effectively suppresses the "distance concentration effect" common in high-dimensional spaces, preventing excessive convergence in distances between samples of different categories. Simultaneously, this design introduces physical interpretability, as energy, vibration amplitude, and damage accumulation in mechanical systems all have physical limits and cannot increase indefinitely.

[0094] Based on this, we define samples. Bounded distance metrics are as follows:

[0095]

[0096] in The absolute difference on the k-th feature dimension; This is an adjustable norm parameter used to control the aggregation method of feature differences across different dimensions; This is a normalization factor to ensure comparability of distances across different dimensions. This metric has the following technical advantages: First, it is a monotonic bounded function. The introduction of this ensures that the distance value is always within the range of 100%. First, it avoids the numerical explosion problem caused by uneven feature scales. Second, the encapsulation mechanism can amplify small but physically significant nonlinear differences, thereby enhancing the sensitivity to early and weak fault detection. Third, the upper bound of the distance reflects the inherent physical constraints of the mechanical system (such as energy limitations caused by stiffness and damping), making the metric reasonable and interpretable in engineering. Finally, the aggregation of normalization factors ensures the balanced contribution of features of different dimensions, improving the discriminability and robustness of the metric space.

[0097] Through the above processing, the unbounded distance in the original feature space is successfully mapped into a similarity metric space that is physically reasonable, numerically bounded, and more discriminative, thus providing a robust foundation for subsequent threshold setting and multi-fault classification.

[0098] S3: Perform metric learning in the bounded similarity metric space, jointly optimize the minimum inter-class margin and the maximum intra-class dispersion to obtain a discriminative embedding representation.

[0099] In the aforementioned monotonically bounded encapsulated function Metric learning is performed within the mapped feature space, employing a strategy of jointly optimizing the minimum inter-class margin and the maximum intra-class dispersion to obtain discriminative embedding representations. Specifically, this includes: First, let the healthy state and fault state samples be denoted as... and In the feature space, for pairs of health status samples Pair of fault state samples .

[0100] It's worth noting that during metric learning, positive sample pairs consist of healthy samples of the same class, ensuring that normal state features are compactly distributed within the embedding space; negative sample pairs consist of samples from different fault modes, and their embedding distance is significantly increased by optimizing this pair, thereby strengthening the separability between different fault categories. The starting point of this design is to address the inter-class confusion problem in multi-fault diagnosis scenarios, rather than simply distinguishing between healthy and fault states. Compared to traditional approaches that only consider health-fault separation, this method achieves stronger discriminative power and robustness in multi-fault identification tasks.

[0101] To construct samples from these samples, and secondly, to enhance the separability between fault categories, a minimum inter-class margin constraint is defined:

[0102]

[0103] in, The margin threshold in the feature space is used to adjust the separability between healthy and faulty sample pairs; this constraint ensures that the embedding distance between different state categories is as large as possible, thereby expanding the inter-class margin. This part is not directly "training the threshold," but rather constructing a bounded metric space and obtaining a discriminative embedding representation by optimizing the minimum inter-class margin and the maximum intra-class dispersion. The threshold setting actually occurs in subsequent S4 (adaptive threshold for fault detection), while the main role of this part is constraint optimization in the metric learning phase. By introducing a bounded metric distance formula, it is ensured that feature difference values ​​are mapped and compressed within a reasonable range. Combining the minimum margin constraint (inter-class separation) and the maximum dispersion control (intra-class compactness), a discriminative embedding representation is formed. The goal of this part is to improve the discriminativity and robustness of the feature representation, rather than directly setting a detection threshold. Relationship with "threshold": At this stage, a "margin constraint parameter" is indeed introduced. This is used to specify the minimum distance between healthy and faulty samples, thereby enhancing classification separability. However, this... This is not the final threshold used for detection; it only serves as a constraint during the metric learning process. The actual detection threshold... It is adaptively determined in subsequent steps (based on the distance distribution between samples, minimizing the false alarm rate and the false negative rate).

[0104] Furthermore, a regularization term for intra-class dispersion is introduced to control the representation compactness among samples of the same class. For healthy sample pairs... The goal is to minimize its embedding distance. For faulty sample pairs Then, by adjusting the intermediate parameters The goal is to control the distribution range of these differences, confining intra-class variations to a physically interpretable bounded interval. The overall optimization objective can be expressed as:

[0105]

[0106] The first item guarantees that the intra-class distance does not exceed the parameter. The second term ensures that the distance between classes is not less than the parameter. This allows for the combined optimization of intra-class compactness and inter-class separation. This means performing an optimization search within the set of all possible feature mapping functions, i.e., by selecting a suitable mapping function. The goal is to minimize the objective function value. In other words, this symbol defines the scope and objective of the optimization process, ensuring that the final embedding representation can not only constrain the excessive dispersion of samples within a class, but also guarantee the discriminability between samples between classes, thereby achieving the construction of a compact embedding space within classes and a separate embedding space between classes.

[0107] Through the above optimizations, the method obtains a discriminative embedding representation, which has the following advantages: enhanced inter-class separability: the interval between feature representations of different fault modes in the embedding space is significantly increased, thereby improving the model's ability to distinguish between multiple fault categories; controlled intra-class scatter: by introducing parameters... This effectively avoids excessive dispersion of samples of the same category, ensuring the compactness and robustness of intra-class features; physical interpretability: margin threshold. With intra-class constraint parameters Consistent with the vibration response characteristics of mechanical systems, it can reflect the separability boundary of rotating machinery under the constraints of structural stiffness and damping characteristics; improved robustness: the optimization mechanism can still maintain discriminative representation under noise interference and signal loss conditions, enhancing the adaptability of the method under complex working conditions.

[0108] S4: Based on the distance distribution obtained by the metric learning, an adaptive threshold is generated; the adaptive threshold is set by minimizing the sum of the false alarm rate and the false negative rate obtained by the distance distribution judgment, and is used for fault detection.

[0109] Based on the distance distribution between samples obtained through the aforementioned bounded metric learning, an adaptive threshold for fault detection is constructed. Specifically, firstly, within the feature space, for healthy state sample pairs... Pair of fault state samples Statistical analysis was performed on the distance distribution to obtain the distribution of the health category. (Representing the distance distribution of healthy state sample pairs) and the distribution of fault classes (This represents the distance distribution of fault state sample pairs). Based on this, a detection threshold is introduced. As an adjustable parameter, this threshold is adaptively set by jointly minimizing the sum of the false positive rate and the false negative rate. The optimization objective is defined as:

[0110]

[0111] Where FPR(τ) represents the false alarm rate when the detection threshold is τ, i.e., the probability that a healthy sample is misclassified as faulty; FNR(τ) represents the false negative rate at that threshold, i.e., the probability that a faulty sample is misclassified as healthy. This represents the optimal detection threshold obtained by minimizing the sum of the false positive probability of healthy samples and the false negative probability of faulty samples. By jointly minimizing the weighted sum of these two types of error probabilities, it is possible to ensure that the detection threshold can adaptively balance the sensitivity and accuracy of fault detection under different operating conditions, thereby avoiding the bias caused by simply pursuing low false positives or low false negatives.

[0112] Furthermore, the determination of the threshold τ depends on the aforementioned bounded metric distance function. Specifically, for any pair of test samples (u, v), if If the sample pair belongs to different categories, then the healthy sample and the faulty sample can be distinguished; if If the condition is met, the sample pair is determined to belong to the same category, either the healthy category or the same type of fault. Guided by this rule, the system can achieve dynamic threshold discrimination based on the learned embedding representation, taking into account the adaptability of distribution differences between different categories, thereby improving the reliability and robustness of fault detection.

[0113] It should be noted that in the previous steps, we only obtained an optimal threshold. And this judgment rule, It ensures that it can be truly applied to fault detection; it plays a key role in the implementation and judgment of this solution, making fault detection operable and complete.

[0114] Therefore, step four not only provides a threshold adaptation mechanism under the metric learning framework, but also effectively improves the robustness and universality of the detection model under complex conditions through joint optimization of false positive rate and false negative rate, so that the method can simultaneously meet the requirements of sensitivity, accuracy and adaptability.

[0115] S5: Within the bounded similarity metric space, call the classifier to identify multiple faults and output the corresponding fault categories.

[0116] Within the bounded metric space, multiple fault identification is achieved based on the aforementioned metric-learned optimized embedding representation, combined with a classifier. Specifically, the sample to be tested is input into the trained bounded metric embedding space, and the distance relationship between it and samples of each known class is calculated. Combined with classification decision rules, the corresponding fault category is output. The classifier may include a nearest neighbor classifier or a support vector machine classifier.

[0117] When using a nearest neighbor classifier, the fault category to which the test sample belongs is determined based on the nearest neighbor distance relationship between the test sample and each category of samples in the training set in the embedding space.

[0118] When a support vector machine classifier is used, different fault category regions are divided in the embedding space using hyperplanes. By determining the relative position of the test sample with each hyperplane, the fault category to which it belongs is determined.

[0119] By integrating the above classifiers within a bounded metric space, it is possible to accurately identify multiple fault types while maintaining clear and separable inter-class distributions and compact clustering of intra-class samples, effectively improving the robustness and reliability of fault classification under complex operating conditions.

[0120] The reason for adopting a design combining threshold detection and a multi-class classifier is that this scheme needs to achieve an organic connection between health / fault detection and multi-fault classification. Relying solely on thresholds, the system can only distinguish between health and faults, which is a binary classification task. However, in real-world mechanical systems, there are often multiple types of faults. Relying solely on binary classification will not meet the needs of engineering diagnosis. Therefore, after threshold detection, a multi-class classifier is introduced to achieve finer-grained multi-fault differentiation. Based on this, the aforementioned bounded metric learning has already optimized the maximization of inter-class margin and minimization of intra-class dispersion in the feature embedding space, making the embedding representation inherently possess the properties of clear distribution of different fault categories and tight clustering of similar samples. Combining the classifier within this discriminatively enhanced space can significantly improve classification performance. Furthermore, this embodiment has a reasonable choice of classifier: the nearest neighbor classifier (k-NN) discriminates based on the nearest neighbor relationship between samples, which is simple, efficient, and suitable for clustered fault modes; the support vector machine (SVM) divides categories through a hyperplane, suitable for handling scenarios with complex category boundaries. This design ensures that different types of fault modes can be accurately identified. Overall, the design ensures a complete closed loop from health / fault detection to multi-fault identification, which improves the practicality and scalability of the method, and enhances the classification robustness and reliability under complex operating conditions.

[0121] Furthermore, a strategy for adaptively selecting hyperparameters involved in the feature space is provided, wherein the hyperparameters include encapsulation functions. Feature space dimension q and norm parameter Specifically, it includes:

[0122] The selection strategy must satisfy the following properties: Monotonicity: When the input value increases, the output value does not decrease, so as to ensure that the order relation in the mapping space is maintained; Boundedness: When the input approaches infinity, the output value is limited by a finite constant. This ensures that the feature distribution is within an interpretable and finite spatial range; differentiability and smoothness: continuous differentiability within the domain, facilitating subsequent optimization and gradient calculation. Therefore, The specific form is selected from a class of monotone bounded functions, preferably including but not limited to the Sigmoid function, the hyperbolic tangent function (Tanh), and the bounded linear correction function (Bounded ReLU). boundary constants The setting corresponds to the physical constraints of vibration amplitude in the actual mechanical system. Its value can be adaptively determined by statistical quantiles (e.g., the 95th or 99th quantile), thereby ensuring that the feature mapping result is consistent with the actual mechanical vibration response.

[0123] The selection strategy for the feature space dimension q: The feature space dimension q is adaptively determined based on the complexity of the fault diagnosis task. The adaptive selection includes minimizing the weighted sum of classification error and model complexity on the validation set.

[0124]

[0125] in, The optimal feature space dimension; Given a dimension q, the model's classification error rate or loss value on the validation set is used to measure the accuracy of the diagnostic results. `q` is a regularization parameter used to adjust the trade-off between classification error and model complexity; `Complexity(q)` is the model complexity, representing the complexity of the model under a given dimension `q`, which can be measured by the number of parameters, computational cost, or structural depth. The goal of this formula is to search among candidate dimensions `q` to minimize the sum of the validation error and the complexity penalty term. This not only ensures the accuracy of fault classification (reducing...) This avoids making the model overly complex (by limiting Complexity(q)), thus achieving a dimension selection result that is optimal in performance and has good generalization ability. .

[0126] norm parameter Selection strategy: The distance metric in the feature space is based on - Norm, whose value is not limited to a single number, but is selected within a given range; The value range is limited to the interval [1, 3] to ensure that the metric method can cover Manhattan distance ( Euclidean distance It can also adapt to cases with higher-order norms, thereby enhancing its adaptability to different fault types; The specific values ​​are adaptively determined through grid search or other parameter optimization methods to match the actual fault mode distribution characteristics of the target mechanical system. Through the aforementioned hyperparameter selection strategy, the method in this embodiment can maintain the rigor of the mathematical expression while possessing flexibility and adaptability, ensuring that the fault diagnosis model has good generalizability and robustness under multiple operating conditions, multiple devices, and uncertain environments.

[0127] This embodiment also provides an intelligent diagnostic system for key equipment of rotating machinery, including:

[0128] The acquisition unit collects the state signals of rotating machinery during the operation of the mechanical system and performs feature extraction in the time domain, frequency domain, and time-frequency domain.

[0129] The transformation unit maps the differences in features using a monotonically bounded encapsulation function; and uses the processing results of the monotonically bounded encapsulation function in each dimension to transform the unbounded distance in the original feature space into a bounded similarity metric space.

[0130] The optimization unit performs metric learning in the bounded similarity metric space, jointly optimizing the minimum inter-class margin and the maximum intra-class dispersion to obtain a discriminative embedding representation; based on the distance distribution obtained by the metric learning, it generates an adaptive threshold; the adaptive threshold is set by minimizing the sum of the false positive rate and the false negative rate obtained by the distance distribution, and is used for fault detection.

[0131] The identification unit calls a classifier within the bounded similarity metric space to identify multiple faults and outputs the corresponding fault categories.

[0132] Reference Figures 2-4 As one embodiment of the present invention, this embodiment provides an intelligent diagnostic method for key equipment of rotating machinery. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0133] The proposed method was first validated using the Case Western Reserve University (CWRU) Bearing Dataset. This dataset, provided by the CWRU Bearing Data Center, is widely considered an authoritative benchmark in rolling bearing fault diagnosis research. The experimental platform included a 2 hp motor, a torque sensor / encoder, and a dynamometer. Different types of faults, including inner ring faults (IR), outer ring faults (OR), and rolling element faults (B), were introduced into the drive-end and fan-end bearings using electro-discharge machining (EDM). Fault diameters ranged from 0.007 inches, 0.014 inches, and 0.021 inches. Vibration signals were acquired under different loads (0–3 hp) and speeds (1720–1797 rpm) at a sampling frequency of 12 kHz.

[0134] To ensure data consistency and comparability, the original vibration signal was divided into 1024 sample segments, and a 50% overlap sampling method was used to augment the data volume. Subsequently, each sample was normalized to zero mean and unit variance to eliminate the influence of dimensional differences and environmental noise.

[0135] All node data was standardized before being input into the model to improve the convergence speed and stability of model training. The dataset was divided into training, validation, and test sets in a ratio of 0.7:0.15:0.15, used for model parameter training, hyperparameter optimization, and final performance evaluation, respectively. The model was trained using the Adam optimizer with an initial learning rate of 0.001, a maximum number of iterations of 200, a batch size of 64, and an early stopping strategy to prevent overfitting.

[0136] Model performance comparison and accuracy verification:

[0137] To comprehensively evaluate the effectiveness and superiority of the proposed method, nine representative state-of-the-art models in the current fault diagnosis and related fields were selected as comparative methods. These models cover graph neural network methods (ACVG), Transformer architecture methods (ST-FD, MFFN, SCLDAN, Swin Transformer), convolutional neural network methods (ConvNext V2, MobileViT V2), and self-supervised learning methods (MAE, DinoV2), demonstrating the latest progress and advantages of existing research in different dimensions.

[0138] Specifically, the comparative methods include specialized fault diagnosis models, which effectively capture the dynamic characteristics, multi-scale features, and cross-domain adaptability of mechanical systems, thereby achieving enhanced fault differentiation capabilities. Simultaneously, to ensure the universality and challenge of the evaluation, several general-purpose deep learning models are introduced, including representative Transformer architectures and convolutional-Transformer hybrid architectures. These methods perform well in visual and temporal tasks and demonstrate potential for modeling vibration signals. In addition, this study incorporates two types of self-supervised learning models, whose core advantage lies in their ability to learn robust feature representations even in the absence of large amounts of labeled data, which is of practical significance for addressing the data scarcity problem in industrial scenarios. To ensure fairness and comparability, all comparative methods are implemented and trained based on a unified experimental protocol. Specifically, this study uses the same data partitioning and preprocessing procedures, and all models run in the same GPU environment; hyperparameter tuning is performed according to the original paper or commonly used configurations, combined with grid search on the validation set to obtain optimal or near-optimal settings. This rigorous experimental design ensures that performance differences primarily reflect the model structure and methodology itself, rather than differences in experimental conditions. This multi-level, cross-paradigm comparison system not only ensures the comprehensiveness and objectivity of the comparative experiments, but also provides a solid benchmark for verifying the robustness and generalization ability of the proposed framework under complex working conditions.

[0139] As shown in Table 1, the first column "Method" indicates the comparison algorithm used, including traditional methods (such as ACVG, ST-FD, MFFN, SCLDAN, etc.), general deep learning models (such as Swin Transformer, ConvNext V2, MobileViT V2, etc.), and self-supervised learning methods (such as MAE, Dinov2). The method (ME) proposed in this invention is also listed. The second column "Accuracy (%)" represents the classification accuracy of each method on the same dataset. The value is given as a percentage, followed by "±", which represents the standard deviation of the results in different experimental runs, reflecting the stability and robustness of the model performance.

[0140] Table 1. Test results of the benchmark algorithm and the proposed method on the dataset.

[0141] Method Accuracy ACVG 96.8 ± 0.3 ST-FD 97.2 ± 0.2 MFFN 97.5 ± 0.2 SCLDAN 97.9 ± 0.1 Swin Transformer 96.5 ± 0.3 ConvNext V2 97.1 ± 0.2 MobileViT V2 96.3 ± 0.3 MAE 96.7 ± 0.2 Dinov2 97.0 ± 0.2 ME 98.4 ± 0.1

[0142] The fault diagnosis method proposed in this invention has been systematically experimentally verified on public datasets. The experimental results are shown in Table 1. The method of this invention demonstrates superior classification accuracy and stability compared to existing technologies in multi-class fault identification tasks. Under the same experimental conditions, the method of this invention achieves an overall classification accuracy of 98.4%, significantly better than the accuracy range of the comparative methods (96.3%–97.9%). Simultaneously, the standard deviation of the method of this invention is only ±0.1, lower than most existing methods, indicating that it maintains high stability under multiple cross-validation and different data distributions. Therefore, this invention has the following beneficial effects: Higher discriminative power: Through optimized feature space construction, this invention can effectively distinguish different types and degrees of faults, avoiding misjudgments caused by blurred boundaries between categories. Stronger robustness: Under complex conditions such as low signal-to-noise ratio and overlapping distributions, this invention can still maintain high recognition accuracy, demonstrating excellent anti-interference ability. Better generalization performance: This invention performs stably across validation folds, indicating that it has strong generalizability and practicality under different equipment operating conditions or data sampling conditions.

[0143] In summary, the experimental results fully verify the significant performance advantages and application value of the method of the present invention compared with the prior art in the field of fault diagnosis, and can provide more reliable early fault identification and health monitoring support for actual industrial systems.

[0144] Comparison of classification accuracy for different fault types:

[0145] like Figure 2As shown, the horizontal axis represents different fault diagnosis methods for comparison, including ACVG, ST-FD, MFFN, SCLDAN, Swin Transformer, ConvNeXt V2, MobileViT V2, MAE, Dinov2, and the ME method proposed in this invention; the vertical axis represents the classification accuracy of each method under different fault categories (from top to bottom: normal, inner ring fault, outer ring fault, and rolling element fault). The method of this invention demonstrates significant advantages in classification tasks for different fault types (normal, inner ring fault, outer ring fault, and rolling element fault). Especially in the identification of rolling element faults, the accuracy of the method of this invention reaches 97.8%, significantly better than the comparison methods (such as SCLDAN's 96.9%). Since rolling element faults typically exhibit more complex and subtle vibration characteristics, the method of this invention can effectively improve the separability of fault features by constructing a more discriminative feature space, thereby achieving accurate identification of complex fault signals.

[0146] Furthermore, such as Figure 3 As shown, the horizontal axis represents the predicted category, and the vertical axis represents the actual category, including normal, inner race faults of different severity (IR007, IR014, IR021), rolling element faults of different severity (B007, B014, B021), and outer race faults of different severity (OR007, OR014, OR021). The values ​​in the matrix represent the recognition accuracy or confusion ratio corresponding to the actual category and the predicted category, where the diagonal elements represent the accuracy of correct classification, and the off-diagonal elements represent the misclassification ratio between different categories.

[0147] Figure 3 The confusion matrix shown verifies the stability and robustness of the proposed method in multi-class fault identification. Overall, the method achieves an accuracy exceeding 97% across all categories, demonstrating excellent classification consistency. Specifically, for similar faults of varying severity (e.g., IR007, IR014, and IR021), the method achieves accuracies of 98.6%, 98.4%, and 98.9%, respectively, with virtually no significant confusion. This indicates that the proposed method can not only effectively distinguish between different types of faults but also accurately identify different degrees of damage within the same type, significantly improving the precision of fault diagnosis.

[0148] Furthermore, the method of this invention also performs exceptionally well in identifying faults under normal operating conditions, achieving an accuracy rate of 99.5%, effectively avoiding misdiagnosis as faults and thus reducing the risk of unnecessary maintenance intervention. The few instances of misdiagnosis mainly occur between similar faults of adjacent severity (e.g., a slight confusion of 0.8% between IR007 and IR014), which is acceptable in practical engineering applications and does not affect the overall diagnostic conclusion. These results fully demonstrate the universality and engineering application value of the method of this invention in fault diagnosis of complex mechanical systems.

[0149] Ablation experiment:

[0150] like Figure 4 As shown, subplot (a) is a comparison of classification accuracy for different fault types. The horizontal axis represents the fault type (including Normal, Inner, Outer, and Ball), and the vertical axis represents the classification accuracy (in %). The bars correspond to the results under the two methods: "Without Encapsulation" and "With Encapsulation". Subplots (b) and (c) are classification confusion matrices, where the horizontal axis represents the predicted category and the vertical axis represents the actual category. Categories include Normal, Inner, Outer, and Ball. Each value in the matrix represents the degree of matching between the predicted result and the actual category. The diagonal elements represent the accuracy of correct classification, and the off-diagonal elements represent the confusion ratio between different categories. (b) shows the result without the bounded encapsulation mechanism, and (c) shows the result with the bounded encapsulation mechanism.

[0151] To fully validate the effectiveness of the proposed method, we conducted a series of ablation experiments on the CWRU bearing dataset. First, we compared the ME method with its variant that does not include a metric encapsulation mechanism, as shown in Fig. 4(a).

[0152] As shown in Fig. 4(a), the classification accuracy of almost all fault types was improved after the introduction of the metric encapsulation mechanism. The improvement was particularly significant for ball faults, with the accuracy increasing from 94.3% to 97.8%, an improvement of 3.5%. Since ball faults are often accompanied by more complex and subtle vibration modes, this improvement fully demonstrates the advantages of the metric encapsulation mechanism in handling complex features.

[0153] Further comparison of the confusion matrices (Fig. 4(b) and Fig. 4(c)) demonstrates the differences between the two models in fault type differentiation. It is evident that the addition of the metric encapsulation mechanism significantly reduces misclassification, particularly in its ability to distinguish between similar faults. For example, the misclassification rate between inner circle faults and spherical faults decreased from 2.7% to 0.9%, showcasing the significant role of the metric encapsulation mechanism in enhancing discriminative ability. This result indicates that the proposed method not only improves overall classification accuracy but also effectively reduces confusion between similar faults, providing solid support for achieving higher-precision mechanical fault diagnosis.

[0154] The impact of boundary values ​​on model performance:

[0155] To further verify the effectiveness of the proposed metric encapsulation mechanism, model performance was tested under different boundary values. The experimental results are shown in Table 2. The first column, “Boundary Value,” indicates the boundary value setting of the metric encapsulation mechanism, where “No encapsulation” indicates no boundary constraint is used; the other values ​​(0.5, 1.0, 1.5, 2.0) correspond to different boundary value sizes. The second column, “Accuracy (%),” represents the model's classification accuracy under this boundary value setting, given as a percentage. The third column, “Convergence Epoch,” represents the number of iterations required for the model to complete training under this boundary value condition, used to measure the training convergence speed. The fourth column, “Training Time (s),” represents the time required for the model to complete training under this boundary value, in seconds, used to reflect the overall training efficiency.

[0156] Table 2. The impact of different boundary values ​​on model performance

[0157] Boundary Value Accuracy (%) Convergence Epoch Training Time (s) No encapsulation 95.7 87 176 0.5 97.2 72 158 1.0 98.4 65 143 1.5 97.9 68 149 2.0 97.1 75 162

[0158] It can be seen that the selection of the boundary value has a significant impact on both the diagnostic accuracy and training efficiency of the model. When the boundary value is set to 1.0, the model achieves optimal performance, with a classification accuracy of 98.4% and the fastest convergence speed, requiring only 65 iterations to complete training. Compared to the case without the metric encapsulation mechanism (87 iterations), the convergence efficiency is improved by 25.3%. Further analysis shows that when the boundary value is too small (e.g., 0.5), the feature space is excessively compressed, limiting the model's expressive power and leading to a decrease in classification accuracy. When the boundary value is too large (e.g., 2.0), the feature space constraint is insufficient, making it difficult for the model to effectively distinguish similar fault types, resulting in a decrease in recognition performance. When the boundary value is set to 1.0, the optimal balance is achieved between constraining the feature space and maintaining the ability to capture complex features, which improves the discriminative ability and ensures the model's generalization performance.

[0159] Therefore, the metric encapsulation mechanism proposed in this invention can effectively improve the accuracy and training efficiency of mechanical fault diagnosis models by reasonably setting boundary values. The optimal boundary value (1.0) not only improves the overall recognition performance but also shortens the model's convergence time, providing a feasible solution for frequent model updates in industrial scenarios. This invention thus achieves: Solving the technical problem: addressing the insufficient fault feature discrimination capability or low training efficiency of existing methods; Adopting the technical means: introducing a metric encapsulation mechanism and setting reasonable boundary values ​​to constrain the feature space; Obtaining the technical effect: improving training speed and diagnostic accuracy while ensuring model discrimination capability, significantly outperforming the control method without the encapsulation mechanism.

[0160] This embodiment also provides a computer device applicable to the intelligent diagnostic method for key equipment of rotating machinery, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent diagnostic method for key equipment of rotating machinery as proposed in the above embodiment.

[0161] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0162] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent diagnostic method for key equipment of rotating machinery as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0163] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent diagnosis of key equipment in rotating machinery, characterized in that: This includes collecting the state signals of rotating machinery during the operation of the mechanical system and extracting features in the time domain, frequency domain, and time-frequency domain; The differences in features are mapped using a monotonically bounded encapsulation function; Using the processing results of the monotonically bounded encapsulation function in each dimension, the unbounded distance in the original feature space is transformed into a bounded similarity metric space. Metric learning is performed in the bounded similarity metric space to jointly optimize the minimum inter-class margin and the maximum intra-class dispersion to obtain a discriminative embedding representation; Based on the distance distribution obtained through the metric learning, an adaptive threshold is generated; The adaptive threshold is set by minimizing the sum of the false alarm rate and the false negative rate obtained by distance distribution determination, and is used for fault detection; Within the bounded similarity metric space, a classifier is invoked to identify multiple faults and output the corresponding fault categories. The monotonically bounded encapsulation function includes, based on the feature set, using a nonlinear feature mapping operator to extend the original d-dimensional features to a q-dimensional high-dimensional feature space, amplifying the complex nonlinear differences of the mechanical system under different working conditions, thereby enhancing the feature representation capability; in the high-dimensional space, arbitrary real-valued differences are compressed and normalized to ensure that feature differences are mapped to a bounded interval.

2. The intelligent diagnostic method for key equipment of rotating machinery as described in claim 1, characterized in that: The status signal uses vibration acceleration signal as the core diagnostic basis; it is acquired by acceleration sensors deployed in key parts of the equipment and sampled under different rotational speeds, loads and environmental interference conditions; Vibration signals cover a variety of typical operating conditions, including but not limited to healthy conditions, outer ring faults, inner ring faults, rolling element faults, and combined faults. The state signal is converted into a discrete time series by an analog-to-digital converter module: ; in, This represents the vibration amplitude at the Lth sampling point, with a sampling frequency of... The total number of sampling points is L.

3. The intelligent diagnostic method for key equipment of rotating machinery as described in claim 2, characterized in that: The original vibration sequence According to fixed window length Divide the data by the sliding step size h to obtain the following results. A window segment, This represents a segment of the i-th window; After standardizing each window, in the standardized signal segment The above uses Fast Fourier Transform to map signal segments to the frequency domain; Time-frequency feature calculation is performed using wavelet packet decomposition: the signal is decomposed into different frequency bands, and sub-band coefficients are obtained. This represents the subband coefficient of the m-th frequency band within the i-th window; and utilizes... Calculate the energy of each sub-band; under the i-th window, the sub-band energy of the m-th frequency band is expressed as: Simultaneously, in each subband, indicators such as energy entropy, envelope spectrum peak value, and high-frequency energy ratio are extracted. For the i-th window, all feature combinations extracted in the time domain, frequency domain, and time-frequency domain are used to form a feature vector. The feature set consists of all window fragments: ; in, The labels correspond to different fault types.

4. The intelligent diagnostic method for key equipment of rotating machinery as described in claim 3, characterized in that: The monotonically bounded encapsulation function The domain is The range is , satisfy: ； in, Show encapsulated functions The maximum output value is the upper bound of the metric mapping, which can restrict the original unbounded feature differences to a finite interval; Let the absolute difference between any two samples be along a certain feature dimension. ; This represents the feature obtained by mapping sample u along the k-th dimension; This represents the feature obtained by mapping sample v along the k-th dimension; Define any two samples Bounded distance measurement: ； in, This represents the adjustable norm parameter, used to control the aggregation method of feature differences across different dimensions; This is a normalization factor to ensure that distances are comparable across different dimensions.

5. The intelligent diagnostic method for key equipment of rotating machinery as described in claim 4, characterized in that: The minimum inter-class interval includes, denoted as follows: Let the sets of healthy state and fault state samples be respectively denoted as... and Construct sample pairs from the set: samples of any two health states. Pair of fault state samples ; To enhance the separability between fault categories, a minimum interval constraint between categories is defined: ； in, The margin threshold in the feature space is used to adjust the separability between healthy sample pairs and faulty sample pairs; the constraint ensures that the embedding distance between different state categories is as large as possible, thereby expanding the inter-class margin. The maximum intra-class dispersion includes, for healthy sample pairs The goal is to minimize the embedding distance. For faulty sample pairs Then, by adjusting the intermediate parameters This controls the distribution range, limiting intraclass differences to a physically interpretable bounded interval; The comprehensive optimization objective can be expressed as: ； The first item guarantees that the intra-class distance does not exceed the parameter. The second term ensures that the distance between classes is not less than the parameter. This allows for the combined optimization of intra-class compactness and inter-class separation. This indicates that an optimization search is performed within the set of all possible feature mapping functions.

6. The intelligent diagnostic method for key equipment of rotating machinery as described in claim 5, characterized in that: The adaptive threshold includes, within the feature space, the health status sample pairs Pair of fault state samples Statistical analysis was performed on the distance distribution to obtain the distribution of the health category. Distribution of fault classes ; Introducing detection thresholds As an adjustable parameter, it is adaptively set by jointly minimizing the sum of the false alarm rate and the false negative rate. The optimization objective is defined as: ; in, This indicates that when the detection threshold is The false alarm rate is the probability that a healthy sample is mistakenly identified as faulty. This represents the false negative rate at that threshold, which is the probability that a faulty sample is mistakenly identified as healthy. This represents the optimal detection threshold obtained by minimizing the sum of the false positive probability of healthy samples and the false negative probability of faulty samples.

7. The intelligent diagnostic method for key equipment of rotating machinery as described in claim 6, characterized in that: The detection threshold includes, for any pair of test samples (u, v), if If the sample pairs belong to different categories, it means that healthy samples and faulty samples can be distinguished; if If the sample pair is determined to belong to the same category, it indicates either the healthy category or the same type of fault category.

8. The intelligent diagnostic method for key equipment of rotating machinery as described in claim 7, characterized in that: The classifier includes inputting the sample to be tested into a pre-trained bounded metric embedding space, calculating the distance relationship between the sample to be tested and each known category sample, and combining the classification decision rules to output the corresponding fault category. The classifier can be either a nearest neighbor classifier or a support vector machine classifier. When using a nearest neighbor classifier, the fault category to which the test sample belongs is determined based on the nearest neighbor distance relationship between the test sample and each category of samples in the training set in the embedding space. When a support vector machine classifier is used, different fault category regions are divided in the embedding space using hyperplanes. By determining the relative position of the test sample to each hyperplane, the fault category to which the test sample belongs can be determined.

9. The intelligent diagnostic method for key equipment of rotating machinery as described in claim 8, characterized in that: In the metric learning, a strategy is employed to adaptively select hyperparameters involved in the feature space; The hyperparameters include wrapper functions. Feature space dimension q and norm parameter Specifically: The selection strategy must satisfy the following properties: monotonicity, meaning the output value does not decrease as the input value increases; and boundedness, meaning the output value is limited by a finite constant as the input approaches infinity. Differentiability and smoothness; continuous differentiability within the domain. The selection strategy for the feature space dimension q: The feature space dimension q is adaptively determined based on the complexity of the fault diagnosis task; it is adaptively set by minimizing the weighted sum of classification error and model complexity on the validation set. ； in, The optimal feature space dimension; Let q be the classification error rate of the model on the validation set given a dimension q; q is the regularization parameter; Complexity(q) is the model complexity, representing the complexity of the model under a given dimension q. norm parameter Selection strategy: The value range is limited to the interval [1,3], and is adaptively determined through grid search to match the actual fault mode distribution characteristics of the target mechanical system.

10. A critical equipment intelligent diagnostic system for rotating machinery, based on the critical equipment intelligent diagnostic method for rotating machinery according to any one of claims 1 to 9, characterized in that: This includes a data acquisition unit that collects the state signals of rotating machinery during the operation of the mechanical system and performs feature extraction in the time domain, frequency domain, and time-frequency domain. The transformation unit maps the differences in features through a monotonically bounded encapsulation function; Using the processing results of the monotonically bounded encapsulation function in each dimension, the unbounded distance in the original feature space is transformed into a bounded similarity metric space. The optimization unit performs metric learning in the bounded similarity metric space, jointly optimizing the minimum inter-class margin and the maximum intra-class dispersion to obtain a discriminative embedding representation; Based on the distance distribution obtained through the metric learning, an adaptive threshold is generated; The adaptive threshold is set by minimizing the sum of the false alarm rate and the false negative rate obtained by distance distribution judgment, and is used for fault detection; the identification unit calls the classifier to realize multiple fault identification within the bounded similarity metric space and outputs the corresponding fault category.

Images