A method and system for identifying faults of industrial rotating machinery by pre-training a time series model

Abstract

This invention provides a method and system for fault identification of industrial rotating machinery using a pre-trained time series model, relating to the field of industrial fault diagnosis technology. The method includes: acquiring raw data of the industrial rotating machinery; preprocessing the raw data and constructing a pre-training dataset based on the preprocessed raw data; constructing a pre-trained time series model; inputting the pre-training dataset into the pre-trained time series model to train the model; acquiring real-time data of the industrial rotating machinery to be detected; preprocessing the real-time data; inputting the preprocessed real-time data into the trained pre-trained time series model to obtain a time series segment to be detected; calculating the time-frequency domain feature loss function between the time series segment to be detected and the predicted time series segment to obtain a time-frequency domain residual value; determining whether the time-frequency domain residual value is greater than an anomaly judgment threshold; if so, determining that the state of the industrial rotating machinery to be detected is faulty; otherwise, determining that the state of the industrial rotating machinery to be detected is normal.

Description

Technical Field

[0001] This invention relates to the field of industrial fault diagnosis technology, and in particular to a method and system for identifying faults in industrial rotating machinery using a pre-trained time series model. Background Technology

[0002] Industrial rotating machinery (such as rolling bearings and gearboxes) are core moving components in industries such as rail transportation and intelligent manufacturing, and their operating status directly determines the safety and reliability of electromechanical systems. Early, hidden faults (such as minor pitting and micro-cracks), if not identified in time, can easily lead to unplanned downtime, huge economic losses, or even major safety accidents. International standards have already imposed mandatory requirements on online vibration monitoring of rotating machinery, and various industries have an increasingly urgent need for highly sensitive and robust fault identification technologies.

[0003] Current fault identification technologies for industrial rotating machinery mainly fall into four categories: First, traditional methods, such as envelope spectrum, wavelet packet energy analysis, and SVM, which rely on manually designed features. Second, deep learning methods, such as CNN and RNN, which can automatically extract features. Third, Transformer-like models, which possess powerful long-sequence modeling capabilities. Fourth, generative time series pre-trained models, which have verified the potential of unsupervised pre-training in time series modeling.

[0004] However, existing technologies have several bottlenecks: early fault samples are extremely scarce, and industrial field conditions are not stable with significant equipment differences, leading to deviations in vibration signal distribution and weak model transferability across operating conditions and equipment. Furthermore, traditional methods rely on manual features, resulting in poor stability and generalization ability. Summary of the Invention

[0005] To address the challenges of extremely scarce early-stage fault samples, non-stationary industrial operating conditions, and significant equipment variations leading to vibration signal distribution shifts and weak model transferability across operating conditions and equipment, as well as the technical issues of traditional methods relying on manual features and exhibiting poor stability and generalization ability, this invention provides a method and system for fault identification of industrial rotating machinery using a pre-trained time-series model.

[0006] The technical solutions provided by the embodiments of the present invention are as follows: The first aspect of this invention provides a method for fault identification of industrial rotating machinery using a pre-trained time series model, comprising: S1: Obtain raw data from industrial rotating machinery.

[0007] S2: Preprocess the original data and construct a pre-training dataset based on the preprocessed original data. The pre-training dataset contains multiple sub-sequence fragments.

[0008] S3: Construct a pre-trained time series model, which includes a fragmented encoder, a decoding Transformer, and a time series prediction output module connected in sequence.

[0009] S4: Input the pre-trained dataset into the pre-trained time series model and train the pre-trained time series model.

[0010] S5: Obtain real-time data of the industrial rotating machinery to be tested.

[0011] S6: Preprocess the real-time data.

[0012] S7: Input the preprocessed real-time data into the pre-trained time series model to obtain the time series segment to be detected.

[0013] S8: Calculate the time-frequency domain residual value based on the time-frequency domain feature loss function in the pre-trained time-series model.

[0014] S9: Determine whether the time-frequency domain residual value is greater than the anomaly detection threshold. If so, classify the time sequence segment to be tested corresponding to the time-frequency domain residual value as an abnormal segment, and determine that the state of the industrial rotating machinery to be tested is faulty. Otherwise, classify the time sequence segment to be tested corresponding to the time-frequency domain residual value as a normal segment, and determine that the state of the industrial rotating machinery to be tested is normal.

[0015] A second aspect of the present invention provides a fault identification system for industrial rotating machinery based on a pre-trained time series model, comprising: processor.

[0016] The memory stores computer-readable instructions, which, when executed by a processor, implement an industrial rotating machinery fault identification method based on a pre-trained timing model as described in the first aspect.

[0017] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an industrial rotating machinery fault identification method with a pre-trained timing model as described in the first aspect.

[0018] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, a pre-trained dataset is constructed, integrating multi-source, multi-device, and multi-condition data for training, fundamentally addressing the weak transfer capability problem caused by signal distribution shift. Simultaneously, a pre-trained time-series model is constructed, using a fragmented encoder to perform local self-attention modeling on high-frequency long sequences, automatically capturing impact features and transient changes. A decoding-style Transformer captures long-range dependencies across segments, achieving global dynamic pattern extraction. The entire process requires no manual intervention; the model automatically learns features possessing both stability and generalization, completely eliminating reliance on human experience. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0020] Figure 1 This is a flowchart illustrating a method for identifying faults in industrial rotating machinery using a pre-trained time series model, as provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of an industrial rotating machinery fault identification system based on a pre-trained time series model, provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0023] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0024] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0025] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0026] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0027] Reference manual attached Figure 1 The diagram shows a flowchart of a method for identifying faults in industrial rotating machinery using a pre-trained time series model, as provided in an embodiment of the present invention.

[0028] This invention provides a method for identifying faults in industrial rotating machinery using a pre-trained time-series model. This method can be implemented using a pre-trained time-series model-based industrial rotating machinery fault identification device, which can be a terminal or a server. The processing flow of the pre-trained time-series model-based industrial rotating machinery fault identification method may include the following steps:

[0029] S1: Obtain raw data from industrial rotating machinery.

[0030] Specifically, the data sources include various publicly available or internal enterprise datasets for industrial rotating machinery vibration and condition monitoring, such as the Huazhong University of Science and Technology Bearing Dataset (HUST), the Mechanical Failure Prevention Technology Society Dataset (MFPT), the Jiangnan University Bearing Dataset (JNU), and the BJTU-RAO Bogie Transmission System Fault Simulation Dataset. These datasets cover bearing faults, gear faults, motor faults, and various normal operating conditions, providing rich time-series samples for the model.

[0031] S2: Preprocess the original data and construct a pre-training dataset based on the preprocessed original data. The pre-training dataset contains multiple sub-sequence fragments.

[0032] The construction of the pre-training dataset IRM-TSD first requires selecting appropriate industrial rotating machinery fault data sources. These data sources typically contain multiple features (such as sensor data, vibration acceleration signals, current signals, etc.) and corresponding fault labels, with the original data mostly in multivariate time series format. To enable data from different sources to be used in the same model, this invention performs unified preprocessing and transformation on the multi-source data, mainly including four steps: data loading and integration, standardization and denoising, time series fragmentation, and transformation from multivariate to univariate time series.

[0033] In one possible implementation, S2 specifically includes sub-steps S201 to S206: S201: Extract numerical features from the original data to obtain a feature set.

[0034] Specifically, numerical feature columns are extracted from the original data, while non-numerical columns such as timestamps and string labels are removed, resulting in multivariate time series data dominated by numerical features. If the data contains multiple variables, each variable typically corresponds to a different feature (e.g., data from different measuring points or vibration signals from different channels).

[0035] It should be noted that extracting numerical features from the original data to obtain a feature set allows for the precise selection of core and effective information directly related to the operating status of industrial rotating machinery. This eliminates non-numerical redundant interference such as timestamps and string labels, and unifies data types to lay a solid foundation of clean data for subsequent standardization, multi-variable to single-variable conversion, and fragmentation processing. At the same time, it ensures that the model focuses only on quantifiable dynamic features of equipment operation, laying a key foundation for the accuracy and stability of subsequent fault identification.

[0036] S202: Standardize the feature set using a standard scaler.

[0037] The calculation formula for standardization is as follows: in, x Represents the standardized features. X Represents the feature set, μ σ represents the mean, and σ represents the standard deviation.

[0038] Specifically, the mean of each feature is converted to 0 and the variance is converted to 1, thereby eliminating the dimensional differences between different features and different data sources.

[0039] Furthermore, bandpass filtering, wavelet denoising, and other methods can be combined before and after standardization to denoise the signal, thereby suppressing the impact of high-frequency interference and power frequency noise on subsequent modeling.

[0040] It should be noted that by standardizing the feature set through a standard scaler, the problem of feature weight imbalance caused by differences in numerical range is eliminated. At the same time, the model training process is more stable and converges faster, clearing scale obstacles for subsequent fusion modeling of cross-device and cross-operating condition data, and significantly improving the model's learning efficiency and generalization ability for fault features.

[0041] S203: Determine whether the original data contains multiple variables. If yes, define the standardized feature set as a multivariate time series and proceed to step S204. Otherwise, define the standardized feature set as a univariate time series and proceed to step S205.

[0042] It should be noted that by determining the number of variables in the original data and defining the time series type accordingly, the core advantage lies in achieving full compatibility with both single-sensor and multi-sensor industrial monitoring scenarios. This avoids processing logic confusion caused by differences in data types and provides a precise classification basis for subsequent differentiated processing of multi-variable to single-variable or directly fragmented data. This significantly improves the versatility and adaptability of the solution and ensures that different data sources can complete preprocessing along the optimal path.

[0043] S204: Convert a multivariate time series into a univariate time series.

[0044] Specifically, let the original multivariate data be a dataset of size ... T × D matrix X ,in T For time steps, D This refers to the number of features (e.g., the number of sensor channels). Then it can be... X The j The column is treated as an independent univariate time series.

[0045] The specific formula for univariate time series is as follows: in, x j Indicates the first j Univariate time series corresponding to each feature Indicates the first time step. j Multivariate time series with 1000 characteristic values Indicates the second time step. j Multivariate time series with 1000 characteristic values Indicates the first T The time step to the first j Multivariate time series with 1000 characteristic values j Indicates the first j One characteristic. x ij Indicates the first i The time step to the first j Multivariate time series with 1000 characteristic values i =1,2,…, T , T This represents the total number of time steps.

[0046] It should be noted that converting multivariate time series into univariate time series adapts to the unified input format requirements of pre-trained models. At the same time, each univariate sequence corresponds to an independent feature after splitting, which can completely preserve the temporal dynamic pattern of a single feature, avoid the interference of multi-feature coupling on the model's capture of local fault features, and provide a standardized data format for subsequent unified fragmentation processing and batch training, which greatly improves the adaptability of the solution to multi-sensor monitoring scenarios and the targeting of fault feature extraction.

[0047] S205: Divide the univariate time series into multiple subsequence segments.

[0048] Specifically, to further convert multivariate time series data into univariate fragment data that can be used for training, this invention targets each feature column... x j Perform fragmentation processing. Let the input data matrix be... X ∈ R T×D , for the j The features yield a univariate sequence x j Then, following the same method as described above, it is cut into multiple pieces of length [missing information]. L A fragment.

[0049] The specific formula for the subsequence segment is as follows: in, Indicates by the first j The first feature generated k Subsequence fragments, L Indicates the length of the subsequence segment. k Indicates the first k Subsequence fragments, j Indicates the first j One characteristic, x Represents a univariate time series. Indicates the first j In the univariate time series corresponding to the feature, the k In the subsequence segment, the first l The value at each time step. l =1,2,…, L , L This indicates the total number of time steps contained in the subsequence segment.

[0050] It should be noted that each feature is converted into a set of separate time series segments, which are treated as independent training samples during the data loading phase. The final training set consists of several univariate input segments, with each input sequence containing only the temporal evolution information of one feature. This allows the model to focus on learning the dynamic patterns of a single sensor channel and provides a unified basic format for subsequent multi-task learning or model ensemble.

[0051] S206: Construct a pre-training dataset based on each sub-sequence fragment.

[0052] It should be noted that a pre-training dataset is constructed based on each sub-sequence fragment, and standardized pre-processed data from single-sensor or multi-sensor scenarios are integrated to form a training corpus that is large in scale and uniform in format. This not only fully preserves the temporal dynamics and potential fault association information of each feature, but also adapts to the batch training requirements of the pre-trained model, providing high-quality data support for the model to learn the general operating rules across devices and working conditions, while ensuring data utilization efficiency and training stability.

[0053] In this embodiment of the invention, the original data is preprocessed and a pre-training dataset containing subsequence fragments is constructed. Through systematic preprocessing (screening numerical features, standardization, variable adaptation, and fragmentation), data noise and format differences are completely eliminated, forming a unified, clean, and standardized corpus that is adapted to the input of the pre-training model. This fully preserves the temporal characteristics of equipment operation and the information related to potential faults, laying a solid data foundation for the model to learn general operating rules across equipment and operating conditions, and significantly improving the accuracy, generalization, and sensitivity to early weak faults in subsequent fault identification.

[0054] S3: Construct a pre-trained time series model, which includes a fragmented encoder, a decoding Transformer, and a time series prediction output module connected in sequence.

[0055] In this embodiment of the invention, a deep fusion of accurate local feature capture and global temporal dependency modeling is achieved through module collaboration: the fragmented encoder adapts to industrial high-frequency long-sequence vibration data, solving the problems of gradient decay and limited receptive field in long-sequence processing; the decoding Transformer mines long-range dynamic patterns across segments through causal self-attention; and the prediction output module ensures that the predicted sequence is aligned with the real sequence format. The overall architecture not only meets the adaptation requirements of unsupervised pre-training to industrial signal characteristics, but also provides stable and efficient model support for subsequent time-frequency domain loss calculation and fault feature learning, significantly improving the integrity and generalization ability of feature representation.

[0056] S4: Input the pre-trained dataset into the pre-trained time series model and train the pre-trained time series model.

[0057] In one possible implementation, S4 specifically includes sub-steps S401 to S409: S401: By using a fragmented encoder, time step embedding and time step position encoding are performed on each subsequence segment to obtain a vector sequence.

[0058] It should be noted that by transforming discrete numerical industrial vibration subsequences into high-dimensional vector representations, time step embedding accurately captures the local transient fault features (such as impacts and subtle fluctuations) of each time step, and time step position encoding fully preserves the temporal logic, avoiding model confusion regarding the temporal order of signals. At the same time, the vector sequence format is adapted to the input requirements of the decoding Transformer, laying a solid foundation of structured features for subsequent global temporal dependency modeling, and significantly improving the feature representation capability of high-frequency long industrial sequences.

[0059] S402: Through a multi-head self-attention mechanism, the correlation and dependency between time steps in each sub-sequence segment are calculated, and the local dynamic features and shock structures within each sub-sequence segment are extracted based on the correlation and dependency.

[0060] It should be noted that it can capture local temporal correlations at different scales in parallel, accurately extract transient impact structures and dynamic change features related to early faults in industrial vibration signals, avoid feature omissions caused by a single attention dimension, and strengthen the weight of key fault information within segments, providing highly identifiable local feature support for subsequent global temporal dependency modeling, and significantly improving the model's sensitivity to capturing weak fault signals.

[0061] S403: By using a feedforward network and normalization operations, local dynamic features are aggregated to obtain fragment-level feature vectors corresponding to the local dynamic features.

[0062] It should be noted that nonlinear integration and dimensional optimization of scattered local temporal features enhance the correlation and compactness of fault-related features. At the same time, normalization operations eliminate feature scale differences, suppress gradient fluctuations, and stabilize the model training process. The generated fragment-level feature vectors not only retain the local impact structure and dynamic laws, but also adapt to the global modeling input requirements of decoding Transformer, greatly improving the effectiveness of feature representation and model training efficiency.

[0063] S404: Add fragment-level positional encoding to the fragment-level feature vectors, and input the fragment-level feature vectors with added positional encoding into the decoding Transformer to obtain fragment-level feature vectors sorted by time.

[0064] It should be noted that the precise preservation of the temporal sequence logic between each sub-sequence segment avoids the loss of segment-level temporal correlations during global modeling. At the same time, the decoding Transformer enhances the capture of long-range dependencies across segments through causal self-attention, allowing the time-ordered feature vectors to fully restore the dynamic evolution of the equipment's operating state. This not only adapts to the temporal continuity characteristics of industrial vibration signals but also provides orderly and coherent global feature support for subsequent global feature fusion and accurate prediction, significantly improving the model's ability to model the global patterns of long sequences.

[0065] S405: By using a causal self-attention mechanism, temporal causal constraints are applied to the time-ordered segment-level feature vectors to obtain segment-level temporal features that satisfy causality.

[0066] It should be noted that by strictly following the temporal evolution of industrial vibration signals, modeling biases caused by future information leaks are completely avoided. At the same time, the causal dependency between segments is strengthened, allowing features to accurately reproduce the continuous evolution of equipment operating status. In particular, it is suitable for the gradual features of early faults from their inception to their manifestation, providing ordered features that conform to physical logic for subsequent global temporal modeling and accurate prediction, and significantly improving the reliability of model prediction and the causal consistency of fault identification.

[0067] S406: Through a feedforward network, nonlinear mapping and feature enhancement are performed on fragment-level temporal features that satisfy causality to obtain high-level features containing global temporal dependency information.

[0068] It should be noted that by deeply mining the complex nonlinear correlations in global time-series dependencies, strengthening the coupling and recognizability of cross-segment fault features, and improving the abstract expression ability of high-level features, the model can better meet the input requirements of the time series prediction output module, making the features more consistent with the dynamic evolution of industrial vibration signals, effectively filtering redundant interference, and significantly improving the model's accuracy in capturing global fault modes and its ability to generalize across operating conditions.

[0069] S407: The time series prediction output module maps high-level features containing global temporal dependency information to the time step dimension to generate predicted time series.

[0070] Specifically, let the target label sequence of a certain training sample be: in, y Represents the target label sequence. y i Indicates the first [number]th [item] in the target label sequence i The value at each time step. i =1,2,…, N , N This represents the total time step length of the target label sequence.N The value is related to the fixed length of the pre-trained subsequence segment. L Consistent.

[0071] The predicted sequence output by the model is denoted as: in, Indicates the predicted sequence, Indicates the first [number]th ... i Predicted values ​​at each time step i =1,2,…, N .

[0072] It should be noted that the accurate restoration of the time sequence format of industrial vibration signals ensures that the prediction results are strictly aligned with the actual sequence in the time dimension. At the same time, it fully preserves the global time sequence dependency and fault correlation characteristics, adapts to the needs of subsequent time-frequency domain joint loss calculation, significantly improves prediction accuracy and sequence consistency, and provides a reliable benchmark for "normal mode" modeling and residual analysis in unsupervised fault identification, which is in line with the physical characteristics of the time sequence continuity of industrial signals.

[0073] S408: Construct the time-frequency domain feature loss function.

[0074] In one possible implementation, S408 specifically includes sub-steps S4081 to S4084: S4081: Construct a temporal loss function based on the predicted time series and the true label series.

[0075] The temporal loss function is specifically as follows: in, L time Indicates time-domain loss. y i Indicates the first i A real label sequence, Indicates the first i A predicted time series, N ∑ represents the length of the time-domain sequence, and ∑ represents the summation operation.

[0076] It should be noted that by precisely constraining the differences between the two in terms of time-domain waveform, transient impact, and trend changes, the model is driven to deeply learn the time-domain dynamic laws of industrial vibration signals under normal operating conditions. This maintains high sensitivity to subtle time-domain fluctuations corresponding to early faults, while laying a solid foundation for time-domain supervision for subsequent joint optimization in the time and frequency domains. This ensures that the predicted sequence is highly consistent with the time-domain physical characteristics of the real signal, significantly improving the accuracy of the model in modeling normal patterns and the reliability of residual analysis.

[0077] S4082: Perform Discrete Fourier Transform on the predicted time series and the real label sequence respectively to obtain the predicted frequency domain sequence and the real frequency domain sequence.

[0078] It should be noted that by performing discrete Fourier transforms on the predicted time series and the real label series respectively to obtain the frequency domain series, frequency domain features that are difficult to capture in the time domain are mined, supplementing the frequency domain physical information of industrial vibration signals, breaking through the limitations of single time domain modeling, providing key support for the subsequent construction of frequency domain loss function, allowing the model to learn the normal laws of both time and frequency domains at the same time, and significantly improving the sensitivity of early fault identification and the completeness of feature expression.

[0079] S4083: Construct a frequency domain loss function based on the predicted frequency domain sequence and the actual frequency domain sequence.

[0080] The frequency domain loss function is specifically as follows: in, L freq Indicates frequency domain loss, Y f ( p ) indicates the first f The first real frequency domain sequence p The complex amplitude of each frequency component, Indicates the first f The first predicted frequency domain sequence p The complex amplitude of each frequency component, M ∑ represents the length of the frequency domain sequence, and ∑ represents the summation operation.

[0081] Specifically, since the amplitudes of different frequency components in the frequency domain may differ by orders of magnitude, this invention selects MAE as the frequency domain error metric to avoid the excessive dominance of low-frequency large amplitude components on the loss, thereby improving the stability of the optimization process.

[0082] It should be noted that a frequency domain loss function is constructed based on the predicted frequency domain sequence and the actual frequency domain sequence. This function precisely constrains the differences between the two in the key frequency domain features of the fault, driving the model to deeply learn the inherent frequency domain laws of industrial vibration signals under normal operating conditions. This compensates for the deficiency of single time domain loss in capturing latent fault information in the frequency domain. Together with the time domain loss, a time-frequency joint supervision mechanism is formed, allowing the model to fully grasp the dual physical characteristics of the normal mode in both time and frequency. In particular, it strengthens the learning of early weak fault features that "appear in the frequency domain before the time domain", significantly improving the accuracy of subsequent fault identification and the completeness of feature expression.

[0083] S4084: The time-domain loss function and the frequency-domain loss function are weighted and combined to construct the time-frequency domain feature loss.

[0084] The time-frequency domain feature loss function specifically includes: the time-domain loss function and the frequency-domain loss function.

[0085] Furthermore, the time-frequency domain feature loss is specifically as follows: in, L TDFL Represents the time-frequency domain feature loss. α Indicates non-negative weight coefficients. L time Indicates time-domain loss. L freq This represents the frequency domain loss.

[0086] It should be noted that by weighting and combining the time-domain loss function and the frequency-domain loss function to construct the time-frequency domain feature loss, a time-frequency joint supervision mechanism is formed. This mechanism retains the sensitivity of the time domain to transient impacts and strengthens the constraint of the frequency domain on the frequency characteristics of latent faults. By flexibly adjusting the weights to adapt to different industrial operating conditions, it makes up for the information limitations of single-domain loss and drives the model to fully learn the dual physical nature of normal mode in both time and frequency. This significantly improves the coverage and accuracy of identification of various faults, especially strengthening the feature discrimination ability of early weak faults.

[0087] S409: Based on the predicted time series, calculate the loss function value of the time-frequency domain feature loss function, and repeat steps S401 to S408 until the loss function value is less than the preset loss function value.

[0088] In this embodiment of the invention, through the collaborative process of "feature encoding - temporal modeling - time-frequency joint supervision", the model deeply integrates the temporal transient features and frequency implicit laws of industrial vibration signals. This not only meets the engineering requirements of long sequence processing, but also comprehensively learns the essential characteristics of normal operation modes across working conditions through multi-module linkage and dual loss constraints. This significantly improves the robustness and generalization ability of the model's feature expression, providing a pre-training foundation with both physical consistency and accurate discrimination for subsequent fault identification.

[0089] S5: Obtain real-time data of the industrial rotating machinery to be tested.

[0090] S6: Preprocess the real-time data.

[0091] The preprocessing here is the same as the preprocessing steps in step S2.

[0092] In this embodiment of the invention, real-time data is preprocessed to strictly align with the data processing standards of the pre-training stage, quickly filter out sudden interference and data noise in the industrial field, unify the format, scale and feature distribution of real-time data, ensure that it is in the same feature space as the normal mode of model learning, fully preserve the potential transient or frequency domain features of abnormal equipment operation, adapt to the input requirements of real-time detection of the model, and lay a solid foundation for data consistency for subsequent accurate anomaly judgment.

[0093] S7: Input the preprocessed real-time data into the pre-trained time series model to obtain the time series segment to be detected.

[0094] In this embodiment of the invention, relying on the inherent compatibility of the preprocessing process and the model input format, real-time vibration data from the industrial site is quickly transformed into standardized detection units. This fully preserves the temporal dynamic characteristics and potential anomaly information of the equipment's real-time operating status, ensuring the timeliness of the detection response to meet the needs of real-time industrial monitoring. It also provides a unified analysis carrier for subsequent anomaly judgment and residual analysis, ensuring that real-time data and the normal learning pattern of the pre-trained model have a comparable basis, thereby improving the consistency and accuracy of fault detection.

[0095] S8: Calculate the time-frequency domain residual value based on the time-frequency domain feature loss function in the pre-trained time-series model.

[0096] In this embodiment of the invention, the difference between real-time data and normal mode is quantified by dual dimensions of time and frequency domain. This captures transient abnormal fluctuations in the time domain and identifies latent fault characteristics in the frequency domain (such as abnormal overtones). The adaptability of the loss function in the pre-training stage ensures the consistency of the physical meaning of the residual values, providing accurate quantitative basis for anomaly judgment. This effectively amplifies the distinguishability between early weak faults and normal mode, avoids missed or misjudged cases caused by single-dimensional comparison, and meets the accuracy requirements of industrial real-time fault monitoring.

[0097] S9: Determine whether the time-frequency domain residual value is greater than the anomaly detection threshold. If so, classify the time sequence segment to be tested corresponding to the time-frequency domain residual value as an abnormal segment, and determine that the state of the industrial rotating machinery to be tested is faulty. Otherwise, classify the time sequence segment to be tested corresponding to the time-frequency domain residual value as a normal segment, and determine that the state of the industrial rotating machinery to be tested is normal.

[0098] The specific formula for determining the time segment to be detected is as follows: in, f anom Indicates the time sequence segment to be detected. L TDFL Represents the time-frequency domain feature loss. τ This represents the anomaly detection threshold. fanom When the value equals 1, it indicates that the predicted sequence segment is an abnormal segment, and the judgment result is a fault. f anom When the value is 0, it indicates that the predicted sequence segment is a normal segment, and the judgment result is normal.

[0099] It should be noted that those skilled in the art can set the anomaly detection threshold according to actual needs, and this invention does not limit this.

[0100] In this embodiment of the invention, the time-frequency domain quantification difference is used as the accurate basis. The feature difference is transformed into a fault conclusion that can be directly implemented through a clear binary judgment logic. It not only covers explicit / implicit fault scenarios by relying on the dual dimensions of time and frequency, avoiding omissions and misjudgments by a single judgment standard, but also adapts to the real-time response needs of industrial sites because the decision logic is simple and quantifiable. At the same time, it clearly defines the boundary between normal and fault, improving the reliability, interpretability and convenience of engineering deployment of the judgment results.

[0101] Reference manual attached Figure 2 The diagram shows a structural schematic of an industrial rotating machinery fault identification system based on a pre-trained time series model provided by the present invention.

[0102] The present invention also provides a pre-trained time series model-based industrial rotating machinery fault identification system 20, applied to the above-mentioned pre-trained time series model-based industrial rotating machinery fault identification method, comprising: Processor 201.

[0103] The memory 202 stores computer-readable instructions, which, when executed by the processor 201, implement the industrial rotating machinery fault identification method based on a pre-trained timing model as described in the method embodiment.

[0104] The industrial rotating machinery fault identification system 20 based on the pre-trained time series model provided by this invention can execute the above-mentioned industrial rotating machinery fault identification method based on the pre-trained time series model and achieve the same or similar technical effects. To avoid repetition, this invention will not elaborate further.

[0105] It should be understood that the processor in the embodiments of the present invention can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0106] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0107] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0108] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0109] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0110] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0111] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0113] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0114] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0115] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0116] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] This invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the industrial rotating machinery fault identification method based on a pre-trained timing model as described in the method embodiment.

[0118] The present invention provides a computer-readable storage medium that can implement the steps and effects of the industrial rotating machinery fault identification method based on the pre-trained time series model of the above method embodiments. To avoid repetition, the present invention will not repeat them.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0120] The following points need to be explained: (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.

[0121] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the invention, i.e., these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.

[0122] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0123] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fault identification of industrial rotating machinery using a pre-trained time series model, characterized in that, include: S1: Obtain raw data from industrial rotating machinery; S2: Preprocess the original data and construct a pre-training dataset based on the preprocessed original data, wherein the pre-training dataset contains multiple sub-sequence segments; S3: Construct a pre-trained time series model, wherein the pre-trained time series model includes a fragmented encoder, a decoding Transformer, and a time series prediction output module connected in sequence; S4: Input the pre-trained dataset into the pre-trained time series model to train the pre-trained time series model; S5: Acquire real-time data of the industrial rotating machinery to be inspected; S6: Preprocess the real-time data; S7: Input the preprocessed real-time data into the pre-trained temporal model after training to obtain the temporal segment to be detected; S8: Calculate the time-frequency domain residual value based on the time-frequency domain feature loss function in the pre-trained time-series model after training; S9: Determine whether the time-frequency domain residual value is greater than the anomaly judgment threshold; if so, classify the time sequence segment to be detected corresponding to the time-frequency domain residual value as an abnormal segment, and determine that the state of the industrial rotating machinery to be detected is faulty; otherwise, classify the time sequence segment to be detected corresponding to the time-frequency domain residual value as a normal segment, and determine that the state of the industrial rotating machinery to be detected is normal.

2. The method for fault identification of industrial rotating machinery using a pre-trained time series model according to claim 1, characterized in that, S2 specifically includes: S201: Extract numerical features from the original data to obtain a feature set; S202: The feature set is standardized using a standard scaler; S203: Determine whether the original data contains multiple variables; if so, define the standardized feature set as a multivariate time series and proceed to step S204; otherwise, define the standardized feature set as a univariate time series and proceed to step S205. S204: Transform the multivariate time series into the univariate time series; S205: Divide the univariate time series into multiple subsequence segments; S206: Construct the pre-training dataset based on each of the said sub-sequence segments.

3. The method for fault identification of industrial rotating machinery using a pre-trained time series model according to claim 1, characterized in that, S4 specifically includes: S401: The fragmentation encoder is used to embed time steps and encode time step positions for each of the subsequence segments to obtain a vector sequence; S402: Calculate the correlation and dependency between time steps in each of the sub-sequence segments using a multi-head self-attention mechanism, and extract local dynamic features and impact structures within each of the sub-sequence segments based on the correlation and dependency. S403: Aggregate the local dynamic features through a feedforward network and normalization operation to obtain the fragment-level feature vector corresponding to the local dynamic features; S404: Add fragment-level position encoding to the fragment-level feature vector, and input the fragment-level feature vector with added position encoding into the decoding Transformer to obtain fragment-level feature vectors sorted by time; S405: By using a causal self-attention mechanism, temporal causal constraints are applied to the time-ordered segment-level feature vectors to obtain segment-level temporal features that satisfy causality; S406: Through a feedforward network, nonlinear mapping and feature enhancement processing are performed on the segment-level temporal features that satisfy causality to obtain high-level features containing global temporal dependency information; S407: The time series prediction output module maps the high-level features containing global temporal dependency information to the time step dimension to generate a predicted time series. S408: Construct the time-frequency domain feature loss function; S409: Based on the predicted time series, calculate the loss function value of the time-frequency domain feature loss function, and repeat steps S401 to S408 until the loss function value is less than the preset loss function value.

4. The method for fault identification of industrial rotating machinery using a pre-trained time series model according to claim 3, characterized in that, The time-frequency domain feature loss function specifically includes: a time-domain loss function and a frequency-domain loss function.

5. The method for fault identification of industrial rotating machinery using a pre-trained time series model according to claim 4, characterized in that, Specifically, S408 includes: S4081: Construct the temporal loss function based on the predicted time series and the real label sequence; S4082: Perform Discrete Fourier Transform on the predicted time series and the real label sequence respectively to obtain the predicted frequency domain sequence and the real frequency domain sequence; S4083: Construct the frequency domain loss function based on the predicted frequency domain sequence and the actual frequency domain sequence; S4084: The time-domain loss function and the frequency-domain loss function are weighted and combined to construct the time-frequency domain feature loss.

6. The method for fault identification of industrial rotating machinery using a pre-trained time series model according to claim 5, characterized in that, The time-domain loss function is specifically: ； in, L time Indicates time-domain loss, y i Indicates the first i A real label sequence, Indicates the first i A predicted time series, N ∑ represents the length of the time-domain sequence, and ∑ represents the summation operation.

7. The method for fault identification of industrial rotating machinery using a pre-trained time series model according to claim 5, characterized in that, The frequency domain loss function is specifically as follows: ； in, L freq Indicates frequency domain loss, Y f ( j ) indicates the first f The first real frequency domain sequence j The complex amplitude of each frequency component, Indicates the first f The first predicted frequency domain sequence j The complex amplitude of each frequency component, M ∑ represents the length of the frequency domain sequence, and ∑ represents the summation operation.

8. The method for fault identification of industrial rotating machinery using a pre-trained time series model according to claim 5, characterized in that, The time-frequency domain feature loss: ； in, L TDFL Represents the time-frequency domain feature loss. α Indicates non-negative weight coefficients. L time Indicates time-domain loss, L freq This represents the frequency domain loss.

9. The method for fault identification of industrial rotating machinery using a pre-trained time series model according to claim 1, characterized in that, The specific formula for determining the time segment to be detected is as follows: ； in, f anom Indicates the time segment to be detected. L TDFL Represents the time-frequency domain feature loss. τ This represents the anomaly detection threshold. f anom When the value equals 1, it indicates that the predicted sequence segment is an abnormal segment, and the judgment result is a fault. f anom When the value is 0, it indicates that the predicted sequence segment is a normal segment, and the judgment result is normal.

10. A fault identification system for industrial rotating machinery based on a pre-trained time series model, characterized in that, include: processor; A memory storing computer-readable instructions, which, when executed by the processor, implement the industrial rotating machinery fault identification method based on a pre-trained timing model as described in any one of claims 1 to 9.

Images