A method, apparatus and equipment for diagnosing idler roller faults

Abstract

This invention discloses a method, apparatus, and equipment for diagnosing idler roller faults, relating to the field of fault diagnosis technology. The method includes: acquiring the original vibration signal of the idler roller under test through each sensing point in a distributed optical fiber sensor; for each sensing point, determining the target statistical feature set of the sensing point based on the original vibration signal acquired by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum; for each sensing point, determining the fault signal impact intensity of the sensing point based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal; forming a fault energy distribution field based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor; and determining the fault information of the idler roller under test based on the fault energy distribution field, ensuring the accuracy of fault diagnosis in high-noise environments.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method, apparatus and equipment for diagnosing idler roller faults. Background Technology

[0002] In industrial settings such as mines and ports, belt conveyors, as core equipment for continuous transportation, rely on the health of tens of thousands of idlers along the line for safe operation. Idler jamming and wear can lead not only to major accidents such as belt tearing but also to unplanned production line downtime, resulting in significant economic losses. Currently, idler health monitoring technology for this scenario faces the following bottlenecks:

[0003] The conflict between monitoring coverage and continuity is a significant issue. Manual inspections and inspection robots cannot achieve continuous, 24 / 7 monitoring across the entire line, resulting in blind spots. Online monitoring solutions based on large-scale sensor networks face extremely high deployment costs on conveyor lines stretching several kilometers or even tens of kilometers, making it difficult to achieve economically feasible full lifecycle health monitoring of idler rollers. Furthermore, the mining port environment presents challenges due to dust, humidity, corrosion, and explosion-proof requirements, posing a severe challenge to the stable operation of electrical equipment.

[0004] Diagnostic algorithms suffer from insufficient engineering adaptability. Intelligent diagnostic algorithms based on deep learning (especially neural networks) rely heavily on large-scale, high-quality fault sample training for accuracy. However, in heavy industrial sites such as mines and ports, early fault samples are extremely scarce, and for safety reasons, it is impossible to expand the dataset by artificially creating faults. This leads to a severe "cold start" problem in the initial deployment of the system, with the model underfitting due to insufficient training. Even during long-term operation, the sporadic nature of fault samples makes it difficult for the model to continuously iterate and optimize, resulting in the common dilemma of "high accuracy in the laboratory, but difficult to implement in the field."

[0005] Feature extraction performance is low in high-noise environments. On-site vibration signals are mixed with a large amount of background noise and unknown interference. Traditional feature extraction methods have limited dimensions and insufficient characterization capabilities, making it difficult to effectively identify fault features from low signal-to-noise ratio signals, resulting in a significant decrease in diagnostic accuracy in practical applications.

[0006] In summary, existing technologies suffer from a significant decrease in diagnostic accuracy under harsh industrial environments due to the scarcity of on-site fault samples and the failure of traditional feature extraction methods in noisy environments. Summary of the Invention

[0007] The present invention provides a method, apparatus and equipment for diagnosing idler roller faults, in order to solve at least one of the above-mentioned problems.

[0008] In a first aspect, embodiments of the present invention provide a method for diagnosing idler roller faults, comprising:

[0009] The raw vibration signal of the roller under test is acquired by each sensing point in the distributed optical fiber sensor.

[0010] For each sensing point, the target statistical feature set of the sensing point is determined based on the original vibration signal collected by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum.

[0011] For each sensing point, the fault signal impact intensity of the sensing point is determined based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal.

[0012] A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor.

[0013] The fault information of the idler roller under test is determined based on the fault energy distribution field.

[0014] Secondly, embodiments of the present invention provide a roller fault diagnosis device, comprising:

[0015] The signal acquisition module is used to acquire the raw vibration signal of the roller under test through each sensing point in the distributed optical fiber sensor.

[0016] The feature processing module is used to determine the target statistical feature set of each sensing point based on the original vibration signal collected by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum.

[0017] The impact intensity determination module is used to determine the impact intensity of the fault signal at each sensing point based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal.

[0018] An energy distribution field determination module is used to determine the fault energy distribution field based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor.

[0019] The fault diagnosis module is used to determine the fault information of the idler roller under test based on the fault energy distribution field.

[0020] Thirdly, embodiments of the present invention provide an electronic device, the electronic device comprising:

[0021] At least one processor;

[0022] and a memory communicatively connected to the at least one processor;

[0023] The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform the idler roller fault diagnosis method according to any embodiment of the present invention.

[0024] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer instructions, which are used to cause a processor to execute the idler roller fault diagnosis method according to any embodiment of the present invention.

[0025] Fifthly, embodiments of the present invention provide a computer program product including a computer program, which, when executed by a processor, implements the idler roller fault diagnosis method described in any embodiment of the present invention.

[0026] The beneficial effects of this invention are:

[0027] The technical solution of this invention involves acquiring the original vibration signal of the idler roller under test through each sensing point in a distributed optical fiber sensor. For each sensing point, a target statistical feature set is determined based on the original vibration signal acquired by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum. For each sensing point, the fault signal impact intensity is determined based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal. A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor. The fault information of the idler roller under test is determined based on the fault energy distribution field. By extracting a high-quality target statistical feature set from a small number of fault samples through mode decomposition and Hilbert transform, the dependence on the amount of fault data is greatly reduced. The fault information of the idler roller under test is determined by the fault energy distribution field of the fault qualitative result reflected by the collaborative fault confidence and the fault quantitative result reflected by the instantaneous amplitude signal. This ensures the accuracy of fault diagnosis in high-noise environments and solves the problem that the existing technology suffers from a serious decrease in diagnostic accuracy in harsh industrial environments due to the scarcity of on-site fault samples and the failure of traditional feature extraction methods in high-noise environments. This is conducive to achieving early, accurate and reliable fault warning.

[0028] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0029] 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.

[0030] Figure 1 This is a flowchart of a roller fault diagnosis method provided in Embodiment 1 of the present invention;

[0031] Figure 2 This is a flowchart of a roller fault diagnosis method provided in Embodiment 2 of the present invention;

[0032] Figure 3 This is a flowchart of a roller fault diagnosis method provided in Embodiment 3 of the present invention;

[0033] Figure 4 This is a flowchart of the execution steps of an adaptive genetic algorithm provided in Embodiment 4 of the present invention;

[0034] Figure 5 This is a schematic diagram of the structure of a roller fault diagnosis device provided in Embodiment 5 of the present invention;

[0035] Figure 6 A schematic diagram of the electronic device used to implement the idler roller fault diagnosis method of this invention. Detailed Implementation

[0036] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0037] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0038] Example 1

[0039] Figure 1 This is a flowchart of a roller fault diagnosis method provided in Embodiment 1 of the present invention. This embodiment is applicable to fault detection of rollers in belt conveyors in industrial environments. The method can be executed by an roller fault diagnosis device, which can be implemented in hardware and / or software and can be configured in electronic equipment. Figure 1 As shown, the method includes:

[0040] S110. The original vibration signal of the roller under test is collected through each sensing point in the distributed optical fiber sensor.

[0041] Among them, distributed optical fiber sensors are sensors that transform an entire optical fiber into a continuous, densely distributed array of "virtual sensing points." By analyzing the modulation information of backscattered light in the optical fiber, they can achieve uninterrupted, high spatial resolution, real-time measurement and precise positioning of vibrations. Using the principle of optical time-domain reflectometry, the optical fiber is logically divided into tens of thousands of continuous spatially resolved units (e.g., one every 1 meter). Each unit is a "sensing point," which is not a physical entity but a measurement position calculated from optical signals.

[0042] The original vibration signal can be considered as the vibration signal of the idler roller under test collected by sensing points in a distributed fiber optic sensor array arranged along the belt conveyor. Generally, the belt conveyor has three idlers in the axial direction, and the distributed fiber optic sensors are arranged in a U-shape on two opposite sides of the belt conveyor (such as the left and right sides of the axial direction).

[0043] In this embodiment, distributed fiber optic sensors are arranged along the belt conveyor, and the original vibration signals of each idler roller to be tested in the belt conveyor are collected through the sensing points of the distributed fiber optic sensors.

[0044] S120. For each sensing point, determine the target statistical feature set of the sensing point based on the original vibration signal collected by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum.

[0045] The Hilbert envelope spectrum is obtained by first extracting the "envelope" of a vibration signal using the Hilbert transform, and then performing spectral analysis on this envelope signal to obtain the spectrum. Intrinsic mode function (IMF) components refer to a class of oscillation mode components in the Hilbert transform that satisfy specific conditions and have definite instantaneous frequencies. A signal can be decomposed into a superposition of several IMF components, each IMF representing the inherent vibrational characteristics of the signal at a specific time scale. The Hilbert envelope spectrum reflects the energy distribution of the vibration signal in the frequency domain. The instantaneous amplitude signal is used to characterize the time-domain variation of the signal energy.

[0046] In this embodiment, modal decomposition is performed on the original vibration signal collected at each sensing point to obtain multiple intrinsic mode function components. A Hilbert transform is then performed on each intrinsic mode function component to obtain the instantaneous amplitude signal and Hilbert envelope spectrum. Feature statistics are then performed on the original vibration signal, instantaneous amplitude signal, and Hilbert envelope spectrum to obtain the target statistical feature set of the sensing point.

[0047] Optionally, the statistical feature set obtained from feature statistics can be filtered to select statistical features that are both important and stable to form the target statistical feature set. For example, statistical features can be filtered based on feature importance, which can be reflected in at least one of information gain, statistical significance, and model performance.

[0048] S130. For each sensing point, the fault signal impact intensity of the sensing point is determined based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal.

[0049] Among them, the fault signal impact intensity can be considered a key quantitative indicator for measuring the severity of transient impact events caused by local defects (such as pitting and spalling). Fault confidence can be considered as the confidence of the fault diagnosis model in predicting the fault category based on the target statistical feature set (such as the fault probability or discrimination score of belonging to a certain type of fault, reflecting the data-driven pattern recognition results).

[0050] In this embodiment, the impact intensity of the fault signal corresponding to the vibration signal collected by a sensing point consists of two parts: one part is the fault confidence level predicted by the fault diagnosis model based on the target statistical feature set, and the other part is the instantaneous amplitude signal of the intrinsic mode function components of the original vibration signal. The fault confidence level provides the qualitative result of the fault in the vibration signal of the sensing point, while the instantaneous amplitude signal provides the quantitative result of the fault in the vibration signal of the sensing point. By separating and coordinating the qualitative and quantitative results of the fault, both highly reliable diagnostic intelligence and high-precision physical quantification capabilities are achieved.

[0051] S140. A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor.

[0052] The fault energy distribution field can be considered as a physical field characterizing the distribution of vibration energy density excited and released by mechanical faults within a multi-dimensional space spanned by spatial location and the impact intensity of the fault signal. The fault energy distribution field provides a direct view of the spatial propagation effect and abnormal energy accumulation effect of the fault, offering reliable observational evidence for fault diagnosis of the idler roller under test.

[0053] In this embodiment, the spatial position of each sensing point of the distributed optical fiber sensor is obtained (for example, the position coordinates of the sensing point in the world coordinate system or the position coordinates in the local coordinate system of the belt conveyor), and a fault energy distribution field is formed based on the fault signal impact intensity and spatial position of each sensing point.

[0054] S150. Determine the fault information of the idler roller under test based on the fault energy distribution field.

[0055] In this embodiment, the fault information of the idler roller under test is determined based on the distribution of energy peaks in the fault energy distribution field (such as the number, peak size and distribution location) and the spatial attenuation distribution of the fault signal impact intensity. The fault information may include the fault category (such as a single fault source or an independent fault source) and / or the fault location.

[0056] The technical solution of this invention involves acquiring the original vibration signal of the idler roller under test through each sensing point in a distributed optical fiber sensor. For each sensing point, a target statistical feature set is determined based on the original vibration signal acquired by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum. For each sensing point, the fault signal impact intensity is determined based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal. A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor. The fault information of the idler roller under test is determined based on the fault energy distribution field. By extracting a high-quality target statistical feature set from a small number of fault samples through mode decomposition and Hilbert transform, the dependence on the amount of fault data is greatly reduced. The fault information of the idler roller under test is determined by the fault energy distribution field, which combines the qualitative fault result reflected by the fault confidence and the quantitative fault result reflected by the instantaneous amplitude signal. This ensures the accuracy of fault diagnosis in high-noise environments and facilitates early, accurate, and reliable fault warning.

[0057] Example 2

[0058] Figure 2 This is a flowchart of a roller fault diagnosis method provided in Embodiment 2 of the present invention. Based on the above embodiments, this embodiment further refines the process of determining the target statistical feature set. Specifically, for each sensing point, determining the target statistical feature set of the sensing point based on the original vibration signal collected by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum includes: for each sensing point, performing modal decomposition on the original vibration signal collected by the sensing point to obtain multiple intrinsic mode function components; performing Hilbert transform on each intrinsic mode function component to obtain an instantaneous amplitude signal and a Hilbert envelope spectrum; determining an initial statistical feature set of the sensing point based on the original vibration signal, the instantaneous amplitude signal corresponding to each intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum; and selecting a target statistical feature set containing a preset number of features from the initial statistical feature set.

[0059] like Figure 2 As shown, the method includes:

[0060] S210. The original vibration signal of the roller under test is collected through each sensing point in the distributed optical fiber sensor.

[0061] S220. For each sensing point, the original vibration signal acquired at the sensing point is subjected to modal decomposition to obtain multiple intrinsic mode function components.

[0062] In this embodiment, the original vibration signal collected by each sensing point in the distributed optical fiber sensor is subjected to mode decomposition to obtain multiple intrinsic mode function components. For example, variational mode decomposition (VMD) is used to adaptively decompose the non-stationary original vibration signal into K quasi-orthogonal intrinsic mode function (IMF) components with limited bandwidth.

[0063] S230. Perform Hilbert transform on each intrinsic mode function component to obtain the instantaneous amplitude signal and Hilbert envelope spectrum.

[0064] In this embodiment, a Hilbert transform is performed on each intrinsic mode function component to extract deep time-frequency features that finely characterize the transient impact and modulation phenomena of the fault. The time-frequency features may include an instantaneous amplitude signal that characterizes the time-domain variation of the signal energy, and a Hilbert envelope spectrum that reflects the distribution of the signal energy in the frequency domain.

[0065] For example, for each eigenmode function component of the original vibration signal Perform Hilbert transforms on each signal to obtain the Hilbert transformed signal. ; The intrinsic mode function components Hilbert transform signal Combined into an analytic signal represented by a complex number. , where j is the imaginary unit of a complex number. The modulus of an analytic signal is its instantaneous amplitude. The instantaneous amplitude signal describes how the amplitude of the IMF component oscillation changes over time. For the instantaneous amplitude signal... Amplitude spectrum obtained by performing Fourier transform , which is the Hilbert envelope spectrum of the IMF component, and f is the fault characteristic frequency; the Hilbert envelope spectrum reveals the frequency components of the amplitude modulation.

[0066] S240. Determine the initial statistical feature set of the sensing point based on the original vibration signal, the instantaneous amplitude signals corresponding to each intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum.

[0067] The initial statistical feature set can be considered as the statistical feature set obtained from preliminary calculations.

[0068] In this embodiment, the time-domain statistical characteristics of the instantaneous amplitude signals of the original vibration signal and each IMF component are calculated, including, for example, mean, peak value, root mean square, kurtosis, and impulse factor. The frequency-domain statistical characteristics of the power spectrum of the original vibration signal and the Hilbert envelope spectrum of each IMF component are calculated, including, for example, centroid frequency, frequency band energy, mean amplitude, variance, and kurtosis. An initial statistical feature set is constructed based on the time-domain and frequency-domain statistical characteristics.

[0069] As an optional implementation of this embodiment, determining the initial statistical feature set of the sensing point based on the original vibration signal, the instantaneous amplitude signals corresponding to each eigenmode function component of the original vibration signal, and the Hilbert envelope spectrum includes:

[0070] A1. Determine the target sliding window parameters based on an adaptive genetic algorithm, and determine the temporal statistical characteristics of the original vibration signal and the instantaneous amplitude signal within the sliding window corresponding to the target sliding window parameters.

[0071] Idler roller failures can be categorized into several types, among which pitting corrosion is significantly reflected in the time-domain characteristics. Pitting corrosion can be understood as damage occurring at a specific point on the idler roller. Assuming the idler roller rotates 10 times per second, it will generate 10 similar impact signals (e.g., each lasting approximately 0.1 seconds). Therefore, the algorithm needs to specifically identify whether the window length is 0.11 seconds, 0.12 seconds, or something else. Other damage types are generally identified through frequency-domain characteristics. Thus, in this problem, the time-domain scales of the impact components are roughly similar.

[0072] For example, the scheme for determining the target sliding window parameters is encoded as an "individual," whose "chromosome" consists of multiple sets of sliding window parameters (which may include window width and sliding step size). During algorithm initialization, a "population" containing several such individuals is randomly generated. The quality of each individual in the population (i.e., each sliding window size partitioning scheme) is quantified by the performance of the diagnostic model based on its extracted features. Specifically, the system divides the signal length according to the individual's chromosome, calculates preset time-domain statistical features (such as mean, peak value, etc.) on each sub-signal segment, and then uses these features to train a lightweight classifier (such as XGBoost). The classifier's classification F1-score on the independent validation set is defined as the individual's fitness. The algorithm iteratively optimizes the population by simulating biological evolution mechanisms, specifically including three steps: selection, crossover, and mutation. In the selection step, high-fitness individuals are chosen as parents. In the crossover step, some boundary point sequences of two parent individuals are exchanged with a certain probability to generate new offspring partitioning schemes, aiming to inherit superior patterns. In the mutation step, the frequency value of a certain boundary point in the offspring individuals is randomly fine-tuned with a low probability to introduce new search possibilities and avoid getting trapped in local optima. The algorithm terminates when the evolution reaches the preset number of iterations or when the fitness converges. The window length and sliding step size corresponding to the individual with the highest fitness in each generation are determined as the target sliding window parameters.

[0073] For example, if the length of the original vibration signal is T, after decomposition, K intrinsic mode function components are obtained; the K intrinsic mode function components plus the original vibration signal can obtain K+1 signals; assuming the sliding window length is L, a total of There are n time-domain statistical features, where n is the number of feature dimensions of the time-domain statistical features.

[0074] A2. Determine the target frequency band based on the adaptive genetic algorithm, and determine the frequency domain statistical characteristics of the power spectrum of the original vibration signal and the Hilbert envelope spectrum within the target frequency band.

[0075] Traditional methods calculate frequency domain statistical features (such as kurtosis) across the entire frequency band. However, the impact response triggered by idler roller faults often concentrates in a specific high-frequency resonant band. If features are calculated directly across the entire frequency band, the energy of the fault feature components will be diluted by the energy of other frequency bands, thereby reducing the distinguishability between fault features and healthy states. To address this problem, this embodiment models the target frequency band partitioning problem as an automated optimization problem and uses an adaptive genetic algorithm to solve it. Its core objective is to accurately locate and isolate the feature sub-bands most sensitive to faults.

[0076] For example, a set of target frequency band partitioning schemes is encoded as an individual, whose chromosome is an increasing sequence of real numbers representing the frequency values ​​of the frequency band boundary points. During algorithm initialization, a population containing several such individuals is randomly generated. The quality of each individual in the population (i.e., each target frequency band partitioning scheme) is quantified by the performance of a diagnostic model based on its extracted features. Specifically, the system partitions the spectrum according to the individual's chromosome, calculates preset frequency domain statistical features (such as centroid frequency, band energy, etc.) in each sub-band, and then uses these features to train a lightweight classifier (such as XGBoost). The classifier's F1 score on the independent validation set is defined as the individual's fitness. The algorithm iteratively optimizes the population by simulating biological evolution mechanisms, specifically including three steps: selection, crossover, and mutation. In the selection step, high-fitness individuals are chosen as parents. In the crossover step, some boundary point sequences of two parent individuals are exchanged with a certain probability to generate new offspring partitioning schemes, aiming to inherit superior patterns. In the mutation step, the frequency value of a certain boundary point in the offspring individuals is randomly fine-tuned with a low probability to introduce new search possibilities and avoid getting trapped in local optima. The algorithm terminates when the evolution reaches the preset number of iterations or when the fitness converges. The frequency band boundary point sequence corresponding to the individual with the highest fitness in each generation is determined as the partitioning scheme for the target frequency band.

[0077] For example, adding the Hilbert envelope spectrum of the K intrinsic mode function components to the Hilbert envelope spectrum of the original vibration signal yields K+1 signals, which are then divided into... The segment then generates a total of There are 1 frequency domain statistical features, where m is the number of feature dimensions of the frequency domain statistical features.

[0078] A3. Construct an initial statistical feature set based on the time-series statistical features and / or the frequency-domain statistical features.

[0079] This embodiment provides a dual adaptive optimization strategy, which simultaneously determines the sliding window for time-domain analysis and the target frequency band for frequency-domain analysis through an adaptive genetic algorithm, and constructs an initial statistical feature set containing time-domain statistical features and / or frequency-domain statistical features, providing high-quality feature input for subsequent intelligent diagnostic models and improving the accuracy, early detection, and reliability of fault diagnosis.

[0080] As an optional embodiment, the execution steps of the adaptive genetic algorithm in the above embodiments include:

[0081] B1. Determine the target population based on the initial parameter set, and set the initialization parameters of the target population; wherein, the initial parameter set includes a sliding window parameter set or a frequency band set; the target parameters in the sliding window parameter set are target sliding window parameters; the target parameters in the frequency band set are target frequency bands.

[0082] In this embodiment, the initial parameter set can be considered as an optional set corresponding to the sliding window parameters in the initial statistical feature set process described in the above embodiments, or an optional set corresponding to the frequency band parameters. For example, a sliding window parameter set can be generated within the optional range of the sliding window parameters, and a frequency band set can be generated within the optional spectrum range. The target parameter can be considered as the optimal or suitable parameter selected from the initial parameter set. For example, the target parameter in the sliding window parameter set is the target sliding window parameter; the target parameter in the frequency band set is the target frequency band.

[0083] The target population can be considered as the population actually participating in the genetic algorithm. The target population can be the initial parameter set, or it can be obtained by screening the initial parameter set. This embodiment does not limit the screening method; for example, it can be determined based on the importance and concentration of parameter features in the initial parameter set. The initialization parameters of the target population may include: population size N, maximum number of iterations G. max Convergence threshold .

[0084] In this embodiment, for the frequency band allocation problem, let the spectrum range be (F min , F max The goal is to divide the spectrum into M consecutive sub-bands and extract frequency domain statistical features (such as centroid frequency, band energy, mean amplitude, variance, kurtosis, etc.) within each sub-band. The partitioning scheme consists of the number of outer segment M and the frequency band sequence. The decision is made jointly, and the following conditions are met: Among them, F min For the minimum spectrum, F max For the maximum spectrum, b q Frequency band sequence The q-th frequency band in the array, where q = {1, 2, ..., M-1}.

[0085] For the time-domain windowing problem, let the vibration signal length be T, and determine the window width L and sliding step size S of the sliding window so that the extracted time-domain statistical features (such as mean, peak value, root mean square, kurtosis, impulse factor, etc.) within each window best reflect the fault information. The windowing scheme must satisfy: ;in The maximum allowable window length (which can be determined by the number of outer layer segments M, usually taken as 1 / 10 to 1 / 5 of the signal length).

[0086] Optionally, determining the target population based on the initial parameter set includes: determining the initial parameter set as the initial population; calculating the inter-class and intra-class distance ratios of the training set corresponding to each individual in the initial population; and filtering the individuals in the population based on the inter-class and intra-class distance ratios to obtain the target population.

[0087] For example, for each individual in the initial population, the inter-class and intra-class distance ratios of the training data target statistical feature set are calculated based on the target parameters extracted from the individual's representation. This can be, for example, as follows:

[0088] ;

[0089] in, This is the inter-class scatter matrix, used to measure the degree of dispersion of samples within each class; The scatter matrix is ​​used to measure the degree of dispersion of each class center relative to the global center; trace(·) is the trace of the matrix.

[0090] In this embodiment, the inter-class and intra-class distance ratio is calculated based solely on the distance in the feature space of the target statistical feature set in the training data. This eliminates the need to train a classifier and is extremely fast. Screening the population based on the inter-class and intra-class distance ratio can eliminate obviously inferior partitioning schemes and reduce the computational overhead of subsequent fine evaluation.

[0091] B2. For the target population of the current generation, randomly generate the outer and inner chromosomes of individuals.

[0092] In this embodiment, a two-layer variable-length chromosome coding structure is adopted. The outer chromosome of an individual in the target population includes the number of outer segments M, which ranges from [2, M]. max ], M max M represents the maximum number of outermost segments, used to control the granularity of segmentation. For example, the number of outermost segments in the outermost chromosome of an individual in the target population corresponding to the sliding window parameter set is M=M1, and the number of outermost segments in the outermost chromosome of an individual in the target population corresponding to the frequency band set is M=M2.

[0093] The inner chromosome of an individual can encode specific partitioning parameters. For example, for an individual in the sliding window parameter set, the inner chromosome can be a binary tuple (L,S) consisting of the window width L and the sliding step size S, with a fixed length of 2. For an individual in the frequency band set, the inner chromosome can be a set of increasing frequency band sequences with a length of M²-1.

[0094] This embodiment employs a two-layer variable-length chromosome coding structure, where the total chromosome length of individuals in the target population varies with the outer chromosome layer, thus achieving variable-length coding and enabling the algorithm to adaptively explore partitioning schemes of different granularities.

[0095] B3. Determine the F1 score of the individual in the validation dataset as its fitness, and select high-quality individuals from the target population as parent individuals to enter the mating pool based on the fitness.

[0096] In this implementation, for each individual, the corresponding training data target statistical feature set is input into a classifier (such as XGBoost) to train the classifier; and the validation dataset is input into the trained classifier to obtain the classification result and F1 score corresponding to the validation dataset. The F1 score is determined as the fitness of the individual. For example, Leave-One-Out Cross-Validation (LOOCV) is used to train the classifier. Leave-One-Out Cross-Validation is the most extreme form of cross-validation method and is a special case of multi-fold cross-validation (when the number of folds equals the total number of samples N). LOOCV can maximize the use of limited samples (especially suitable for scenarios with only tens to hundreds of samples) and has low sensitivity to overfitting.

[0097] Then, based on fitness, several high-quality individuals (the individuals with the highest fitness) are selected from the target population as parent individuals to enter the mating pool. For example, a tournament selection method (tournament_size = 3) is used, where three individuals are randomly selected from the target population each time, and the one with the highest fitness is chosen to enter the mating pool. Simultaneously, the top 10% of elite individuals are retained and directly enter the next generation to ensure that the optimal solution is not lost.

[0098] B4. Randomly pair parent individuals with different outer chromosomes from the mating pool, and perform crossover operation on the inner chromosomes of each pair of parent individuals according to the adaptive crossover probability to obtain two offspring individuals.

[0099] The adaptive crossover probability can be considered as a crossover probability that can change adaptively.

[0100] In this embodiment, two parent individuals, P1 and P2, are randomly selected from the mating pool. The outer chromosome of P1 is M1, and the outer chromosome of P2 is M2. Based on the adaptive crossover probability, a common position is randomly selected to truncate the inner chromosomes of the two parent individuals, and the right-hand portions of the truncated chromosomes are exchanged to realize the crossover operation on the inner chromosomes of the parent individuals, thereby generating two offspring individuals.

[0101] For example, in the target population corresponding to the sliding window parameter set, the common location can be the common time period (t) between two parent individuals. min , t max A random public time point t is selected from ) public ;t min t is the lower limit of the public time period. maxThis represents the upper limit of the common time period. For the target population corresponding to the frequency band set, the common location can be the common spectrum range (f) between two parent individuals. min , f max A random common frequency band location r is selected from f. min f is the lower limit value of the common frequency band. max This is the upper limit of the public frequency band.

[0102] B5. Perform mutation operation on the offspring individuals according to the adaptive mutation probability, and calculate the fitness of the offspring individuals.

[0103] In this embodiment, mutation operations can include the following three types, which act on outer chromosomes and / or inner chromosomes respectively:

[0104] (1) Boundary point perturbation: Gaussian noise is added to a parameter (such as frequency band, window width L, or sliding step size S) in the inner chromosome. And ensure that it is still within the effective range (e.g., the frequency domain boundary remains increasing and the time domain window satisfies S≤L).

[0105] (2) Inserting a new boundary point: with insertion probability p ins A new boundary point is randomly inserted, and the number of outer segment M is increased by 1 (frequency band division only).

[0106] (3) Delete boundary points: with deletion probability p del Randomly delete an existing boundary point, and at the same time reduce the number of outer segment M by 1. M ≥ 2 must be guaranteed (frequency band division only).

[0107] In this embodiment, at least one of the three types of mutation operations described above can be used. The combined effect of these three types of mutations allows the algorithm to perform both fine-grained local searches and to change the granularity of the partitioning, thus avoiding getting trapped in local optima.

[0108] In this embodiment, the fitness of offspring individuals is calculated in the same way as the fitness of individuals in the target population, and will not be described again in this embodiment.

[0109] B6. Based on the fitness of individuals, determine a preset number of individuals from the population after merging the parent and offspring individuals as the new generation target population.

[0110] In this embodiment, the parent individuals are merged with the generated offspring individuals to obtain a new population. A preset number of individuals are determined from the new population as the next generation target population according to the individual fitness and elite selection strategy.

[0111] B7. Return to the execution of the steps related to selecting superior individuals from the target population as parent individuals to enter the mating pool based on the fitness of the individual, until the termination condition is met, output the best individual among all generations, and determine the best individual as the target parameter in the initial parameter set.

[0112] The optimal individual can be the one with the highest fitness.

[0113] In this embodiment, the process returns to step B2 to continue the next iteration until the termination condition is met. The best individual from all generations is then output, and this best individual is designated as the target parameter in the initial parameter set. Specifically, for the target population corresponding to the sliding window parameter set, the sliding window parameter corresponding to the best individual from all generations is determined as the target sliding window parameter; for the target population corresponding to the frequency band set, the frequency band corresponding to the best individual from all generations is determined as the target frequency band. For example, the termination condition could be that the optimal fitness has not improved for a consecutive preset number of generations or that the maximum number of iterations has been reached.

[0114] Optionally, the execution steps of the adaptive genetic algorithm further include:

[0115] B8. The ratio of the standard deviation of the fitness of each individual in the current generation of the target population to the maximum fitness is determined as the diversity level of the current generation of the target population.

[0116] In this embodiment, the diversity of the target population can be defined as the ratio of the standard deviation of an individual's fitness to its maximum fitness, for example:

[0117] ;

[0118] in, The standard deviation of the fitness of each individual in the current generation of the target population. The maximum fitness of individuals in the current generation of the target population is represented by D. The value of D ranges from [0,1]. A larger value indicates higher population diversity, while a smaller value indicates that population diversity tends to converge.

[0119] B9. Determine the adaptive crossover probability based on the preset minimum crossover probability, the preset maximum crossover probability, and the degree of diversity.

[0120] Wherein, the minimum crossover probability P is preset. c,min and the preset maximum crossover probability P c,max The settings can be customized according to requirements; this embodiment does not impose any restrictions on this.

[0121] In this embodiment, the adaptive crossover probability P c It can be defined as:

[0122] .

[0123] B10. Determine the adaptive mutation probability based on the preset minimum mutation probability, the preset maximum mutation probability, and the degree of diversity. Wherein, the preset minimum mutation probability P... m,min and the preset maximum mutation probability P m,max The settings can be customized according to requirements; this embodiment does not impose any restrictions on this.

[0124] In this embodiment, the adaptive mutation probability P m It can be defined as:

[0125] .

[0126] In traditional genetic algorithms, the crossover and mutation probabilities are two fixed values. The crossover probability controls the frequency of gene exchange, i.e., the ability to explore new combinations; the mutation probability controls the frequency of gene mutation, i.e., the ability to introduce entirely new genes, and is crucial to preventing the algorithm from getting trapped in local optima. Fixed crossover and mutation probabilities can cause the algorithm to linger near a local optimum in the later stages of iteration.

[0127] In this embodiment, when the population diversity is high (i.e., when D is large), the above formula will ensure the adaptive crossover probability P. c and adaptive mutation probability P m The crossover probability is automatically reduced to approach a minimum, meaning it is appropriately lowered to avoid excessively disrupting existing favorable patterns. This minimizes the disruption to existing patterns and allows the algorithm to perform a stable search within the discovered regions. When population diversity is low (i.e., when D is small), the population may get stuck in local optima. The above formula ensures that the adaptive crossover probability P... c and adaptive mutation probability P m The algorithm automatically increases its value, tending towards the maximum value. At this point, the crossover probability is increased to enhance the exploration capability. Through high-intensity crossover and mutation, new genes are forcibly introduced, helping the algorithm escape local optima and open up new search directions.

[0128] Compared to the fixed crossover and mutation probabilities in traditional genetic algorithms, this embodiment provides an adaptive crossover probability P... c The adaptive mutation mechanism and the adaptive mutation probability P m The adaptive mutation mechanism has the following advantages:

[0129] When individuals in a population generally have a small number of outer segments M, fine-tuning of boundary points is required. The adaptive mutation mechanism automatically reduces the adaptive mutation probability P when population diversity D is high. m This perfectly corresponds to the need for such fine-tuning, allowing for slight perturbation of the boundary points;

[0130] When individuals in a population generally have a large number of outer segments M, it is necessary to explore the combination of boundary points. The adaptive crossover mechanism automatically increases the adaptive crossover probability P when the population diversity D is low. c This facilitated the recombination of different boundary point sequences.

[0131] Traditional genetic algorithms require repeated adjustments to crossover and mutation probabilities for different problems, often relying on experience and being time-consuming to find suitable values. Adaptive crossover and mutation mechanisms transform parameter adjustment into setting maximum and minimum boundary values. Setting a reasonable range is much easier than finding a perfect fixed value and is more robust to the problem. For example, setting P... c,min =0.5, P c,max =0.9 means that the algorithm can automatically and adaptively adjust between 0.5 and 0.9 during the evolution process.

[0132] In a specific example, for the sliding window partitioning problem, high-quality individuals are selected from the target population corresponding to the sliding window parameter set based on fitness to enter the mating pool as parent individuals. From the mating pool, parent individuals P1 and P2 with different outer chromosomes are randomly paired. P1's outer chromosome (i.e., the number of outer segments) is M1, and P2's outer chromosome (i.e., the number of outer segments) is M2. The inner chromosomes are binary pairs P1=(L1, S1) and P2=(L2, S2), satisfying the constraint 1≤S≤L≤L max Where L1 is the window width of the sliding window corresponding to parent individual P1; S1 is the sliding step size of the sliding window corresponding to parent individual P1; L2 is the window width of the sliding window corresponding to parent individual P2; and S2 is the sliding step size of the sliding window corresponding to parent individual P2.

[0133] Two offspring individuals are generated through crossover:

[0134] The first offspring individual is ,but , ;

[0135] The second offspring individual is ,but , ;

[0136] Where α∈(0,1) is a random number corresponding to a common time point, which can usually be α=0.5 or randomly generated according to a uniform distribution; L c1 S is the window width of the sliding window corresponding to offspring individual C1; c1 L is the sliding step size of the sliding window corresponding to offspring individual C1; c2 S is the window width of the sliding window corresponding to offspring individual C2; c2The sliding step size is the sliding window corresponding to the offspring individual C2.

[0137] Since arithmetic cross is a convex combination, if all parent generations satisfy S≤L, the offspring generation will automatically satisfy the inequality because linear combinations preserve inequalities. Therefore, for time-domain window parameters, if a cross that preserves convexity is used (such as arithmetic cross or analog binary cross SBX), the constraints are naturally preserved, requiring no fine-tuning.

[0138] In another specific example, for the frequency band partitioning problem, high-quality individuals are selected from the target population corresponding to the frequency band set based on fitness to serve as parent individuals and enter the mating pool; from the mating pool, parent individuals P1 and P2 with different outer chromosomes are randomly paired. The outer chromosome number (i.e., the number of outer segments) of P1 is M1, and the outer chromosome number (i.e., the number of outer segments) of P2 is M2. The inner chromosome of P1 is a monotonically increasing frequency band sequence. , must meet ;in, For the q1th frequency band in frequency band sequence B1, Similarly, the inner chromosomes of P2 have monotonically increasing frequency band sequences. , must meet ;in, This refers to the q2th frequency band in frequency band sequence B2. .

[0139] In the public spectrum range (f min , f max A real number r is randomly selected from within the range as the common frequency band position. For the parent individual P1, B1 is divided into two parts at the common frequency band position: the left side... Right side Similarly, for parent individual P2, B2 is divided into two parts at the common frequency band position: the left side... Right side .

[0140] Crossing off parent individuals P1 and P2 produces two offspring individuals:

[0141] The first offspring individual is (That is, splicing L1 and R2 while maintaining the internal order of L1 and R2 respectively).

[0142] The second offspring individual is (That is, splicing L2 and R1 while maintaining the internal order of L2 and R1 respectively).

[0143] Update the outer segment number corresponding to the first offspring individual C1 to... Update the outer segment number corresponding to the second offspring individual C2 to... .

[0144] For example, suppose F min = 0 Hz, F max = 0 Hz; For parent individual P1, the number of outer segments M1 = 4, and the frequency band sequence is B1 = [200, 400, 600]; For parent individual P2, the number of outer segments M2 = 3, and the frequency band sequence is B1 = [150, 500]. A common frequency band position r = 350Hz is randomly selected from the common frequency range [200, 400]. The frequency band sequence of the parent individuals is segmented at this common frequency band position, resulting in frequency bands less than 350Hz for parent individual P1 (including 200Hz) and frequency bands greater than 350Hz (including 400Hz and 600Hz); frequency bands less than 350Hz for parent individual P2 (including 150Hz) and frequency bands greater than 350Hz (including 500Hz). Crossover operations are performed on parent individuals P1 and P2 to obtain two offspring individuals C1 and C2: The corresponding number of outer layer segments is ; The corresponding number of outer layer segments is .

[0145] The frequency band sequences corresponding to the two offspring individuals both satisfy monotonically increasing. Since all points on the left are less than the common frequency band position r, and all points on the right are greater than the common frequency band position r, the concatenated sequence naturally satisfies increasing order, requiring no additional fine-tuning. If the common frequency band position r happens to be equal to a boundary point, the common frequency band position r can be reselected.

[0146] S250. Select a target statistical feature set from the initial statistical feature set that contains a preset number of statistical features.

[0147] In this embodiment, candidate statistical feature sets are selected from the initial statistical feature set based on two aspects: the average information gain of statistical features and the probability of significant differences in statistical features. A preset number of features corresponding to the target statistical feature set is determined, and the target statistical feature set is composed of the preset number of statistical features selected from the candidate statistical feature set.

[0148] For example, the way to select a preset number of statistical features from the candidate statistical feature set can be to select a preset number of statistical features with a low probability of significant difference from the candidate statistical feature set; or it can be to select a preset number of statistical features with a high average information gain from the candidate statistical feature set.

[0149] As an optional implementation of this embodiment, the step of selecting a target statistical feature set containing a preset number of statistical features from the initial statistical feature set includes:

[0150] C1. Based on the decision tree model, determine the average information gain corresponding to each statistical feature in the initial statistical feature set, and determine the set of statistical features in the initial statistical feature set whose average information gain is greater than the gain threshold as the potential statistical feature set.

[0151] The average information gain (Gain) measures the average loss reduction resulting from a statistical feature when it serves as a split point across all decision trees. In this embodiment, it is used to evaluate the contribution of statistical features to fault diagnosis. Features with high average information gain can effectively distinguish fault categories and reduce diagnostic uncertainty; features with low average information gain contribute little to classification decisions and can be considered for removal.

[0152] The latent statistical feature set can be considered as the set of statistical features whose average information gain meets the requirements. The gain threshold can be considered as the lower limit of the average information gain corresponding to the desired statistical feature. The specific value of the gain threshold can be set according to experience and needs, or the statistical features in the initial statistical feature set can be sorted in descending order according to the average information gain, and the average information gain corresponding to the statistical feature with the preset ranking can be determined as the gain threshold. This embodiment does not limit the setting method of the gain threshold.

[0153] In this embodiment, a decision tree model, such as the XGBoost model, is trained based on the training set. The average information gain of each statistical feature when used as a split node is determined according to the trained decision tree model. Statistical features with an average information gain greater than a gain threshold are selected from the initial statistical feature set to form a potential statistical feature set.

[0154] For example, in XGBoost, for a split in a decision tree, suppose the node before the split contains the sample set I, and after the split it is divided into left child nodes. and right child node The formula for calculating the gain value of splitting is:

[0155] ;

[0156] ; ;

[0157] in, Let be the first derivative of the loss function with respect to the model's predicted value for the i-th sample. It is the second derivative of the loss function with respect to the model prediction of the i-th sample (generally optimized using a second-order Taylor expansion). This is the sum of the first derivatives corresponding to the left child node. It is the sum of the second derivatives corresponding to the left child node; This is the sum of the first derivatives corresponding to the right child nodes. It is the sum of the second derivatives corresponding to the right child node; The L2 regularization coefficient is... This is a splitting penalty term (used to control the complexity of the tree).

[0158] The average information gain is calculated as follows: For each feature, its gain value is calculated when it is used as a splitting feature in all splits of all decision trees; each gain value is weighted according to the number of samples covered by that split (or directly accumulated) to obtain the total gain of that feature; the total gain is divided by the number of splits (or according to the model settings) to obtain the average information gain. For example, the statistical features in the initial statistical feature set can be sorted in descending order according to the average information gain, and the ranking of the statistical feature corresponding to the average information gain with the smallest difference from the gain threshold can be determined. The set of all statistical features ranked before this one is determined as the potential statistical feature set.

[0159] C2. Evaluate the probability of significant differences between prior fault samples and normal samples on each statistical feature in the potential statistical feature set based on nonparametric statistical test methods; determine the set of statistical features in the potential statistical feature set whose probability of significant difference is less than the set significance level as the candidate statistical feature set.

[0160] The probability of significant difference can be considered as the probability that a prior faulty sample and a normal sample differ significantly in a certain statistical feature. The candidate statistical feature set can be considered as the set of statistical features that satisfy both the requirements of average information gain and probability of significant difference.

[0161] In limited training data, some features may exhibit spurious "differences" between fault / normal groups due purely to random sampling fluctuations, and these differences lack reproducibility. Nonparametric statistical tests act as filters; for example, the p-value of the Mann-Whitney U test quantifies the probability that the observed inter-group differences are purely random. Setting a significance level, such as 0.05, means that only features with a ≤5% probability of being randomly generated are allowed to pass through. This is equivalent to using a sieve with a statistical confidence level ≥95% to filter out features whose differences are most likely just data coincidences. Ultimately, only statistically significant features remain. This means that their inter-group differences are more likely to reflect the essential difference between fault and normal states, rather than specific noise in the training set. Models built based on these features are more likely to generalize their learned patterns to new, unseen data.

[0162] In this embodiment, to avoid overfitting of the fault model to specific patterns in the training data, statistical hypothesis testing is performed on the statistical features ranked highest in importance (average information gain), i.e., the statistical features in the potential statistical feature set. This assesses the probability of significant differences between prior fault samples and normal samples on each statistical feature in the potential statistical feature set. For example, for each statistical feature in the potential statistical feature set, the fault feature numerical sequence for each fault sample and the positive feature numerical sequence for each normal sample on that statistical feature are determined. The specific process of statistical hypothesis testing can be as follows: A null hypothesis H0 and an alternative hypothesis H1 are proposed. The null hypothesis H0 is that the feature has the same distribution in the fault group and the normal group; the alternative hypothesis H1 is that the distributions of the two groups are different (or have positional shifts). The two groups of data are merged and sorted, and the rank sum is calculated. Then, the statistic U is used to measure the rank advantage of one group of data relative to the other group of data, and the exact or asymptotic probability of significant difference (i.e., p-value) is calculated based on the U value and the sample size.

[0163] This embodiment ensures that the selected candidate statistical feature set has both high discriminative power and statistical robustness through two stages: model-driven (i.e. feature selection based on decision tree model) and statistical-driven (i.e. feature selection based on the probability of significant difference).

[0164] C3. Obtain the preset number of features corresponding to the target statistical feature set, sort the statistical features in the candidate statistical feature set in ascending order according to the probability of difference significance, and determine the set of statistical features with the highest preset number of features as the target statistical feature set.

[0165] The preset number of features can be considered as the number of statistical features ultimately retained. In this embodiment, the preset number of features can be set according to requirements or determined based on the contribution of statistical features to fault classification.

[0166] In this embodiment, after the above-mentioned double screening, the number of retained features may still be as high as several hundred dimensions, which poses a risk of overfitting when the training samples are limited. Therefore, the statistical features in the candidate statistical feature set are further sorted in ascending order according to the probability of significant difference, and the set of statistical features with the first preset number of features in the sorted set is determined as the target statistical feature set, so that the number of features in the final target statistical feature set is the preset number of features.

[0167] Optionally, obtaining the preset number of features corresponding to the target statistical feature set includes:

[0168] C31. Obtain historical vibration signals collected by the distributed optical fiber sensor, and determine the training dataset and validation dataset based on the historical vibration signals.

[0169] C32. Determine the candidate statistical feature set of the training data corresponding to the training dataset.

[0170] In this embodiment, for each training data in the training dataset, modal decomposition is performed on the training data to obtain multiple intrinsic mode function components. Hilbert transform is performed on each of the intrinsic mode function components to obtain the instantaneous amplitude signal and Hilbert envelope spectrum. Based on the training data, the instantaneous amplitude signal and Hilbert envelope spectrum corresponding to each intrinsic mode function component of the training data, the candidate statistical feature set of the training data corresponding to the training dataset is determined.

[0171] C33. Sort the statistical features in the candidate statistical feature set of the training data in ascending order according to the probability of significant difference of each statistical feature in the candidate statistical feature set of the training data.

[0172] The candidate statistical feature set of the training data can be considered as the candidate statistical feature set obtained by performing feature statistics and feature filtering on the training data. It should be noted that the method for determining the candidate statistical feature set corresponding to the original vibration signal is the same as the method for determining the candidate statistical feature set of the training data.

[0173] In this embodiment, the probability of significant difference between prior fault samples and normal samples on each statistical feature in the candidate statistical feature set of the training data is evaluated based on a nonparametric statistical test method, and the statistical features in the candidate statistical feature set of the training data are sorted in ascending order according to the probability of significant difference.

[0174] C34. According to multiple different preset extraction quantities, the target statistical feature set of the training data is extracted from the candidate statistical feature set of the training data in sequence according to the sorting result.

[0175] The target statistical feature set of the training data can be considered as the target statistical feature set obtained by feature filtering from the candidate statistical feature set of the training data. The preset extraction quantity can be a set of different extraction quantities, or it can be determined according to different proportions of the number of features in the target statistical feature set of the training data (such as 5%, 10%, 15%, 20%, etc.).

[0176] In this embodiment, multiple different preset extraction quantities are set. For each preset extraction quantity, a preset number of training data target statistical features are extracted from the training data candidate statistical feature set according to the sorting result. The set of extracted training data target statistical features is determined as the training data target statistical feature set.

[0177] C35. For each training data target statistical feature set, a classifier is trained based on the training data target statistical feature set, and the classifier is validated based on the validation dataset to obtain the F1 score of the validation dataset.

[0178] In this embodiment, for the training data target statistical feature set obtained according to each preset extraction quantity, the training data target statistical feature set is input into the classifier to train the classifier; and the verification dataset is input into the trained classifier to obtain the classification result and F1 score corresponding to the verification dataset.

[0179] C36. Determine the preset number of features by the preset number of extractions corresponding to the target statistical feature set of the training data used by the classifier corresponding to the validation set with the highest F1 score.

[0180] In this embodiment, the target statistical feature set of the training data used by the classifier corresponding to the validation set with the highest F1 score is determined during training, and the preset extraction quantity corresponding to the determined target statistical feature set of the training data is determined as the preset feature quantity.

[0181] This embodiment presents a data-driven, adaptive optimization method for determining the number of preset features. By using historical data verification and a model performance-oriented search strategy, it solves the core problem of "relying on experience and lacking quantitative basis" in the traditional selection of the number of features, and achieves the optimal balance between the number of features and model performance, as well as maximizing the contribution of a limited number of features.

[0182] S260. For each sensing point, the fault signal impact intensity of the sensing point is determined based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal.

[0183] S270. A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor.

[0184] S280. Determine the fault information of the idler roller under test based on the fault energy distribution field.

[0185] The technical solution of this invention involves acquiring the original vibration signal of the idler roller under test through each sensing point in a distributed optical fiber sensor. For each sensing point, modal decomposition is performed on the acquired original vibration signal to obtain K intrinsic mode function (IMF) components. A Hilbert transform is then performed on each IMF component to obtain the instantaneous amplitude signal and Hilbert envelope spectrum. Based on the original vibration signal, the instantaneous amplitude signals corresponding to the K IMF components, and the Hilbert envelope spectrum, an initial statistical feature set for the sensing point is determined. A target statistical feature set containing a preset number of features is selected from the initial statistical feature set. For each sensing point, the fault signal impact intensity is determined based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the IMF components of the original vibration signal. A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor. The fault information of the idler roller under test is determined based on the fault energy distribution field. By extracting a high-quality target statistical feature set from a small number of fault samples through mode decomposition and Hilbert transform, the dependence on the amount of fault data is greatly reduced. Furthermore, through model-driven, statistical-driven, and fixed-number feature selection, the selected target statistical feature set is ensured to have high discriminative power, statistical robustness, and an optimal balance between feature quantity and model performance. By determining the fault energy distribution field of the fault qualitative and quantitative results, the fault information of the idler roller under test is determined, ensuring the accuracy of fault diagnosis in high-noise environments and facilitating early, accurate, and reliable fault warning.

[0186] Example 3

[0187] Figure 3 This is a flowchart of a roller fault diagnosis method provided in Embodiment 3 of the present invention. Based on the above embodiments, this embodiment further refines the method for determining the fault energy distribution field and the fault identification method based on the fault energy distribution field. Specifically, for each sensing point, determining the fault signal impact intensity of the sensing point based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal includes: for each sensing point, predicting the fault confidence of the target statistical feature set based on the fault diagnosis model; for the k-th order intrinsic mode function component of the original vibration signal, determining the instantaneous amplitude mean based on the instantaneous amplitude signal of the intrinsic mode function component, k≤K; and determining the fault signal impact intensity of the sensing point by the weighted sum of the fault confidence and the instantaneous amplitude mean corresponding to the first K order intrinsic mode function components.

[0188] And, determining the fault information of the idler roller under test based on the fault energy distribution field includes: searching for local maximum points in the fault energy distribution field within a single detection cycle, and identifying the local maximum points as potential fault points; clustering active sensor points in the spatial neighborhood of the potential fault points to obtain clusters; wherein the fault signal impact intensity of the active sensor points is greater than an intensity threshold; fitting the spatial attenuation distribution of the fault signal impact intensity of the sensor points within the clusters; and determining the fault type of the potential fault points based on the number of potential fault points and the spatial attenuation distribution.

[0189] like Figure 3 As shown, the method includes:

[0190] S310: Collect the raw vibration signal of the roller under test through each sensing point in the distributed optical fiber sensor.

[0191] S320. For each sensing point, determine the target statistical feature set of the sensing point based on the original vibration signal collected by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum.

[0192] In this embodiment, the specific implementation of S320 can be referred to S130, or S230~S250, and will not be described in detail here.

[0193] S330. For each sensing point, predict the fault confidence of the target statistical feature set based on the fault diagnosis model; determine the mean instantaneous amplitude of the instantaneous amplitude signal corresponding to the k-th intrinsic mode function component of the original vibration signal; determine the fault signal impact intensity of the sensing point by weighting the fault confidence and the mean instantaneous amplitude of the first K intrinsic mode function components respectively, where k ≤ K and k and K are both positive integers.

[0194] The instantaneous amplitude mean of the k-th intrinsic mode function (IMF) refers to the arithmetic mean of the instantaneous amplitude envelope signal obtained by performing a Hilbert transform on the k-th IMF component.

[0195] In this embodiment, to accurately quantify the fault impact intensity at each sensing point, instead of relying on a single feature, a fault signal impact intensity is constructed that integrates the fault confidence and the instantaneous amplitude mean of the first K eigenmode function components.

[0196] Specifically, the fault signal impulse intensity can be expressed as:

[0197] ;

[0198] in, Let be the fault signal impact intensity of the r-th sensing point; For fault confidence, The weights for the fault confidence. The instantaneous amplitude mean of the k-th eigenmode function component reflects the physical impact energy of the signal itself. , where is the weighting coefficient of the instantaneous amplitude mean of the k-th eigenmode function component; The weighting coefficients are the weighted sum of the instantaneous amplitude mean values ​​corresponding to the first K eigenmode function components. and Used to balance the contribution of data-driven prior knowledge and physical signal evidence (typically configurable) ); It can be configured based on its correlation with typical failure impacts.

[0199] In this embodiment, the fault signal impact intensity simultaneously captures both the "probability of the fault as perceived by the model" and the "impact intensity displayed by the signal itself," serving as the final signal intensity output for each sensing point, thus combining high sensitivity and strong noise immunity.

[0200] S340. A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor.

[0201] S350. During the detection period, search for local maximum points in the fault energy distribution field and identify the local maximum points as potential fault points.

[0202] Among them, potential fault points can be considered as sensing points that may be faulty.

[0203] In a spatial distribution field, a real local fault will form a clear spatial energy peak. Therefore, in this embodiment, by searching for the local maximum point in the fault energy distribution field within each detection cycle, potential fault points can be preliminarily determined, and their spatial locations can be obtained.

[0204] S360. Cluster the active sensing points in the spatial neighborhood of the potential fault point to obtain a cluster; fit the spatial attenuation distribution of the fault signal impact intensity of the sensing points in the cluster; wherein, the fault signal impact intensity of the active sensing point is greater than the intensity threshold.

[0205] Among them, active sensing points can be considered as sensing points where the fault signal impact intensity is greater than the intensity threshold.

[0206] Since vibration is transmitted through the structure, a single fault source will form a signal intensity distribution that attenuates with distance within its spatial neighborhood. In this embodiment, active sensing points with fault signal impact intensity greater than an intensity threshold within the spatial neighborhood (which can be a preset spatial range) of a potential fault point are clustered to obtain a cluster; and the spatial attenuation distribution of fault signal impact intensity of sensing points within the cluster is fitted.

[0207] S370. Determine the target fault point based on the number of potential fault points and the spatial attenuation distribution.

[0208] In this embodiment, the target fault point among the potential fault points is determined based on the number of potential fault points and the spatial attenuation distribution of the potential fault points.

[0209] As an optional implementation of this embodiment, determining the target fault point based on the number of potential fault points and the spatial attenuation distribution includes:

[0210] C1. If the number of potential fault points is one, and the spatial attenuation distribution of the potential fault point conforms to the single-source attenuation model, then the potential fault point is determined as the target fault point.

[0211] Among them, the single-source attenuation model can be considered as a distribution of signal strength attenuation with distance within the spatial neighborhood of a single fault point.

[0212] In this embodiment, if there is only one potential fault point and the spatial attenuation distribution of the potential fault point conforms to the single-source attenuation model, the potential fault point can be considered as the target fault point and the target fault point is a single fault source.

[0213] C2. If the number of potential fault points is greater than one, then each of the potential fault points that does not belong to the symmetrical potential fault point group and whose spatial attenuation distribution conforms to the single-source attenuation model shall be determined as the target fault point; the potential fault point in the symmetrical potential fault point group with a larger fault signal impact intensity attenuation ratio in the spatial attenuation distribution shall be determined as the target fault point.

[0214] A symmetrical potential fault point group can be considered as a group of fault points symmetrically distributed along the axial direction of the belt conveyor. A symmetrical potential fault point group may include two, three, or more fault points distributed axially.

[0215] In this embodiment, if the number of potential fault points is greater than one, the existence of a symmetrically distributed group of potential fault points is determined based on the spatial location of each potential fault point.

[0216] For potential fault points in a symmetrical potential fault point group, the spatial attenuation distributions of each potential fault point in the group are compared. The potential fault point with the larger attenuation ratio of the fault signal impulse intensity in the spatial attenuation distribution is identified as the target fault point. Since a sensing point on one side can detect fault signals from both its own side and the opposite side, and the fault signal sensed on its own side is more pronounced, the side in the symmetrical potential fault point group with a significantly higher fault signal impulse intensity and closer to the theoretical source characteristics can be identified as the side where the actual fault occurs. For each potential fault point not belonging to the symmetrical potential fault point group, if the spatial attenuation distribution of a single potential fault point conforms to the single-source attenuation model, then that potential fault point is identified as the target fault point.

[0217] As an optional embodiment, determining the fault information of the idler roller under test based on the fault energy distribution field further includes:

[0218] S380. Within multiple consecutive detection cycles, count the target fault points at the same spatial location where the impact intensity of the fault signal is greater than the intensity threshold; determine the target fault points with count values ​​greater than the count threshold as the actual fault points of the idler roller to be tested, and determine the spatial location of the actual fault points.

[0219] In this embodiment, to distinguish between persistent faults and transient interference, multi-round time verification is introduced. Within multiple consecutive detection cycles (e.g., N detection cycles), target fault points at the same spatial location are counted, and the number of times the fault signal impact intensity of the target fault point exceeds the intensity threshold is recorded (i.e., counting target fault points at the same spatial location whose fault signal impact intensity exceeds the intensity threshold is recorded). Target fault points with count values ​​exceeding the counting threshold are identified as the actual fault points of the idler roller under test, and the spatial location of the actual fault points is determined.

[0220] In addition, after obtaining the fault signal impact intensity and spatial location of the actual fault point, the fault severity level can be determined based on the fault signal impact intensity, generating structured fault warning information containing the fault severity level and spatial location. The fault warning information can be displayed, or the operation and maintenance response can be directly driven based on the fault warning information to form a complete diagnostic closed loop.

[0221] This embodiment is based on a multi-round time verification mechanism to distinguish between the persistence of real faults and the sporadic nature of interference, thereby improving the robustness of the system.

[0222] The technical solution of this invention involves acquiring the original vibration signal of the idler roller under test through each sensing point in a distributed optical fiber sensor. For each sensing point, a target statistical feature set is determined based on the original vibration signal acquired by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum. For each sensing point, the fault confidence level of the target statistical feature set is predicted based on a fault diagnosis model. The mean instantaneous amplitude corresponding to the instantaneous amplitude signal of the k-th order intrinsic mode function component of the original vibration signal is determined, where k ≤ K. The weighted sum of the fault confidence level and the mean instantaneous amplitude corresponding to the first K order intrinsic mode function components is determined as the fault signal impact intensity of the sensing point. A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor. Within the detection period... The process involves: searching for local maxima in the fault energy distribution field and identifying these local maxima as potential fault points; clustering active sensor points within the spatial neighborhood of potential fault points to obtain clusters; fitting the spatial attenuation distribution of fault signal impact intensity of sensor points within each cluster, where the fault signal impact intensity of active sensor points is greater than an intensity threshold; counting target fault points at the same spatial location with fault signal impact intensity greater than an intensity threshold within multiple consecutive detection cycles; identifying target fault points with count values ​​greater than a count threshold as actual fault points of the idler roller under test and determining the spatial location of the actual fault points. By extracting a high-quality target statistical feature set from a small number of fault samples through mode decomposition and Hilbert transform, the dependence on the amount of fault data is greatly reduced. By determining the fault information of the idler roller under test through the fault energy distribution field of the combined qualitative and quantitative fault results, the accuracy of fault diagnosis in a high-noise environment is guaranteed, which is conducive to achieving early, accurate and reliable fault warning.

[0223] Example 4

[0224] Figure 4 This is a schematic diagram illustrating the execution steps of an adaptive genetic algorithm provided in Embodiment 4 of the present invention. It should be noted that the adaptive genetic algorithm in this embodiment is not only used to determine the target sliding window parameters and target frequency band in Embodiment 3 above, but can also be used as an independent algorithm to determine the optimal partitioning scheme in other application scenarios.

[0225] Genetic algorithms, as an efficient global optimization method, can avoid subjective biases caused by manual parameter settings and have been widely used in parameter optimization problems. However, existing genetic algorithms have the following shortcomings when applied to this scenario: (1) Fixed segment number limitation, specifically, traditional frequency band division methods usually preset the number of segments and cannot adaptively determine the optimal number of segments. (2) Fixed parameter limitation, specifically, the crossover probability and mutation probability of traditional genetic algorithms are usually fixed and can easily fall into local optima in the later stages of evolution, making it difficult to stably search for the global optimal solution under conditions of few samples.

[0226] To address the shortcomings of existing technologies, a two-layer adaptive genetic algorithm is proposed. Its core lies in optimizing the number of segments in the outer layer and optimizing boundary points in the inner layer. Combined with an adaptive parameter adjustment mechanism based on population diversity, the partitioning scheme is optimized under conditions of few samples.

[0227] like Figure 4 As shown, the execution steps of the adaptive genetic algorithm include:

[0228] S410. Obtain the target population and set the initialization parameters of the target population;

[0229] S420. For the target population of the current generation, randomly generate the outer and inner chromosomes of individuals;

[0230] S430. Determine the fitness of individuals and select high-quality individuals from the target population as parent individuals to enter the mating pool based on their fitness.

[0231] S440. Randomly pair parent individuals with different outer chromosomes from the mating pool, and perform crossover operation on the inner chromosomes of each pair of parent individuals according to the adaptive crossover probability to obtain two offspring individuals.

[0232] S450. Perform mutation operations on the offspring individuals according to the adaptive mutation probability, and calculate the fitness of the offspring individuals.

[0233] Optionally, mutation operations can include the following three types, acting on outer chromosomes and / or inner chromosomes respectively:

[0234] (1) Boundary point perturbation: Gaussian noise is added to a certain parameter in the inner chromosome. And ensure that it remains within its effective range;

[0235] (2) Inserting a new boundary point: with insertion probability p ins A new boundary point is randomly inserted, and the number of outer segments M is increased by 1.

[0236] (3) Delete boundary points: with deletion probability p del Randomly delete an existing boundary point, and at the same time reduce the number of outer segments M by 1, while ensuring that M≥2.

[0237] S460. Based on the fitness of individuals, determine a predetermined number of individuals from the population after merging parent and offspring individuals as the target population for the next generation.

[0238] S470, Return to the execution of the steps related to selecting superior individuals from the target population as parent individuals to enter the mating pool based on the individual's fitness, until the termination condition is met, and output the best individual among all generations.

[0239] Optionally, obtain the target population, including:

[0240] Obtain the initial population;

[0241] Calculate the inter-class and intra-class distance ratios of the feature sets corresponding to each individual in the initial population;

[0242] Individuals in the population are selected by comparing inter-class and intra-class distances to obtain the target population.

[0243] Optionally, the execution steps of the adaptive genetic algorithm may also include:

[0244] The ratio of the standard deviation of fitness to the maximum fitness of each individual in the current generation of the target population is defined as the diversity level of the current generation of the target population.

[0245] The adaptive crossover probability is determined based on the preset minimum crossover probability, the preset maximum crossover probability, and the degree of diversity.

[0246] The adaptive mutation probability is determined based on the preset minimum mutation probability, the preset maximum mutation probability, and the degree of diversity.

[0247] It should be understood that the specific implementation of the above steps can refer to steps B1 to B10 in Embodiment 3, and the population and individuals can be determined in combination with the specific application scenario. This embodiment will not repeat the details.

[0248] This embodiment provides an adaptive genetic algorithm execution step including: obtaining a target population and setting initialization parameters for the target population; for the current generation of the target population, randomly generating outer and inner chromosomes for individuals; determining the fitness of individuals and selecting high-quality individuals from the target population as parent individuals to enter the mating pool based on their fitness; randomly pairing parent individuals with different outer chromosomes from the mating pool, performing crossover operations on the inner chromosomes of each pair of parent individuals according to adaptive crossover probability to obtain two offspring individuals; performing mutation operations on the offspring individuals according to adaptive mutation probability and calculating the fitness of the offspring individuals; determining a preset number of individuals from the population after merging parent and offspring individuals according to their fitness as the new generation of the target population; returning to execute the relevant steps of selecting high-quality individuals from the target population as parent individuals to enter the mating pool based on their fitness, until the termination condition is reached, and outputting the best individual among all generations. This paper presents a two-layer adaptive genetic algorithm. Through a two-layer variable-length encoding, the partitioning problem is modeled as a two-layer optimization problem: the outer layer optimizes the number of segments, and the inner layer optimizes boundary points. A variable-length chromosome encoding scheme is designed, enabling the algorithm to adaptively explore partitioning schemes of different granularities, solving the problem of manual pre-setting required by traditional fixed-segmentation methods. A crossover operator and three types of mutation operations (boundary point perturbation, insertion, and deletion) are designed, allowing the algorithm to search simultaneously in both the segment number and boundary point dimensions, balancing local fine-tuning and global granularity changes, effectively addressing the evolutionary imbalance problem under variable-length encoding. An adaptive adjustment mechanism based on population diversity indices, using adaptive crossover and mutation probabilities, regulates the evolutionary process at a macro level, effectively avoiding population diversity loss due to premature dominance by a few elite individuals under conditions of small sample size. This adaptive adjustment mechanism is naturally adapted to the two-layer variable-length encoding structure, automatically sensing the distribution of individuals with different segment numbers and adaptively adjusting the intensity of genetic operations.

[0249] Example 5

[0250] Figure 5 This is a schematic diagram of the structure of a roller fault diagnosis device provided in Embodiment 5 of the present invention. Figure 5 As shown, the device includes: a signal acquisition module 510, a feature processing module 520, an impact intensity determination module 530, an energy distribution field determination module 540, and a fault diagnosis module 550; wherein:

[0251] The signal acquisition module 510 is used to acquire the raw vibration signal of the roller under test through each sensing point in the distributed optical fiber sensor.

[0252] The feature processing module 520 is used to determine the target statistical feature set of each sensing point based on the original vibration signal collected by the sensing point and the instantaneous amplitude signal and Hilbert envelope spectrum corresponding to the intrinsic mode function components of the original vibration signal.

[0253] The impact intensity determination module 530 is used to determine the impact intensity of the fault signal at each sensing point based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal.

[0254] The energy distribution field determination module 540 is used to determine the fault energy distribution field based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor.

[0255] The fault diagnosis module 550 is used to determine the fault information of the idler roller under test based on the fault energy distribution field.

[0256] This invention provides a roller fault diagnosis device that collects the original vibration signal of the roller under test through each sensing point in a distributed optical fiber sensor. For each sensing point, a target statistical feature set is determined based on the original vibration signal collected by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum. For each sensing point, the fault signal impact intensity of the sensing point is determined based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal. A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor. The fault information of the roller under test is determined based on the fault energy distribution field. By extracting a high-quality target statistical feature set from a small number of fault samples through mode decomposition and Hilbert transform, the dependence on the amount of fault data is greatly reduced. The fault information of the roller under test is determined by the fault energy distribution field of the combined qualitative and quantitative fault results, ensuring the accuracy of fault diagnosis in high-noise environments and facilitating early, accurate, and reliable fault warning.

[0257] Optionally, the feature processing module 520 includes:

[0258] The mode decomposition unit is used to perform mode decomposition on the original vibration signal acquired by the sensing point for each sensing point to obtain multiple intrinsic mode function components.

[0259] The Hilbert transform unit is used to perform Hilbert transform on each of the intrinsic mode function components to obtain the instantaneous amplitude signal and the Hilbert envelope spectrum.

[0260] The initial feature set determination unit is used to determine the initial statistical feature set of the sensing point based on the original vibration signal, the instantaneous amplitude signals corresponding to each intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum.

[0261] The target feature set determination unit is used to select a target statistical feature set containing a preset number of statistical features from the initial statistical feature set.

[0262] Optionally, the initial feature set determination unit is specifically used for:

[0263] The target sliding window parameters are determined based on an adaptive genetic algorithm, and the temporal statistical characteristics of the original vibration signal and the instantaneous amplitude signal within the sliding window corresponding to the target sliding window parameters are determined.

[0264] The target frequency band is determined based on an adaptive genetic algorithm, and the power spectrum of the original vibration signal and the frequency domain statistical characteristics of the Hilbert envelope spectrum within the target frequency band are determined.

[0265] An initial statistical feature set is constructed based on the time-series statistical features and / or the frequency-domain statistical features.

[0266] Optionally, the execution steps of the adaptive genetic algorithm include:

[0267] The target population is determined based on the initial parameter set, and the initialization parameters of the target population are set; wherein, the initial parameter set includes a sliding window parameter set or a frequency band set; the target parameters in the sliding window parameter set are the target sliding window parameters; the target parameters in the frequency band set are the target frequency bands;

[0268] For the current generation of the target population, the outer and inner chromosomes of individuals are randomly generated.

[0269] The individual's F1 score in the validation dataset is determined as its fitness, and high-quality individuals are selected from the target population as parent individuals to enter the mating pool based on the fitness.

[0270] From the mating pool, parent individuals with different outer chromosomes are randomly paired, and crossover operation is performed on the inner chromosomes of each pair of parent individuals according to the adaptive crossover probability to obtain two offspring individuals.

[0271] The offspring individuals are subjected to mutation operations based on adaptive mutation probabilities, and the fitness of the offspring individuals is calculated.

[0272] A predetermined number of individuals are selected from the population resulting from the merging of the parent and offspring individuals, based on their fitness, to form the target population for the next generation.

[0273] Return to the execution of the steps related to selecting superior individuals from the target population as parent individuals to enter the mating pool based on the fitness of the individual, until the termination condition is met, output the best individual among the individuals of each generation, and determine the best individual as the target parameter in the initial parameter set.

[0274] Optionally, determining the target population based on the initial parameter set includes:

[0275] The initial parameter set is determined as the initial population;

[0276] Calculate the inter-class and intra-class distance ratios of the training set for each individual in the initial population;

[0277] Individuals in the population are screened based on the inter-class and intra-class distance ratios to obtain the target population.

[0278] Optionally, the execution steps of the adaptive genetic algorithm further include:

[0279] The ratio of the standard deviation of the fitness of each individual in the current generation of the target population to the maximum fitness is determined as the diversity level of the current generation of the target population.

[0280] The adaptive crossover probability is determined based on the preset minimum crossover probability, the preset maximum crossover probability, and the degree of diversity.

[0281] The adaptive mutation probability is determined based on the preset minimum mutation probability, the preset maximum mutation probability, and the degree of diversity.

[0282] Optionally, the target feature set determination unit includes:

[0283] The latent feature set determination subunit is used to determine the average information gain corresponding to each statistical feature in the initial statistical feature set based on the decision tree model; and to determine the set of statistical features in the initial statistical feature set whose average information gain is greater than the gain threshold as the latent statistical feature set.

[0284] The candidate feature set determination subunit is used to evaluate the probability of significant difference between prior fault samples and normal samples on each statistical feature in the potential statistical feature set based on nonparametric statistical test methods; the set of statistical features in the potential statistical feature set whose probability of significant difference is less than a set significance level is determined as the candidate statistical feature set;

[0285] The target feature set determination subunit is used to obtain the preset number of features corresponding to the target statistical feature set, sort the statistical features in the candidate statistical feature set in ascending order according to the difference significance probability, and determine the set of statistical features with the highest preset number of features as the target statistical feature set.

[0286] Optionally, the target feature set determining subunit is specifically used for:

[0287] The historical vibration signals collected by the distributed optical fiber sensor are acquired, and the training dataset and validation dataset are determined based on the historical vibration signals.

[0288] Determine the candidate statistical feature set of the training data corresponding to the training dataset;

[0289] The statistical features in the candidate statistical feature set of the training data are sorted in ascending order according to the probability of significant difference of each statistical feature in the candidate statistical feature set of the training data.

[0290] According to multiple different preset extraction quantities, the target statistical feature set of the training data is extracted sequentially from the candidate statistical feature set of the training data based on the sorting result.

[0291] For each training data target statistical feature set, a classifier is trained based on the training data target statistical feature set, and the classifier is validated based on the validation dataset to obtain the F1 score of the validation dataset;

[0292] The preset number of features is determined by the preset number of extractions of the target statistical feature set of the training data used by the classifier corresponding to the validation set with the highest F1 score.

[0293] Optionally, the impact strength determination module 530 is specifically used for:

[0294] For each sensing point, the fault confidence of the target statistical feature set is predicted based on the fault diagnosis model;

[0295] Determine the mean instantaneous amplitude of the instantaneous amplitude signal corresponding to the k-th eigenmode function component of the original vibration signal;

[0296] The fault signal impact intensity at the sensing point is determined by the weighted sum of the instantaneous amplitude mean values ​​corresponding to the fault confidence and the first K eigenmode function components, respectively; k≤K, and both k and K are positive integers.

[0297] Optionally, the fault diagnosis module 550 includes:

[0298] The potential fault point determination unit is used to search for local maximum points in the fault energy distribution field within the detection period, and determine the local maximum points as potential fault points.

[0299] A clustering unit is used to cluster active sensing points in the spatial neighborhood of the potential fault point to obtain a cluster; wherein the fault signal impact intensity of the active sensing point is greater than an intensity threshold.

[0300] A spatial attenuation distribution fitting unit is used to fit the spatial attenuation distribution of the fault signal impact intensity of the sensing points within the cluster.

[0301] The target fault point determination unit is used to determine the target fault point based on the number of potential fault points and the spatial attenuation distribution.

[0302] Optionally, the target fault point determination unit is specifically used for:

[0303] If the number of potential fault points is one, and the spatial attenuation distribution of the potential fault point conforms to the single-source attenuation model, then the potential fault point is determined as the target fault point.

[0304] If the number of potential fault points is greater than one, then each of the potential fault points that does not belong to the symmetrical potential fault point group and whose spatial attenuation distribution conforms to the single-source attenuation model shall be determined as the target fault point; the potential fault point in the symmetrical potential fault point group with a larger fault signal impact intensity attenuation ratio in the spatial attenuation distribution shall be determined as the target fault point.

[0305] Optionally, the fault diagnosis module 550 further includes:

[0306] The target fault point counting module is used to count target fault points in the same spatial location and whose fault signal impact intensity is greater than the intensity threshold within multiple consecutive detection cycles.

[0307] The actual fault point determination unit is used to determine the target fault point with a count value greater than the count threshold as the actual fault point of the idler roller under test, and to determine the spatial location of the actual fault point.

[0308] The idler roller fault diagnosis device provided in the embodiments of the present invention can execute the idler roller fault diagnosis method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0309] Example 6

[0310] Figure 6 A schematic diagram of an electronic device 10 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers, and may also be DAS devices, fiber optic sensing systems, fault monitoring systems, etc. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0311] like Figure 6As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0312] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0313] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as idler roller fault diagnosis methods.

[0314] In some embodiments, the idler roller fault diagnosis method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded into and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the idler roller fault diagnosis method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the idler roller fault diagnosis method by any other suitable means (e.g., by means of firmware).

[0315] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0316] In some embodiments, the idler roller fault diagnosis method may be implemented as a computer program, which is implicitly included in a computer program product. When executed by a processor, the computer program implements the idler roller fault diagnosis method of the present invention. The computer program product can be understood as a software product that primarily implements its solution through a computer program. The computer program used to implement the method of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer program causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer program may be executed entirely on a machine, partially on a machine, partially on a remote machine as a standalone software package, or entirely on a remote machine or server.

[0317] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0318] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0319] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0320] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0321] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0322] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for diagnosing idler roller faults, characterized in that, include: The raw vibration signal of the roller under test is acquired by each sensing point in the distributed optical fiber sensor. For each sensing point, the target statistical feature set of the sensing point is determined based on the original vibration signal collected by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum. For each sensing point, the fault signal impact intensity of the sensing point is determined based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal. A fault energy distribution field is formed based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor. The fault information of the idler roller under test is determined based on the fault energy distribution field. For each sensing point, based on the original vibration signal acquired by the sensing point and the instantaneous amplitude signal and Hilbert envelope spectrum corresponding to the intrinsic mode function components of the original vibration signal, the target statistical feature set of the sensing point is determined, including: For each sensing point, the original vibration signal collected at the sensing point is subjected to modal decomposition to obtain multiple intrinsic mode function components; Perform Hilbert transform on each of the intrinsic mode function components to obtain the instantaneous amplitude signal and the Hilbert envelope spectrum; The initial statistical feature set of the sensing point is determined based on the original vibration signal, the instantaneous amplitude signals corresponding to each intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum. From the initial statistical feature set, a target statistical feature set containing a preset number of statistical features is selected; For each sensing point, the fault signal impact intensity of the sensing point is determined based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function components of the original vibration signal, including: For each sensing point, the fault confidence of the target statistical feature set is predicted based on the fault diagnosis model, and the mean instantaneous amplitude of the instantaneous amplitude signal corresponding to the k-th intrinsic mode function component of the original vibration signal is determined. The weighted sum of the fault confidence and the mean instantaneous amplitude of the first K intrinsic mode function components is determined as the fault signal impact intensity of the sensing point; k≤K, and k and K are both positive integers.

2. The method according to claim 1, characterized in that, The step of determining the initial statistical feature set of the sensing point based on the original vibration signal, the instantaneous amplitude signals corresponding to each eigenmode function component of the original vibration signal, and the Hilbert envelope spectrum includes: The target sliding window parameters are determined based on an adaptive genetic algorithm, and the temporal statistical characteristics of the original vibration signal and the instantaneous amplitude signal within the sliding window corresponding to the target sliding window parameters are determined. The target frequency band is determined based on an adaptive genetic algorithm, and the power spectrum of the original vibration signal and the frequency domain statistical characteristics of the Hilbert envelope spectrum within the target frequency band are determined. An initial statistical feature set is constructed based on the time-series statistical features and / or the frequency-domain statistical features.

3. The method according to claim 2, characterized in that, The execution steps of the adaptive genetic algorithm include: The target population is determined based on the initial parameter set, and the initialization parameters of the target population are set; wherein, the initial parameter set includes a sliding window parameter set or a frequency band set; the target parameters in the sliding window parameter set are the target sliding window parameters; the target parameters in the frequency band set are the target frequency bands; For the current generation of the target population, the outer and inner chromosomes of individuals are randomly generated. The individual's F1 score on the validation dataset is determined as the individual's fitness, and high-quality individuals are selected from the target population as parent individuals to enter the mating pool based on the individual's fitness. From the mating pool, parent individuals with different outer chromosomes are randomly paired, and crossover operation is performed on the inner chromosomes of each pair of parent individuals according to the adaptive crossover probability to obtain two offspring individuals. The offspring individuals are subjected to mutation operations based on adaptive mutation probabilities, and the fitness of the offspring individuals is calculated. A predetermined number of individuals are selected from the population resulting from the merging of the parent and offspring individuals, based on their fitness, to form the target population for the next generation. Return to the execution of the steps related to selecting superior individuals from the target population as parent individuals to enter the mating pool based on the fitness of the individual, until the termination condition is met, output the best individual among the individuals of each generation, and determine the best individual as the target parameter in the initial parameter set.

4. The method according to claim 3, characterized in that, The step of determining the target population based on the initial parameter set includes: determining the initial parameter set as the initial population, calculating the inter-class and intra-class distance ratio of the training set corresponding to each individual in the initial population, and filtering the individuals in the population based on the inter-class and intra-class distance ratio to obtain the target population; And / or, The execution steps of the adaptive genetic algorithm further include: determining the ratio of the standard deviation of the fitness of each individual in the current generation to the maximum fitness as the diversity level of the current generation target population; determining the adaptive crossover probability based on the preset minimum crossover probability, the preset maximum crossover probability and the diversity level; and determining the adaptive mutation probability based on the preset minimum mutation probability, the preset maximum mutation probability and the diversity level.

5. The method according to claim 1, characterized in that, The step of selecting a target statistical feature set containing a preset number of statistical features from the initial statistical feature set includes: The average information gain corresponding to each statistical feature in the initial statistical feature set is determined based on the decision tree model; the set of statistical features in the initial statistical feature set whose average information gain is greater than the gain threshold is determined as the potential statistical feature set; The probability of significant differences between prior fault samples and normal samples on each statistical feature in the potential statistical feature set is evaluated based on nonparametric statistical test methods; the set of statistical features in the potential statistical feature set whose probability of significant difference is less than a set significance level is determined as the candidate statistical feature set; Obtain the preset number of features corresponding to the target statistical feature set, sort the statistical features in the candidate statistical feature set in ascending order according to the difference significance probability, and determine the set of statistical features with the highest preset number of features as the target statistical feature set.

6. The method according to claim 5, characterized in that, The step of obtaining the preset number of features corresponding to the target statistical feature set includes: The historical vibration signals collected by the distributed optical fiber sensor are acquired, and the training dataset and validation dataset are determined based on the historical vibration signals. Determine the candidate statistical feature set of the training data corresponding to the training dataset; The statistical features in the candidate statistical feature set of the training data are sorted in ascending order according to the probability of significant difference of each statistical feature in the candidate statistical feature set of the training data. According to multiple different preset extraction quantities, the target statistical feature set of the training data is extracted sequentially from the candidate statistical feature set of the training data based on the sorting result. For each training data target statistical feature set, a classifier is trained based on the training data target statistical feature set, and the classifier is validated based on the validation dataset to obtain the F1 score of the validation dataset; The preset number of features is determined by the preset number of extractions of the target statistical feature set of the training data used by the classifier corresponding to the validation set with the highest F1 score.

7. The method according to claim 1, characterized in that, Determining the fault information of the idler roller under test based on the fault energy distribution field includes: During the detection period, local maximum points in the fault energy distribution field are searched, and these local maximum points are identified as potential fault points. Active sensor points in the spatial neighborhood of the potential fault points are clustered to obtain clusters, wherein the fault signal impact intensity of the active sensor points is greater than an intensity threshold. The spatial attenuation distribution of the fault signal impact intensity of the sensor points in the clusters is fitted, and the target fault point is determined based on the number of potential fault points and the spatial attenuation distribution.

8. A roller fault diagnosis device, characterized in that, include: The signal acquisition module is used to acquire the raw vibration signal of the roller under test through each sensing point in the distributed optical fiber sensor. The feature processing module is used to determine the target statistical feature set of each sensing point based on the original vibration signal collected by the sensing point, the instantaneous amplitude signal corresponding to the intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum. The impact intensity determination module is used to determine the impact intensity of the fault signal at each sensing point based on the fault confidence predicted by the fault diagnosis model for the target statistical feature set and the instantaneous amplitude signal of the intrinsic mode function component of the original vibration signal. An energy distribution field determination module is used to determine the fault energy distribution field based on the fault signal impact intensity and spatial location corresponding to each sensing point of the distributed optical fiber sensor. The fault diagnosis module is used to determine the fault information of the idler roller under test based on the fault energy distribution field. The feature processing module includes: The mode decomposition unit is used to perform mode decomposition on the original vibration signal acquired by the sensing point for each sensing point to obtain multiple intrinsic mode function components. The Hilbert transform unit is used to perform Hilbert transform on each of the intrinsic mode function components to obtain the instantaneous amplitude signal and the Hilbert envelope spectrum. The initial feature set determination unit is used to determine the initial statistical feature set of the sensing point based on the original vibration signal, the instantaneous amplitude signals corresponding to each intrinsic mode function component of the original vibration signal, and the Hilbert envelope spectrum. The target feature set determination unit is used to select a target statistical feature set containing a preset number of statistical features from the initial statistical feature set; The impact strength determination module is specifically used for: For each sensing point, the fault confidence of the target statistical feature set is predicted based on the fault diagnosis model; Determine the mean instantaneous amplitude of the instantaneous amplitude signal corresponding to the k-th eigenmode function component of the original vibration signal; The fault signal impact intensity at the sensing point is determined by the weighted sum of the instantaneous amplitude mean values ​​corresponding to the fault confidence and the first K eigenmode function components, respectively; k≤K, and both k and K are positive integers.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform the idler roller fault diagnosis method according to any one of claims 1-7.

Images