Device lubrication poor intelligent identification method and device for device integrity management

Abstract

The application provides a device lubrication poor intelligent identification method and device for device integrity management, relates to the technical field of signal processing, and comprises the following steps: obtaining a vibration signal of a key rotating device of a coal mine, inputting the vibration signal into a pretreatment model, performing noise reduction and standardization processing, and obtaining a standardized signal; wherein the pretreatment model is used for adaptive noise reduction strategy based on dynamic wavelet packet decomposition to perform noise reduction processing on the vibration signal; inputting the standardized signal into a state identification model to obtain a lubrication state of the key rotating device of the coal mine; the state identification model comprises a multiscale convolution module, a space-time attention weighting module, a feature fusion module and a classifier; the state identification model is used for selectively enhancing and fusing the multiscale signal features of the standardized signal in the time domain and the frequency domain, and classifying based on the fused signal features to obtain the lubrication state of the key rotating device of the coal mine. The embodiment of the application can improve the model stability and state identification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and data processing, in particular to a device lubrication poor intelligent identification method and device for device integrity management. BACKGROUND

[0002] With the continuous improvement of the mechanization and intelligentization of coal mines, key rotating devices such as coal mining machines, scraper conveyors, and hydraulic supports are facing various fault risks such as insufficient lubrication, accelerated wear, and structural loosening under high-strength and continuous operation conditions. If these faults cannot be identified in time, they will seriously threaten the safety of coal mine production and the service life of the equipment. Traditional device fault detection methods mostly rely on manual inspection or threshold alarm based on experience rules, which have problems such as response lag, inaccurate fault identification, poor adaptability to complex working conditions, and are difficult to meet the needs of modern coal mines for real-time, intelligent, and accurate operation and maintenance.

[0003] In the prior art, traditional vibration signal denoising methods mostly rely on fixed thresholds, which can easily lead to the loss of key signal features when facing different device working conditions and non-stationary noise, and cannot accurately reflect the actual state of the device. Existing fault diagnosis methods often only focus on time domain or frequency domain features, ignoring the interaction between the two, and often fail to capture the coupling relationship between multiple features. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a device lubrication poor intelligent identification method and device for device integrity management, which effectively filters out non-stationary noise through dynamic wavelet packet denoising, while retaining the unique high-frequency transient impact and harmonic features of poor lubrication, avoiding the misfiltration of high-frequency fault signals, and providing high-quality data for subsequent feature extraction. The multi-scale convolution module covers features of different time and frequency scales, and combines with the enhancement of key features by the spatiotemporal attention mechanism to avoid missed judgment due to fault features being submerged.

[0005] In the first aspect, the present application embodiment provides a device lubrication poor intelligent identification method for device integrity management, comprising:

[0006] obtaining a vibration signal of a key rotating device of a coal mine;

[0007] inputting the vibration signal into a preprocessing model to perform denoising and standardization processing on the vibration signal, and obtaining a standardized signal; wherein the preprocessing model is used for adaptive denoising strategy based on dynamic wavelet packet decomposition to perform denoising processing on the vibration signal;

[0008] input the standardized signal into a state recognition model to obtain the lubrication state of the coal mine key rotating equipment; the state recognition model comprises a multi-scale convolution module, a space-time attention weighting module, a feature fusion module and a classifier; the state recognition model is used for selectively enhancing and fusing multi-scale signal features of the standardized signal in the time domain and the frequency domain, and classifying based on the fused signal features to obtain the lubrication state of the coal mine key rotating equipment.

[0009] In the preferred embodiment of the present application, the above-mentioned inputting the standardized signal into the state recognition model to obtain the lubrication state of the coal mine key rotating equipment comprises:

[0010] input the standardized signal into the multi-scale convolution module to extract multi-scale signal features; wherein the multi-scale convolution module is constructed based on a multi-branch structure of mixed dilated convolution and adaptive kernel width;

[0011] input the multi-scale signal features into the space-time attention weighting module to obtain a time domain attention weight matrix and a frequency domain attention weight matrix corresponding to the multi-scale signal features;

[0012] input the multi-scale signal features, the time domain attention weight matrix and the frequency domain attention weight matrix into the feature fusion module to selectively enhance and fuse the multi-scale signal features in the time domain and the frequency domain, and obtain a fusion feature matrix;

[0013] input the fusion feature matrix into the classifier to obtain the lubrication state of the coal mine key rotating equipment; wherein the classifier is constructed based on a pseudo-inverse weight matrix and a residual convolution.

[0014] In the preferred embodiment of the present application, the above-mentioned inputting the multi-scale signal features into the space-time attention weighting module to obtain the time domain attention weight matrix and the frequency domain attention weight matrix corresponding to the multi-scale signal features comprises:

[0015] input the multi-scale signal features into a time domain attention weight matrix calculation formula to obtain the time domain attention weight matrix;

[0016] input the multi-scale signal features into a frequency domain attention weight matrix calculation formula to obtain the frequency domain attention weight matrix.

[0017] In the preferred embodiment of the present application, the above-mentioned inputting the multi-scale signal features, the time domain attention weight matrix and the frequency domain attention weight matrix into the feature fusion module to selectively enhance and fuse the multi-scale signal features in the time domain and the frequency domain, and obtaining a fusion feature matrix, comprises:

[0018] determine a time-frequency domain gating weight matrix according to the time domain attention weight matrix and the frequency domain attention weight matrix;

[0019] weight the multi-scale signal features through the time domain attention weight matrix and the frequency domain attention weight matrix respectively to obtain a weighted time domain attention weight matrix and a weighted frequency domain attention weight matrix;

[0020] determine a fusion feature matrix according to the weighted time domain attention weight matrix, the weighted frequency domain attention weight matrix and the time-frequency domain gating weight matrix.

[0021] In the preferred embodiment of the present application, the above-mentioned inputting the standardized signal into the multi-scale convolution module to extract multi-scale signal features comprises:

[0022] determine the kernel width of the dynamic kernel width convolution according to the entropy value of the standardized signal;

[0023] determine the first signal feature according to the kernel width of the dynamic kernel width convolution;

[0024] determine the second signal feature according to the preset kernel width of the fundamental frequency convolution;

[0025] determine the third signal feature according to the preset expansion rate, the preset kernel width of the expansion convolution;

[0026] splice the first signal feature, the second signal feature and the third signal feature to obtain the multi-scale signal feature.

[0027] In the preferred embodiment of the present application, the above-mentioned inputting the vibration signal into the preprocessing model to perform noise reduction and standardization processing on the vibration signal to obtain a standardized signal comprises:

[0028] obtain a dynamic threshold corresponding to the vibration signal; the dynamic threshold refers to a dynamic threshold corresponding to each of a plurality of sub-signals obtained by wavelet decomposition on the vibration signal;

[0029] perform noise reduction on the vibration signal according to the dynamic threshold to obtain a noise-reduced vibration signal;

[0030] perform normalization on the noise-reduced vibration signal according to a preset sliding window to obtain a standardized signal.

[0031] In the preferred embodiment of the present application, the above-mentioned performing noise reduction on the vibration signal according to the dynamic threshold to obtain a noise-reduced vibration signal comprises:

[0032] perform inverse discrete wavelet transform according to the dynamic threshold and each of the sub-signals to obtain a time domain vibration signal;

[0033] acquire an adaptive weight corresponding to the vibration signal;

[0034] de-noise the time-domain vibration signal according to the adaptive weight to obtain a de-noised vibration signal.

[0035] In a preferred embodiment of the present application, before the standardized signal is input into the state recognition model, the method further comprises:

[0036] acquiring an initial weight of an initial state recognition model;

[0037] acquiring a plurality of sample signals and a classification label corresponding to each sample signal;

[0038] inputting each sample signal and the corresponding classification label into the initial state recognition model to train the initial state recognition model and obtain a loss function result of the initial state recognition model; wherein the loss function is determined according to a wavelet basis in the pre-processing model, a pseudo-inverse Jacobian stability term, and a time-frequency contrast term;

[0039] updating the initial weight according to the loss function result to obtain an updated weight;

[0040] counting the number of updates of the updated weight and returning to the step of inputting each sample signal and the corresponding classification label into the initial state recognition model to train the initial state recognition model and obtain a loss function result of the initial state recognition model until the number of updates is greater than or equal to a preset number of updates;

[0041] setting the initial state recognition model according to the updated weight to obtain a state recognition model.

[0042] In a preferred embodiment of the present application, the updating of the initial weight according to the loss function result to obtain an updated weight comprises:

[0043] updating the initial weight according to the loss function result to obtain a first weight;

[0044] for each sample signal, acquiring a sample lubrication state corresponding to the output of the initial state recognition model;

[0045] for each sample signal, determining a fuzzy membership degree of the sample signal according to the sample lubrication state;

[0046] adjusting the first weight according to the fuzzy membership degree of each sample signal to obtain an updated weight.

[0047] In a second aspect, the embodiments of the present application also provide a device lubrication poor intelligent identification device for device integrity management, comprising:

[0048] A signal acquisition module is configured to acquire a vibration signal of a key rotating device of a coal mine.

[0049] A preprocessing module is configured to input the vibration signal into a preprocessing model, perform noise reduction and standardization processing on the vibration signal, and obtain a standardized signal.

[0050] An identification module is configured to input the standardized signal into a state identification model, and obtain a lubrication state of the key rotating device of the coal mine.

[0051] In a third aspect, the embodiments of the present application also provide an electronic device, comprising a processor and a memory, wherein the memory stores computer executable instructions capable of being executed by the processor, and the processor executes the computer executable instructions to implement the device lubrication poor intelligent identification method for device integrity management of the first aspect.

[0052] In a fourth aspect, the embodiments of the present application also provide a computer readable storage medium, wherein the computer readable storage medium stores computer executable instructions, and the computer executable instructions, when called and executed by a processor, cause the processor to implement the device lubrication poor intelligent identification method for device integrity management of the first aspect.

[0053] The embodiments of the present application have the following beneficial effects:

[0054] The embodiments of the present application provide a device lubrication poor intelligent identification method, device, equipment and medium for device integrity management, which effectively filters out non-steady-state noise through dynamic wavelet packet noise reduction, while retaining the unique high-frequency transient impact and harmonic characteristics of poor lubrication, avoiding the misfiltering of high-frequency fault signals, and providing high-quality data for subsequent feature extraction.

[0055] Other features and advantages of the present application will be set forth in the descriptions that follow, and in part will be apparent from the description, or can be learned by practice of the application as hereinafter described, or can be learned by practice of the application.

[0056] In order to make the above objectives, features and advantages of the present application more apparent, the following will specifically describe a preferred embodiment in combination with the accompanying drawings, and the detailed description is as follows. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or the prior art description. Obviously, the drawings described below are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0058] Figure 1 A flowchart of a device lubrication poor intelligent identification method for device integrity management provided by an embodiment of the present application is shown in the figure.

[0059] Figure 2a A flowchart of another device lubrication poor intelligent identification method for device integrity management provided by an embodiment of the present application is shown in the figure.

[0060] Figure 2b A base frequency convolution feature extraction result schematic diagram provided by an embodiment of the present application is shown in the figure.

[0061] Figure 2c A fixed expansion convolution feature extraction result schematic diagram provided by an embodiment of the present application is shown in the figure.

[0062] Figure 2d A multi-scale convolution module feature extraction result schematic diagram provided by an embodiment of the present application is shown in the figure.

[0063] Figure 2e A traditional time attention distribution result schematic diagram provided by an embodiment of the present application is shown in the figure.

[0064] Figure 2f A traditional frequency domain attention distribution result schematic diagram provided by an embodiment of the present application is shown in the figure.

[0065] Figure 2g A frequency domain attention weight matrix result schematic diagram provided by an embodiment of the present application is shown in the figure.

[0066] Figure 2h A time domain attention weight matrix result schematic diagram provided by an embodiment of the present application is shown in the figure.

[0067] Figure 2iA dynamic core width adaptive mechanism performance verification result schematic diagram provided for the embodiment of the present application is shown in the figure.

[0068] Figure 3a A flow chart of another device lubrication poor intelligent identification method for device integrity management provided for the embodiment of the present application is shown in the figure.

[0069] Figure 3b A vibration signal schematic diagram is shown in the figure.

[0070] Figure 3c A layer 1 wavelet coefficient schematic diagram is shown in the figure.

[0071] Figure 3d A layer 2 wavelet coefficient schematic diagram is shown in the figure.

[0072] Figure 3e A layer 3 wavelet coefficient schematic diagram is shown in the figure.

[0073] Figure 3f A layer 4 wavelet coefficient schematic diagram is shown in the figure.

[0074] Figure 3g A layer 5 wavelet coefficient schematic diagram is shown in the figure.

[0075] Figure 3h A layer 6 wavelet coefficient schematic diagram is shown in the figure.

[0076] Figure 3i A denoised vibration signal schematic diagram is shown in the figure.

[0077] Figure 3j A standardized signal schematic diagram is shown in the figure.

[0078] Figure 4a A flow chart of another device lubrication poor intelligent identification method for device integrity management provided for the embodiment of the present application is shown in the figure.

[0079] Figure 4b A different initialization method optimization schematic diagram for state recognition model training process is shown in the figure.

[0080] Figure 5 A device lubrication poor intelligent identification device structure schematic diagram for device integrity management provided for the embodiment of the present application is shown in the figure.

[0081] Figure 6 A structure schematic diagram of an electronic device provided for the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0082] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0083] With the continuous improvement of the degree of mechanization and intelligence of coal mines, key rotating equipment such as coal mining machines, scraper conveyors, hydraulic supports and the like face various fault risks such as insufficient lubrication, aggravated wear and tear, structural loosening and the like under the working conditions of high strength and continuous operation. If these faults cannot be identified in time, they will seriously threaten the safety of coal mine production and the service life of equipment. The traditional equipment fault detection method relies on manual inspection or threshold alarm based on experience rules, and has problems such as response lag, inaccurate fault identification, poor adaptability to complex working conditions and the like, and is difficult to meet the needs of modern coal mines for real-time, intelligent and accurate operation and maintenance.

[0084] In the prior art, the traditional vibration signal denoising method relies on a fixed threshold, which easily leads to the loss of key signal features when facing different working conditions and non-steady noise of equipment, and cannot accurately reflect the actual state of the equipment. The existing fault diagnosis method often only focuses on time domain or frequency domain features, ignoring the interaction between the two. The present application can effectively combine the features of time domain and frequency domain through the time-space dual attention mechanism, enhance the expression of key fault features, and solve the shortcomings of traditional methods in feature modeling. The traditional convolutional neural network often has difficulty in capturing the coupling relationship between multiple features when processing complex time-varying signals, especially when processing transient impact and long-term wear and tear features in the vibration signal.

[0085] Based on this, the device lubrication poor intelligent identification method for device integrity management provided by the embodiment of the application can effectively eliminate strong noise (such as mechanical vibration and electromagnetic interference) in a complex coal mine environment through the adaptive noise reduction of dynamic wavelet packet decomposition, and retain weak fault signals (such as vibration characteristics of early bearing wear), compared with a traditional fixed threshold method, the method avoids false filtering of high-frequency fault signals, and provides high-quality data for subsequent feature extraction. Through standardization processing, the signal amplitude difference under different devices and different working conditions can be eliminated, so that the model can be applicable to multiple devices of the same type (such as fans of the same type in different mines), and the repeated training cost is reduced. The multi-scale convolution module covers features of different time and frequency scales, and combines with the enhancement of key features by the time-space attention mechanism, so as to accurately locate the key position of the fault in the time domain and the frequency domain, enhance the identification ability of the “time-frequency coupling feature” caused by poor lubrication, and avoid missed judgment (such as high-frequency features of slight pitting of a gear) due to the fault features being submerged.

[0086] In order to facilitate the understanding of the embodiment, first, a device lubrication poor intelligent identification method for device integrity management disclosed by the embodiment of the application is introduced in detail.

[0087] Embodiment 1

[0088] The embodiment of the application provides a device lubrication poor intelligent identification method for device integrity management, Figure 1 A flowchart of the device lubrication poor intelligent identification method for device integrity management provided by the embodiment of the application is shown as shown in the figure. Figure 1 The device lubrication poor intelligent identification method for device integrity management can include the following steps:

[0089] In step S101, the vibration signal of the key rotating device of the coal mine is acquired.

[0090] The vibration signal is the vibration signal of the key rotating device of the coal mine, such as the vibration signal of the bearing of the cutting part of the coal mining machine, the gear box of the reducer of the scraper conveyor and the like. The vibration signal generated during the operation of the device is collected through the vibration sensor installed on the key rotating device of the coal mine.

[0091] In the embodiment of the present application, the vibration sensor can be a three-axis acceleration sensor installed on the bearing seat or the gear box shell surface of the equipment, the sampling frequency is set to 10 kHz, covering the base frequency (0-500 Hz) and its harmonic frequency band of the equipment; the vibration signal can be collected in real time through the edge computing gateway, preprocessed by an anti-aliasing filter (cut-off frequency 5 kHz), and stored as an original time domain waveform in HDF5 format. The original time domain waveform contains the normal vibration characteristics of the equipment, but also mixes noise, and the signal may be affected by environmental interference, unstable equipment state and other factors.

[0092] In step S102, the vibration signal is input into a preprocessing model for noise reduction and standardization processing to obtain a standardized signal.

[0093] The preprocessing model is used for adaptive noise reduction strategy based on dynamic wavelet packet decomposition to perform noise reduction processing on the vibration signal. The vibration signal has multi-frequency noise interference and dynamic amplitude fluctuation characteristics, and the conventional fixed threshold filtering method cannot adapt to the noise distribution difference under different working conditions, resulting in loss of feature information.

[0094] The embodiment of the present application adopts an adaptive noise reduction strategy based on dynamic wavelet packet decomposition, combined with a sliding window standardization method, to eliminate the dimensional difference and non-stationary noise interference of the vibration signal. Specifically, the vibration signal can be decomposed into sub-signals of different frequency bands through wavelet packet decomposition. The noise components in each sub-signal are removed through adaptive threshold processing (such as setting the noise fluctuations below the threshold to zero), and then the noise-reduced signal is reconstructed to retain the effective vibration features related to faults. For example, the adaptive threshold can be dynamically adjusted by a sliding window with a preset width to statistically analyze the local noise energy, thereby avoiding the false filtering of high-frequency transient impact features caused by fixed threshold, and solving the problem of signal amplitude variation caused by load fluctuation of coal mine equipment. For example, when the equipment is suddenly loaded, the noise is enhanced, and the threshold is automatically increased to retain the useful signal; when the load is stable, the threshold is reduced to fine noise reduction.

[0095] During the decomposition process, the weights of the sub-signals of each frequency band can be dynamically adjusted by the signal-to-noise ratio (SNR), the fault feature frequency band (such as the harmonic frequency band of poor lubrication) with high SNR is enhanced, and the noise frequency band with low SNR is suppressed. Compared with the fixed threshold, the high-frequency transient impact signal of the bearing jamming (which is easily misjudged as noise by traditional methods) can be retained, and at the same time, the low-frequency mechanical interference in the coal mine field can be filtered out.

[0096] The noise-reduced signal is converted into a standardized signal with a mean of 0 and a variance of 1, the influence of signal amplitude differences in different devices and different working conditions (such as the vibration amplitude of the same device under light load and heavy load) is eliminated, the consistency of subsequent model input is ensured, the signal is smoother in the time dimension, and the influence of local abnormal fluctuations (such as transient voltage interference) on subsequent feature extraction is avoided.

[0097] In step S103, the standardized signal is input into the state recognition model to obtain the lubrication state of the coal mine key rotating equipment.

[0098] The state recognition model includes a multi-scale convolution module, a space-time attention weighting module, a feature fusion module, and a classifier; the state recognition model is used for selective enhancement and fusion of multi-scale signal features of the standardized signal in the time domain and the frequency domain, and classification based on the fused signal features to obtain the lubrication state of the coal mine key rotating equipment.

[0099] The application adopts a convolutional neural network as a state recognition model, including a multi-scale convolution module, a space-time attention weighting module, an attention feature fusion module, and a multi-layer fully connected neural network classifier module. The standardized signal is input into the state recognition model, and through multi-scale feature extraction, space-time attention enhancement, feature fusion, and classification, the lubrication state of the device (such as normal state, insufficient lubrication, grease contamination, abnormal wear, etc.) is output.

[0100] Specifically, the multi-scale convolution module can be composed of multiple convolution layers with different kernel sizes (such as 1x3, 1x5, and 1x7 convolution kernels), which respectively extract features of the signal in different time scales. For example, a small convolution kernel captures high-frequency short-time features (such as transient impact), and a large convolution kernel captures low-frequency long-time features (such as periodic vibration of rotor imbalance). The space-time attention weighting module is divided into time domain attention weighting and frequency domain attention weighting, wherein the time domain attention weighting is used to assign weights to the features output by the multi-scale convolution module at different time points, highlighting the signal at the moment of failure (such as the sudden vibration of bearing jamming). The frequency domain attention weighting is used to assign weights to the features output by the multi-scale convolution module in different frequency bands, enhancing the frequency components related to the fault (such as the meshing frequency harmonics of gear tooth breakage). The feature fusion module is used to fuse the multi-scale convolution features after space-time attention weighting, integrating the global features in the time domain (time variation) and the frequency domain (frequency distribution), and avoiding the one-sidedness of a single feature. The classifier usually adopts a fully connected neural network or a softmax layer, and outputs the classification results of the lubrication state (such as normal state, insufficient lubrication, grease contamination, and abnormal wear) based on the fused features, and gives the confidence (such as 95% probability of insufficient lubrication).

[0101] The device lubrication poor intelligent identification method for device integrity management provided by the embodiment of the application can effectively eliminate strong noise (such as mechanical vibration and electromagnetic interference) in a complex coal mine environment and retain weak fault signals (such as vibration characteristics of early bearing wear) through adaptive noise reduction of dynamic wavelet packet decomposition, avoids false filtering of high-frequency fault signals compared with a traditional fixed threshold method, and provides high-quality data for subsequent feature extraction. Through standardization processing, signal amplitude differences under different devices and different working conditions can be eliminated, the model can be applicable to multiple devices of the same type (such as fans of the same type in different mines), and repeated training costs are reduced. The multi-scale convolution module covers features of different time and frequency scales, and combines with enhancement of key features by a time-space attention mechanism to accurately locate the key position of the fault in the time domain and the frequency domain, enhance the identification ability of the "time-frequency coupling feature" caused by poor lubrication, and avoid missed judgment (such as high-frequency features of slight pitting of a gear) due to the fault features being submerged. The classifier outputs specific fault types (such as distinguishing between "poor lubrication" and "bearing wear") based on the fused global features, avoids fuzzy judgment of a traditional method (such as artificial auscultation), and improves the pertinence of maintenance (such as determining a maintenance scheme without disassembly and inspection).

[0102] Embodiment 2

[0103] The embodiment of the application further provides another device lubrication poor intelligent identification method for device integrity management; the method is implemented on the basis of the method in the above embodiment; and the method mainly describes a specific implementation manner of inputting the standardized signal into a state recognition model to obtain a lubrication state of the coal mine key rotating device.

[0104] Figure 2a As shown in the flowchart of another device lubrication poor intelligent identification method for device integrity management provided by the embodiment of the application, Figure 2a the device lubrication poor intelligent identification method for device integrity management can include the following steps:

[0105] In step S201, a vibration signal of a coal mine key rotating device is acquired.

[0106] In step S202, the vibration signal is input into a preprocessing model, noise reduction and standardization processing are performed on the vibration signal, and a standardized signal is obtained.

[0107] In step S203, the standardized signal is input into a multi-scale convolution module, and multi-scale signal features are extracted.

[0108] The multi-scale convolution module is constructed based on a mixed dilated convolution and an adaptive kernel width multi-branch structure.

[0109] The fixed convolution kernel of the conventional convolutional neural network is difficult to capture the coupling relationship between the sudden transient characteristics and the long-term wear characteristics in the vibration signal. The multi-branch structure of the mixed dilated convolution and the adaptive kernel width convolution is adopted, the signal characteristics in different time scales are extracted through the combination of multi-scale convolution, and then the multi-level feature fusion of the vibration signal is realized.

[0110] Specifically, the standardized signal is input into the multi-scale convolution module to extract multi-scale signal characteristics, including: determining the kernel width of the dynamic kernel width convolution according to the entropy value of the standardized signal; determining the first signal feature according to the kernel width of the dynamic kernel width convolution; determining the second signal feature according to the preset kernel width of the fundamental frequency convolution; determining the third signal feature according to the preset dilated rate, the preset kernel width of the dilated convolution; and splicing the first signal feature, the second signal feature and the third signal feature to obtain the multi-scale signal feature.

[0111] The multi-scale convolution module is represented by the following formula:

[0112]

[0113] The multi-scale signal feature combines the signal features under different convolution kernel sizes, couples the multi-scale features, including the fundamental frequency harmonic, the long-term wear and the transient impact, and solves the problem of insufficient modeling of the time-varying fault characteristics in the traditional single-scale convolution.

[0114] The fundamental frequency convolution is used to calculate the second signal feature, The vibration signal after noise reduction is the standardized signal, The kernel width of the fundamental frequency convolution is set to 64.

[0115] The dilated convolution is used to calculate the third signal feature, and d is the dilated rate, The kernel width of the dilated convolution is set to 32. l is the layer number, The dilated rate exponentially increases with the layer number, and the receptive field is expanded to capture the long-term wear characteristics, such as the periodic amplitude modulation caused by the wear of the gear box.

[0116] The dynamic kernel width convolution is used to calculate the first signal feature, The kernel width of the dynamic kernel width convolution is set to 32. The calculation method of the dynamic kernel width convolution is represented as:

[0117]

[0118] ​​ The kernel width can be dynamically adjusted according to the entropy value of the vibration signal. When the entropy value is high, for example, the complexity of the transient impact signal is high, the kernel width is increased to capture the characteristics of a wider time window, and vice versa, the kernel width is reduced to focus on details. Wherein, represents the calculation of the entropy value of the vibration signal, which is used to measure the complexity of the signal. is a kernel width mapping function, and the calculation method is represented as:

[0119]

[0120] wherein, is a hyperbolic tangent function, is a rounding operation.

[0121] represents that different convolution outputs are spliced on the channel dimension to form a multi-channel feature, and a multi-scale signal feature is obtained. When splicing, in order to unify the output channel dimension of the multi-branch structure, the embodiment of the application adopts 1*1 convolution for feature channel alignment, that is, through 1*1 convolution, the number of channels of each branch is mapped to a unified dimension, ensuring the effectiveness of subsequent feature fusion.

[0122] In step S204, the multi-scale signal feature is input into a space-time attention weighting module to obtain a time domain attention weight matrix and a frequency domain attention weight matrix corresponding to the multi-scale signal feature.

[0123] The time domain attention weight matrix is used for selectively enhancing the multi-scale signal feature in the time domain. The frequency domain attention matrix is used for selectively enhancing the multi-scale signal feature in the frequency domain.

[0124] In the process of processing the multi-scale signal feature, the traditional pseudo-inverse learning lacks attention to the local time-frequency characteristics of the vibration signal. The embodiment of the application adopts a space-time double attention weight matrix, dynamically weights the features through the space-time attention weight matrix, and realizes the selective enhancement of the key fault features in the time domain and the frequency domain.

[0125] Specifically, the multi-scale signal feature is input into a space-time attention weighting module to obtain a time domain attention weight matrix and a frequency domain attention weight matrix corresponding to the multi-scale signal feature, including: inputting the multi-scale signal feature into the time domain attention weight matrix calculation formula to obtain the time domain attention weight matrix; inputting the multi-scale signal feature into the frequency domain attention weight matrix calculation formula to obtain the frequency domain attention weight matrix.

[0126] The calculation formula of the time domain attention weight matrix can be represented as:

[0127]

[0128] is a time-domain attention weight matrix, representing the correlation of each time step in the time series, and strengthening the fault-sensitive time points, such as the starting time of the impact event; is a Softmax function; is a scaling factor; is a time attention query matrix, and the calculation method is represented as:

[0129]

[0130] is a time-domain attention query weight matrix, which is a trainable parameter, is a multi-scale convolution module output feature.

[0131] is a time attention key matrix, and the calculation method is represented as:

[0132]

[0133] is the transpose of is a time-domain attention key weight matrix, which is a trainable parameter.

[0134] The calculation formula of the frequency domain attention weight matrix can be represented as:

[0135]

[0136] is a frequency domain attention weight matrix, representing the correlation of the frequency domain features, and selectively enhancing the fault-related frequency band by using a diagonal matrix, such as the feature frequency sideband caused by poor lubrication; represents the diagonalization of the matrix, that is, the diagonal line elements are retained, and other elements are set to 0; is a Sigmoid activation function; is a frequency domain attention query weight matrix, which is a trainable parameter; is a multi-scale signal feature output by a multi-scale convolution module.

[0137] In step S205, the multi-scale signal feature, the time-domain attention weight matrix, and the frequency domain attention weight matrix are input into a feature fusion module to selectively enhance and fuse the multi-scale signal feature in the time domain and the frequency domain, and a fusion feature matrix is obtained.

[0138] ​Specifically, in the feature fusion model, first, the multi-scale signal features are weighted by the time domain attention weight matrix and the frequency domain attention weight matrix respectively, and the weighted multi-scale signal features are fused to obtain a fusion feature matrix. Based on the time domain attention weight matrix and the frequency domain attention weight matrix, the feature fusion can consider the characteristics of the equipment lubrication failure, and by combining the time domain and frequency domain signal to reconstruct the fusion feature, the extraction ability of the non-stationary impact and harmonic coupling characteristics caused by poor lubrication can be enhanced.

[0139] Further, the multi-scale signal features, the time domain attention weight matrix and the frequency domain attention weight matrix are input into the feature fusion module to selectively enhance and fuse the multi-scale signal features in the time domain and the frequency domain, and a fusion feature matrix is obtained, including: determining a time-frequency domain gating weight matrix according to the time domain attention weight matrix and the frequency domain attention weight matrix; weighting the multi-scale signal features by the time domain attention weight matrix and the frequency domain attention weight matrix respectively to obtain a weighted time domain attention weight matrix and a weighted frequency domain attention weight matrix; determining a fusion feature matrix according to the weighted time domain attention weight matrix, the weighted frequency domain attention weight matrix and the time-frequency domain gating weight matrix.

[0140] For example, the fusion feature matrix can be calculated by the following formula:

[0141]

[0142] For the fusion feature matrix, the coupling of frequency domain filtering enhancement and time domain gating is realized to achieve feature resonance, solving the problem of insufficient modeling of time-frequency coupling characteristics in traditional methods. The multi-scale signal features output by the multi-scale convolution module; The time domain attention weight matrix; The frequency domain attention weight matrix; The Fourier transform function; The inverse Fourier transform function; The Fourier transform is used to convert the feature to the frequency domain, the diagonal matrix is used to realize the selective reinforcement of the frequency band, and then the inverse Fourier transform is used to reconstruct the time domain signal, and then the frequency domain filtering enhancement is realized; The ReLU activation function; and the Hadamard product operator symbol; and the outer product operator symbol.

[0143] The time-frequency domain gating weight matrix combines time-frequency attention, dynamically modulates the feature channel weight, and enhances the representation of non-stationary impact characteristics; It can be calculated by the following formula:

[0144]

[0145] is a gating weight matrix, which is a trainable parameter; is a time domain attention weight matrix; is a frequency domain attention weight matrix; is a Sigmoid activation function.

[0146] In step S206, the fusion feature matrix is input into a classifier to obtain the lubrication state of the key rotating equipment of the coal mine.

[0147] In the embodiment of the present application, in order to verify the effectiveness of the technical scheme in the embodiment of the present application, the experimental analysis is as follows:

[0148] Figure 2b is a base frequency convolution feature extraction result schematic diagram provided by the embodiment of the present application. Figure 2c is a fixed dilated convolution feature extraction result schematic diagram provided by the embodiment of the present application. Figure 2d is a multi-scale convolution module feature extraction result schematic diagram provided by the embodiment of the present application. Figure 2b 、 Figure 2c and Figure 2d are feature extraction results of the same vibration signal. As shown in Figure 2b 、 Figure 2c and Figure 2d , the time-frequency feature extraction capabilities of different convolution structures are compared through three-dimensional spectrum thermodynamic diagrams, the capturing effect of the multi-scale convolution module on the transient impact and long-term wear coupling features in the vibration signal is verified, and the feature response distribution of the base frequency convolution, the fixed dilated convolution and the multi-scale convolution of the present application can be seen. The base frequency convolution presents Gaussian distribution characteristics in a specific frequency band, but the response to time domain mutation is insufficient. The fixed dilated convolution has a response in a wide frequency band, but the time domain resolution is limited. The technical scheme of the embodiment of the present application forms a high-resolution response peak near the impact feature in the time domain 1 second through dynamic kernel width adjustment and multi-branch fusion, and simultaneously exhibits a wide-band enhancement feature in the 200 to 400 Hz frequency band. In the time and frequency dimensions, better feature coverage is achieved, proving that the multi-scale convolution can effectively decouple the time-frequency coupling features.

[0149] Figure 2e is a traditional time attention distribution result schematic diagram provided by the embodiment of the present application. Figure 2f is a traditional frequency domain attention distribution result schematic diagram provided by the embodiment of the present application. Figure 2g is a frequency domain attention weight matrix result schematic diagram provided by the embodiment of the present application. Figure 2h is a time domain attention weight matrix result schematic diagram provided by the embodiment of the present application. Wherein, Figure 2e 、 Figure 2f 、 Figure 2g andFigure 2h The result of attention mechanism processing for the same vibration signal. By adopting the double-flow heat map to visualize the weight distribution of the space-time attention mechanism, the enhancement effect of the attention module on the key characteristics of the poor lubrication fault is verified. By comparing the attention distribution in the time domain and the frequency domain of the traditional method and the technical solution provided by the embodiment of the application, it can be seen that the traditional method presents a single Gaussian distribution in the time dimension, and only focuses on the main cycle stage of the vibration signal; the frequency domain attention is concentrated in the fundamental frequency band, and the response to the harmonic component is insufficient. The time domain attention of the technical solution provided by the embodiment of the application forms a double-peak distribution at 0.4 seconds (the moment of impact) and 1.1 seconds (the continuous wear stage), accurately locates the time domain position of the fault characteristics, and the frequency domain attention is simultaneously enhanced at the fundamental frequency and three times the frequency, effectively capturing the harmonic resonance phenomenon caused by poor lubrication, proving that the space-time attention mechanism can realize the selective enhancement of the fault characteristics in two dimensions.

[0150] Figure 2i It is a performance verification result schematic diagram of the dynamic kernel width adaptive mechanism provided by the embodiment of the application. Since the multi-scale convolution module adopts dynamic kernel width convolution in the construction process, as shown in Figure 2i The effectiveness of the dynamic kernel width adaptive mechanism is verified by the double-axis curve. The performance of fixed kernel width and dynamic kernel width under different signal complexity is compared to reveal the adaptability of the kernel width adjustment strategy to feature extraction. The experimental results show that when the signal complexity is low, the dynamic kernel width is automatically reduced to enhance the time domain resolution, and when the complexity is high, the kernel width is increased to capture the long-term dependence feature. The accuracy curve shows that the dynamic kernel width is better than the fixed kernel width in each complexity interval, especially in the medium complexity region, the performance is significantly improved. The kernel width adjustment mechanism driven by entropy effectively realizes the dynamic matching of the convolution receptive field and the signal feature, proving the adaptability of the adaptive convolution kernel design to complex working conditions.

[0151] The device poor lubrication intelligent identification method for device integrity management provided by the embodiment of the application adopts a space-time dual attention weighting mechanism, which can dynamically and selectively enhance key fault characteristics in the time domain and the frequency domain. The time attention matrix strengthens the weight of the fault sensitive time point, and the frequency domain attention focuses on the characteristic frequency band, improving the detection ability of the harmonic resonance phenomenon caused by poor lubrication. In the multi-scale convolution module, the combination of multi-scale convolution and adaptive kernel width can simultaneously capture the sudden transient impact feature and the long-term wear feature in the signal, solving the problem of insufficient modeling of time-varying fault characteristics in the traditional convolutional neural network, so that the time-frequency features of the vibration signal can be more comprehensively extracted.

[0152] Embodiment 3

[0153] The embodiment of the present application also provides another device lubrication poor intelligent identification method for device integrity management.

[0154] Figure 3a The flow chart of another device lubrication poor intelligent identification method for device integrity management provided by the embodiment of the present application is shown in Figure 3a The device lubrication poor intelligent identification method for device integrity management can include the following steps:

[0155] Step S301, the vibration signal of the key rotating device of the coal mine is acquired.

[0156] Step S302, the dynamic threshold corresponding to the vibration signal is acquired.

[0157] The dynamic threshold refers to the dynamic threshold corresponding to each of the plurality of sub-signals (wavelet coefficients) obtained by wavelet decomposition of the vibration signal. Specifically, for the non-stationary noise of the vibration signal, such as the amplitude fluctuation caused by the load change of the device, the local noise energy is statistically calculated through a sliding window to determine the dynamic threshold, thereby avoiding the false filtering of the high-frequency transient impact characteristics caused by the fixed threshold. The dynamic threshold can be determined by the following formula:

[0158]

[0159] Wherein, is the standard deviation of the k-th layer wavelet coefficient in the window [t-5, t+5] at the t-th moment, and N is the window length; k is a positive integer.

[0160] Step S303, the vibration signal is denoised according to the dynamic threshold to obtain a denoised vibration signal.

[0161] Specifically, the vibration signal is denoised according to the dynamic threshold to obtain a denoised vibration signal, including: performing inverse discrete wavelet transform according to the dynamic threshold and each of the sub-signals to obtain a time-domain vibration signal; acquiring an adaptive weight corresponding to the vibration signal; and denoising the time-domain vibration signal according to the adaptive weight to obtain the denoised vibration signal.

[0162] The denoised vibration signal can be represented by the following formula:

[0163]

[0164] is the denoised vibration signal, This represents the denoised vibration signal at time t, which is the smoothed signal obtained by performing wavelet packet decomposition and dynamic hard thresholding on the vibration signal. It is a vibration signal. The vibration signal at time t represents the vibration characteristics of the equipment as well as various noise interferences; Let k be the wavelet basis function of the k-th layer; Represents the discrete wavelet transform; Let represent the inverse discrete wavelet transform, used to recover the time-domain signal from the multiple sub-signals obtained after wavelet decomposition, thus obtaining the time-domain vibration signal; K is the number of wavelet packet decomposition levels, i.e., the number of wavelet basis functions selected during the signal decomposition process; let For v, If we consider a dynamic hard threshold function, then the formula for the dynamic hard threshold function is as follows:

[0165]

[0166] Let be the dynamic threshold of the k-th layer at time t.

[0167] The adaptive weight at time t is calculated as follows:

[0168]

[0169] Adaptive weights are used to selectively enhance high signal-to-noise ratio subbands in the frequency domain, such as the fundamental harmonic components of equipment, suppress low signal-to-noise ratio subbands such as environmental noise, and highlight harmonic distortion characteristics caused by poor lubrication.

[0170] in, It is a smoothing factor; For the first Time of the first Layer signal-to-noise ratio, calculated as follows:

[0171]

[0172] For the first Time of the first The energy of the noise signal in the layer; The vibration signal obtained after wavelet packet decomposition is the first... Layer signal at the first The energy of a moment.

[0173] Step S304: Normalize the noise-reduced vibration signal according to the preset sliding window to obtain a standardized signal.

[0174] Specifically, a fixed length sliding window is first set, and every time the signal sequence moves one step forward, the data points in the window range are extracted with the current data point as the center. Then, the mean and standard deviation of all data points in the window are calculated. Then, the current data point is subtracted from the mean of the window and then divided by the standard deviation of the window to obtain the normalized value. The process is repeated, and the window slides sequentially until the end of the signal, thereby realizing the local normalization of each data point in its adjacent range, making the signal more stable in time sequence, which is conducive to subsequent feature extraction. The preset sliding window length can be the same as the sliding window length set when determining the dynamic threshold.

[0175] In step S305, the standardized signal is input into the state recognition model to obtain the lubrication state of the coal mine key rotating equipment.

[0176] Exemplarily, Figure 3b is a schematic diagram of a vibration signal. Figure 3c is a schematic diagram of layer 1 wavelet coefficients. Figure 3d is a schematic diagram of layer 2 wavelet coefficients. Figure 3e is a schematic diagram of layer 3 wavelet coefficients. Figure 3f is a schematic diagram of layer 4 wavelet coefficients. Figure 3g is a schematic diagram of layer 5 wavelet coefficients. Figure 3h is a schematic diagram of layer 6 wavelet coefficients. As Figures 3c to 3h shown, the six diagrams show the vibration signal (original signal) shown in Figure 3b , the coefficients of each layer after wavelet packet decomposition, the wavelet transform decomposes the signal into multiple frequency bands, and each layer of wavelet coefficients represents the information of the signal in different frequency ranges. The low-frequency part contains the main trend and slow change of the signal, while the high-frequency part contains the detail information and noise. Through wavelet packet decomposition, the different frequency band components of the signal can be separated, which helps to identify the noise part and useful features in the signal. In the noise reduction process, it is usually necessary to remove the high-frequency noise components and retain the main signal part of the low frequency.

[0177] Figure 3i is a schematic diagram of the denoised vibration signal. As Figure 3i shown, the denoised signal obtained by wavelet packet decomposition and hard threshold processing, after hard threshold processing, the noise components are suppressed, and the main features of the signal are retained. As can be seen, the noise is significantly reduced. By removing the noise components in the signal, the signal becomes more stable, which is convenient for subsequent feature extraction and analysis. The denoised signal retains the normal vibration characteristics of the equipment and removes environmental noise and other interference, which helps to improve the diagnosis accuracy.

[0178] Figure 3j is a schematic diagram of the standardized signal. As Figure 3jAs shown, the noise-reduced signal after the sliding window standardization processing, in the standardization process, each data point of the signal is normalized according to its adjacent window data, eliminating the dimensional difference and the non-steady fluctuation of the signal, and the purpose of the standardization processing is to eliminate the influence of the dimensional difference or the different time scales in the signal, so that the signals under different devices or different working conditions are compared and analyzed under the same standard, and in addition, the standardized signal is smoother, which helps to eliminate local abnormal fluctuations, so that the subsequent feature extraction and model training are more stable and efficient.

[0179] The device lubrication poor intelligent identification method for device integrity management provided by the embodiment of the application adopts dynamic wavelet packet decomposition combined with hard threshold processing for adaptive noise reduction of the vibration signal, dynamically adjusts the threshold to adapt to the signal noise characteristics under different working conditions, can effectively remove non-steady noise, and retains the vibration characteristics of the device, solves the shortcomings of the traditional fixed threshold method under different working conditions, and avoids the misfiltration of high-frequency transient impact characteristics.

[0180] Embodiment 4

[0181] The embodiment of the application further provides another device lubrication poor intelligent identification method for device integrity management; the method is implemented on the basis of the method of the above-mentioned embodiment; the method mainly describes the specific implementation mode of inputting the vibration signal into the preprocessing model, performing noise reduction and standardization processing on the vibration signal, and obtaining a standardized signal.

[0182] Figure 4a The flowchart of another device lubrication poor intelligent identification method for device integrity management provided by the embodiment of the application is shown in Figure 4a The device lubrication poor intelligent identification method for device integrity management can include the following steps:

[0183] Step S401, acquiring the vibration signal of the key rotating device of the coal mine.

[0184] Step S402, inputting the vibration signal into a preprocessing model, performing noise reduction and standardization processing on the vibration signal, and obtaining a standardized signal.

[0185] Step S403, acquiring the initial weight of the initial state recognition model.

[0186] The initial state recognition model refers to the state recognition model before training. The initial weight refers to the initial value set for the initial state recognition model. In the embodiment of the application, an initialization method based on pseudo-inverse learning is adopted to ensure that the initial weight conforms to the time-frequency characteristics of the data. Specifically, the initial weight can be represented in the following manner:

[0187]

[0188] is an initial weight; is a feature matrix, and , is the first denoised vibration signal, is the Nth denoised vibration signal; is the transpose of ; N is the number of sample signals; is a one-hot encoding matrix of the lubrication state; I is an identity matrix; is a ridge regression coefficient, such as . Wherein, for sample signals, the lubrication state is determined by a domain expert combined with time-frequency analysis and labeled, for example, the lubrication state includes: normal state, insufficient lubrication, grease pollution, abnormal wear. The format of the label is represented by one-hot encoding to represent multi-classification labels, such as [1, 0, 0, 0] for normal state, [0, 1, 0, 0] for insufficient lubrication, etc.

[0189] Step S404, a plurality of sample signals and classification labels corresponding to each of the sample signals are obtained.

[0190] The classification label refers to the labeling result obtained after labeling the vibration signal, that is, one-hot encoding, which is determined by a domain expert combined with time-frequency analysis and labeled. The sample signal can be a plurality of vibration signals collected by a coal mine key rotating equipment in a historical time period.

[0191] Step S405, each of the sample signals and the corresponding classification labels are input into an initial state recognition model, the initial state recognition model is trained, and a loss function result of the initial state recognition model is obtained.

[0192] Wherein, the loss function is determined according to the wavelet basis in the pre-processing model, the pseudo-inverse Jacobian stability term and the time-frequency contrast term.

[0193] Since the coal mine equipment failure has multi-scale time-frequency feature coupling and pseudo-inverse weight stability requirements, the embodiment of the present application adopts a multi-domain contrast regularization loss, learns through time-frequency contrast and pseudo-inverse Jacobian constraint, realizes feature decoupling reinforcement and model generalization ability enhancement, and thereby constructs a loss function. Specifically, the loss function can be represented by the following formula:

[0194]

[0195] In the formula, As a wavelet domain feature alignment term, the wavelet basis of the denoising process is embedded into the loss function as prior knowledge, realizing end-to-end optimization and feature space consistency constraint, and The L1 norm is used for the dynamic wavelet packet feature extraction operator to enhance the robustness of the abnormal impact feature.

[0196] As a pseudo-inverse Jacobian stability term, the numerical instability of the pseudo-inverse weight is improved by Jacobian matrix regularization to enhance the convergence of the iteration, and The Jacobian matrix representing the pseudo-inverse weight of the input feature is constrained to suppress the gradient explosion.

[0197] As a time-frequency contrast term, a multi-modal contrast learning strategy is used to force the model to establish a time-frequency consistent representation, solving the problem of insufficient modeling of time-frequency coupled features by traditional cross-entropy loss.

[0198] Wherein, is the influence factor of the pseudo-inverse Jacobian stability term, and the example is ;

[0199] is the influence factor of the time-frequency contrast term, and the example is ;

[0200] is the similarity threshold, and the example is ;

[0201] is the loss function of the convolutional neural network;

[0202] is a dynamic wavelet packet feature extraction operator, which reuses the weight in S1 , forcing the network to learn a representation that aligns with the wavelet denoising feature space;

[0203] is a logarithmic function, and the default base is the natural constant;

[0204] is the L1 norm;

[0205] is the feature representation of the i-th sample after the convolutional neural network;

[0206] is the true fault class label of the i-th sample;

[0207] is the total number of layers of the convolutional neural network;

[0208] is the partial derivative symbol;

[0209] is the Frobenius norm;

[0210] is the feature representation of the i-th sample in the batch;

[0211] is a feature representation of the jth sample in the batch;

[0212] is a similarity scaling factor, exemplary, ;

[0213] is a cosine similarity function;

[0214] is a batch size, i.e., the number of sample signals in a batch input to the initial state recognition model.

[0215] Specifically, a plurality of sample signals and classification labels corresponding to each sample signal are input into the initial state recognition model, the initial state recognition model is trained, and a loss function result of the initial state recognition model is calculated.

[0216] Step S406, according to the loss function result, the initial weight is updated to obtain an updated weight.

[0217] Since the traditional gradient descent method does not consider the differentiated influence of the importance of space-time features on parameter updating, the embodiment of the present application proposes a parameter updating strategy of fusing space-time attention images, adjusts the gradient update direction through a space-time attention matrix, and realizes adaptive optimization of the initial weight. Specifically, the initial weight can be updated according to the following formula:

[0218]

[0219] In the formula, is the updated weight obtained by the ith iteration;

[0220] is the updated weight obtained by the ith iteration; is the gradient of the loss function of the initial state recognition model with respect to the initial weight of the ith iteration;

[0221] is a learning rate of the initial state recognition model, exemplary, ;

[0222] ;

[0223] is a label matrix.

[0224] ​​Step S407, the number of updates of the updated weight is counted, and the step of inputting the sample signal and the corresponding classification label into the initial state recognition model to train the initial state recognition model to obtain the loss function result of the initial state recognition model is returned until the number of updates is greater than or equal to a preset number of updates.

[0225] Specifically, steps S405 and S406 are repeatedly executed until a preset stop iteration condition is met, that is, the model training is completed. Exemplarily, the preset stop iteration condition is that a preset maximum number of iterations is reached, and preferably, the preset maximum number of iterations is set to 1000 times.

[0226] Step S408, the initial state recognition model is set according to the updated weight to obtain a state recognition model.

[0227] Specifically, the updated weight is set as a parameter of the initial state recognition model to obtain the state recognition model.

[0228] Step S409, the standardized signal is input into the state recognition model to obtain the lubrication state of the coal mine key rotating equipment.

[0229] Further, the initial weight is updated according to the loss function result to obtain an updated weight, including: the initial weight is updated according to the loss function result to obtain a first weight; for each sample signal, a sample lubrication state corresponding to the output of the initial state recognition model is obtained; for each sample signal, the fuzzy membership degree of the sample signal is determined according to the sample lubrication state; and the first weight is adjusted according to the fuzzy membership degrees of the sample signals to obtain the updated weight.

[0230] The first weight refers to the updated weight obtained after the first iteration in the training process of the initial state recognition model. The fuzzy membership degree is used to describe the importance of the sample signal to the training of the initial state recognition model.

[0231] To cope with the problem of coal mine equipment working condition drift and alleviate the influence of equipment working condition drift on model performance, such as the influence of vibration mode offset caused by environmental temperature change on model performance. In the process of training the initial state recognition model, the fuzzy membership degree can be set for each sample signal after the first iteration, and the weight of the sample signal is dynamically adjusted by using the fuzzy membership degree, thereby increasing the adaptability of the state recognition model obtained after training to different working conditions and improving the adaptability of the state recognition model to changing data. It can be understood that for each sample signal, the corresponding fuzzy membership degree can be determined by a fuzzy membership degree function.

[0232] The fuzzy membership function can be expressed by the following formula:

[0233]

[0234] In the formula, For the first The fuzzy membership degree of each sample is dynamically adjusted based on the confidence level of the sample fault state and lubrication state output by the initial state recognition model. For sample signals with low confidence (which may be new operating conditions), the weight is reduced, that is, the fuzzy membership degree is reduced, thereby reducing the false updates of the initial state recognition model.

[0235] Let be the prediction confidence level for the i-th sample;

[0236] It is an exponential function with the natural constant as its base;

[0237] For example, the slope factor. ;

[0238] For example, a confidence threshold. ;

[0239] Furthermore, based on the fuzzy membership degree of each sample signal, the first weight is incrementally updated to adjust the first weight, resulting in the updated weight, thus achieving adaptive adjustment of the model, as shown below:

[0240]

[0241]

[0242] In the formula, These are the updated weights of the convolutional neural network.

[0243] It is a fuzzy membership matrix;

[0244] The fuzzy membership degree of the first sample;

[0245] Let be the fuzzy membership degree of the Nth sample.

[0246] Figure 4b This diagram illustrates the optimization of the state recognition model training process using different initialization methods. For example... Figure 4bAs shown, the optimization effect of pseudo-inverse initialization on the model training process is analyzed through the parameter space contour map, the parameter update paths of random initialization and pseudo-inverse initialization are compared, and the convergence characteristics of different initialization methods in non-convex optimization are revealed. The traditional random initialization path presents severe oscillation and repeatedly hesitates in the loss surface saddle point area. The technical scheme of the embodiment of the application directly points the initialization path to the global optimal neighborhood, the trajectory is smooth and converges rapidly, and the parameter space visualization shows that the pseudo-inverse weight initial value is located in the low gradient flat area of the loss surface, which is highly consistent with the time-frequency feature distribution of the vibration signal. In theory, it ensures that the optimization process avoids local minimum traps and verifies the improvement of the model training stability of the pseudo-inverse initialization.

[0247] The technical scheme of the embodiment of the application uses pseudo-inverse initialization and pseudo-inverse Jacobian stability terms to optimize the training process of the state recognition model. Pseudo-inverse initialization directly makes the network weight more consistent with the time-frequency features of the vibration signal, avoiding the slow convergence and instability problems that may be caused by traditional random initialization. The pseudo-inverse Jacobian stability term further ensures the stability of the gradient and the convergence of the model in the non-convex optimization process. The fuzzy membership function and the incremental learning strategy are used to address the working condition drift problem of coal mine equipment, dynamically adjust the weight of the training sample, enhance the adaptability of the model to data under different working conditions, reduce the weight of low confidence samples, reduce model misupdates, and improve the stability and reliability of the model.

[0248] Embodiment 5

[0249] Corresponding to the above method embodiment, the embodiment of the application provides a device lubrication poor intelligent recognition device for equipment integrity management, Figure 5 A structure diagram of a device lubrication poor intelligent recognition device for equipment integrity management provided by the embodiment of the application is shown in Figure 5 As shown, the device lubrication poor intelligent recognition device for equipment integrity management can include:

[0250] The signal acquisition module 501 is configured to acquire the vibration signal of the coal mine key rotating equipment.

[0251] The preprocessing module 502 is configured to input the vibration signal into a preprocessing model, perform noise reduction and standardization processing on the vibration signal, and obtain a standardized signal. The preprocessing model is configured to perform noise reduction processing on the vibration signal based on an adaptive noise reduction strategy of dynamic wavelet packet decomposition.

[0252] The identification module 503 is configured to input the standardized signal into a state identification model to obtain the lubrication state of the coal mine key rotating equipment; the state identification model comprises a multi-scale convolution module, a time-space attention weighting module, a feature fusion module and a classifier; the state identification model is configured to selectively enhance and fuse the multi-scale signal features of the standardized signal in the time domain and the frequency domain, and classify based on the fused signal features to obtain the lubrication state of the coal mine key rotating equipment.

[0253] The device lubrication poor intelligent identification device for device integrity management provided by the embodiment of the application can effectively eliminate strong noise (such as mechanical vibration and electromagnetic interference) in a complex coal mine environment and retain weak fault signals (such as vibration features of early bearing wear) through adaptive noise reduction of dynamic wavelet packet decomposition, avoids misfiltering of high-frequency fault signals compared with a traditional fixed threshold method, and provides high-quality data for subsequent feature extraction. Through standardization processing, signal amplitude differences under different devices and different working conditions can be eliminated, the model can be applicable to multiple devices of the same type (such as fans of the same type in different mines), and repeated training costs are reduced. The multi-scale convolution module covers features of different time and frequency scales, the key features are enhanced in combination with a time-space attention mechanism, the fault is accurately positioned at a key position in the time domain and the frequency domain, the identification capability of “time-frequency coupling features” caused by poor lubrication is enhanced, and misjudgment caused by fault features being submerged (such as high-frequency features of slight gear pitting) is avoided. The classifier outputs specific fault types (such as distinguishing between “poor lubrication” and “bearing wear”) based on the fused global features, avoids fuzzy judgment of a traditional method (such as artificial auscultation), and improves the pertinence of maintenance (such as determining a maintenance scheme without disassembly and inspection).

[0254] In some embodiments, the identification module 503 is further configured to:

[0255] input the standardized signal into the multi-scale convolution module to extract multi-scale signal features; wherein the multi-scale convolution module is constructed based on a multi-branch structure of a mixed dilated convolution and an adaptive kernel width;

[0256] input the multi-scale signal features into the time-space attention weighting module to obtain a time domain attention weight matrix and a frequency domain attention weight matrix corresponding to the multi-scale signal features;

[0257] input the multi-scale signal features, the time domain attention weight matrix and the frequency domain attention weight matrix into the feature fusion module to selectively enhance and fuse the multi-scale signal features in the time domain and the frequency domain, and obtain a fused feature matrix;

[0258] Input the fusion feature matrix into a classifier to obtain the lubrication state of the key rotating equipment of the coal mine; wherein the classifier is obtained based on a pseudo-inverse weight matrix and a residual error convolution.

[0259] In some embodiments, the inputting the multi-scale signal feature into a spatio-temporal attention weighting module to obtain a time domain attention weight matrix and a frequency domain attention weight matrix corresponding to the multi-scale signal feature comprises:

[0260] Input the multi-scale signal feature into a time domain attention weight matrix calculation formula to obtain a time domain attention weight matrix;

[0261] Input the multi-scale signal feature into a frequency domain attention weight matrix calculation formula to obtain a frequency domain attention weight matrix.

[0262] In some embodiments, the multi-scale signal feature, the time domain attention weight matrix and the frequency domain attention weight matrix are input into a feature fusion module to selectively enhance and fuse the multi-scale signal feature in the time domain and the frequency domain to obtain a fusion feature matrix, comprising:

[0263] According to the time domain attention weight matrix and the frequency domain attention weight matrix, a time-frequency domain gating weight matrix is determined;

[0264] The multi-scale signal feature is weighted by the time domain attention weight matrix and the frequency domain attention weight matrix respectively to obtain a weighted time domain attention weight matrix and a weighted frequency domain attention weight matrix;

[0265] According to the weighted time domain attention weight matrix, the weighted frequency domain attention weight matrix and the time-frequency domain gating weight matrix, a fusion feature matrix is determined.

[0266] In some embodiments, the normalized signal is input into a multi-scale convolution module to extract a multi-scale signal feature, comprising:

[0267] According to the entropy value of the normalized signal, a kernel width of dynamic kernel width convolution is determined;

[0268] According to the kernel width of the dynamic kernel width convolution, a first signal feature is determined;

[0269] According to a preset kernel width of fundamental frequency convolution, a second signal feature is determined;

[0270] According to a preset expansion rate, a preset kernel width of expansion convolution, a third signal feature is determined;

[0271] The first signal feature, the second signal feature and the third signal feature are spliced to obtain a multi-scale signal feature.

[0272] In some embodiments, the preprocessing module 502 is further configured to:

[0273] obtain a dynamic threshold corresponding to the vibration signal; the dynamic threshold refers to a dynamic threshold corresponding to each of a plurality of sub-signals obtained by wavelet decomposition of the vibration signal;

[0274] de-noise the vibration signal according to the dynamic threshold to obtain a de-noised vibration signal;

[0275] normalize the de-noised vibration signal according to a preset sliding window to obtain a standardized signal.

[0276] In some embodiments, de-noising the vibration signal according to the dynamic threshold to obtain a de-noised vibration signal comprises:

[0277] performing inverse discrete wavelet transform according to the dynamic threshold and each of the sub-signals to obtain a time-domain vibration signal;

[0278] obtain an adaptive weight corresponding to the vibration signal;

[0279] de-noise the time-domain vibration signal according to the adaptive weight to obtain a de-noised vibration signal.

[0280] In some embodiments, the device further comprises:

[0281] an initialization module configured to obtain an initial weight of an initial state recognition model;

[0282] a sample acquisition module configured to acquire a plurality of sample signals and a classification label corresponding to each of the sample signals;

[0283] a training module configured to input each of the sample signals and the corresponding classification label into the initial state recognition model, train the initial state recognition model, and obtain a loss function result of the initial state recognition model; wherein the loss function is determined according to a wavelet basis in the preprocessing model, a pseudo-inverse Jacobian stability term, and a time-frequency contrast term;

[0284] an updating module configured to update the initial weight according to the loss function result to obtain an updated weight;

[0285] a feedback module configured to count the number of updates of the updated weight and return to perform the step of inputting each of the sample signals and the corresponding classification label into the initial state recognition model, training the initial state recognition model, and obtaining a loss function result of the initial state recognition model until the number of updates is greater than or equal to a preset number of updates;

[0286] The setting module is configured to set the initial state recognition model according to the updated weight, to obtain a state recognition model.

[0287] The device provided by the embodiment of the present application has the same implementation principle and generated technical effects as the foregoing method embodiment, and for brevity of description, the part not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiment.

[0288] Embodiment 5

[0289] The embodiment of the present application further provides an electronic device for running the device lubrication poor intelligent identification method for device integrity management; referring to Figure 6 The electronic device shown in the structural schematic diagram of the electronic device, the electronic device includes a memory 600 and a processor 601, wherein the memory 600 is used to store one or more computer instructions, and one or more computer instructions are executed by the processor 601 to realize the device lubrication poor intelligent identification method for device integrity management.

[0290] Further, Figure 6 The electronic device shown further includes a bus 602 and a communication interface 603, and the processor 601, the communication interface 603 and the memory 600 are connected through the bus 602.

[0291] Wherein, the memory 600 can contain a high-speed random access memory (RAM, Random Access Memory), and can also include a non-volatile memory (non-volatile memory), such as at least one disk memory. Through at least one communication interface 603 (may be wired or wireless), the communication connection between the system network element and at least one other network element can be realized, and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used. The bus 602 can be an ISA bus, a PCI bus or an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of expression, Figure 6 In the figure, only one bidirectional arrow is used to represent, but only one bus or one type of bus is not represented.

[0292] The processor 601 can be an integrated circuit chip having a signal processing capability. In the implementation process, each step of the above method can be completed by the integrated logic circuit or the instruction in the form of software in the processor 601. The processor 601 described above can be a general processor, including a central processing unit (CPU), a network processor (NP), and the like; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. Each method, step and logic block diagram disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, and the like storage medium mature in the art. The storage medium is located in the memory 600, and the processor 601 reads the information in the memory 600, and combines the hardware to complete the steps of the method of the above-mentioned embodiments.

[0293] The embodiment of the present application also provides a computer readable storage medium, the computer readable storage medium stores computer executable instructions, when the computer executable instructions are called and executed by a processor, the computer executable instructions cause the processor to implement the above-mentioned device integrity management-oriented device poor lubrication intelligent identification method, and specific implementation can be referred to the method embodiment, and will not be repeated here.

[0294] The computer program product for implementing the device integrity management-oriented device poor lubrication intelligent identification method provided by the embodiment of the present application includes a computer readable storage medium storing non-volatile program codes executable by a processor, and the instructions included in the program codes can be used to execute the method described in the foregoing method embodiment, and specific implementation can be referred to the method embodiment, and will not be repeated here.

[0295] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiment, and will not be repeated here.

[0296] In several embodiments provided by the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. The described device embodiments are merely schematic, and for example, the division of the units is only a logical function division, and there can be another division manner in actual implementation; for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, or the among different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electric, mechanical or other forms.

[0297] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. In actual implementation, some or all of the units can be selected according to the actual needs to achieve the purposes of the embodiments of the present application.

[0298] In addition, each function unit in the various embodiments of the present application can be integrated in one processing unit, or each unit can exist physically as a separate unit, or two or more units can be integrated in one unit.

[0299] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a non-volatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art, or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0300] Finally, it should be noted that the above-described embodiments are merely specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit the present application, and the protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still make modifications or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features therein, within the technical range disclosed by the present application. The modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for intelligent identification of poor equipment lubrication for equipment integrity management, characterized in that, include: Acquire vibration signals from key rotating equipment in coal mines; The vibration signal is input into a preprocessing model to perform noise reduction and standardization on the vibration signal to obtain a standardized signal; wherein, the preprocessing model is used to perform noise reduction on the vibration signal based on an adaptive noise reduction strategy of dynamic wavelet packet decomposition. The standardized signal is input into the state recognition model to obtain the lubrication state of the key rotating equipment in the coal mine. The state recognition model includes a multi-scale convolution module, a spatiotemporal attention weighting module, a feature fusion module, and a classifier. The state recognition model is used to selectively enhance and fuse the multi-scale signal features of the standardized signal in the time and frequency domains, and to classify the fused signal features to obtain the lubrication state of the key rotating equipment in the coal mine. Before inputting the standardized signal into the state recognition model, the method further includes: obtaining the initial weights of the initial state recognition model; obtaining multiple sample signals and classification labels corresponding to each sample signal; inputting each sample signal and the corresponding classification label into the initial state recognition model to train the initial state recognition model and obtain the loss function result of the initial state recognition model; wherein, the loss function is determined based on the wavelet basis, pseudo-inverse Jacobian stability term, and time-frequency contrast term in the preprocessing model; updating the initial weights according to the loss function result to obtain the updated weights; counting the number of updates of the updated weights, and returning to execute the step of inputting the sample signal and the corresponding classification label into the initial state recognition model to train the initial state recognition model and obtain the loss function result of the initial state recognition model, until the number of updates is greater than or equal to the preset number of updates; setting the initial state recognition model according to the updated weights to obtain the state recognition model; The loss function is expressed by the following formula: in, This is the loss function for the convolutional neural network; The influence factor for the pseudo-inverse Jacobian stability term; The influence factor for the time-frequency comparison term; The similarity threshold; For dynamic wavelet packet feature extraction operators; It is a logarithmic function; It is an L1 norm; The true fault category label for the i-th sample; This represents the total number of layers in the convolutional neural network. The sign for partial derivatives; It is the Frobenius norm; This represents the feature representation of the i-th sample in the batch; Let j be the feature representation of the j-th sample in the batch; This is a similarity scaling factor; The cosine similarity function; This refers to the batch size.

2. The method according to claim 1, characterized in that, The step of inputting the standardized signal into the state recognition model to obtain the lubrication state of the key rotating equipment in the coal mine includes: The standardized signal is input into a multi-scale convolution module to extract multi-scale signal features; wherein, the multi-scale convolution module is constructed based on a multi-branch structure with hybrid dilated convolution and adaptive kernel width; The multi-scale signal features are input into the spatiotemporal attention weighting module to obtain the temporal domain attention weight matrix and the frequency domain attention weight matrix corresponding to the multi-scale signal features. The multi-scale signal features, the time-domain attention weight matrix, and the frequency-domain attention weight matrix are input into the feature fusion module to selectively enhance and fuse the multi-scale signal features in the time and frequency domains to obtain a fused feature matrix. The fused feature matrix is ​​input into a classifier to obtain the lubrication status of the key rotating equipment in the coal mine; wherein the classifier is constructed based on a pseudo-inverse weight matrix and residual convolution.

3. The method according to claim 2, characterized in that, The multi-scale signal features are input into the spatiotemporal attention weighting module to obtain the temporal domain attention weight matrix and frequency domain attention weight matrix corresponding to the multi-scale signal features, including: The multi-scale signal features are input into the formula for calculating the temporal attention weight matrix to obtain the temporal attention weight matrix; The multi-scale signal features are input into the frequency domain attention weight matrix calculation formula to obtain the frequency domain attention weight matrix.

4. The method according to claim 2, characterized in that, The multi-scale signal features, the time-domain attention weight matrix, and the frequency-domain attention weight matrix are input into the feature fusion module to selectively enhance and fuse the multi-scale signal features in the time and frequency domains, resulting in a fused feature matrix, including: The time-frequency domain gating weight matrix is ​​determined based on the time-domain attention weight matrix and the frequency-domain attention weight matrix. The multi-scale signal features are weighted by the time-domain attention weight matrix and the frequency-domain attention weight matrix, respectively, to obtain the weighted time-domain attention weight matrix and the weighted frequency-domain attention weight matrix; The fusion feature matrix is ​​determined based on the weighted time-domain attention weight matrix, the weighted frequency-domain attention weight matrix, and the time-frequency-domain gating weight matrix.

5. The method according to claim 2, characterized in that, The standardized signal is input into a multi-scale convolution module to extract multi-scale signal features, including: The kernel width of the corresponding dynamic kernel width convolution is determined based on the entropy value of the standardized signal. The first signal feature is determined based on the kernel width of the dynamic kernel-width convolution; The second signal feature is determined based on the preset kernel width of the fundamental frequency convolution; The third signal feature is determined based on the preset dilation rate and the preset kernel width of the dilated convolution. The first signal feature, the second signal feature, and the third signal feature are concatenated to obtain multi-scale signal features.

6. The method according to claim 1, characterized in that, The step of inputting the vibration signal into a preprocessing model to perform noise reduction and normalization processing on the vibration signal to obtain a normalized signal includes: Obtain the dynamic threshold corresponding to the vibration signal; the dynamic threshold refers to the dynamic threshold corresponding to each of the multiple sub-signals obtained after wavelet decomposition of the vibration signal. The vibration signal is denoised according to the dynamic threshold to obtain the denoised vibration signal; The noise-reduced vibration signal is normalized according to a preset sliding window to obtain a standardized signal.

7. The method according to claim 6, characterized in that, The step of denoising the vibration signal according to the dynamic threshold to obtain the denoised vibration signal includes: Based on the dynamic threshold and each of the sub-signals, an inverse discrete wavelet transform is performed to obtain the time-domain vibration signal; Obtain the adaptive weights corresponding to the vibration signal; The time-domain vibration signal is denoised according to the adaptive weights to obtain the denoised vibration signal.

8. The method according to claim 1, characterized in that, The step of updating the initial weights based on the loss function result to obtain the updated weights includes: Based on the loss function result, the initial weights are updated to obtain the first weights; For each sample signal, obtain the sample lubrication state output by the initial state recognition model; For each sample signal, the fuzzy membership degree of the sample signal is determined based on the sample lubrication state; The first weight is adjusted based on the fuzzy membership degree of each sample signal to obtain the updated weight.

9. A smart device for identifying poor lubrication in equipment for equipment integrity management, characterized in that, include: The signal acquisition module is used to acquire vibration signals from key rotating equipment in coal mines. The preprocessing module is used to input the vibration signal into the preprocessing model, and to perform noise reduction and standardization on the vibration signal to obtain a standardized signal; wherein, the preprocessing model is used to perform noise reduction on the vibration signal based on an adaptive noise reduction strategy of dynamic wavelet packet decomposition. The identification module is used to input the standardized signal into the state identification model to obtain the lubrication state of the key rotating equipment in the coal mine. The state identification model includes a multi-scale convolution module, a spatiotemporal attention weighting module, a feature fusion module, and a classifier. The state identification model is used to selectively enhance and fuse the multi-scale signal features of the standardized signal in the time and frequency domains, and to classify the fused signal features to obtain the lubrication state of the key rotating equipment in the coal mine. The preprocessing module is further configured to: obtain initial weights of the initial state recognition model; obtain multiple sample signals and corresponding classification labels for each sample signal; input each sample signal and the corresponding classification label into the initial state recognition model to train the initial state recognition model and obtain the loss function result of the initial state recognition model; wherein the loss function is determined based on the wavelet basis, pseudo-inverse Jacobian stability term, and time-frequency contrast term in the preprocessing model; update the initial weights according to the loss function result to obtain updated weights; count the number of updates of the updated weights and return to execute the step of inputting the sample signals and the corresponding classification labels into the initial state recognition model to train the initial state recognition model and obtain the loss function result of the initial state recognition model, until the number of updates is greater than or equal to the preset number of updates; and set the initial state recognition model according to the updated weights to obtain the state recognition model. The loss function is expressed by the following formula: in, This is the loss function for the convolutional neural network; The influence factor for the pseudo-inverse Jacobian stability term; The influence factor for the time-frequency comparison term; The similarity threshold; For dynamic wavelet packet feature extraction operators; It is a logarithmic function; L1 norm The true fault category label for the i-th sample; This represents the total number of layers in the convolutional neural network. The sign for partial derivatives; It is the Frobenius norm; This represents the feature representation of the i-th sample in the batch; Let j be the feature representation of the j-th sample in the batch; This is a similarity scaling factor; The cosine similarity function; This refers to the batch size.

Images