A Spectrum-Based Bearing Wear Warning Method and System

Abstract

The present invention discloses a bearing wear warning method and system based on spectrum. The method includes: acquiring data of equipment vibration, processing the data and converting it into a first spectrogram; setting a step size, performing a sliding window on the first spectrogram according to the step size to obtain a plurality of first spectrogram segments; labeling the first spectrogram segments according to the original data and using them as training samples; using triplet loss as the loss function and resnet50 as the feature extraction network to train the feature extraction network to obtain benchmark features; processing the test data into a second spectrogram to obtain second spectrogram segments, and using the trained feature extraction network to obtain the vibration features in the corresponding time period of the second spectrogram segments; calculating the Euclidean distance between the vibration features of each period and the benchmark features, and judging whether the vibration is abnormal according to the threshold.

Description

Technical Field

[0001] The present invention relates to the technical field of early warning, and in particular, to a bearing wear early warning method and system based on spectrum. Background Art

[0002] Vibration is an important characteristic during the operation of rotating machinery. By using a data collector to collect vibration information on the operating state of mechanical equipment (such as bearings), and then through vibration spectrum analysis, the causes of faults such as rotor imbalance, shaft bending, bearing damage and looseness, and misalignment of the shafting can be quickly and accurately diagnosed, so as to achieve the purpose of early fault discovery, rapid and timely diagnosis, fixed-point and quantitative conclusion, and clear mechanism.

[0003] However, this solution can only judge whether there is a fault during the detection of traditional rotating machinery, and cannot perform quantitative analysis on the loss and give an early warning. Summary of the Invention

[0004] In view of the above technical problem that the bearing loss cannot be warned, the present invention provides a bearing wear early warning method and system based on spectrum.

[0005] In a first aspect, an embodiment of the present application provides a bearing wear early warning method based on spectrum, including:

[0006] Data conversion step: Using a sensor to obtain the original data of the equipment vibration, processing the original data, and converting the processed original data into a first spectrogram;

[0007] Spectrogram segment obtaining step: Setting a step size, and performing a sliding window on the first spectrogram according to the step size in chronological order to obtain a plurality of first spectrogram segments;

[0008] Training sample obtaining step: Labeling the first spectrogram segments according to the original data. If there is a fault, the label is 1, and if there is no fault, the label is 0. The labeled first spectrogram segments are used as training samples;

[0009] Training step: Using triplet loss as the loss function and resnet50 as the feature extraction network, training the feature extraction network to obtain benchmark features;

[0010] Vibration feature obtaining step: Processing the test data into a second spectrogram according to the data conversion step, cutting the second spectrogram according to the step size to obtain second spectrogram segments, and using the trained feature extraction network with the second spectrogram segments as the input to obtain the vibration features during the time period corresponding to the second spectrogram segments;

[0011] Judgment step: Calculate the Euclidean distance between the vibration characteristics and the reference characteristics for each period of time. If the Euclidean distance is less than the warning value, it is normal vibration; if the Euclidean distance is greater than the fault value, it is a fault; if the Euclidean distance is greater than the warning value and less than the fault value, a warning is issued.

[0012] The above-mentioned bearing wear warning method based on spectrum, wherein the data conversion step includes:

[0013] Time series data acquisition step: Use the sensor to acquire the time series data of the vibration of the equipment;

[0014] Frequency domain data acquisition step: Perform Fourier transform on the time series data to convert the time series data into frequency domain data;

[0015] Low-frequency data filtering step: Use a high-pass filter to filter the low-frequency data in the frequency domain data to obtain high-frequency data;

[0016] First spectrogram acquisition step: Convert the high-frequency data into the first spectrogram.

[0017] The above-mentioned bearing wear warning method based on spectrum, wherein the training step includes:

[0018] Data preprocessing step: Form P and N pairs between the first spectrogram segments, and make the ratio of negative:positive range from 1:1 to 1:4;

[0019] Model training step: Train the feature extraction network by weighting the loss respectively, and obtain the reference characteristics based on the trained feature extraction network.

[0020] The above-mentioned bearing wear warning method based on spectrum, wherein the data preprocessing step includes:

[0021] Size adjustment step: Adjust the size of the first spectrogram segments to a suitable size;

[0022] P and N pair acquisition step: Shuffle the order of the first spectrogram segments, and randomly select one from different categories of first spectrogram segments in a random order to form P and N pairs with any one of the first spectrogram segments;

[0023] Overfitting prevention step: Initially, the ratio of negative:positive = 1:1, and multiply it by 1.01 after each epoch until the ratio of negative:positive = 1:4.

[0024] The above-mentioned bearing wear warning method based on spectrum, wherein the model training step includes:

[0025] Weighting step: Calculate the gradients of the three losses respectively and perform weighting.

[0026] Feature extraction step: Remove the softmax layer from the trained feature extraction network, and use the output of the fully connected layer as the extracted feature.

[0027] Benchmark feature obtaining step: Calculate the mean value of the features of the normal data in the training set to obtain the benchmark feature.

[0028] In a second aspect, an embodiment of the present application provides a bearing wear warning system based on spectrum, including:

[0029] Data conversion module: Use a sensor to obtain the original data of the device vibration, process the original data, and convert the processed original data into a first spectrogram.

[0030] Spectrogram segment obtaining module: Set a step size, and perform sliding window on the first spectrogram according to the step size in chronological order to obtain a number of first spectrogram segments.

[0031] Training sample obtaining module: Label the first spectrogram segments according to the original data. If there is a fault, the label is 1. If there is no fault, the label is 0. Use the labeled first spectrogram segments as training samples.

[0032] Training module: Use triplet loss as the loss function and resnet50 as the feature extraction network to train the feature extraction network to obtain the benchmark feature.

[0033] Vibration feature obtaining module: Process the test data into a second spectrogram according to the data conversion module, cut the second spectrogram according to the step size to obtain second spectrogram segments, and use the trained feature extraction network with the second spectrogram segments as the input to obtain the vibration features in the corresponding time period of the second spectrogram segments.

[0034] Judgment module: Calculate the Euclidean distance between the vibration features of each period and the benchmark feature. If the Euclidean distance is less than the warning value, it is normal vibration. If the Euclidean distance is greater than the fault value, it is a fault. If the Euclidean distance is greater than the warning value and less than the fault value, a warning is issued.

[0035] For the above bearing wear warning system based on spectrum, where the data conversion module includes:

[0036] Time series data obtaining unit: Use the sensor to obtain the time series data of the device vibration.

[0037] Frequency domain data acquisition unit: Perform Fourier transform on the timing data to convert the timing data into frequency domain data;

[0038] Low-frequency data filtering unit: Use a high-pass filter to filter the low-frequency data in the frequency domain data to obtain high-frequency data;

[0039] First spectrogram acquisition unit: Convert the high-frequency data into the first spectrogram.

[0040] The above-mentioned bearing wear warning system based on spectrum, wherein, the training module includes:

[0041] Data preprocessing unit: Form P and N pairs between the first spectrogram segments, and make the ratio of negative:positive range from 1:1 to 1:4;

[0042] Model training unit: Train the feature extraction network by weighting the losses respectively, and obtain benchmark features based on the trained feature extraction network.

[0043] The above-mentioned bearing wear warning system based on spectrum, wherein, the data preprocessing unit includes:

[0044] Size adjustment unit: Adjust the size of the first spectrogram segments to a suitable size;

[0045] P and N pair acquisition unit: Shuffle the order of the first spectrogram segments, and randomly select one from the first spectrogram segments of different categories in a random order to form P and N pairs with any one of the first spectrogram segments;

[0046] Overfitting prevention unit: Initially, negative:positive = 1:1, and multiply by 1.01 after each epoch until negative:positive = 1:4.

[0047] The above-mentioned bearing wear warning system based on spectrum, wherein, the model training unit includes:

[0048] Weighting unit: Calculate the gradients of the three losses respectively and perform weighting;

[0049] Feature extraction unit: Remove the softmax layer from the trained feature extraction network, and use the output of the fully connected layer as the extracted feature;

[0050] Benchmark feature acquisition unit: Calculate the mean of the features of the normal data in the training set to obtain benchmark features.

[0051] Compared with the prior art, the advantages and positive effects of the present invention are:

[0052] The present invention can not only judge whether a fault occurs during the detection of traditional rotating machinery, but also quantitatively analyze the losses and give an early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the steps of a bearing wear warning method based on spectrum provided by the present invention;

[0054] Figure 2 Based on what the present invention provides Figure 1 It is a flowchart of step S1;

[0055] Figure 3 Based on what the present invention provides Figure 1 It is a flowchart of step S4;

[0056] Figure 4 Based on what the present invention provides Figure 3 It is a flowchart of step S41;

[0057] Figure 5 Based on what the present invention provides Figure 3 It is a flowchart of step S42;

[0058] Figure 6 It is a vibration data acquisition diagram of the sensor in the first embodiment of the present invention;

[0059] Figure 7 It is a diagram of converting time series data into frequency domain data in the first embodiment of the present invention;

[0060] Figure 8 It is a diagram of the filtered vibration signal in the first embodiment of the present invention;

[0061] Figure 9 It is a converted spectrogram in the first embodiment of the present invention;

[0062] Figure 10 It is a framework diagram of a bearing wear warning system based on spectrum provided by the present invention.

[0063] Among them, the reference numerals are:

[0064] 1. Data conversion module; 11. Time series data acquisition unit; 12. Frequency domain data acquisition unit; 13. Low-frequency data filtering unit; 14. First spectrogram acquisition unit; 2. Spectrogram segment acquisition module; 3. Training sample acquisition module; 4. Training module; 41. Data preprocessing unit; 411. Size adjustment unit; 412. P, N pair acquisition unit; 413. Overfitting prevention unit; 42. Model training unit; 421. Weighting unit; 422. Feature extraction unit; 423. Benchmark feature acquisition unit; 5. Vibration feature acquisition module; 6. Judgment module. DETAILED DESCRIPTION OF THE INVENTION

[0065] In order to make the objectives, technical solutions and advantages of this application more clear and understandable, the following describes and explains this application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application. Based on the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of this application.

[0066] Obviously, the accompanying drawings in the following description are only some examples or embodiments of this application. For those of ordinary skill in the art, without making creative efforts, this application can also be applied to other similar scenarios based on these drawings. In addition, it can also be understood that although the efforts made in this development process may be complex and lengthy, for those of ordinary skill in the art related to the content disclosed in this application, some designs, manufacturing or production changes made on the basis of the technical content disclosed in this application are only conventional technical means and should not be understood as the content disclosed in this application being insufficient.

[0067] Referring to "embodiments" in this application means that the specific features, structures or characteristics described in combination with the embodiments can be included in at least one embodiment of this application. The appearance of this phrase in various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. It is explicitly and implicitly understood by those of ordinary skill in the art that the embodiments described in this application can be combined with other embodiments without conflict.

[0068] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the ordinary meanings understood by those with ordinary skills in the technical field to which this application belongs. The words such as "a", "an", "one kind", "the" and the like involved in this application do not indicate a limitation in quantity and may represent a singular or plural number. The terms "include", "comprise", "have" and any variations thereof involved in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product or device including a series of steps or modules (units) is not limited to the listed steps or units, but may further include unlisted steps or units, or may further include other steps or units inherent to these processes, methods, products or devices. The words such as "connect", "be connected", "couple" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "plurality" involved in this application means two or more. "And / or" describes the association relationship of associated objects and indicates that three relationships may exist. For example, "A and / or B" may represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the front and rear associated objects. The terms "first", "second", "third" and the like involved in this application are only used to distinguish similar objects and do not represent a specific order for the objects.

[0069] The present invention will be described in detail below in conjunction with the various embodiments shown in the accompanying drawings. However, it should be noted that these embodiments are not limitations on the present invention, and any equivalent transformation or substitution in terms of function, method, or structure made by those of ordinary skill in the art based on these embodiments shall fall within the protection scope of the present invention.

[0070] Before elaborating on each embodiment of the present invention in detail, an overview of the core inventive concept of the present invention is given and will be elaborated in detail through the following several embodiments.

[0071] The present invention provides a bearing wear warning method based on spectrum. By converting the high-frequency data generated by the device into a spectrogram, training samples of the model are obtained. By extracting the feature of the spectrogram segment to represent the vibration information during this time period, it is judged whether the vibration is abnormal according to the threshold.

[0072] Spectrogram: A spectrogram is a voice spectrogram. Generally, a spectrogram is obtained by processing the received time-domain signal. The abscissa of the spectrogram is time, the ordinate is frequency, and the coordinate point value is the energy of the voice data. Since three-dimensional information is expressed in a two-dimensional plane, the magnitude of the energy value is represented by color. The darker the color, the stronger the voice energy at this point. Since the vibration signal is close to the voice information to a certain extent, this patent converts the vibration signal into a spectrogram for analyzing the vibration state.

[0073] Embodiment 1:

[0074] Referring to Figure 1 as shown, Figure 1 it is a schematic diagram of the steps of a bearing wear warning method based on spectrum provided by the present invention. As Figure 1 shown, this embodiment discloses a specific implementation manner of a bearing wear warning method based on spectrum (hereinafter referred to as "the method").

[0075] The vibration caused by the bearing is called bearing tone. All rolling bearings generate a certain degree of tone. The more severe the bearing wear, the higher the degree of bearing tone. Therefore, removing the low-frequency vibration components and obtaining the remaining vibration signal can be used to judge the bearing wear degree.

[0076] Specifically, the method disclosed in this embodiment mainly includes the following steps:

[0077] Referring to Figure 2 , perform step S1: Use a sensor to obtain the original data of the device vibration, process the original data, and convert the processed original data into a first spectrogram.

[0078] Among them, step S1 specifically includes the following contents:

[0079] Step S11: Use the sensor to obtain the time-series data of the device vibration;

[0080] Step S12: Perform a Fourier transform on the time-series data to convert the time-series data into frequency-domain data;

[0081] Step S13: Use a high-pass filter to filter the low-frequency data in the frequency-domain data to obtain high-frequency data;

[0082] Step S14: Convert the high-frequency data into the first spectrogram.

[0083] Then perform step S2: Set a step size, and perform a sliding window on the first spectrogram according to the step size in chronological order to obtain a number of first spectrogram segments.

[0084] Perform step S3: Label the first spectrogram segments according to the original data. If there is a fault, the label is 1. If there is no fault, the label is 0. Use the labeled first spectrogram segments as training samples.

[0085] Referring to Figure 3 , perform step S4: Use triplet loss as the loss function and resnet50 as the feature extraction network to train the feature extraction network to obtain benchmark features.

[0086] Specifically, the input is a triple<a,p,n> ,in,

[0087] a: normal anchor data;

[0088] p: positive, a sample of the same category as a;

[0089] n: negative, which is a sample of a different category.

[0090] The formula is:

[0091] L = max(d(a,p)-d(a,n)+margin)

[0092] Therefore, the ultimate optimization goal is to shorten the distance between a and p and increase the distance between a and n.

[0093] Wherein, step S4 specifically includes the following contents:

[0094] Step S41: forming P and N pairs between the first spectrogram segments, and making the negative:positive ratio range from 1:1 to 1:4.

[0095] Reference Figure 4 , step S41 specifically includes:

[0096] Step S411: adjusting the size of the first spectrogram segment to a suitable size;

[0097] Step S412: shuffling the order of the first spectrogram segments, and selecting one of the first spectrogram segments from different categories in a random order to form a P, N pair with any of the first spectrogram segments;

[0098] Step S413: Initially negative:positive=1:1, multiply by 1.01 after each epoch until negative:positive=1:4.

[0099] Step S42: training the feature extraction network by weighting the losses respectively, and obtaining the benchmark features based on the trained feature extraction network.

[0100] Reference Figure 5 , step S42 specifically includes:

[0101] Step S421: Calculate the gradients of the three losses respectively and perform weighting;

[0102] Among them, the initial learning rate is 0.001 at the beginning and 0.0001 at the last 5% of the epochs. The gradients are calculated separately for the three losses and then weighted. Among them, the weight of the verification loss is 1, and the weights of the other two identification losses are 0.5.

[0103] Step S422: Remove the softmax layer from the trained feature extraction network, and use the output of the fully connected layer as the extracted features. The fully connected layer has 512 nodes, corresponding to 512-dimensional vibration features as the output.

[0104] Step S423: Calculate the mean of the features of the normal data in the training set to obtain the reference features. The reference features are a 512-dimensional vector.

[0105] Then, execute Step S5: Process the test data into a second spectrogram according to Step S1, and cut the second spectrogram according to the step size to obtain second spectrogram segments. Use the trained feature extraction network, take the second spectrogram segments as inputs, and obtain the vibration features within the corresponding time periods of the second spectrogram segments. Among them, the vibration features are 512-dimensional vectors.

[0106] Finally, execute Step S6: Calculate the Euclidean distance between the vibration features of each period and the reference features. If the Euclidean distance is less than the warning value, it is normal vibration; if the Euclidean distance is greater than the fault value, it is a fault; if the Euclidean distance is greater than the warning value and less than the fault value, a warning is issued.

[0107] Next, please refer to Figures 6 - 9 , and the application process of this method is specifically described as follows:

[0108] 1. Use sensors to obtain device vibration data. The data graph is referred to Figure 6 .

[0109] 2. Perform Fourier transform on the data to convert the time series data into the frequency domain, as Figure 7 shown.

[0110] 3. Use a high-pass filter to filter the low-frequency (below 500 Hz) data, and the obtained high-frequency data is used for fault warning, as Figure 8 shown.

[0111] 4. Convert the filtered data into a spectrogram, as Figure 9 shown.

[0112] 5. Slide a window on the spectrogram according to the time order with a step size of 0.5 seconds to obtain n spectrogram segments with a length of 0.5 seconds. And label these segments according to the original data. The fault label is 1, and the non-fault label is 0. These data are used as training samples.

[0113] 6. Training Method

[0114] This method uses triplet loss as the loss function and resnet50 as the feature extraction network.

[0115] The input is a triplet <a, p, n>

[0116] a: anchor normal data

[0117] p: positive, a sample of the same category as a

[0118] n: negative, a sample of a different category from a

[0119] The formula is:

[0120] L = max(d(a, p) - d(a, n) + margin)

[0121] So the final optimization goal is to narrow the distance between a and p and widen the distance between a and n.

[0122] 1) Data preprocessing:

[0123] Adjust the cut spectrogram image segments to a size of 256 * 256;

[0124] Shuffle the dataset, use a random image order, and then select one image from different categories to form P and N pairs with it;

[0125] Initial negative: positive = 1:1, multiply by 1.01 after each epoch until 1:4 to prevent overfitting.

[0126] 2) Training:

[0127] Initial learning rate: start at 0.001 and end at 0.0001 for the last 5% of epochs;

[0128] Calculate the gradients for the three losses separately and then weight them. The weight of the verification loss is 1, and the weights of the other two identification losses are 0.5;

[0129] Extract the spectrogram image segment features to represent the vibration information during that time period. After training, we can obtain the feature extraction network. Remove the last softmax layer and use the output of the second-to-last fully connected layer as the extracted feature. This layer has 512 nodes, corresponding to 512-dimensional vibration features output.

[0130] 7. Determine whether the vibration is abnormal according to the threshold

[0131] Calculate the mean of the features of the normal data in the training set to obtain a 512-dimensional vector as the reference feature.

[0132] Process the test data into spectrograms (repeat steps 1-4). Use the model trained in step 6, with each 0.5-second segment as the input, to obtain the vibration features (512-dimensional vector) within that time period.

[0133] Calculate the Euclidean distance d between the features of each time period and the reference feature. According to the test, when the distance d < 0.7, it is normal vibration; when d > 0.9, it is a fault; when 0.7 < d < 0.9, a warning can be issued.

[0134] Example 2:

[0135] Combined with a method for bearing wear warning based on spectrum disclosed in Example 1, this example discloses a specific implementation example of a bearing wear warning system based on spectrum (hereinafter referred to as "the system").

[0136] Refer to Figure 10 As shown, the system includes:

[0137] Data conversion module 1: Use a sensor to obtain the original data of the device vibration, process the original data, and convert the processed original data into the first spectrogram;

[0138] Spectrogram segment acquisition module 2: Set a step size, and perform a sliding window on the first spectrogram according to the step size in chronological order to obtain a number of first spectrogram segments;

[0139] Training sample acquisition module 3: Label the first spectrogram segments according to the original data. If there is a fault, the label is 1; if there is no fault, the label is 0. Use the labeled first spectrogram segments as training samples;

[0140] Training module 4: Use triplet loss as the loss function and resnet50 as the feature extraction network to train the feature extraction network to obtain the reference feature;

[0141] Vibration feature acquisition module 5: Process the test data into the second spectrogram through the data conversion module 1, cut the second spectrogram according to the step size to obtain second spectrogram segments, and use the trained feature extraction network with the second spectrogram segments as the input to obtain the vibration features within the time period corresponding to the second spectrogram segments;

[0142] Judgment module 6: Calculate the Euclidean distance between the vibration characteristics and the reference characteristics for each period of time. If the Euclidean distance is less than the warning value, it is normal vibration; if the Euclidean distance is greater than the fault value, it is a fault; if the Euclidean distance is greater than the warning value and less than the fault value, a warning is issued.

[0143] Specifically, the data conversion module 1 includes:

[0144] Time series data acquisition unit 11: Use the sensor to acquire the time series data of the vibration of the device;

[0145] Frequency domain data acquisition unit 12: Perform Fourier transform on the time series data to convert the time series data into frequency domain data;

[0146] Low-frequency data filtering unit 13: Use a high-pass filter to filter the low-frequency data in the frequency domain data to obtain high-frequency data;

[0147] First spectrogram acquisition unit 14: Convert the high-frequency data into the first spectrogram.

[0148] Specifically, the training module 4 includes:

[0149] Data preprocessing unit 41: Form P and N pairs between the first spectrogram segments, and make the ratio of negative to positive range from 1:1 to 1:4;

[0150] Model training unit 42: Train the feature extraction network by weighting the losses respectively, and obtain the reference characteristics based on the trained feature extraction network.

[0151] Among them, the data preprocessing unit 41 includes:

[0152] Size adjustment unit 411: Adjust the size of the first spectrogram segments to an appropriate size;

[0153] P and N pair acquisition unit 412: Shuffle the order of the first spectrogram segments, and randomly select one from different categories of first spectrogram segments in a random order to form P and N pairs with any one of the first spectrogram segments;

[0154] Overfitting prevention unit 413: Initially, the ratio of negative to positive is 1:1, and it is multiplied by 1.01 after each epoch until the ratio of negative to positive is 1:4.

[0155] Among them, the model training unit 42 includes:

[0156] Weighting unit 421: Calculate the gradients of the three losses respectively and perform weighting;

[0157] Feature extraction unit 422: Remove the softmax layer from the trained feature extraction network, and use the output of the fully connected layer as the extracted feature;

[0158] Reference feature acquisition unit 423: Calculate the mean value of the features of the normal data in the training set to obtain the reference feature.

[0159] For the technical solutions of the same parts in a bearing wear warning system based on spectrum disclosed in this embodiment and a bearing wear warning method based on spectrum disclosed in Embodiment 1, please refer to Embodiment 1 and will not be elaborated here.

[0160] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0161] In summary, the beneficial effects of the present invention are that the present invention can not only judge whether a failure occurs during the detection of traditional rotating machinery, but also quantitatively analyze the loss and give an early warning.

[0162] The above embodiments only represent several implementation manners of the present application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A bearing wear warning method based on spectrum, characterized in that Including: Data conversion step: Using a sensor to obtain the original data of the device vibration, processing the original data, and converting the processed original data into a first spectrogram; Spectrogram segment obtaining step: Setting a step size, and performing a sliding window on the first spectrogram according to the step size in chronological order to obtain a number of first spectrogram segments; Training sample obtaining step: Labeling the first spectrogram segments according to the original data. If there is a fault, the label is 1. If there is no fault, the label is 0. Using the labeled first spectrogram segments as training samples; Training step: Using triplet loss as the loss function and resnet50 as the feature extraction network to train the feature extraction network to obtain benchmark features; Vibration feature obtaining step: Processing the test data into a second spectrogram according to the data conversion step, cutting the second spectrogram according to the step size to obtain second spectrogram segments, and using the trained feature extraction network with the second spectrogram segments as input to obtain the vibration features within the corresponding time period of the second spectrogram segments; Judgment step: Calculating the Euclidean distance between the vibration features of each period and the benchmark features. If the Euclidean distance is less than the warning value, it is normal vibration; if the Euclidean distance is greater than the fault value, it is a fault; if the Euclidean distance is greater than the warning value and less than the fault value, a warning is issued.

2. The bearing wear warning method according to claim 1, wherein The data conversion step includes: Time series data obtaining step: Using the sensor to obtain the time series data of the device vibration; Frequency domain data obtaining step: Performing a Fourier transform on the time series data to convert the time series data into frequency domain data; Low-frequency data filtering step: Using a high-pass filter to filter the low-frequency data in the frequency domain data to obtain high-frequency data; First spectrogram obtaining step: Converting the high-frequency data into the first spectrogram.

3. The bearing wear warning method according to claim 1, characterized in that The training step includes: Data preprocessing step: Forming P and N pairs between the first spectrogram segments, and making negative: positive from 1:1 to 1:4; Model training step: Training the feature extraction network based on the triplet loss, and obtaining benchmark features based on the trained feature extraction network.

4. The bearing wear warning method according to claim 3, wherein The data preprocessing step includes: Size adjustment step: Adjusting the size of the first spectrogram segments to a size of 256*256; P and N pair obtaining step: Shuffling the order of the first spectrogram segments, and randomly selecting one from the first spectrogram segments of different categories in a random order to form P and N pairs with any one of the first spectrogram segments; Overfitting prevention step: Initially, negative: positive = 1:1, and multiplying by 1.01 after each epoch until negative: positive = 1:

4.

5. The bearing wear warning method according to claim 3, wherein, The model training step includes: Feature extraction step: Removing the softmax layer from the trained feature extraction network and using the output of the fully connected layer as the extracted feature; Reference feature acquisition step: Calculate the mean of the features of the normal data in the training set to obtain the reference feature.

6. A bearing wear warning system based on spectrum, characterized in that, Including: Data conversion module: Use a sensor to obtain the original data of the device vibration, process the original data, and convert the processed original data into a first spectrogram; Spectrogram segment acquisition module: Set a step size, and perform a sliding window on the first spectrogram according to the step size in chronological order to obtain a number of first spectrogram segments; Training sample acquisition module: Label the first spectrogram segments according to the original data. If there is a fault, the label is 1. If there is no fault, the label is 0. Use the labeled first spectrogram segments as training samples; Training module: Use triplet loss as the loss function and resnet50 as the feature extraction network to train the feature extraction network to obtain the reference feature; Vibration feature acquisition module: Process the test data into a second spectrogram according to the data conversion module, cut the second spectrogram according to the step size to obtain second spectrogram segments, and use the trained feature extraction network. Use the second spectrogram segments as input to obtain the vibration features in the corresponding time period of the second spectrogram segments; Judgment module: Calculate the Euclidean distance between the vibration features of each period and the reference feature. If the Euclidean distance is less than the warning value, it is normal vibration; if the Euclidean distance is greater than the fault value, it is a fault; if the Euclidean distance is greater than the warning value and less than the fault value, give a warning.

7. The bearing wear warning system according to claim 6, wherein, The data conversion module includes: Time series data acquisition unit: Use the sensor to obtain the time series data of the device vibration; Frequency domain data acquisition unit: Perform Fourier transform on the time series data to convert the time series data into frequency domain data; Low-frequency data filtering unit: Use a high-pass filter to filter the low-frequency data in the frequency domain data to obtain high-frequency data; First spectrogram acquisition unit: Convert the high-frequency data into the first spectrogram.

8. The bearing wear warning system according to claim 6, characterized in that, The training module includes: Data preprocessing unit: Form P and N pairs between the first spectrogram segments, and make negative: positive from 1:1 to 1:4; Model training unit: Train the feature extraction network based on the triplet loss, and obtain the reference feature based on the trained feature extraction network.

9. The bearing wear warning system according to claim 8, wherein The data preprocessing unit includes: Size adjustment unit: Adjust the size of the first spectrogram segments to a size of 256*256; P, N pair acquisition unit: Shuffle the order of the first spectrogram segments, and randomly select one from the first spectrogram segments of different categories in a random order to form P and N pairs with any one of the first spectrogram segments; Overfitting prevention unit: Initially negative: positive = 1:1, multiply by 1.01 after each epoch until negative: positive = 1:

4.

10. The bearing wear warning system according to claim 8, characterized in that, The model training unit includes: Feature extraction unit: Remove the softmax layer from the trained feature extraction network, and use the output of the fully connected layer as the extracted feature; Benchmark feature acquisition unit: Calculate the mean of the features of the normal data in the training set to obtain the benchmark feature.

Images