A machine learning-based electromechanical equipment fault diagnosis system and method

Abstract

The present invention relates to the technical field of motor testing, and specifically to a fault diagnosis system and method for electromechanical equipment based on machine learning. The system includes: a high-frequency signal injection module that injects high-frequency signals in the range of 10 kHz to 50 kHz into the motor through the power supply line, analyzes the electromagnetic response of the signals, and obtains electromagnetic parameter monitoring data; based on the electromagnetic parameter monitoring data, statistical features are extracted and pattern recognition is performed to generate early fault feature signals. In the present invention, through high-frequency signal injection and differential signal amplification, early identification and precise positioning of motor faults are achieved. By injecting high-frequency signals, changes in electromagnetic parameters can be captured, and these changes usually occur before physical fault manifestations. Differential amplification further enhances the ability to extract useful signals from these minute changes, enabling identification and response before the fault develops to a more severe stage.

Description

Technical Field

[0001] The present invention relates to the technical field of motor testing, and particularly to an electromechanical equipment fault diagnosis system and method based on machine learning. Background Art

[0002] The technical field of motor testing focuses on the performance analysis, fault diagnosis, and maintenance of motors and their related systems. This technical field utilizes various testing methods and tools to evaluate the electrical and mechanical characteristics of motors, including resistance testing, insulation testing, vibration analysis, temperature monitoring, and more advanced diagnostic techniques such as motor current signature analysis (MCSA) and thermography. The main purpose of motor testing is to ensure equipment operation efficiency, prevent faults from occurring, extend equipment lifespan, and improve safety. And the electromechanical equipment fault diagnosis system aims to identify and analyze potential faults of electromechanical equipment (such as electric motors, generators, and other rotating equipment) in real time.

[0003] Although existing motor testing technologies can evaluate and maintain motor performance, they have deficiencies in quickly responding to environmental changes and early fault diagnosis. Existing methods rely on physical tests and parameter monitoring, and these methods can only detect problems when the faults have already started to affect the operation of the machine. For example, vibration analysis and temperature monitoring only show obvious abnormalities after the problem occurs, resulting in insufficiently timely responses to rapidly developing faults. This lagged fault response not only increases maintenance costs but also may lead to greater equipment damage or safety risks due to the inability to resolve small problems in a timely manner. Summary of the Invention

[0004] The purpose of the present invention is to solve the drawbacks existing in the prior art, and to propose an electromechanical equipment fault diagnosis system and method based on machine learning.

[0005] To achieve the above purpose, the present invention adopts the following technical solution: An electromechanical equipment fault diagnosis system based on machine learning includes:

[0006] A high-frequency signal injection module that injects high-frequency signals in the range of 10 kHz to 50 kHz into the motor through the power supply line, analyzes the electromagnetic response of the signals, and obtains electromagnetic parameter monitoring data; based on the electromagnetic parameter monitoring data, extracts statistical features and performs pattern recognition to generate early fault feature signals;

[0007] A dynamic feature analysis module that transforms the early fault feature signals, extracts envelope analysis data of the signals, and establishes an envelope analysis result of the fault signals; uses differential signal amplification to process the envelope analysis result of the fault signals to obtain differential amplified fault features;

[0008] A fault diagnosis module that constructs a feature vector based on the differential amplified fault features and inputs it into a support vector machine classifier to calculate and obtain a fault recognition result;

[0009] A fault threshold dynamic adjustment module adjusts the fault diagnosis threshold in real time according to the motor type and operating conditions, and establishes dynamic fault threshold parameters.

[0010] Preferably, the steps for obtaining the electromagnetic parameter monitoring data are as follows:

[0011] Inject a high-frequency signal in the range of 10 kHz to 50 kHz into the motor through the power line, collect the obtained electromagnetic waveform data as the analysis basis, and form initial electromagnetic response data;

[0012] Perform spectrum analysis on the initial electromagnetic response data, calculate the electromagnetic parameter intensity, and the calculation formula is:

[0013]

[0014] Where, represents the electromagnetic parameter intensity at frequency , represents the signal intensity at the th time point at frequency , is the acquisition time point of the th signal, is the total number of signal acquisitions;

[0015] Based on the electromagnetic parameter intensity, analyze the electromagnetic response characteristics of the signal, and obtain the electromagnetic parameter monitoring data.

[0016] Preferably, the steps for obtaining the early fault characteristic signal are as follows:

[0017] Based on the electromagnetic parameter monitoring data, remove outliers and perform normalization processing to obtain preprocessed electromagnetic data;

[0018] Based on the preprocessed electromagnetic data, extract the statistical metrics of the fault characteristics, and the calculation formula is:

[0019]

[0020] Where, represents the statistical metric of the fault characteristic, is the th data point after processing, is the average value of is the standard deviation of is the total number of samples;

[0021] Based on the statistical metrics, apply a decision tree for pattern recognition to discriminate the fault characteristic signal and obtain the early fault characteristic signal.

[0022] Preferably, the steps for obtaining the envelope analysis result of the fault signal are as follows:

[0023] Based on the early fault feature signal, perform Hilbert transform to map the real signal at each time point to the complex plane to obtain an analytic signal;

[0024] Based on the analytic signal, calculate the envelope of the signal. The calculation formula is:

[0025]

[0026] where represents the envelope of the signal, is the signal strength at time point j, is the phase angle at time point j, is the total number of sample points;

[0027] Based on the envelope of the signal, apply continuous wavelet transform for multi-scale analysis to identify abnormal changes in the signal and form the envelope analysis result of the fault signal.

[0028] Preferably, the steps for obtaining the differential amplification fault feature are as follows:

[0029] Based on the envelope analysis result of the fault signal, perform differential signal amplification to obtain an amplified fault signal;

[0030] Based on the amplified fault signal, apply a high-pass filter to filter out the low-frequency components in the signal and retain the high-frequency part to generate a differential amplification fault signal;

[0031] Based on the differential amplification fault signal, perform a feature extraction process to identify fault features by detecting and analyzing the peaks and waveform patterns in the signal and form the differential amplification fault feature.

[0032] Preferably, the steps for obtaining the fault recognition result are as follows:

[0033] Collect the differential amplification fault features and perform normalization processing to obtain a preprocessed feature vector;

[0034] Based on the preprocessed feature vector, configure and train a support vector machine classifier. The support vector machine classifier uses a radial basis function as the kernel function to obtain a configured and trained SVM classifier;

[0035] Based on the configured and trained SVM classifier, perform fault classification on the test data to identify and classify the fault types and form the fault recognition result.

[0036] Preferably, the steps for obtaining the dynamic fault threshold parameter are as follows:

[0037] Collect the current load level, operating speed, temperature, and type information of the motor to form an input data set for fault diagnosis;

[0038] Based on the input data set for fault diagnosis, calculate a new fault diagnosis threshold, and the calculation formula is:

[0039]

[0040] where, is the new fault diagnosis threshold, is the initial threshold, , , respectively represent the load level, speed, and temperature, , , are adjustment coefficients, is the normalization factor;

[0041] According to the new fault diagnosis threshold, update the fault diagnosis threshold setting to form dynamic fault threshold parameters.

[0042] The present invention provides a method for fault diagnosis of electromechanical equipment based on machine learning, including the following steps:

[0043] Based on the electromagnetic parameter monitoring data, extract the statistical features of the electromagnetic parameters, use the statistical features to identify the fault modes, obtain the abnormal signal patterns through pattern recognition, and generate early fault feature signals;

[0044] Perform frequency domain transformation on the early fault feature signals, extract the envelope data of the signals, conduct envelope analysis of the signals, generate the envelope analysis results of the signals, and use differential amplification to process the envelope analysis results of the signals to obtain differential amplification fault features;

[0045] Use the differential amplification fault features as input data, construct a feature vector, input it into a support vector machine classifier, perform classification processing on the input data through the support vector machine, calculate and obtain the fault probability and type, and output the fault recognition result;

[0046] According to the fault recognition result, with reference to the type and operating conditions of the motor, dynamically adjust the fault diagnosis threshold, perform real-time update of the fault threshold, reset the parameters for fault judgment, and establish and update dynamic fault threshold parameters.

[0047] Compared with the prior art, the advantages and positive effects of the present invention are:

[0048] In the present invention, through high-frequency signal injection and differential signal amplification, early identification and precise positioning of motor faults are achieved. By injecting high-frequency signals, changes in electromagnetic parameters can be captured, which usually occur before the manifestation of physical faults. Differential amplification further enhances the ability to extract useful signals from these minute changes, enabling identification and response before the fault develops to a more severe stage. In addition, the method of dynamically adjusting the fault diagnosis threshold automatically adjusts the threshold according to real-time operation data, improving adaptability and response speed in a changing operating environment, thus having advantages in ensuring equipment safety and operating efficiency. Brief Description of the Drawings

[0049] Figure 1 It is a system flowchart of the present invention. Detailed Embodiment

[0050] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0051] Please refer to Figure 1 , the present invention provides a technical solution: A machine-electronic equipment fault diagnosis system based on machine learning includes:

[0052] A high-frequency signal injection module injects high-frequency signals in the range of 10 kHz to 50 kHz into the motor through the power supply line, analyzes the electromagnetic response of the signals, and obtains electromagnetic parameter monitoring data; based on the electromagnetic parameter monitoring data, statistical features are extracted and pattern recognition is performed to generate early fault feature signals;

[0053] A dynamic feature analysis module transforms the early fault feature signals, extracts envelope analysis data of the signals, and establishes an envelope analysis result of the fault signals; the envelope analysis result of the fault signals is processed by differential signal amplification to obtain differential amplification fault features;

[0054] A fault diagnosis module constructs a feature vector based on the differential amplification fault features and inputs it into a support vector machine classifier to calculate and obtain a fault recognition result;

[0055] A fault threshold dynamic adjustment module adjusts the fault diagnosis threshold in real time according to the motor type and operating conditions, and establishes dynamic fault threshold parameters.

[0056] The steps for obtaining the electromagnetic parameter monitoring data are as follows:

[0057] Inject high-frequency signals in the range of 10 kHz to 50 kHz into the motor through the power line, collect the obtained electromagnetic waveform data as the analysis basis to form initial electromagnetic response data;

[0058] Perform spectral analysis on the initial electromagnetic response data and calculate the electromagnetic parameter intensity. The calculation formula is as follows:

[0059]

[0060] Wherein, represents the electromagnetic parameter intensity at frequency , represents the signal intensity at the th time point at frequency , is the acquisition time point of the th signal, is the total number of signal acquisitions;

[0061] Based on the electromagnetic parameter intensity, analyze the electromagnetic response characteristics of the signal to obtain electromagnetic parameter monitoring data.

[0062] Specifically, the process of injecting a high-frequency signal through the power line includes setting a signal generating device with a transmission frequency in the range of 10 kHz to 50 kHz at the power line port of the motor, ensuring segmented coverage of different frequency bands within the injection range, recording the electromagnetic waveform data at the corresponding moment and synchronously recording information such as injection power, injection duration, and sampling moment. Subsequently, compare each record with the pre-established frequency upper and lower limit ranges for verification. For example, compare the injection frequency with the range of 10 kHz to 50 kHz, and compare the injection power with the RF power range of 0 W to 5 W. Record all data sets that meet the above range requirements and store them as the original electromagnetic waveform data. To check for drift or distortion of the injected signal, monitor the working stability of the injection device and perform frequency recovery when deviation occurs. Temporarily mark the data with obvious drift in the statistical record. Combine the information of all sampling moments to concatenate the complete electromagnetic waveform data set in chronological order and summarize it into the initial electromagnetic response data.

[0063] The advantage of the formula is that by integrating time-domain and frequency-domain information, it can quantitatively analyze the electromagnetic response characteristics of the motor at a given frequency, so as to identify fault characteristics for different frequency components in subsequent steps;

[0064] The acquisition steps of are as follows: At each sampling moment k, analyze the amplitude of the component with frequency . Extract the amplitude at this moment according to the previously acquired electromagnetic waveform data and map it to . The specific value needs to be calculated according to the sampling rate and the amplitude. For example, under the condition of a sampling rate of 400 kS / s, if the signal amplitude at a certain moment is 1.2 V, then record

[0065] The acquisition step is to read the time value from the time-base record at sampling time k. For example, at the k-th sampling point, the time record is , where is provided by the sampling rate. For example, a sampling rate of 400 kS / s corresponds to ;

[0066] The acquisition step is to obtain the total amount of data by multiplying the sampling duration by the sampling rate and use it as the total number of signal acquisitions. For example, when sampling for 1.28 ms, points can be obtained;

[0067] Calculation process:

[0068] Extract all at frequency and calculate the exponential term according to . Assume that here Hz, , then at a certain k value and s, there is:

[0069]

[0070] Multiply this value by and accumulate;

[0071] Divide the accumulated result by , then take the square of the complex modulus as the value of ;

[0072] Repeat the above calculation process, traverse different f values once and obtain a series of ;

[0073] In a specific example, if at k = 50 , the exponential term is approximately 0.997 after calculation. Accumulate this product to the partial sum. After accumulating all 512 sampling points, divide by 512 and then take the modulus square to obtain ;

[0074] This result indicates that at f = 30000 Hz, the electromagnetic parameter intensity is approximately 1.14, which is used to further determine whether there is a possibility of a significant increase in amplitude at the corresponding frequency point.

[0075] The execution process based on the electromagnetic parameter intensity requires, after traversing at different frequencies, sorting all the obtained intensity values according to the frequency index, collecting the sections with higher intensities and comparing whether their values exceed the intensity threshold set in advance with reference to historical fault data. For example, the intensity threshold is set between 0.9 and 1.2. If the If it exceeds 1.2, it will be marked in the intensity anomaly list. If it is lower than 0.9, it is in the low-intensity range and will be recorded routinely. At the same time, for the recorded intensity values, the control data collected under the same conditions in the early stage needs to be retrieved for comparison. For example, compare the intensity recorded at 30 kHz this time with the intensity distribution in the historical record range from 20 kHz to 40 kHz. If this intensity shows a continuous increasing trend, it is classified as a potential anomaly. Synchronously analyze the time series waveform of this section and identify whether there are stray spikes or bandwidth mutations. Finally, summarize the above-extracted intensity and time series characteristics to form electromagnetic response characteristic data, and summarize to obtain electromagnetic parameter monitoring data.

[0076] The steps for obtaining the early fault characteristic signals are as follows:

[0077] Based on the electromagnetic parameter monitoring data, remove the outliers and perform normalization processing to obtain the preprocessed electromagnetic data;

[0078] Based on the preprocessed electromagnetic data, extract the statistical metrics of the fault characteristics. The calculation formula is:

[0079]

[0080] where represents the statistical metric of the fault characteristic, is the th data point after processing, is the average value of is the standard deviation of is the total number of samples;

[0081] Based on the statistical metrics, apply a decision tree for pattern recognition to discriminate the fault characteristic signals and obtain the early fault characteristic signals.

[0082] Specifically, based on the electromagnetic parameter monitoring data, it is necessary to remove outliers and perform normalization processing to obtain preprocessed electromagnetic data. During the execution process, first, all the previously collected electromagnetic parameter monitoring data are arranged in chronological order, and the original values of each sampling point are read in sequence. Then, they are compared with the established multi-dimensional detection ranges. For example, the current is compared with the range from 0A to 5A, the voltage is compared with the range from 0V to 24V, the temperature is compared with the range from 0°C to 90°C, etc. For the values outside the above ranges, they are marked as abnormal, and the sampling points and corresponding timestamps where these abnormal marks appear are recorded. Subsequently, according to the abnormal records accumulated during multiple sampling processes, the overall abnormal distribution of this section of data is statistically analyzed. Samples that deviate significantly from the mainstream distribution are excluded or stored separately to compare the abnormal characteristics in different situations. After removing the outliers, the remaining data is numerically standardized using the normalization method. This processing process needs to first calculate the maximum and minimum values of each electromagnetic parameter and confirm the distribution range, and then use the linear normalization method to map the data to the range from 0 to 1. Multiple rounds of inspections are carried out in combination with the distribution characteristics of each electromagnetic parameter to prevent a very small number of abnormal isolated points from being missed and causing distortion of the normalization result. During the process, by cross-comparing the minimum, maximum, and average values of the samples before and after normalization, if the values show significant deviation or unreasonable concentration trends, the recorded abnormal points and their surrounding sampling values are checked again to identify the parts that may be mislabeled, and a relatively clean and well-distributed coherent normalization data set is obtained. To facilitate subsequent fault feature extraction, the normalized data is smoothed in the time dimension. The moving average is calculated within each sampling window, and the positions with large fluctuations are observed at different levels, so as to obtain a relatively stable electromagnetic waveform sequence. Finally, the time-series data obtained by integrating the above processes of removing outliers and normalization processing is output as preprocessed electromagnetic data.

[0083] The benefit of the formula is that it can reflect the sharpness change of the electromagnetic data in the distribution form through the cubic ratio between the absolute deviation and the standard deviation, and cooperate with the total number of samples to identify the mutation situation of the fault characteristics in the numerical distribution;

[0084] The acquisition steps of are as follows: Define according to the value of the th sampling point in the preprocessed data sequence, and each point corresponds to a specific electromagnetic amplitude; each point corresponds to a specific electromagnetic amplitude; each point corresponds to a specific electromagnetic amplitude;

[0085] The acquisition steps of are as follows: Sum among all sampling points and divide by, and obtain the value from the measured data; Sum among all sampling points and divide by, and obtain the value from the measured data; Sum among all sampling points and divide by, and obtain the value from the measured data;

[0086] The acquisition steps of are as follows: First calculate according to the statistical method The sum is then divided by , and then the square root is taken to obtain the standard deviation;

[0087] The acquisition steps of are as follows: count the number of sampling points of the currently preprocessed data, which is the result of the time series segmentation of the aforementioned electromagnetic parameter monitoring data acquisition;

[0088] Calculation process:

[0089] Set sample points in the actual measurement, and obtain , ;

[0090] Traverse from 1 to 400, calculate point by point and divide by 0.12, then cube the result of this ratio and denote it as ;

[0091] Sum all and divide by 400 again to obtain , and the process can be expressed as:

[0092]

[0093] Select several values in the sample for illustration. For example, when , then , 0.06 divided by 0.12 gives 0.5, and the cube of 0.5 is 0.125. After successively accumulating all sample points, if the result is approximately 162, then ; ;

[0094] This result indicates that the kurtosis measure of the electromagnetic data at the current sample size is approximately 0.405, which is in the medium dispersion range. If the subsequent obtained is greater than 0.8, it reflects that the distribution has a relatively obvious peak component, while less than 0.2 indicates that the data is relatively smooth and stable.

[0095] Based on the statistical measure, a decision tree needs to be applied for pattern recognition to distinguish the fault characteristic signal. During the execution process, first associate and compare the statistical measure with the corresponding electromagnetic data sequence, extract a set of feature values for each sampling window. For example, record various parameters such as the T value, current amplitude range, and voltage fluctuation amplitude within each time slice. Subsequently, by querying the training data set used when defining the decision tree and comparing the splitting thresholds configured on the decision tree nodes, the statistical measure Compare the combined values of each parameter with the pre - summarized threshold ranges one by one. For example, divide the range of T into three levels: 0 to 0.3, 0.3 to 0.6, and 0.6 to 1.0, so as to judge the branch path through the tree nodes. Set corresponding ranges for some other parameters and compare them item by item at each tree node. When the corresponding conditions are triggered, select a certain branch and continue to drill down to the next - level node. When finally reaching the leaf node through layer - by - layer splitting, obtain the fault determination label. Then, merge the multiple sampling windows that are assigned to the same leaf node and generate a fault area statistics list. If it is found that the labels in a certain time period show abnormal characteristics, mark it as the time slice of the fault characteristic signal. Combine the high - frequency and low - frequency characteristic distributions of the electromagnetic data to list the distribution positions of each abnormal characteristic in the time series. Then, integrate these abnormal characteristics and confirm their cumulative occurrence times within the continuous sampling period, and analyze whether they continuously reach the specified condition threshold. For example, when the statistical metric is higher than 0.7, the current amplitude exceeds 3.5A, and it continues for multiple samplings, it can be directly classified as the fault characteristic signal. Finally, summarize these determination results into the early fault characteristic signal.

[0096] The steps to obtain the result of the fault signal envelope analysis are as follows:

[0097] Based on the early fault characteristic signal, perform the Hilbert transform to map the real signal at each time point to the complex plane to obtain the analytic signal;

[0098] Based on the analytic signal, calculate the envelope of the signal. The calculation formula is:

[0099]

[0100] where, represents the envelope of the signal, is the signal intensity at time j, is the phase angle at time j, is the total number of sample points;

[0101] Based on the envelope of the signal, apply the continuous wavelet transform for multi - scale analysis to identify the abnormal changes in the signal and form the result of the fault signal envelope analysis.

[0102] Specifically, based on the early fault characteristic signals, combined with the electromagnetic timing data and fault characteristic information obtained previously, first, read the corresponding real signal values one by one according to the recorded time index and check their amplitude ranges and sampling moments during each reading. By comparing with the pre-established effective ranges, such as the current range from 0A to 5A, the voltage range from 0V to 24V, etc., mark the sampling values to ensure no missed or misread phenomena. Then, map all the real signal values point by point in the order of the time series and map them to the complex plane. At this time, it is necessary to determine the phase data at the same time for each sampling point. If it is found during the phase measurement process that the phases of some sampling points deviate significantly from the main distribution, list these sampling points separately in an index and recheck several adjacent sampling points around them. Next, deduce the coordinate positions of each sampling point in the complex plane from the amplitude and phase data of each sampling point, use the amplitude as the modulus length and the phase to define the polar angle, and then merge them on all sampling points to obtain a signal sequence in complex form. Check whether there are obvious jumps or discontinuous records in the entire signal sequence. By arranging these complex sequences in order and counting the distances between adjacent sampling points, check whether there is a discontinuous phase phenomenon caused by sudden increases or decreases in sampling points. After completing the coordinate mapping of all sampling points, establish a data structure on the complex plane, generate an analytic signal and store it in the analytic signal list for subsequent use in the form of an index corresponding to the original time series. Finally, associate this analytic signal with the previously saved early fault characteristic signals and record the corresponding relationship between the two at the same time series position for further processing of these analytic signals in subsequent steps to obtain the analytic signal.

[0103] The advantage of the formula is that by means of the cumulative average processing of the components of the real signal in the complex plane, the amplitudes and phases of different sampling points can be unified and integrated into a concise envelope amplitude metric, thus providing a more intuitive reference for judging high-frequency and low-frequency components in subsequent steps;

[0104] The acquisition step of is to read the amplitude value at the moment The amplitude value is provided by the amplitude field in the analytic signal previously mapped to the complex plane, and the specific value is obtained by the amplitude detection instrument in the previous test. For example, the amplitude is recorded as 0.83 at the moment ;

[0105] The acquisition step of is to read the phase information from the same moment and uniformly measure it within the range of 0 to The value comes from the previous phase detection result. For example, the phase is measured as radians at the moment;

[0106] The acquisition steps are as follows: count all valid sampling points and exclude the number of abnormal points marked as deviating from the main distribution before, and then use the obtained count as , for example, after detection, there are approximately 300 remaining time points, that is ;

[0107] Calculation process:

[0108] Sum the at each time point and divide by to obtain A, in the form of:

[0109]

[0110] Sum the at each time point and divide by to obtain B, in the form of:

[0111]

[0112] Let , calculate according to the above definition. If in a given actual example, there is , and through statistics, it is obtained that:

[0113]

[0114] Then:

[0115]

[0116] Furthermore:

[0117]

[0118] This result indicates that the numerical value of the current signal envelope amplitude is approximately 0.80. If the same operation is performed on different time segments in the future and similar or higher amplitude values are obtained, it means that there are envelope components with strong amplitudes in those time segments;

[0119] Based on the envelope of the signal, continuous wavelet transform is applied for multi-scale analysis to identify abnormal changes in the signal. During the execution process, the previously obtained signal envelope sequence and the amplitude values at the corresponding moments are listed item by item in the same data table, and the phase information, electromagnetic vibration frequency, and possibly associated current reference values at that moment are retained in the table. At the same time, before the continuous wavelet transform operation, this part of the signal envelope should be divided into frequency bands and the frequency range should be marked in segments from the lower frequency to the higher frequency. For each segment, the amplitude changes in the envelope sequence are read sequentially and compared with the corresponding range of the set multi-scale frequency bands. For example, 10 Hz to 500 Hz is called the low-frequency band, 500 Hz to 2000 Hz is called the medium-frequency band, and 2000 Hz to 10000 Hz is called the high-frequency band. Then, the envelope amplitude distribution is smoothed in each segment and the local average amplitude and central frequency are calculated. Comparing with the recorded statistical distribution data, if it is found that the envelope amplitude continuously increases beyond the set segment threshold in a certain segment, it is marked. The threshold is set based on the quantiles obtained by summarizing the envelope amplitudes in the same frequency band from the vibration and noise data collected during 300 hours of continuous operation of the machine. The signal segments at these marked positions are input into the wavelet transform process for multi-scale expansion. At different scales, the continuous evolution of the amplitude on the time axis is observed. The possible transient spikes are compared multiple times through the envelope distributions at adjacent scales, and the parallel records of current and voltage in the reference time series are combined to check for synchronous anomalies. Through further inspection of these abnormal points or sections, the fault signal envelope analysis results are formed.

[0120] The steps for obtaining the differential amplification fault characteristics are as follows:

[0121] Based on the fault signal envelope analysis results, differential signal amplification is performed to obtain the amplified fault signal;

[0122] Based on the amplified fault signal, a high-pass filter is applied to filter out the low-frequency components in the signal and retain the high-frequency part to generate the differential amplification fault signal;

[0123] Based on the differential amplification fault signal, the feature extraction process is executed. By detecting and analyzing the peaks and waveform patterns in the signal, the fault characteristics are identified to form the differential amplification fault characteristics.

[0124] Specifically, based on the results of the fault signal envelope analysis, first read the fault signal envelope analysis data formed previously one by one according to the time index, and check the amplitude information and phase indicators corresponding to each time point while reading. Compare this information with the previously recorded current and voltage reference values one by one. For example, compare the current reference value with the range from 0 A to 5 A, and compare the voltage reference value with the range from 0 V to 24 V to determine whether the corresponding time point belongs to the normal range or the deviation range. When there are multiple deviations or continuous fluctuations, mark the interval and enter the differential operation sequence. Then, when performing differential processing on these marked intervals, select the difference between the amplitude values of adjacent sampling points within the same time window and record the difference result. If the difference exceeds the specified value multiple times within the same window, it is necessary to query the origin of the specified value. For example, perform percentile statistics on the amplitude difference distribution data collected after the device has been continuously running for 300 hours and set a certain percentile as the threshold. Mark the difference values that exceed this threshold and add them to the amplification list. Subsequently, perform amplification operations on the marked difference values one by one. Obtain the amplified fault signal by multiplying the difference value by the amplification factor. The selection of the amplification factor needs to combine the differential amplitude distribution range obtained previously and determine a suitable gain through segmented cumulative calculation, so that it still maintains appropriate recognizability in the medium-amplitude scenario after amplification, and does not saturate in the large-amplitude scenario. After completing the differential amplification, update the differential amplification result set and save the final amplified fault signal.

[0125] Based on the amplified fault signal and combined with the previously determined amplification factor distribution, perform a high-pass filtering operation on each amplified data segment. First, classify the time-domain signal into segments before filtering and give the specific center frequency of each segment. Define the range from 0 Hz to 50 Hz as the low-frequency band, the range from 50 Hz to 500 Hz as the middle-frequency band, and 500 Hz and above as the high-frequency band, etc. Then, when filtering, retain the high-frequency band components and filter out the energy components below the center frequency of the segment. During the filtering process, it is necessary to determine a cut-off frequency based on the long-term accumulated statistics of the motor running vibration frequency and use this cut-off frequency as the main basis for the filtering parameter. For example, if it is statistically found from the previously extracted vibration data samples that most of the normal vibrations of the device are concentrated below 200 Hz, then set the cut-off frequency of the high-pass filter to be between 200 Hz and 300 Hz, and fine-tune the specific threshold by judging whether it will frequently cross this frequency point during continuous operation. After the low-frequency components in the signal are attenuated, then observe whether the amplitude of the high-frequency part remains within the normal range. For example, compare the high-frequency amplitude with the previously recorded high-frequency amplitude distribution range. If there are multiple obvious increases and the amplitude exceeds the reference range, mark it. Finally, summarize the results retained and output by the high-pass filter as the differential amplified fault signal and record it.

[0126] Based on the differential amplified fault signal, when performing feature extraction, first match the previously obtained differential amplified fault signal with the time axis, read the amplitude peak points and waveform change trends in several sampling windows one by one, and at the same time record the corresponding current amplitude or voltage amplitude during this period and compare it with the current range of 0A to 5A and voltage range of 0V to 24V during the normal operation of the device for multiple times. If peaks exceeding the reference peak threshold calculated from the previous sample distribution are detected in multiple places, these peak points are regarded as fault feature candidates, and further compare the peak change amplitude and occurrence frequency in adjacent sampling windows. If relatively concentrated and large-amplitude peaks appear multiple times in the same window, mark them in the fault feature list. At the same time, combine the extreme value positions in the rising and falling segments of the waveform to detect whether they conform to the fault features of waveform mutation or jitter type. Summarize each candidate peak and its surrounding waveform pattern features and evaluate the repeated distribution of peaks in the cross-window scenario. If similar intensity waveform patterns are found in multiple time windows, further confirm that this pattern belongs to stable fault features. Finally, record all the peaks and waveform pattern sets that meet the fault feature recognition conditions to generate differential amplified fault features.

[0127] The steps to obtain the fault recognition result are as follows:

[0128] Collect differential amplified fault features, perform normalization processing to obtain the preprocessed feature vector;

[0129] Based on the preprocessed feature vector, configure and train a support vector machine classifier. The support vector machine classifier uses the radial basis function as the kernel function to obtain the configured and trained SVM classifier;

[0130] Based on the configured and trained SVM classifier, perform fault classification on the test data, identify and classify the fault types to form the fault recognition result.

[0131] Specifically, to collect differential amplification fault features, first read the timing data and corresponding amplitude information of each differential amplification fault feature obtained previously one by one. Aggregate these amplitude information together according to the same time index and summarize them in the form of a list. Then, check the value range of each record in the summary list. For example, compare the current amplitude with the range of 0A to 5A, and compare the voltage amplitude with the range of 0V to 24V, so as to identify whether there is data beyond the range. If it is found that the data is beyond the range, mark it in the summary list and record the amplitude status of this sample point separately. Next, calculate the statistics of the entire differential amplification fault feature set, including arithmetic mean, maximum value, minimum value, variance, etc. Use these statistics to evaluate the gap and distribution concentration degree between each fault feature. Then, set a distribution threshold interval according to experience or historical analysis results, and mark the samples that are obviously in the extreme value position as candidate anomalies for further inspection. After completing the preliminary marking, start to execute the standardization process. By subtracting the corresponding mean value from each feature data and then dividing it by the corresponding standard deviation, map each sample to the standard normal distribution interval with a mean of 0 and a standard deviation of 1. For example, when processing the current amplitude, first subtract its mean value and then divide it by the standard deviation. The same method is also used for the voltage amplitude and vibration amplitude. After the process ends, a set of numerical sets with a more compact range and not affected by the original dimension is formed. At the same time, retain the time index of each record in each record to maintain its referenceability in the subsequent stage. Finally, merge all the differential amplification fault features after this standardization process and arrange them in a fixed order to obtain the preprocessed feature vector.

[0132] Based on the preprocessed feature vector, first assign the corresponding label information in the training set to each feature vector record. If the true labels of the device in different fault scenarios have been collected previously, retrieve the corresponding records from the label set and append them to the end of the training data table. Then, compare each feature vector one by one before training to check whether the numerical distribution in fields such as current amplitude, voltage amplitude, and vibration parameters is consistent with the common feature intervals in the label. If there is an inconsistency and the value has exceeded the set detection range, mark this record in the training data table and check whether the attributed label needs to be corrected. After all samples are verified, split them into a training set and a validation set according to the previously collected differential amplification fault feature data and corresponding label information. Then, call the support vector machine classifier with the radial basis function as the kernel function and gradually input the feature vectors and labels in the training set, and use the batch gradient method for iterative training. During the iterative process, update the classification decision boundary multiple times and evaluate the classification accuracy with the validation set at the end of each iteration. Stop training when the evaluation result is stable, record the final classification accuracy to confirm the classification effect of the model in fault type recognition. Finally, package the support vector machine classifier with stable classification performance uniformly, record the configuration information, and regard it as the configured and trained SVM classifier.

[0133] Based on the configured and trained SVM classifier, first perform the same normalization preprocessing on the feature vectors of each record in the test dataset with the previously generated standardized parameters, including adjusting the differential feature terms such as current amplitude, voltage amplitude, and vibration difference with the same mean and standard deviation. Then, send these processed vector data into the trained SVM classifier, use the previously trained radial basis function kernel to identify each test data item by item, summarize the classification label results generated for each identification, and compare them with the true labels of the test samples recorded in advance. If the classification label is inconsistent with the true label, mark it in the test dataset for subsequent review. After all test samples are identified, perform statistics based on the comparison results to check the accuracy rate and confusion situation of the classifier for various fault types. If the accuracy rate exceeds the preset threshold, include this classifier in the subsequent available list and apply it to the real-time monitoring of the device. Otherwise, it is necessary to trace back to the training set distribution and feature processing method again to find the reason. Finally, classify the fault classification results of each test sample to display different fault types and their distributions on the time axis, forming the fault identification results.

[0134] The steps to obtain the dynamic fault threshold parameter are as follows:

[0135] Collect the current load level, running speed, temperature of the motor, and the type information of the motor to form the input dataset for fault diagnosis;

[0136] Based on the input dataset for fault diagnosis, calculate the new fault diagnosis threshold, and the calculation formula is:

[0137]

[0138] where, is the new fault diagnosis threshold, is the initial threshold, 、 、 respectively represent the load level, speed, and temperature, 、 、 are the adjustment coefficients, is the normalization factor;

[0139] According to the new fault diagnosis threshold, update the fault diagnosis threshold setting to form the dynamic fault threshold parameter.

[0140] Specifically, information such as the current load level, operating speed, temperature, and type of the motor is collected. During the execution process, first, the indicators related to the motor load are read separately and compared with the established load range list. For example, the load value is compared with the range of 0% to 120%. If the load data exceeds 120%, the sample point is marked in the record and archived for subsequent analysis. Then, the operating speed of the motor is recorded and compared with the range of 0 rpm to 10,000 rpm, the temperature is compared within the range of 0°C to 90°C, and the type information of the motor is verified to match the pre - confirmed type list in the database. If a type code is not in the list, it is listed as a type to be confirmed. Then, the above data items that meet the range or are to be confirmed are integrated into the input data set for fault diagnosis in sequence. To ensure the correspondence of information such as speed, load, and temperature among different time periods, it is necessary to combine and record the data items at each data sampling moment to form a complete record. After completion of the summary, each record is checked line by line for the mutual relationship between its load, operating speed, and temperature. For example, the average load and operating speed conditions corresponding to a temperature not exceeding 40°C are statistically analyzed, and the load and speed at temperatures higher than 40°C are compared accordingly. If the load continuously increases or the operating speed changes too fast under high - temperature conditions, the time period is marked for subsequent correlation analysis in combination with the calculation of diagnostic thresholds in the following steps, and finally, an input data set for fault diagnosis is formed.

[0141] The benefit of the formula is that by combining the three elements of load level, speed, and temperature and introducing multiple adjustment coefficients to dynamically regulate the threshold, the threshold for fault judgment can be made more flexible and suitable for the actual working conditions;

[0142] The acquisition step of is the initial threshold set when the device leaves the factory or is initially deployed. This value is obtained through industrial verification data and allows for minor fine - tuning in the later stage;

[0143] The acquisition step of is that the actual load monitoring device records the average load level of the motor over a period of time and compares this value with the rated load to obtain a proportional value. For example, if the measured load is 80% of the rated load, i.e., ;

[0144] The acquisition step of is to collect the operating speed of the motor through a speed sensor, compare it with the maximum speed of 10,000 rpm, and convert the result into a quantization value between 0 and 1. For example, when the operating speed is 5,800 rpm at this stage, let ;

[0145] The acquisition steps are as follows: measure the motor temperature with a temperature sensor. When the current temperature is 35°C, take the digit of the actual temperature and directly assign it;

[0146] and and The acquisition steps of the adjustment coefficient are as follows: sort out the industrial test data in multiple rounds, statistically analyze the sensitivity of the failure rate per unit time to the above three items of load, speed, and temperature, and obtain the corresponding adjustment coefficient through weighted average processing;

[0147] The acquisition steps of

[0148] are formed by evaluating and sorting out the threshold correction amplitude in multiple operation results. For example, during 300 consecutive hours of equipment operation, record the fluctuation range of the threshold before and after each failure, and calculate the correction coefficient;

[0149] Calculation process:

[0150] Given an example, let , ;

[0151] Load level , speed quantization , temperature ;

[0152] Obtain , , through the failure sensitivity analysis sorted out above and record it as a reference value;

[0153] Learn from the existing normalization standard ;

[0154] Calculate , calculate , and then calculate ;

[0155] Accumulate the numerator, , , , the total sum ;

[0156] Divide by , the result , and then multiply by , to obtain ;

[0157] Finally, add it to to obtain ;

[0158] The result shows that under the current load, speed and temperature conditions, the new fault diagnosis threshold is calculated to be 1.908. If the fault index obtained in subsequent monitoring exceeds 1.908, it can be determined as abnormal.

[0159] According to the new fault diagnosis threshold, when updating the fault diagnosis threshold setting, all registered motor operation data need to be verified again. During the execution process, the sampled data is read item by item and compared with the threshold of 1.908. If the corresponding fault index value exceeds 1.908, the data item will be marked as suspected abnormal in the summary record. At the same time, cross-reference is made with the earlier fault threshold to identify whether there were also fault symptoms in a similar range before. If it is found that there is a difference between the current mark and the previous result, the difference record is split and the abnormal mark is retained under the current threshold scheme. If the fault index continuously exceeds the threshold multiple times for the same type of equipment or the same type of information, the number of overlimit times is counted and the current fault symptom conclusion is reported. To ensure the coherence of the next diagnosis, information such as the speed, load, and temperature of the motor at the time of overlimit also needs to be recorded for subsequent further analysis. Finally, all the latest determined situations are saved into the updated fault diagnosis threshold parameters.

Claims

1. A mechanical and electrical equipment fault diagnosis system based on machine learning, characterized in that: The system comprises: A high-frequency signal injection module injects a high-frequency signal in the range of 10kHz to 50kHz into the motor through the power supply line, analyzes the electromagnetic response of the signal, and obtains electromagnetic parameter monitoring data; based on the electromagnetic parameter monitoring data, extracts statistical features and performs pattern recognition to generate early fault characteristic signals; The dynamic feature analysis module transforms the early fault feature signal, extracts the envelope analysis data of the signal, and establishes the fault signal envelope analysis result; processes the fault signal envelope analysis result by differential signal amplification to obtain the differential amplification fault feature; A fault diagnosis module, based on the differential amplification fault feature, constructs a feature vector and inputs it into a support vector machine classifier to calculate and obtain a fault identification result; The fault threshold dynamic adjustment module adjusts the fault diagnosis threshold in real time according to the motor type and operating conditions, and establishes dynamic fault threshold parameters.

2. The electromechanical equipment fault diagnosis system based on machine learning according to claim 1 is characterized in that: The steps for acquiring the electromagnetic parameter monitoring data are as follows: A high-frequency signal in the range of 10kHz to 50kHz is injected into the motor through the power line, and the collected electromagnetic waveform data is used as the basis for analysis to form initial electromagnetic response data; The initial electromagnetic response data is subjected to spectrum analysis to calculate the electromagnetic parameter intensity, and the calculation formula is: in, Represents the frequency The electromagnetic parameter intensity under Representative time point at frequency The signal strength, For the The acquisition time point of each signal, is the total number of signal acquisitions; Based on the electromagnetic parameter intensity, the electromagnetic response characteristics of the signal are analyzed to obtain electromagnetic parameter monitoring data.

3. The electromechanical equipment fault diagnosis system based on machine learning according to claim 1 is characterized in that: The steps of obtaining the early fault characteristic signal are: Based on the electromagnetic parameter monitoring data, outliers are removed and normalized to obtain preprocessed electromagnetic data; Based on the pre-processed electromagnetic data, the statistical measure of the fault characteristics is extracted, and the calculation formula is: in, A statistical measure that characterizes the fault, After processing data points, for The average value of for The standard deviation of is the total number of samples; Based on the statistical measurement, a decision tree is applied to perform pattern recognition, identify fault characteristic signals, and obtain early fault characteristic signals.

4. The electromechanical equipment fault diagnosis system based on machine learning according to claim 1, characterized in that: The steps for obtaining the fault signal envelope analysis result are: Based on the early fault characteristic signal, a Hilbert transform is performed to map the real signal at each time point to a complex plane to obtain an analytical signal; Based on the analytical signal, the envelope of the signal is calculated, and the calculation formula is: in, represents the envelope of the signal, is the signal strength at time j, is the phase angle at time j, is the total number of sample points; Based on the signal envelope, continuous wavelet transform is applied to perform multi-scale analysis to identify abnormal changes in the signal and form the fault signal envelope analysis results.

5. The electromechanical equipment fault diagnosis system based on machine learning according to claim 1, characterized in that: The steps of obtaining the differential amplifier fault characteristics are as follows: Based on the fault signal envelope analysis result, differential signal amplification is performed to obtain an amplified fault signal; Based on the amplified fault signal, a high-pass filter is applied to generate a differential amplified fault signal by filtering out low-frequency components in the signal and retaining high-frequency components; Based on the differential amplification fault signal, a feature extraction process is performed to identify fault features by detecting and analyzing peak values ​​and waveform patterns in the signal to form differential amplification fault features.

6. The electromechanical equipment fault diagnosis system based on machine learning according to claim 1, characterized in that: The steps of obtaining the fault identification result are: Collecting the differential amplification fault features, performing standardization processing, and obtaining a preprocessed feature vector; Based on the preprocessed feature vector, a support vector machine classifier is configured and trained, wherein the support vector machine classifier uses a radial basis function as a kernel function to obtain a configured and trained SVM classifier; Based on the configuration and the trained SVM classifier, the test data is fault classified, the fault type is identified and classified, and a fault identification result is formed.

7. The electromechanical equipment fault diagnosis system based on machine learning according to claim 1, characterized in that: The steps for obtaining the dynamic fault threshold parameter are: Collect information about the motor's current load level, operating speed, temperature, and motor type to form an input data set for fault diagnosis; Based on the input data set of the fault diagnosis, a new fault diagnosis threshold is calculated, and the calculation formula is: in, is the new fault diagnosis threshold, is the initial threshold, , , represent load level, speed and temperature respectively, , , is the adjustment factor, is the normalization factor; According to the new fault diagnosis threshold, the fault diagnosis threshold setting is updated to form a dynamic fault threshold parameter.

8. A method for diagnosing electromechanical equipment faults based on machine learning, characterized in that: The electromechanical equipment fault diagnosis system based on machine learning according to any one of claims 1 to 7 comprises the following steps: Based on the electromagnetic parameter monitoring data, the statistical characteristics of the electromagnetic parameters are extracted, the fault mode is identified using the statistical characteristics, the abnormal signal mode is obtained through pattern recognition, and the early fault characteristic signal is generated; Perform frequency domain transformation on the early fault characteristic signal, extract the signal envelope data, perform signal envelope analysis, generate signal envelope analysis results, use differential amplification to process the signal envelope analysis results, and obtain differential amplification fault characteristics; The differential amplifier fault feature is used as input data to construct a feature vector, which is input into the support vector machine classifier. The support vector machine classifies the input data, calculates and obtains the fault probability and type, and outputs the fault identification result. According to the fault identification result, referring to the type and operating conditions of the motor, the fault diagnosis threshold is dynamically adjusted, the fault threshold is updated in real time, the fault judgment parameters are reset, and the dynamic fault threshold parameters are established and updated.

Images