Mobile power supply fault detection method and system based on intelligent diagnosis

Abstract

The invention relates to the technical field of electrical fault detection, in particular to a mobile power supply fault detection method and system based on intelligent diagnosis, and the method comprises the following steps: collecting voltage and current data of a mobile power supply in a normal operation state, and generating an electrical performance baseline data set through recording the timestamp of the data and a corresponding electrical performance index; and based on the electrical performance baseline data set, extracting time sequence features of voltage and current to obtain electrical performance feature data. According to the invention, based on voltage and current data acquisition in a normal state of the mobile power supply, a baseline data set covering timestamps and electrical performance indexes is established, and detail information of electrical performance changes is captured by extracting time sequence features such as peak values, mean values and standard deviations; meanwhile, frequency spectrum features are accurately extracted by means of fast Fourier transform, the intensity and phase information of frequency components is mastered, then signals are decomposed through wavelet transform, and richer electrical performance anomaly information under multiple scales is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical fault detection, and particularly to a mobile power supply fault detection method and system based on intelligent diagnosis. Background Art

[0002] The mobile power supply fault detection method refers to the detection and analysis methods adopted for portable energy storage devices such as mobile power supplies (portable chargers) when they exhibit abnormal charging, insufficient power, abnormal discharging, unstable output, or other functional faults.

[0003] The existing technologies only focus on the diagnosis of abnormal phenomena after the mobile power supply fails, mainly performing local analysis through manual or simple instruments, lacking deeper feature mining for abnormal voltage and current manifestations, with insufficient accuracy in data feature extraction, resulting in difficulty in effectively detecting subtle or early abnormal signals; at the same time, the existing methods lack the analysis of electrical performance characteristics from the frequency domain or multi-scale perspective, often being unable to effectively process the hidden and intermittent abnormal fluctuations in the signals, causing the neglect of potential hazards and prone to false negatives or false positives in fault diagnosis. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies existing in the prior art and propose a mobile power supply fault detection method and system based on intelligent diagnosis.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions. A mobile power supply fault detection method based on intelligent diagnosis includes the following steps:

[0006] Collect voltage and current data of the mobile power supply in the normal operating state, generate an electrical performance baseline dataset by recording the timestamps of the data and the corresponding electrical performance indicators; based on the electrical performance baseline dataset, extract the time series features of voltage and current to obtain electrical performance feature data;

[0007] Based on the electrical performance feature data, use the fast Fourier transform to analyze the intensity and phase of each frequency component, extract key spectral features, and generate spectral feature data; perform multi-scale decomposition on the spectral feature data, and use the wavelet transform method to decompose the signal into details and approximation data on multiple time scales to obtain multi-scale spectral data;

[0008] By comparing the multi-scale spectral data with the corresponding multi-scale spectral features in the electrical performance baseline dataset, identify fluctuations and non-periodic anomalies deviating from the normal state, and generate a preliminary fault diagnosis result; perform geometric curvature analysis on the preliminary fault diagnosis result, calculate the geometric deformation degree between data points, and obtain fault identification feature data;

[0009] Based on the fault identification feature data, a fault prediction model is trained through a support vector machine, and the fault prediction model is applied to continuously monitor the mobile power supply to obtain a fault monitoring and early warning result.

[0010] Preferably, the steps for obtaining the electrical performance baseline data set are as follows: continuously monitor and record the current and voltage data of the mobile power supply, add a time stamp to each piece of data to form a preliminary current-voltage time series record;

[0011] Based on the preliminary current-voltage time series record, calculate the instantaneous power, energy consumption, and efficiency of the current and voltage to obtain electrical performance index data;

[0012] Based on the electrical performance index data, integrate each index with the time stamp to construct an electrical performance baseline data set.

[0013] Preferably, the steps for obtaining the electrical performance feature data are as follows: extract the voltage data and current data from the electrical performance baseline data set, calculate the change rate between adjacent data points, smooth the change rate sequence, and at the same time combine the extreme points of the voltage data and current data to calibrate the fluctuation amplitude and stable interval of the current and voltage, and generate processed voltage-current time series data;

[0014] Based on the processed voltage-current time series data, calculate the time series feature values of the voltage and current. The calculation formula is:

[0015]

[0016] where P represents the time series feature value, Q i is the voltage data at time point i, Q i-1 is the voltage data at time point i - 1, I j is the current data at time point j, I j-1 is the current data at time point j - 1, m is the total number of data in the time series, Q k is the voltage data at time point k, I k is the current data at time point k;

[0017] Based on the time series feature values, combine the extreme values, mean values, and change trends of the current and voltage to obtain electrical performance feature data.

[0018] Preferably, the steps for obtaining the spectrum feature data are as follows: preprocess the voltage and current time series in the electrical performance feature data, including removing noise and outliers, and applying a window function to reduce spectrum leakage to obtain processed time series data;

[0019] Based on the processed time series data, calculate the comprehensive feature of each frequency. The calculation formula is:

[0020]

[0021] Among them, S(f) represents the comprehensive feature of frequency f, X(f) is the result of the fast Fourier transform at frequency f, is the conjugate of X(f), Re(X(f)) and Im(X(f)) respectively represent the real part and the imaginary part of X(f), represents the phase, α, β, γ are adjustment coefficients, and σ is a smoothing parameter;

[0022] Based on the comprehensive feature, key spectral features are screened, and the key spectral features are combined to generate spectral feature data.

[0023] Preferably, the step of obtaining the multi-scale spectral data is: based on the spectral feature data, calculate the wavelet coefficients at each scale, and the calculation formula is:

[0024]

[0025] Among them, W(a, b) represents the wavelet coefficient at scale a and position b, S n is the nth point of the spectral data, ψ is the wavelet mother function, N is the total number of data points, n - b represents the time delay, and gw is the Gaussian window parameter for adjusting the window width;

[0026] Based on the wavelet coefficients, the signal is reconstructed to separate and identify the details and approximation parts at each time scale, and multi-scale spectral data is obtained.

[0027] Preferably, the step of obtaining the preliminary fault diagnosis result is: obtain the multi-scale spectral data and the electrical performance baseline data set, perform standardization processing on the multi-scale spectral data and the electrical performance baseline data set to make the data formats consistent, and obtain the standardized multi-scale spectral data and electrical performance baseline data set;

[0028] Based on the standardized multi-scale spectral data and the electrical performance baseline data set, compare the spectral features at each scale, identify the deviations in frequency response, energy distribution, and time series stability, and obtain the deviation analysis result;

[0029] Based on the deviation analysis result, evaluate and classify each identified anomaly, sort and mark the anomalies according to the severity of the impact and the frequency of occurrence, and define the estimated fault type for each anomaly to obtain the preliminary fault diagnosis result.

[0030] Preferably, the step of obtaining the fault identification feature data is: based on the preliminary fault diagnosis result, calculate the geometric deformation degree between data points, obtain the curvature value, and the calculation formula is:

[0031]

[0032] Among them, K(s) represents the curvature value, and Δx and Δy are the horizontal and vertical distance differences of adjacent data points in the time series, respectively;

[0033] Based on the curvature value, analyze the curvature change between each data point, and obtain the fault identification feature data by integrating the curvature change results.

[0034] Preferably, the steps for obtaining the fault monitoring and early warning result are as follows: Based on the fault identification feature data and multi-scale spectrum data, configure and train a support vector machine model, select a radial basis function and adjust the model penalty coefficient and kernel parameter to obtain a trained fault prediction model;

[0035] Based on the trained fault prediction model, deploy the trained fault prediction model to the mobile power supply monitoring, analyze the collected power operation data in real time, evaluate the fault risk, and obtain the fault monitoring and early warning result.

[0036] The present invention provides a mobile power supply fault detection system, including:

[0037] A data acquisition module that records the voltage and current of the mobile power supply during operation, attaches a time stamp, and generates an electrical performance baseline data set;

[0038] A feature extraction module that extracts the peak value, mean value, and standard deviation of the voltage and current from the electrical performance baseline data set to generate electrical performance feature data;

[0039] A spectrum analysis module that applies a fast Fourier transform to the electrical performance feature data, analyzes the intensity and phase of each frequency component, generates spectrum feature data, and performs multi-scale decomposition on the spectrum feature data through wavelet transform to obtain multi-scale spectrum data;

[0040] A fault pre-diagnosis module that compares the multi-scale spectrum data with the electrical performance baseline data set, identifies fluctuations and non-periodic anomalies, generates a preliminary fault diagnosis result, performs geometric curvature analysis on the preliminary fault diagnosis result, calculates the deformation degree between data points, and obtains fault feature data;

[0041] A fault prediction module that trains a support vector machine model based on the fault feature data, performs continuous monitoring, predicts and issues a fault monitoring and early warning, and generates a fault early warning result.

[0042] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0043] In the present invention, based on the voltage and current data acquisition of the mobile power supply in the normal state, a baseline data set covering time stamps and electrical performance indicators is established. By extracting time series features such as peaks, means, and standard deviations, the detailed information of electrical performance changes is captured. At the same time, the spectral features are accurately extracted by means of the fast Fourier transform to master the intensity and phase information of the frequency components. Then, the wavelet transform is used to decompose the signal to obtain richer electrical performance anomaly information at multiple scales. And through the comparison of the multi-scale data with the baseline state, the non-periodic fault fluctuations deviating from the normal state are accurately identified, improving the sensitivity and accuracy of fault location. Further, the geometric curvature analysis is carried out to further finely quantify the degree of data anomaly, excavate more subtle anomaly feature changes, and enhance the detection ability of hidden faults inside the mobile power supply. At the same time, based on the support vector machine, fault prediction and real-time monitoring are realized, early warning of faults is achieved, the potential safety hazards caused by delayed fault diagnosis are reduced, and the overall service life and operation stability of the mobile power supply are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the steps of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] 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 accompanying 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.

[0046] Please refer to Figure 1 , the present invention provides a technical solution, a method for detecting faults in a mobile power supply based on intelligent diagnosis, including the following steps:

[0047] Collect the voltage and current data of the mobile power supply in the normal operating state, and generate an electrical performance baseline data set by recording the time stamps of the data and the corresponding electrical performance indicators; based on the electrical performance baseline data set, extract the time series features of the voltage and current to obtain electrical performance feature data;

[0048] Based on the electrical performance feature data, use the fast Fourier transform to analyze the intensity and phase of each frequency component, extract the key spectral features, and generate spectral feature data; perform multi-scale decomposition on the spectral feature data, and by means of the wavelet transform method, decompose the signal into details and approximate data on multiple time scales to obtain multi-scale spectral data;

[0049] By comparing the multi-scale spectral data with the corresponding multi-scale spectral features in the electrical performance baseline data set, identify the fluctuations and non-periodic anomalies deviating from the normal state, and generate a preliminary fault diagnosis result; perform geometric curvature analysis on the preliminary fault diagnosis result, calculate the geometric deformation degree between data points, and obtain fault identification feature data;

[0050] Based on the fault identification feature data, a fault prediction model is trained through a support vector machine, and the fault prediction model is applied to continuously monitor the mobile power supply to obtain a fault monitoring and warning result.

[0051] The steps for obtaining the electrical performance baseline data set are as follows: continuously monitor and record the current and voltage data of the mobile power supply, add a time stamp to each piece of data to form a preliminary current-voltage time series record;

[0052] Based on the preliminary current-voltage time series record, calculate the instantaneous power, energy consumption, and efficiency of the current and voltage to obtain electrical performance index data;

[0053] Based on the electrical performance index data, integrate each index with the time stamp to construct an electrical performance baseline data set.

[0054] Specifically, refer to the acquisition device installed on the mobile power supply to obtain current and voltage information, and attach a time identifier corresponding to the recording moment to each record. First, set a current reference range of 0A to 5A and a voltage reference range of 0V to 24V. If the measured value of a certain record exceeds this current or voltage range, then in subsequent processing, mark this record as an abnormal measurement and temporarily not include it in the analysis scope. These range values can be set through prior tests or empirical values. For example, first use a typical mobile power supply sample for continuous measurement, count the minimum and maximum values of the current and voltage, and then expand a certain proportion at both ends of this range to obtain the current range values used currently. Then, record the real-time status of the mobile power supply at a frequency of once per second, and associate the actually collected data with the corresponding time identifier. During the association process, store the collected current and voltage measurement points as a sequentially arranged sequence data according to the time sequence. If it is found that the adjacent sampling time interval exceeds the pre-specified 1-second error range (this range can be obtained from the internal clock accuracy verification, for example, set ±0.05 seconds as the allowable deviation), then mark the occurrence of time sequence fluctuations in this sampling in the record and continue with subsequent records. Finally, after summarizing several sampling records, arrange the current and voltage information in time sequence uniformly to obtain a preliminary current-voltage time series record.

[0055] Based on the preliminary current-voltage-time series records, the current and voltage data at each moment are calculated to obtain the instantaneous power, and during the statistical stage, the cumulative duration of the recording interval is combined to deduce the corresponding energy consumption and efficiency. First, the instantaneous power is calculated by P(t) = U(t) × I(t), where U(t) is the voltage value at that moment and I(t) is the current value at that moment. To determine the reasonable fluctuation range of the power, multiple sets of measurement data under typical operating conditions can be selected first, and the average power is calculated. Then, a temporary power threshold is defined with a 20% floating range added and subtracted from this average value. If the calculated instantaneous power exceeds this range at a certain time, it is regarded as a possible outlier and temporarily stored for future use. Subsequently, in the energy consumption calculation, the power at each moment is multiplied by the time interval to obtain the corresponding energy increment, and the total energy consumption is obtained by summing all the increments. For the determination of efficiency, a reference efficiency benchmark can be set according to the existing specifications of the mobile power supply, such as 85%. Then, the measured energy output and energy input are compared with this benchmark. If there is a significant deviation, an additional mark is made and subsequent verification is carried out, and finally, the electrical performance index data are obtained.

[0056] Based on the electrical performance index data, each target value is combined with the recording time. First, in the summary stage, time series tags are added to the power, energy consumption, and efficiency of each record. Then, the tagged data are sorted in sequence. If it is found that the time identifiers of some records are repeated or missing, corrections or supplements are made at this stage. To determine whether the corrected data is continuous, a time continuity determination index can be introduced, and the time difference between adjacent records is compared with the set allowable deviation threshold. For example, the allowable deviation threshold is defined as ±0.1 seconds. If it exceeds, it is recorded separately and excluded during subsequent analysis. After completing this step, the power and efficiency data are checked again using the previously determined power threshold and efficiency benchmark value. If there is a large range of exceedance, these records are separately placed in the abnormal archive area and these data are excluded during the final integration. Finally, an electrical performance baseline dataset is constructed based on the obtained time and the valid records corresponding to each index.

[0057] The steps for obtaining the electrical performance characteristic data are as follows: Extract the voltage data and current data from the electrical performance baseline dataset, calculate the change rate between adjacent data points, smooth the change rate sequence, and at the same time, combine the extreme points of the voltage data and current data to calibrate the fluctuation range and stable interval of the current and voltage, and generate the processed voltage-current time series data;

[0058] Based on the processed voltage-current time series data, calculate the time series characteristic values of the voltage and current. The calculation formula is:

[0059]

[0060] Among them, P represents the time series characteristic value, Qi is the voltage data at time point i, Q i-1 is the voltage data at time point i - 1, I j is the current data at time point j, I j-1 is the current data at time point j - 1, m is the total number of data in the time series, Q k The voltage data at time point k, I k is the current data at time point k;

[0061] Based on the time series eigenvalue, combined with the extreme values, mean values and change trends of current and voltage, the electrical performance characteristic data are obtained.

[0062] Specifically, voltage data and current data are extracted from the electrical performance baseline dataset, and the voltage and current values of each sampling point are compared in combination with the order of adjacent acquisition times, and their change rates are calculated. The calculation of the change rate can first count the difference between every two consecutive points and then divide by the time interval. In order to avoid high-frequency jitter, an empirical smoothing threshold can be set and multiple moving average operations can be performed. For example, every 10 change rate values in each batch are averaged once and the average result is used as the smoothed output. Each time smoothing is performed, the pre-set allowable range such as the voltage range from 0V to 24V and the current range from 0A to 5A needs to be compared. If the local change rate continuously deviates from this range, it is determined whether to exclude this section of data according to the actual recording situation, and then the positions of the maximum and minimum voltage values and current values are marked. By observing the span between the high value and the low value, the fluctuation amplitude can be divided into several levels. For example, 0V to 8V is used as the low amplitude, 8V to 16V is used as the medium amplitude, and 16V to 24V is used as the high amplitude. The current can be divided into segments such as 0A to 1.5A, 1.5A to 3.0A, and 3.0A to 5.0A in the same way. In this process, the stable interval is identified by continuously comparing the floating trends of voltage and current. The specific method is to mark the region as relatively stable if the adjacent smoothed change rates are all at relatively small values, and summarize them in different time periods by continuous recording. Finally, all the smoothed change rate data, the corresponding extreme value information and the stable interval identification are summarized to generate the processed voltage and current time series data.

[0063] The benefit of the formula is that it comprehensively considers the change difference between voltage and current and the absolute measure of the product of voltage and current, so that the time series eigenvalue not only reflects the fluctuation degree of electrical performance over time, but also includes the power level.

[0064] Q i The acquisition steps of Q are: Q iDenote the voltage data at the \(i\)-th time point, which is obtained by actual measurement at the external port of the mobile power supply and can be acquired by a voltage acquisition device with a measurement range of 0V to 30V and a resolution of 0.01V. When the measured voltage at a certain point is 12.12V and the calibration deviation is 0.02V, \(Q\) is obtained. i = 12.10V.

[0065] Q i-1 The acquisition steps of \(Q\) are as follows: \(Q\) i-1 represents the voltage data at the previous moment adjacent to \(Q\) i When the measured voltage at the 3rd second is 11.96V, the voltage measured at the 2nd second is regarded as \(Q\) 3-1 = \(Q_2\). If the voltage obtained after calibration at this moment is 11.90V, then \(Q_2\) can be substituted into the differential operation, and finally the voltage difference of \((Q_3 - Q_2)\) is formed.

[0066] I j The acquisition steps of \(I\) are as follows: \(I\) j is used to represent the current value collected at the \(j\)-th time point and is obtained by a current sensor with a rated measurement range of 0A to 5A and an accuracy of 0.001A. When the device measures a current of 3.002A at the 5th second and the deviation calibration is -0.002A, then \(I_5 = 3.000A\).

[0067] I j-1 The acquisition steps of \(I\) are as follows: \(I\) j-1 i.e., the current value at the previous moment adjacent to \(I\) j In the sampling sequence, the calibrated current at the 4th second is recorded as 2.980A, and the difference of 0.020A is obtained by subtracting it from 3.000A at the 5th second.

[0068] The acquisition steps of \(m\) are as follows: \(m\) is the number of times of differential operation or product operation in the time series and is directly related to the number of valid data points collected. If a total of \(m + 1\) pieces of real voltage and current data are obtained during the monitoring period, then when performing \((Q\) i - \(Q\) i-1 ) or \((I\) j - \(I\) j-1 ), a total of \(m\) differential calculations can be performed. If 10 valid data points are obtained in a complete acquisition cycle, then \(m = 9\).

[0069] Q k and \(I\) k The acquisition steps of \(Q\) are as follows: \(Q\) k and \(I\) k respectively represent the voltage and current data at the \(k\)-th time point.

[0070] Calculation process: Set the sampling frequency to 1Hz, collect valid data at 4 moments in total, that is, m+1=4 and m=3, and the voltage data measured and calibrated at each moment are Q1=12.10V, Q2=11.90V, Q3=12.20V, Q4=12.30V, and the corresponding current data are I1=2.100A, I2=2.050A, I3=2.200A, I4=2.250A. First, calculate the voltage difference and average:

[0071]

[0072] Here Q0 temporarily refers to the starting time of this sequence. If the voltage value obtained before the monitoring starts is recorded as 12.00V, then:

[0073]

[0074] Then calculate the current difference and average:

[0075]

[0076] If the starting current data at the 0th second is recorded as 2.000A, then:

[0077]

[0078] Then calculate the voltage-current product term:

[0079]

[0080] Substituting the above results into the formula, we get:

[0081]

[0082] The results show that within the selected sampling interval, the differential combined effect of voltage and current and the product level are ultimately converted into a characteristic value of approximately 5.06. When combined with subsequent electrical performance analysis tools, it can be used to determine the relative relationship between the fluctuation amplitude and the typical power. When the value is too large, it indicates a large fluctuation and a high product. When the value is relatively small, it indicates a relatively gentle fluctuation and a low power.

[0083] Based on the time - series eigenvalues, combined with the extreme values, mean values, and change trends of current and voltage, compare and sort the previously obtained eigenvalue sequence. First, screen out the maximum and minimum values from the collected eigenvalues and check the corresponding moments to determine when the current or voltage reaches the highest and lowest points. Then, continuously observe the change trend of the eigenvalues. If it is found that the eigenvalues at multiple moments are concentrated in a relatively narrow numerical range, such as comparing the eigenvalues in the ranges of 0 - 10, 10 - 20, and 20 - 30 with the known reference ranges respectively and marking the cumulative number of times when each recording point is in which section. By referring to the previously set stable benchmark, for example, if in the early sampling, the eigenvalues of general equipment operation fall within the range of 10 - 15 more frequently, then the range of 10 - 15 can be recorded as an empirical range. If it is detected that most values in the current sequence are concentrated in this empirical range, it indicates that the overall state remains stable. Subsequently, continue to observe the continuous trend of the eigenvalues in chronological order. Set a threshold for trend judgment obtained from the statistical calculation of past data. For example, in industrial production, the average values of the maximum and minimum eigenvalues in the previously recorded 50 maintenance data can be calculated respectively, and then combined with their floating ratios to calculate a determination line for distinguishing the rising or falling trend. For example, the average value plus or minus 1.5 can be set as a boundary. When the latest eigenvalue is greater than the average value plus 1.5, it is determined to be rising, and when it is less than the average value minus 1.5, it is determined to be falling. Finally, based on these screening and distinguishing actions, correspond the eigenvalues with the moments and current - voltage data to obtain electrical performance characteristic data.

[0084] The steps for obtaining the spectral - feature data are as follows: Pre - process the voltage and current time series in the electrical performance characteristic data, including removing noise and outliers, and applying a window function to reduce spectral leakage to obtain the processed time - series data;

[0085] Based on the processed time - series data, calculate the comprehensive feature for each frequency. The calculation formula is:

[0086]

[0087] where \(S(f)\) represents the comprehensive feature of frequency \(f\), \(X(f)\) is the result of the fast Fourier transform at frequency \(f\), is the conjugate of \(X(f)\), \(Re(X(f))\) and \(Im(X(f))\) represent the real part and the imaginary part of \(X(f)\) respectively, represents the phase, \(\alpha\), \(\beta\), \(\gamma\) are adjustment coefficients, and \(\sigma\) is a smoothing parameter;

[0088] Based on the comprehensive features, screen out the key spectral features and combine the key spectral features to generate the spectral - feature data.

[0089] Specifically, for the preprocessing of the voltage and current time series in the electrical performance characteristic data, it is necessary to first check whether there is a time misalignment for each record by comparing the sampling frequency. If the recording interval is found to exceed 0.05 seconds, it is marked as a time series anomaly and excluded subsequently. Then, the measured values are compared item by item within the voltage range of 0V to 24V and the current range of 0A to 5A. Any record that does not fall within this range is regarded as an invalid measurement and excluded. To remove noise, the statistical result of the maximum amplitude disturbance in the previously stored 100 running records is retrieved. The mean value plus a fixed increment such as 0.02V or 0.01A is used as the noise rejection threshold, and the difference between the current sample and this threshold is monitored. If the difference exceeds the allowable deviation, the point is determined to be a noise point and deleted. After that, a window function operation needs to be performed on the remaining data. The window function length is set to 64 in advance because it is found that this length can bring better smoothness for power spectrum estimation when comparing multiple measurement results before. A partial overlap can be retained between adjacent sections to maintain the transitional connection. The same window function is applied to each section and the entire time series is cropped and spliced in a sequential processing manner. Next, in the processing link, the amplitude performance of the voltage and current is checked section by section. If the amplitude difference within a certain section exceeds a fixed threshold such as 2V or 0.5A, it is recorded as a strong fluctuation section and additional investigation is carried out in the subsequent steps. If the difference does not reach this threshold, it is directly retained. After completing all the above links, the updated sequence data obtained is integrated and output, which is the processed time series data.

[0090] The advantage of the formula lies in integrating the amplitude information, logarithmic amplitude information, and the characteristic quantity of the larger value of the real part and the imaginary part, and considering the influence of the phase on the overall characteristics in the form of an exponential factor, so as to take into account the multi-faceted information of the amplitude and phase in the frequency analysis link.

[0091] The steps to obtain α are as follows: α is a non-negative number, mainly used to measure the proportion of the amplitude term in the comprehensive characteristics, and needs to be determined according to the actual operation data. First, during the 120-minute continuous operation of the device, the FFT is performed on multiple time series to obtain their amplitude value ranges, then the standard deviation of the amplitude distribution is calculated and compared with the mean value, and then referring to the amplitude statistics generated by the typical working voltage and current listed in the product manual, these data are summarized into a set of fluctuation metric values, denoted as A1, A2, …, A n , after calculating the mean value A avg of these fluctuation metrics, then a coefficient is selected to reflect the importance of the amplitude term, and thus α is determined. For example, α is set in the form of . Finally, an actual measurement is performed once and the stable value is recorded. For example, when A avg = 2.4 and max(A i ) = 3.0, α = 0.8×(2.4 / 3.0) = 0.64 can be calculated.

[0092] The steps to obtain β are as follows: The β coefficient mainly controls the proportion of the logarithmic magnitude term |ln(X(f))| in the entire calculation result. It is necessary to measure the order-of-magnitude distribution of the amplitude values in multiple operating stages. To quantify the order-of-magnitude range of the amplitude values, each FFT amplitude value can be converted into the natural logarithm form and its minimum and maximum logarithmic magnitudes can be recorded. If the span of the logarithmic magnitude distribution is large, β should be appropriately increased to emphasize the difference degree of this term. If the distribution is relatively concentrated, β can be relatively small. Let where D ln represents the range of the logarithmic magnitude within the measurement period, and D sum represents the sum of the total amplitude ranges of voltage and current. Finally, this preliminary value is compared with the statistics of multiple experiments to be corrected to a constant value. For example, after multiple tests, record D ln = 3.2, D sum = 15.6, then

[0093] The steps to obtain γ are as follows: γ is used to balance the contribution of the maximum value terms of the real part and the imaginary part to the comprehensive feature. First, retrieve the maximum amplitudes that Re(X(f)) and Im(X(f)) may reach in a batch of monitoring results, select a more prominent part of them for summarization, and calculate the average range difference E avg , and set γ to where E peak represents the maximum range difference that the real part or the imaginary part appears in all observations. To obtain a fixed and available value, it needs to be determined after multiple actual measurements. This value can be taken as a value within an amplitude range interval such as 0.15 or 0.2, and then a secondary correction is performed based on the accurate data. For example, when E avg = 5.0, E peak = 10.0, then γ = 0.3×(5.0 / 10.0) = 0.15.

[0094] The steps to obtain X(f) are as follows: X(f) represents the FFT result at frequency f. This value must be a complex number and can be obtained by performing a discrete Fourier transform on the processed time series. The specific procedure is to first determine the sampling length N, and then perform the FFT operation. The complex number output for the f-th frequency is denoted as X(f).

[0095] The steps to obtain is the complex conjugate of X(f). If X(f) = a + bi, then where a and b are retrieved from the complex part of the aforementioned FFT output respectively. When obtaining this parameter, only find the complex number corresponding to f in the same sequence, keep its real part unchanged, and take the negative of its imaginary part. In the actual program, the real part and the imaginary part of X(f) can be stored separately, and then directly change the sign of the imaginary part in the code to form a new quantity

[0096] The steps to obtain Re(X(f)) are as follows: Re(X(f)) is the real part component of X(f). It is necessary to extract the real part of the complex number from the complex numbers output by the FFT. The specific process is to first record the result of the FFT operation, then read the real part value from the corresponding index and convert it into a floating-point form that can be used for numerical analysis. In order to obtain a reliable real part range, the original signal can be averaged and filtered before the FFT, and the real part value is recorded after eliminating the DC offset, and finally an input term that can participate in the formula calculation together with other parameters is formed.

[0097] The steps to obtain Im(X(f)) are as follows: Im(X(f)) is the imaginary part component of X(f). Similar to extracting the real part, it is directly read from the imaginary part field in the FFT result. If there is an obvious phase shift during the device measurement period, relatively large positive or negative values may be observed in the imaginary part. Therefore, this kind of abnormality will be verified during data post-processing. If there is no abnormality, Im(X(f)) will be recorded according to the normal process and brought into various operations.

[0098] The steps to obtain φ(f) are as follows: It is the phase value at the current frequency. It is necessary to put the previously obtained Re(X(f)) and Im(X(f)) into the arctangent function for calculation in a certain order. For the case where Re(X(f)) is too small or close to 0, a logical judgment will be added at the code level to avoid the abnormality when the denominator is close to 0. After obtaining this phase, it can be stored in the phase list, and finally the phase values corresponding to each f are respectively brought into the subsequent calculation process.

[0099] The steps to obtain σ are as follows: σ represents the parameter used to smooth the phase effect. It can be quantitatively evaluated according to the severity of the phase change at different frequencies. First, record the variation range of the phase with frequency within a relatively long period and calculate its mean square deviation, and denote this mean square deviation as V phase , and then let where c is a constant used to adjust the attenuation rate, which can be selected by referring to the results of multiple spectrum analyses. When is relatively large, it indicates that the phase dispersion is obvious, and c can be appropriately increased to smooth the phase effect. If the measured V phase = 0.36 and c = 1.2, then

[0100] Calculation process: Suppose at a certain frequency f = 50 Hz, the FFT complex value X(50) obtained is 2.0 + 1.2i, then From this, we can get Re(X(50)) = 2.0, Im(X(50)) = 1.2. After determining α = 0.64, β = 0.10, γ = 0.15, σ = 0.72 through the above methods, and then calculate the phase radian

[0101] First, calculate the amplitude

[0102] Then, obtain the logarithmic amplitude |ln(X(f))| = |ln(2.0 + 1.2i)|. For simplicity, substitute |2.0 + 1.2i| ≈ 2.33 into the logarithmic calculation. ln(2.33) ≈ 0.84, and the absolute value remains 0.84.

[0103] Next, calculate max(Re(X(f)), Im(X(f))) = max(2.0, 1.2) = 2.0.

[0104] Substitute these results into the parentheses term of the formula:

[0105] α·2.33 + β·0.84 + γ·2.0 = 0.64×2.33 + 0.10×0.84 + 0.15×2.0 = 1.4912 + 0.084 + 0.3 = 1.8752;

[0106] Phase correction factor

[0107] Thus, S(50) = 1.8752×0.4724 ≈ 0.886; this result indicates that at the 50 Hz position, this comprehensive eigenvalue is approximately 0.886. The larger the value, the more significant the amplitude, logarithmic amplitude, real part, and imaginary part features are at that frequency, and the relatively smaller the phase influence is. By repeating the above operations at other frequencies f, the distribution of S(f) across the entire frequency band can be obtained.

[0108] After calculating the comprehensive eigenvalue at different frequencies one by one based on the comprehensive features, it is necessary to check whether there are some eigenvalues corresponding to frequencies that are significantly higher than those in their adjacent frequency segments within the full frequency band range, and it is also necessary to verify whether the actual voltage and current amplitudes or phases corresponding to these high value points are above the abnormal determination threshold defined previously. The threshold setting can select an overlimit standard by combining the range of eigenvalues recorded in 30 consecutive tests. For example, first calculate the mean and standard deviation of all eigenvalues, and then set the mean plus twice the standard deviation as the critical standard. If the eigenvalue at a certain frequency exceeds this standard, it is regarded as a suspicious frequency point. Then, confirm whether these suspicious points are concentrated frequency by frequency. If the number of high values concentrated in a frequency segment exceeds a certain counting threshold, then this frequency segment is determined as a key area. At the same time, it is necessary to further screen by combining the fluctuation of the eigenvalues after phase correction. If most of the high values are concentrated on one side with a smaller or larger phase, they can be classified separately and marked as a group with an obvious phase tendency. Finally, separate this type of eigenvalue from the values in other normal intervals, and retain several of the most critical frequency points according to the set screening strategy. Record the features corresponding to these high value points together to form key spectral feature data, and the spectral feature data can be obtained after summarization.

[0109] The steps for obtaining multi-scale spectral data are as follows: Based on the spectral feature data, calculate the wavelet coefficients at each scale. The calculation formula is:

[0110]

[0111] where W(a, b) represents the wavelet coefficient at scale a and position b, S n is the nth point of the spectral data, ψ is the wavelet mother function, N is the total number of data points, n - b represents the time delay, and gw is the Gaussian window parameter used to adjust the window width;

[0112] Based on the wavelet coefficients, reconstruct the signal to separate and identify the details and approximation parts at each time scale, and obtain multi-scale spectral data.

[0113] Specifically, the advantage of the formula is that on the basis of the wavelet mother function ψ, a Gaussian window is additionally introduced to regulate the local area. Thus, it can not only reflect the multi-scale analysis characteristics of wavelet transform for time and frequency, but also take into account the local smoothness, so as to better balance the details and overall trends in the multi-scale analysis of spectral data.

[0114] The steps to obtain a are as follows: a is a scale parameter representing the scaling ratio. The larger the value, the more obvious the stretching degree, which means a wider coverage range of the wavelet function. In practice, the length of the previously obtained spectral data can be estimated first. For example, it is determined that a certain number of discrete points are collected in the range of 0 Hz to 500 Hz. Then, a batch of representative scales are selected, such as a ∈ {1, 2, 4, 8, 16}, etc. For example, when consulting industrial tests, it is found that when the data is mainly concentrated in the range of 30 Hz to 120 Hz, using a = 2 and a = 4 can better capture local details. Then, the peak-valley range of the overall waveform recorded in the initial test table is converted into a certain multiple, and the corresponding a list can be determined and substituted into the calculation.

[0115] The steps to obtain b are as follows: b represents the time (or position) offset selected on the signal sequence. It is necessary to traverse the entire sequence point by point from beginning to end to locate the local structure at different times. The value range of b can be from 0 to N - 1, where N is the total number of data points. Each b corresponds to a local alignment point, which will cause the item (n - b) in to produce a relative displacement. To make the setting of the b value reasonable, the previously obtained time or frequency index can be combined, and each index point can be scanned. For example, if the data length N = 128, b can take an integer value from 0 to 127 to cover the entire range, so as to capture the wavelet matching degree at different positions. When giving an example, b = 0 can be substituted into the calculation to obtain W(a, 0) first, and then b = 1 is used to find W(a, 1), and so on until b = 127, finally forming a comprehensive description of the entire signal at a certain scale a.

[0116] S n The steps to obtain are as follows: S n is the value of the spectral data at the nth discrete point, which is extracted from the previously performed frequency domain analysis or FFT results. Each S n corresponds to a specific amplitude or some characteristic measure. To obtain these data, the original voltage and current information needs to be subjected to FFT first to obtain the characteristic values of several frequency points, and then sorted into {S0, S1, …, S N-1}.

[0117] The steps to obtain are as follows: ψ is the wavelet mother function, represents a dimensionless variable obtained by combining the sequence index with the scale and position. When obtaining it, it is necessary to first determine which wavelet mother function to use and substitute into the corresponding analytical expression. If the Morlet wavelet is selected, ψ(x) = exp(-x 2 / 2)cos(5x) and other forms can be set. The Mexican hat wavelet or other types can also be used, but each mother function has its own analytical expression. During the measurement, first, a and b are used to give The specific value is then sent into ψ(·) to complete the operation. To avoid unnecessary truncation errors, the ψ values can be pre-computed and stored in a table, and then read by index. For example, when the Morlet wavelet is selected, is discretized step by step in the interval from -10 to 10, and obtained by table lookup or interpolation

[0118] The steps to obtain N are as follows: N is the total number of data points, which refers to the length of the spectral data participating in the calculation in wavelet transform. The specific value can be directly obtained by statistically analyzing the previously generated spectral feature sequence. If the sampling length of 1024 was selected in the initial FFT stage, and there may be 512 valid points left after intercepting the useful frequency points, then N can be recorded as 512.

[0119] (n - b) is obtained as follows: (n - b) is the term in the formula used to represent the time delay (or displacement) relationship, and is obtained by subtracting the current counting index n from the position index b. The acquisition process: in the program loop, n traverses from 0 to N - 1, b is fixed at the current position for scanning and calculation. When b changes, (n - b) also changes to ensure that each calculated corresponds to different relative translations. If (n - b) is negative, it can still be substituted into the wavelet mother function for calculation. Therefore, a sufficient domain needs to be set to ensure that the wavelet mother function can still be correctly evaluated in the negative value region. For a specific example, when n = 2 and b = 5, there is (2 - 5) = -3. When interpolating or looking up in the table, it is necessary to ensure that the ψ value at index -3 is legally defined.

[0120] The steps to obtain gw are as follows: gw is the Gaussian window parameter used to adjust the window width. When the value is larger, the window function decays more slowly and covers a wider range. The smaller it is, the faster it decays and the stronger its limitation in the time domain. The "required smoothness" can be calculated in combination with the statistics of the actual distribution of the signal and corresponding to the value of gw. For example, first analyze the high-frequency noise of hundreds of industrial measurement samples, quantify the jitter degree in its high-frequency band, then compare this value with the variance of the center frequency to obtain an average attenuation range, and finally determine gw. The value can also be fixed as some representative values within the interval such as 1.0 or 2.0, and then through repeated experiments in the actual scenario, find which gw can bring more stable transformation results. In the example, if it is summarized after multiple tests that gw = 1.5 meets the requirements.

[0121] Calculation process:

[0122] In a specific example, take N = 5, that is, the data of five frequency points {S0, S1, S2, S3, S4} are used for calculation. Let S0 = 0.5, S1 = 1.2, S2 = 0.9, S3 = 1.5, and S4 = 1.1. Let the scale a = 2 and the position b = 2, and select the Morlet wavelet mother function ψ(x) = exp(-x 2 / 2)cos(5x), and the Gaussian window parameter gw = 1.5.

[0123] According to the formula:

[0124]

[0125] First, perform step-by-step calculations for each item:

[0126] When n = 0, (n - 2) / 2 = -1, (n - 2) 2 = 4, so Therefore, the contribution of this item is approximately S0×0.172×0.411 = 0.5×0.172×0.411≈0.0354.

[0127] When n = 1, 0.801, and the contribution is S1×(-0.624)×0.801 = 1.2×(-0.624)×0.801≈ -0.599.

[0128] When n = 2, The contribution is S2×1×1 = 0.9.

[0129] When n = 3, The contribution is S3×(-0.624)×0.801 = 1.5×(-0.624)×0.801≈ -0.749.

[0130] When n = 4, The contribution is S4×0.172×0.411 = 1.1×0.172×0.411≈0.078.

[0131] Add up these contribution values to get the total: Then multiply by to get W(2, 2)≈ -0.236.

[0132] The result shows that at the wavelet scale a = 2 and position b = 2, the weighted response of the signal to the Morlet mother function and the Gaussian window is approximately -0.236. If a distribution map is formed after scanning at various locations, it can be further determined at which positions the local structure of the signal is the most significant or the weakest. When the value is greater than a certain threshold, it can be regarded that there are obvious characteristic patterns at that location, and when it is less than another threshold, it can be regarded as stable or without significant structure.

[0133] Based on the wavelet coefficients, it is necessary to stack or distribute the corresponding coefficient values point by point at each scale to the corresponding positions and reconstruct the obtained wavelet coefficient matrices layer by layer. For this purpose, first arrange the wavelet coefficients in the order from left to right within the same scale, and judge whether the coefficient change at adjacent positions exceeds the specified range. Here, a range detection method can be set, such as obtaining the average value and standard deviation of the coefficients at the same scale in the recent ten measurements in advance, and then regarding the average value plus twice the standard deviation as the upper limit. If the current coefficient greatly exceeds the upper limit, it is positioned as an abnormal coefficient and recorded in a separate list, and then continue to make the same judgment for other positions. After completing the screening at this scale, the same method can be used to process the next scale. Combine all the coefficients into a unified structure in the multi-scale dimension. Each scale has its own corresponding coefficient distribution graph. By arranging the detail layer and the approximation layer separately during the reconstruction process, it is possible to distinguish the special fluctuation regions that may exist in different time regions and frequency regions. Subsequently, compare the coefficient distribution of each scale on the same time axis. If there are frequency bands with excessive coefficients at multiple scales, it indicates that there are more obvious differences at that section of the position. In order to avoid missing judgments, it is also necessary to check the coefficient positions that are close to the edge but have not exceeded the limit, and accumulate them point by point to identify the coefficient group with a certain amplitude locally but not reaching the abnormal determination threshold. If this group appears continuously in time, it can be summarized as a weaker detail component. In this way, the time-frequency structure diagrams corresponding to different resolutions are finally formed, and then read the detail information of several key paragraphs from these data and merge them into a complete spectrum reconstruction path, so as to merge the performances at different time scales in the subsequent analysis and obtain multi-scale spectrum data.

[0134] The steps to obtain the preliminary fault diagnosis result are as follows: obtain the multi-scale spectrum data and the electrical performance baseline data set, perform standardization processing on the multi-scale spectrum data and the electrical performance baseline data set to make the data formats consistent, and obtain the standardized multi-scale spectrum data and the electrical performance baseline data set;

[0135] Based on the standardized multi-scale spectrum data and the electrical performance baseline data set, compare the spectrum characteristics at each scale, identify the deviations in frequency response, energy distribution, and time series stability, and obtain the deviation analysis result;

[0136] Based on the deviation analysis results, evaluate and classify each identified anomaly, sort and label the anomalies according to the severity of the impact and the frequency of occurrence, define the estimated failure type for each anomaly, and obtain the preliminary fault diagnosis result.

[0137] Specifically, obtain the multi-scale spectrum data and the electrical performance baseline dataset. First, compare the recording methods of the two and check whether there are differences in the acquisition frequency and sampling length. Correlate the index frequency points in the multi-scale spectrum data with the time or voltage / current indicators in the electrical performance baseline dataset. If it is found that it is difficult to correspond the index frequency points and the time axis one by one, a unified index system needs to be introduced. Use a preset reference sequence to uniformly label multiple records, check whether there are repeated or jump moments in the records and fill them in or remove them. Then, set an alignment accuracy threshold according to the previous operating conditions or application scenarios of the device. For example, the accuracy threshold can be determined by multiplying the average interval obtained by dividing the overall recording duration by the number of sampling points by a ratio. For example, when the average interval is 1 second and the ratio is 0.05, the accuracy threshold is 0.05 seconds. Then, traverse all the recording points in the two types of data and check whether the alignment deviation is less than the accuracy threshold. When the deviation exceeds, mark the record as an abnormal index and exclude it. Subsequently, perform numerical normalization operations on the respective parameter columns in the multi-scale spectrum data and the electrical performance baseline dataset. Usually, first calculate the minimum and maximum values of each column of data and transform them to the range of 0 to 1 according to the following way. If there are negative values in the numerical values of some columns, an offset needs to be performed before normalization to correct the minimum value to 0 so that the normalized result will not have underflow or invalid results. After normalization, check the field order of the two sets of data to make the frequency items in the multi-scale spectrum data correspond one by one with the time period or the corresponding measurement number in the electrical performance baseline dataset. If it is detected that there is a missing matching relationship in some sections, fill in null values and mark them as missing. After the entire process is completed, a set of standardized multi-scale spectrum data and electrical performance baseline dataset can be obtained.

[0138] Based on the standardized multi-scale spectrum data and the electrical performance baseline dataset, traverse the spectral feature values within each scale range and compare these feature values with the voltage and current records during the corresponding period in the baseline. Considering the changes in frequency response and the differences in energy distribution, first divide the multi-scale spectrum data into several segments according to frequency bands, calculate the statistical values of the feature values within each segment, and check whether these statistical values are within the baseline intervals determined previously. For example, when the baseline data shows that the main energy during the operation of most devices is concentrated in the range of 20 Hz to 80 Hz, the average spectral amplitude within this range can be compared with the baseline average amplitude to see if the deviation is greater than twice the mean square deviation. If it is greater, record this segment as a suspicious band, and continue to perform synchronous comparison on the electrical performance information during the same period to check whether the voltage exceeds the threshold value estimated in advance based on historical data. For example, define the voltage threshold within the interval obtained by extending 5% from both ends according to the device rating, or extend the current threshold in the same way. If it is detected that the voltage or current exceeds these thresholds continuously for multiple times, it is recorded as a deviation warning. After summarizing the results, it can be found which scales and frequency bands have significant deviations. Then, check the stability of the time series. When comparing the fluctuations of feature values in adjacent periods, set a fluctuation limit, such as adding or subtracting several times the standard deviation based on the average of the existing baseline to determine whether the fluctuation is significant. Through this series of comparison operations, the deviation results corresponding to each scale can be obtained, forming the deviation analysis result.

[0139] Based on the deviation analysis result, list each detected abnormal event separately and number it. First, associate the occurrence time of all abnormalities with the multi-scale spectral feature values at the same time, and identify whether there are abnormal high points or continuous deviations in the frequency distribution or energy amplitude when these abnormalities occur. Then, compare it with the operating parameters calibrated in the electrical performance baseline dataset to check whether the occurrence of this abnormality corresponds to the situation where the voltage is too high, the current fluctuates excessively, or other indicators exceed the preset limits. These preset limit values are usually calculated by statistically averaging the device during stable operation and combining the maximum fluctuation amplitude. For example, if the highest voltage recorded in the past 100 measurements is 22 V, then a severity judgment upper limit of 23 V can be set by increasing this value by 1 V. Take into account the occurrence frequency and the influence range, classify the abnormalities as those with greater influence or frequent occurrence and place them in the priority marking position, forming a list of abnormalities sorted according to the severity of influence. For each abnormality, an estimated fault label can be given, such as identifying it as line instability, power supply aging, or short-term load overlimit, etc. Through this list, each type of abnormality can be classified, and finally, all abnormalities and their classification information are integrated into the corresponding fault types to obtain the preliminary fault diagnosis result.

[0140] The steps for obtaining the fault identification feature data are as follows: Based on the preliminary fault diagnosis result, calculate the geometric deformation degree between data points to obtain the curvature value. The calculation formula is:

[0141]

[0142] Among them, K(s) represents the curvature value, and Δx and Δy are the horizontal and vertical distance differences of adjacent data points in the time series, respectively;

[0143] Based on the curvature value, analyze the curvature changes between data points, and obtain the fault identification feature data by integrating the curvature change results.

[0144] Specifically, the advantage of the formula is that it comprehensively considers the horizontal and vertical distance differences between adjacent data points and non-linearly characterizes the curvature, so that the geometric deformation amplitude between data points can be sensitively detected. Even when there are differences in orders of magnitude between Δx and Δy, the formula can still present an accurate estimate of the curvature.

[0145] The acquisition step of Δx is as follows: Δx represents the horizontal difference of adjacent data points in the time series, and the value is obtained by operating on the time index or sequence position index of adjacent samplings. It is necessary to first clarify the label or timestamp of each data point in the sequence, and then calculate the difference between adjacent points. If the device parameters are recorded at a sampling interval of 1 second in actual monitoring, the time indices of the i-th point and the (i - 1)-th point can be denoted as t i and t i-1 , and then let Δx = t i -t i-1 . If the possible fluctuations of sampling need to be considered, it is necessary to calibrate the sampling time first, and then extract Δx after confirming no duplicates and jump deletions. For example, in a monitoring, the current and voltage of the device operation are recorded once per second, and the device runs continuously for 300 seconds. Then the time difference Δx between adjacent indices is mostly 1. However, if the sampling delay at a certain moment is more than 0.2 seconds, an anomaly will be marked and retested, and finally a corrected sequence will be formed to make Δx consistent with the actual average sampling period.

[0146] The acquisition step of Δy is as follows: Δy represents the difference of adjacent data points in the vertical direction, which can be regarded as the parameter value difference between adjacent points on a time series curve. For example, in the scenario of mobile power supply fault diagnosis, each data point may contain a real-time monitoring value. For example, according to the preliminary fault diagnosis results generated previously, the fault indicators (such as power fluctuation amount or spectrum intensity) at adjacent moments are recorded as Y i and Y i-1 , and then let Δy = Y i -Y i-1 . If these Y i are to be obtained, the power or other indicators can be deduced after collecting voltage and current data in the early stage, and then retained for fault diagnosis after several steps of analysis. When Y i and Y i-1After obtaining, the direct difference of Δy can be used to form the final input value. To ensure numerical accuracy, extreme outliers in the original data can be screened first, and records that significantly exceed the rated range of existing equipment can be verified again. This way, the authenticity can be maintained as much as possible when calculating Δy. If the equipment runs continuously for 30 minutes and 1800 power values are collected, then any two adjacent records can be used for the operation of Δy.

[0147] Calculation process:

[0148] In a specific example, let the time series index difference Δx between two adjacent points be 1, and the fault index value difference Δy be 0.3. Substitute these two quantities into the formula:

[0149]

[0150] First calculate (1.09) -3 / 2 =(1.09) -1.5 ≈(1.09) -1 ×(1.09) -0.5 ≈0.9174×0.9574≈0.878, so the value of the whole term is 0.3×0.878 = 0.2634. When K(s)=0.2634, this curvature value can be understood as the magnitude of the geometric bending degree of this time series between adjacent points.

[0151] The result shows that when the vertical difference and horizontal difference between adjacent points are at such an order of magnitude, the bending condition of the corresponding curve is about 0.2634. If it is found in subsequent monitoring that the curvature continuously increases to more than 1.0, it means that there are more significant steep changes between adjacent data points on this section of the curve. This value can be further compared with the curvature threshold established in advance. Assuming that, for example, above 0.8 belongs to the high curvature area, it can be determined that there is an obvious jump in the fault index here, thus assisting in the fine positioning of fault symptoms.

[0152] Based on the curvature values, it is necessary to calculate the curvature for all adjacent points in the entire fault diagnosis curve and form a curvature sequence in chronological order. Then, check whether there are paragraphs with continuous high curvature values in the entire sequence. First, determine an effective range by referring to previous statistical data. For example, set the high curvature value range between 0.5 and 2.0. When the curvature value enters this interval, record the point position at the corresponding moment. If its duration exceeds 5 sampling periods or longer, mark it as a major fluctuation. Then, combine the previously calculated frequency spectrum analysis results and voltage and current characteristics to confirm whether the abnormality reappears. If there are multiple high curvature value segments distributed in different time windows, classify them segment by segment to distinguish the blocks with more concentrated fluctuations and occasional convex points. By comparing these high value blocks with the curvature distribution of the device's baseline state, it is possible to check whether the device has shown similar curvature characteristics during normal operation. If the previously statistically obtained normal curvature distribution is usually concentrated between 0.05 and 0.20, then once it exceeds 0.5, it indicates that the curve shape has a significant difference from the baseline. Finally, after summarizing the curvature changes at each moment, associate them with the fault events, and obtain the fault identification feature data by inducing possible fault modes.

[0153] The steps to obtain the fault monitoring and early warning results are as follows: Based on the fault identification feature data and multi-scale frequency spectrum data, configure and train a support vector machine model, select the radial basis function and adjust the model penalty coefficient and kernel parameters to obtain a trained fault prediction model;

[0154] Based on the trained fault prediction model, deploy the trained fault prediction model to the mobile power supply monitoring, analyze the collected power operation data in real time, evaluate the fault risk, and obtain the fault monitoring and early warning results.

[0155] Specifically, based on the fault identification feature data and multi-scale spectrum data, first read each corresponding record in both, and compare parameters such as the voltage, current, and the previously obtained fault type label contained in the record. Organize all the data into a sample set that can be used for support vector machine training. To determine the values of the penalty coefficient and kernel parameter, candidate values will be selected within a limited grid range and tested one by one. For example, increase the penalty coefficient from 0.1 to 10 at multiple interval points, and increase the kernel parameter from 0.01 to 1 at multiple interval points. Subsequently, perform cross-validation for each combination of the penalty coefficient and kernel parameter, calculate the accuracy and misjudgment rate of each category in the training set, and compare these statistical indicators with the previously set qualified range. For example, an accuracy lower than 80% can be regarded as not meeting the standard, and the balance point of the accuracy and misjudgment rate can be used as the optimal value according to the requirements of known operation and maintenance cases. Then, when the number of samples is sufficient, divide the data set in the way of 80% for training and 20% for validation. Use the radial basis function to establish a support vector machine classifier and input the values of each dimension in the fault identification feature data as input features. If the accuracy in the validation set exceeds the aforementioned qualified range and the misjudgment rate remains within 5% under different fault types, then regard the current parameter combination as the optimal one, record these parameter values and conduct a complete training run. After completion, check whether the compliance of the classification result of the model with all training samples and the actual fault type label still meets the requirements. If it meets, stop the search and solidify this model as the trained fault prediction model.

[0156] Based on the trained fault prediction model, it is necessary to deploy the previously obtained model parameters and the support vector machine classifier in the mobile power supply monitoring scenario and access the real-time sampled voltage and current data. Therefore, when the device is running continuously, collect operation information such as voltage and current at a frequency of once per second or higher, and use the existing preprocessing rules to calibrate the instantaneous sampled values. If it exceeds the acquisition calibration threshold range, record this sampling as abnormal and continue to compare it subsequently. Then, convert the calibrated data into the same form of fault identification features as during training and send it into the support vector machine model. The model will output a decision score according to the classification hyperplane and kernel function determined through training before. If the fault possibility corresponding to the score exceeds the previously set risk threshold, for example, define the threshold as a decision value critical line obtained by analyzing previous training samples, or different fault types correspond to their respective score thresholds. If the real-time score reaches or exceeds this threshold, record the risk level at the current moment on the monitoring interface and make a fault mark. When high scores are generated multiple times within the same period, it is regarded as a higher-level fault risk. Finally, organize the decision situations of each period and form a fault monitoring and early warning result.

[0157] The above are only the preferred embodiments of the present invention and do not limit the present invention in other forms. Any person skilled in the relevant art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A mobile power supply fault detection method based on intelligent diagnosis, characterized in that: The following steps are involved: Collect the voltage and current data of the mobile power supply in normal operation, and generate an electrical performance baseline data set by recording the timestamp of the data and the corresponding electrical performance indicators; Based on the electrical performance baseline data set, extracting time series characteristics of voltage and current to obtain electrical performance characteristic data; Based on the electrical performance characteristic data, the intensity and phase of each frequency component are analyzed by fast Fourier transform, key spectrum characteristics are extracted, and spectrum characteristic data are generated; the spectrum characteristic data is multi-scaled decomposed, and the signal is decomposed into details and approximate data on multiple time scales by wavelet transform method to obtain multi-scale spectrum data; By comparing the multi-scale spectrum data with the corresponding multi-scale spectrum features in the electrical performance baseline data set, fluctuations and non-periodic anomalies that deviate from the normal state are identified, and a preliminary fault diagnosis result is generated; Performing geometric curvature analysis on the preliminary fault diagnosis results, calculating the degree of geometric deformation between data points, and obtaining fault identification feature data; Based on the fault identification feature data, a fault prediction model is trained by a support vector machine, and the fault prediction model is applied to continuously monitor the mobile power supply to obtain a fault monitoring and early warning result.

2. The mobile power supply fault detection method based on intelligent diagnosis according to claim 1 is characterized in that: The steps of acquiring the electric performance baseline data set are: real-time monitoring and recording the current and voltage data of the mobile power supply, adding a timestamp to each piece of data, and forming a preliminary current and voltage time series record; Based on the preliminary current and voltage time series records, calculating the instantaneous power of current and voltage as well as energy consumption and efficiency to obtain electrical performance index data; Based on the electrical performance indicator data, each indicator is integrated with a timestamp to construct an electrical performance baseline data set.

3. The mobile power supply fault detection method based on intelligent diagnosis according to claim 1 is characterized in that: The step of acquiring the electrical performance characteristic data is as follows: extracting voltage data and current data from the electrical performance baseline data set, calculating the change rate between adjacent data points, smoothing the change rate sequence, and combining the extreme points of the voltage data and the current data to calibrate the fluctuation amplitude and stable range of the current and voltage, and generate processed voltage and current time series data; Based on the processed voltage and current time series data, the time series characteristic values ​​of voltage and current are calculated, and the calculation formula is: Among them, P represents the time series characteristic value, Q i is the voltage data at time point i, Q i-1 is the voltage data at time point i-1, I j is the current data at time point j, I j-1 is the current data at time point j-1, m is the total number of data in the time series, Q k Voltage data at time point k, I k is the current data at time point k; Based on the time series characteristic values, combined with the extreme values, mean values ​​and change trends of current and voltage, electrical performance characteristic data are obtained.

4. The mobile power supply fault detection method based on intelligent diagnosis according to claim 1 is characterized in that: The step of acquiring the spectrum characteristic data is: preprocessing the voltage and current time series in the electrical performance characteristic data, including removing noise and outliers, applying window function processing to reduce spectrum leakage, and obtaining processed time series data; Based on the processed time series data, the comprehensive characteristics of each frequency are calculated using the following formula: Among them, S(f) represents the comprehensive characteristics of frequency f, X(f) is the fast Fourier transform result at frequency f, is the conjugate of X(f), Re(X(f)) and Im(X(f)) represent the real and imaginary parts of X(f), respectively. represents the phase, α, β, γ are adjustment coefficients, and σ is a smoothing parameter; Based on the comprehensive features, key spectrum features are screened, and the key spectrum features are combined to generate spectrum feature data.

5. The mobile power supply fault detection method based on intelligent diagnosis according to claim 1, characterized in that: The step of acquiring the multi-scale spectrum data is: based on the spectrum feature data, calculating the wavelet coefficients of each scale, the calculation formula is: Among them, W(a,b) represents the wavelet coefficient at scale a and position b, S n is the nth point of the spectrum data, ψ is the wavelet mother function, N is the total number of data points, nb represents the time delay, and gw is the Gaussian window parameter used to adjust the window width; Based on the wavelet coefficients, the signal is reconstructed to separate and identify details and approximate parts at each time scale, and multi-scale spectrum data is obtained.

6. The mobile power supply fault detection method based on intelligent diagnosis according to claim 1, characterized in that: The step of obtaining the preliminary fault diagnosis result is: obtaining the multi-scale spectrum data and the electrical performance baseline data set, standardizing the multi-scale spectrum data and the electrical performance baseline data set to make the data format consistent, and obtaining standardized multi-scale spectrum data and electrical performance baseline data set; Based on the standardized multi-scale spectrum data and the electrical performance baseline data set, the spectrum characteristics at each scale are compared, the deviations in frequency response, energy distribution and time series stability are identified, and the deviation analysis results are obtained; Based on the deviation analysis results, each identified anomaly is evaluated and classified, the anomalies are sorted and marked according to the severity of the impact and the frequency of occurrence, an estimated fault type is defined for each anomaly, and a preliminary fault diagnosis result is obtained.

7. The mobile power supply fault detection method based on intelligent diagnosis according to claim 1, characterized in that: The step of acquiring the fault identification feature data is: based on the preliminary fault diagnosis result, the degree of geometric deformation between data points is calculated to obtain the curvature value, and the calculation formula is: Among them, K(s) represents the curvature value, Δx and Δy are the horizontal and vertical distance differences between adjacent data points in the time series, respectively; Based on the curvature value, the curvature change between the data points is analyzed, and the fault identification feature data is obtained by integrating the curvature change results.

8. The mobile power supply fault detection method based on intelligent diagnosis according to claim 1, characterized in that: The steps of obtaining the fault monitoring and early warning results are: based on the fault identification feature data and the multi-scale spectrum data, configuring and training a support vector machine model, selecting a radial basis function and adjusting the model penalty coefficient and kernel parameter to obtain a trained fault prediction model; Based on the trained fault prediction model, the trained fault prediction model is deployed to the mobile power supply monitoring, the collected power supply operation data is analyzed in real time, the fault risk is evaluated, and the fault monitoring and early warning results are obtained.

9. The mobile power supply fault detection system according to any one of claims 1 to 8, characterized in that: include: The data acquisition module records the voltage and current of the mobile power supply during operation, adds a timestamp, and generates an electrical performance baseline data set; The feature extraction module extracts the peak value, mean value, and standard deviation of voltage and current from the electrical performance baseline data set to generate electrical performance feature data; The spectrum analysis module applies fast Fourier transform to the electrical performance characteristic data, analyzes the intensity and phase of each frequency component, generates spectrum characteristic data, and performs multi-scale decomposition of the spectrum characteristic data through wavelet transform to obtain multi-scale spectrum data; The fault diagnosis module compares the multi-scale spectrum data with the electrical performance baseline data set, identifies fluctuations and non-periodic anomalies, generates preliminary fault diagnosis results, performs geometric curvature analysis on the preliminary fault diagnosis results, calculates the degree of deformation between data points, and obtains fault feature data; The fault prediction module trains the support vector machine model based on the fault feature data, conducts continuous monitoring, predicts and issues fault monitoring warnings, and generates fault warning results.

Images