A method and system for detecting faults in a vehicle-mounted starting power device

Abstract

The application relates to the technical field of fault detection, in particular to a vehicle-mounted starting power device fault detection method and system; the method comprises the following steps: collecting an output current signal of a vehicle-mounted starting power device to obtain a discrete current signal; constructing a differential sign statistical differential inversion number, calculating a waveform response factor; then calculating a bottom noise obvious factor, calculating the proportion of odd harmonic energy in total harmonic energy, and combining the waveform response factor and the bottom noise obvious factor to obtain a sawtooth response factor; based on the sawtooth response factor, the final sampling weight of the output current signal is obtained; based on the final sampling weight, an isolated tree of an isolated forest algorithm is constructed, a new output current signal is detected by using the isolated tree, the vehicle-mounted starting power device is subjected to fault detection, and the efficiency of the fault detection of the vehicle-mounted starting power device is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of fault detection, in particular to a vehicle-mounted starting power device fault detection method and system. BACKGROUND

[0002] The vehicle-mounted starting power is applied to automobile starting, power supply and electric energy management, and generally includes key devices such as a battery, a DC-DC converter, a MOSFET and an inductor. The core function thereof is to realize two key tasks of large-current cold starting of an engine (providing a peak current at an instant of engine ignition to overcome mechanical static friction) and continuous power supply of vehicle-mounted equipment (stabilizing battery voltage through a DC-DC converter to supply power to ECU, vehicle lamps, audio and video systems and other loads) through battery energy storage and efficient electric energy conversion. However, the working process thereof is accompanied by severe electrical characteristics, and in actual application, these power devices are prone to abnormalities such as poor contact, component aging and EMI interference under high-frequency and large-current working conditions. Traditional fault detection methods mostly rely on threshold monitoring of voltage and current, lack comprehensive analysis of abnormal waveforms and spectral characteristics, and often cannot timely discover early potential faults, and have high false positive rates and poor reliability.

[0003] The isolation forest algorithm is an efficient unsupervised anomaly detection method based on a random tree structure (Isolation Tree), and the core idea thereof is to isolate abnormal points through a random partition strategy. Abnormal data points are usually few in number and have feature values significantly different from normal points, and are more likely to be "isolated" by randomly selected features and partition values. It recursively partitions data points by randomly selecting features and partition values to construct multiple "isolated trees". In each isolated tree, abnormal points are isolated by fewer partition steps to reach the leaf nodes of the tree because they are "different from others". The traditional isolation forest algorithm randomly extracts sub-samples from the training set when constructing isolated trees, which may cause some key features sensitive to faults to be used at low frequencies, so that abnormal points are buried in normal samples, thereby affecting the detection capability. SUMMARY

[0004] In order to solve the problem of high false positive rate and poor reliability in the fault detection process of the vehicle-mounted starting power device, the application provides a vehicle-mounted starting power device fault detection method and system.

[0005] In the first aspect, the application provides a vehicle-mounted starting power device fault detection method, which adopts the following technical scheme:

[0006] An output current signal of the vehicle-mounted starting power device is collected, a sliding window is set on the output current signal, and discrete current signals are obtained; for each discrete current signal, a differential symbol statistical difference inversion number is constructed, and a waveform response factor representing the degree of sawtooth jitter of the waveform is calculated in combination with a linear change trend;

[0007] The average power spectrum density of each discrete current signal in a preset continuous frequency band is taken as a bottom noise obvious factor, the proportion of the odd harmonic energy in the total harmonic energy is calculated, and the abnormal degree of the output current signal is represented by a sawtooth response factor combined with the waveform response factor and the bottom noise obvious factor; and the final sampling weight of the output current signal is obtained based on the sawtooth response factor.

[0008] The isolated tree of the isolation forest algorithm is constructed based on the final sampling weight, and the new output current signal is detected by using the isolated tree, and when the obtained abnormal score is greater than a preset abnormal threshold, it is determined that the vehicle-mounted starting power supply device has a fault.

[0009] The waveform response factor in the time domain is constructed to quantify the sawtooth jitter, the bottom noise obvious factor in the frequency domain is constructed to detect the non-harmonic noise lifting, the high fluctuation segment is given a higher sampling weight by the dynamic weight of the sawtooth response factor, the selection probability of the abnormal feature in the isolated tree is improved, and the sensitive detection of the fault of the vehicle-mounted starting power supply device is realized.

[0010] Further, the method for obtaining the discrete current signal is that the output current signal is discretely sampled by using a time discretization method to obtain the discrete current signal at each sampling time.

[0011] Further, the method for obtaining the difference symbol is that ; in the formula ， the difference symbol of the discrete current signal at each sampling time, represents the first-order difference between the discrete current signal at each sampling time and the discrete current signal at the previous adjacent sampling time.

[0012] The beneficial effect is that the current change trend is discretized into three states by defining a segmented function, the limitation that the traditional continuous difference is sensitive to noise is broken through, the anti-noise performance is improved, the physical behavior of the power supply is symbolized, and the jitter fault of the power supply is accurately obtained.

[0013] Further, the method for counting the difference inversion times is that for each sliding window, when the difference symbol at the sampling time is opposite to the difference symbol at the previous adjacent sampling time, it is regarded as a difference inversion.

[0014] Further, the calculation formula of the waveform response factor is that the mean value of the absolute value difference between the discrete current signals at each sampling time and the previous adjacent sampling time in each sliding window is calculated, and the product of the mean value and the difference inversion times is taken as the waveform response factor.

[0015] Beneficial effects are: the application fuses the difference in the number of reversals and the average of the current mutation amplitude through the design of the waveform response factor. A single reversal cannot distinguish between slight jitter and serious failure, while the product form simultaneously captures the jitter frequency and distortion intensity, quantifying the severity of the fault.

[0016] Further, the method for obtaining the harmonic energy is: windowing the discrete current signal in each sliding window using a window function, processing the windowed discrete current signal using a fast Fourier transform to obtain the amplitudes of all harmonics of the discrete current signal, and calculating the harmonic energy based on the harmonic amplitudes.

[0017] Further, the final sampling weight of the output current signal is obtained based on the sawtooth response factor, including: assigning a preset initial sampling weight of the output current signal to a value of 1, and normalizing the sawtooth response factor of all sliding windows as an abnormal weight of the output current signal of each sliding window; for the output current signal of each sliding window, the sum of the abnormal weight and the preset initial sampling weight is taken as the final sampling weight of the output current signal of the sliding window.

[0018] Beneficial effects are: through the superposition design of the initial weight and the abnormal weight, a dynamic sampling weight mechanism is constructed, key features are strengthened, and redundant features are suppressed.

[0019] Further, the isolated tree of the isolation forest algorithm is constructed based on the final sampling weight, including: when the isolated tree is constructed by sampling the output current signal using the isolation forest, the mean of the final sampling weights of the output current signals of all sliding windows is calculated as a split threshold, and the isolated tree is constructed by randomly sampling the output current signals greater than or equal to the split threshold and the output current signals less than the split threshold, respectively.

[0020] Beneficial effects are: by using the weight mean as the split threshold to reconstruct the isolated tree sampling mechanism, the probability of using abnormal training data for training is increased, the probability of selecting key features as the partition dimension is increased, and the isolated tree can better distinguish abnormal points.

[0021] Further, the steps of calculation are all based on the output current signal within the sliding window.

[0022] In a second aspect, the application provides a vehicle-mounted starting power supply device fault detection system, which adopts the following technical scheme:

[0023] A vehicle-mounted starting power supply device fault detection system includes a processor and a memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, a vehicle-mounted starting power supply device fault detection method according to the above is realized.

[0024] The computer program of the vehicle-mounted starting power supply device fault detection method is generated and stored in the memory to be loaded and executed by the processor, so that the system is made of the memory and the processor, and convenient use is achieved.

[0025] The application has the following technical effects:

[0026] The sawtooth response factor is proposed, three types of features of time domain waveform, frequency domain noise floor and odd harmonic energy ratio are fused, the sawtooth waveform feature is quantified by the average value of the difference inversion number and the current mutation amplitude, the current distortion caused by the jitter or failure of the switching device is directly captured, the average power spectral density of the continuous frequency band above the fundamental wave is used to detect the lifting of non-harmonic high-frequency noise, and the dependence on a single harmonic peak is avoided, the sudden increase of the odd harmonic ratio is combined, the fault features are verified twice, the small current sawtooth fluctuation and high-frequency noise are sensitively captured, the early potential fault ignored by the traditional method is identified, the performance degradation problem of the isolated forest in the high-dimensional redundant data is solved, the dynamic weight mechanism is introduced, the training data sampling weight is adjusted based on the sawtooth response factor value, the output current signal with a high sawtooth response factor value is given a high weight, the probability of selecting the key feature in the isolated tree construction is increased, the sampling probability of the stable redundant feature (such as the smooth current data) is inhibited, the noise division is reduced, the separation efficiency of the isolated tree for the abnormal point is improved, and the path length calculation time is shortened; the robustness of the model to the high-dimensional power supply signal is enhanced, the false detection and missed detection rates are reduced, through the double breakthrough of the sawtooth response factor and the weighted isolated forest, the feature with large fluctuation and significant deviation in the history is given a higher weight, the probability of selecting the key feature as the division dimension is increased, the isolated tree can better distinguish the abnormal points, the selection probability of the relatively stable and redundant features is reduced, the false detection and missed detection risks are reduced, and the efficiency of the fault detection of the vehicle-mounted starting power supply device is improved. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 It is a method flowchart of the vehicle-mounted starting power supply device fault detection method of the application. DETAILED DESCRIPTION

[0028] The application discloses a vehicle-mounted starting power supply device fault detection method, collects the output current signal of the vehicle-mounted starting power supply device, acquires a discrete current signal, constructs a difference symbol to calculate the difference inversion number, calculates a waveform response factor, further calculates a noise floor obvious factor, calculates the proportion of the odd harmonic energy in the total harmonic energy, and combines the waveform response factor and the noise floor obvious factor to acquire a sawtooth response factor; the final sampling weight of the output current signal is acquired based on the sawtooth response factor; the isolated tree of the isolated forest algorithm is constructed based on the final sampling weight, the new output current signal is detected by using the isolated tree, the vehicle-mounted starting power supply device is detected for faults, and the efficiency of the fault detection of the vehicle-mounted starting power supply device is improved.

[0029] Referring to Figure 1 A vehicle-mounted starting power supply device fault detection method includes steps S1-S3.

[0030] S1: Collect the output current signal of the vehicle-mounted starting power supply device.

[0031] To reduce the interference of noise, a low-pass filter is used to eliminate high-frequency noise in the signal, and a DC offset is used to eliminate constant components in the signal, so as to better analyze the changing part of the signal.

[0032] S2: Set a sliding window on the output current signal to obtain discrete current signals; for each discrete current signal, construct a difference symbol statistical difference inversion number, and combine a linear change trend to calculate a waveform response factor representing the degree of sawtooth jitter of the waveform; for each discrete current signal, perform frequency domain transformation, and take the average power spectrum density in the preset continuous frequency band as a bottom noise obvious factor, calculate the proportion of the energy of the odd harmonic wave in the total harmonic wave energy, and combine the waveform response factor and the bottom noise obvious factor to obtain a sawtooth response factor representing the abnormality degree of the output current signal.

[0033] Isolation forest is an unsupervised anomaly detection algorithm based on tree structure. When the traditional isolation forest algorithm detects anomalies in power supply related signals of power supply devices, in the tree construction process, sample points need to be randomly selected from training data, which may cause some key features that are highly sensitive to faults to be used at low frequency, while some very stable power supply related data has little contribution to anomalies. If it frequently participates in tree construction, it will introduce "noise division", which will "bury" the abnormal points in normal samples, thereby affecting the detection ability; therefore, based on the characteristics of the power supply related signals, the application introduces a weight mechanism. For training data, features with large fluctuations and significant offsets in history are given higher weights to increase the probability of key features being selected as division dimensions, so that the isolation tree can more quickly and accurately distinguish abnormal points, while reducing the selection probability of relatively stable and redundant features, thereby reducing the risk of false positives and false negatives.

[0034] A sliding window is constructed. In one embodiment of the application, the length of the sliding window is 10 ms, and the implementer can select other values based on actual conditions.

[0035] The power supply related signals in each sliding window are analyzed. When the starting power supply appears abnormal conditions, i.e., the switching device in the power supply fails to drive or jitters, the current signal will appear irregular sawtooth shape.

[0036] Based on the above analysis, for the output current signal of each sliding window, the output current signal is discretely sampled to obtain a discrete current signal at each sampling time. The discrete sampling method selected by the application is a time discretization method, which samples the signal at a fixed time interval to form a discrete current signal. In order to prevent signal distortion during sampling, in one embodiment of the application, the sampling frequency is five times the Nyquist frequency of the output current signal, and the implementer can select other values based on actual conditions.

[0037] For the discrete current signal at each sampling time of each sliding window, the first-order difference between the discrete current signal and the discrete current signal at the previous adjacent sampling time is calculated, and the difference symbol at each sampling time is constructed, and the formula is: ; In the formula, ， is the difference symbol of the discrete current signal at each sampling time, represents the first-order difference between the discrete current signal at each sampling time and the discrete current signal at the previous adjacent sampling time.

[0038] It should be noted that the first-order difference can reflect the linear trend of the data sequence. For the discrete current signal at each sampling time, when the first-order difference is greater than the value 0, the difference symbol amplitude of the discrete current signal is 1, when the first-order difference is less than the value 0, the difference symbol amplitude of the discrete current signal is negative 1, and in other cases, the difference symbol amplitude of the discrete current signal is 0.

[0039] That is, when the current signal appears irregular sawtooth shape, the difference symbol of the discrete current signal at adjacent sampling times is opposite, therefore, for each sliding window, when the difference symbol at the sampling time is opposite to the difference symbol at the previous adjacent sampling time, the first-order difference is reversed, and the number of difference reversals of the sliding window is finally counted, and the number of difference reversals of all sliding windows is normalized; wherein the normalization method selected by the application is the Z-Score normalization algorithm; further, the waveform response factor of the discrete current signal of each sliding window is constructed, and the calculation formula is: ; In the formula, is the waveform response factor of the discrete current signal of each sliding window; is the normalized difference reversal number of the discrete current signal of each sliding window, is the average of the absolute values of the difference between the discrete current signals at all sampling times and the previous adjacent sampling times in each sliding window.

[0040] It should be noted that when the current waveform of the starting power supply appears irregular sawtooth shape, that is, the value of the obtained difference reversal number is larger, and the difference between the discrete current signal at the sampling time and the discrete current signal at the previous adjacent time is larger, that is, The larger the value of , the more obvious the sawtooth shape of the current. In this case, the larger the value of the obtained waveform response factor is, and vice versa.

[0041] Furthermore, when the current signal exhibits an irregular sawtooth shape, the harmonic content in the frequency domain will increase significantly. Even harmonics will be eliminated in actual equipment operation, while the remaining odd harmonics will cause greater harm to the equipment. At the same time, repeated abrupt changes in the high-frequency ranges in the frequency domain will create background noise, resulting in a raised noise horizon.

[0042] Based on the above analysis, in order to reduce the "spectral leakage" phenomenon caused by truncating a finite-length sequence before performing the Fast Fourier Transform (FFT), the discrete current signal of each sliding window is windowed using a window function. In one embodiment of this application, the window function used is the Hanning window, and implementers can select other window functions based on actual conditions. The windowed discrete current signal is processed using Fast Fourier Transform (FFT) to extract the amplitude of all harmonics and the frequency of the fundamental wave of the discrete current signal. The harmonics include odd harmonics and even harmonics. The energy of all odd harmonics and the energy of the total harmonics are calculated based on the amplitude of the harmonics. The process of using Fast Fourier Transform to extract the amplitude and calculate the energy is a well-known technique and will not be described in detail in this application.

[0043] Furthermore, for the discrete current signals of each sliding window after the Fast Fourier Transform (FFT), a preset continuous frequency band is selected. In one embodiment of this application, the lower limit of the preset continuous frequency band is twice the frequency of the fundamental wave, and the upper limit is the Nyquist frequency. The implementer can select other values ​​for the upper and lower limits of the continuous frequency band based on actual conditions, and obtain the power spectral density of the discrete current signal of each sliding window. The method for obtaining the power spectral density is a well-known technique and will not be elaborated upon in this application. The mean of the power spectral density within the continuous frequency band is calculated as the noise floor significance factor of the discrete current signal of each sliding window. To facilitate subsequent processing, the noise floor significance factor of all sliding windows is normalized.

[0044] It should be noted that the noise floor factor quantifies the average power spectral density of continuous frequency bands above the fundamental frequency. When the discrete current signal exhibits an irregular sawtooth shape in the time domain, these non-harmonic components will manifest as an overall increase in noise floor in the spectrum. The noise floor factor can quickly and sensitively reflect high-frequency anomalies without relying on discrete harmonic peak values.

[0045] Based on the above analysis, and combining the waveform response factor with the frequency domain characteristics of the sawtooth shape of the current waveform, the sawtooth response factor for each sliding window is constructed, and the calculation formula is as follows: In the formula, The aliasing response factor for each sliding window. a normalized floor factor of the discrete current signal of each sliding window, a normalized waveform response factor of the discrete current signal of each sliding window, a sum of energy of all odd harmonics of the discrete current signal of each sliding window, an energy of total harmonics of the discrete current signal of each sliding window.

[0046] It should be noted that for the discrete current signal of each sliding window, when the switching device in the power supply fails to drive or dithers, the discrete current signal will have an irregular sawtooth shape in the time domain, at this time, the waveform response factor obtained is larger, at the same time, the content of odd harmonics of the sawtooth-shaped discrete current signal in the frequency domain will increase, that is, the value of is larger, at the same time, a floor noise will be formed in the frequency domain, the value of the floor factor is larger, at this time, the value of the sawtooth response factor obtained is larger, that is, the possibility of failure of the starting power device is larger; on the contrary, the value of the sawtooth response factor obtained is smaller, that is, the possibility of failure of the starting power device is smaller.

[0047] At this point, the sawtooth response factor of each sliding window is obtained.

[0048] S3: obtaining a final sampling weight of the output current signal based on the sawtooth response factor; constructing an isolation tree of the isolation forest algorithm based on the final sampling weight, and detecting a new output current signal by using the isolation tree, when an abnormal score obtained is greater than a preset abnormal threshold, determining that the vehicle-mounted starting power device has a failure.

[0049] Based on the above analysis, when the traditional isolation forest algorithm detects the power supply related signal of the power device, in the construction process of the isolation tree, the sample points need to be randomly selected from the training data, and the present application assigns different weights to different sample points based on the sawtooth response factor, specifically:

[0050] assigning a preset initial sampling weight of the output current signal to a numerical value 1, and normalizing the sawtooth response factor of all sliding windows as an abnormal weight of the output current signal of each sliding window; for the output current signal of each sliding window, the sum of the abnormal weight and the preset initial sampling weight is taken as the final sampling weight of the output current signal of the sliding window.

[0051] In the process of constructing the isolation tree by sampling the output current signal using the isolation forest, the average of the final sampling weight of the output current signal of all sliding windows is calculated as a split threshold, and the isolation tree is constructed by randomly sampling the output current signal greater than or equal to the split threshold and the output current signal less than the split threshold, respectively, so that the output current data of the key features with high sensitivity to faults can be obtained, and the isolation tree can distinguish the abnormal points more quickly and accurately. The construction process of the isolation tree is a known technology, and thus will not be described herein. In an embodiment of the present application, the selected subsampling size is 1024, and the number of trees is 100 in the process of constructing the isolation tree, and other values can be selected by the implementer based on the actual situation.

[0052] Based on the trained isolation tree, the new output current signal of the vehicle-mounted starting power supply device is detected for abnormality, and an abnormal score is calculated. When the abnormal score is greater than a preset abnormal threshold, it is determined that the current signal is abnormal data, and it is determined that the vehicle-mounted starting power supply device has a fault. Otherwise, it is determined that the vehicle-mounted starting power supply device has no fault. In an embodiment of the present application, the value of the preset abnormal threshold is 0.7, and other values can be selected by the implementer based on the actual situation.

[0053] The embodiment of the present application also discloses a vehicle-mounted starting power supply device fault detection system, which comprises a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, a vehicle-mounted starting power supply device fault detection method according to the present application is realized.

[0054] The above system also comprises a communication bus and a communication interface and other components familiar to those skilled in the art. The settings and functions thereof are known in the art, and thus will not be described herein.

[0055] The above are preferred embodiments of the present application, and do not limit the protection scope of the present application. Therefore, any equivalent changes made on the basis of the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method of detecting a failure of a vehicle-mounted starting power supply device, characterized by, The method comprises the steps of: collecting an output current signal of a vehicle-mounted starting power supply device, setting a sliding window on the output current signal, and obtaining discrete current signals; for each discrete current signal, constructing a difference symbol statistical difference inversion number, and combining a linear change trend to calculate a waveform response factor representing the degree of waveform sawtooth jitter; For each discrete current signal, a frequency domain transformation is performed, the average power spectral density of the discrete current signal in a preset continuous frequency band is taken as a bottom noise obvious factor, the proportion of the energy of an odd harmonic wave in the total harmonic energy is calculated, and the waveform response factor and the bottom noise obvious factor are combined to obtain a sawtooth response factor representing the abnormality degree of the output current signal; Based on the sawtooth response factor, a final sampling weight of the output current signal is obtained, including: assigning a preset initial sampling weight of the output current signal to a numerical value 1, and normalizing the sawtooth response factors of all sliding windows as the abnormality weight of the output current signal of each sliding window; for the output current signal of each sliding window, the sum of the abnormality weight and the preset initial sampling weight is taken as the final sampling weight of the output current signal of the sliding window; Based on the final sampling weight, an isolated tree of an isolated forest algorithm is constructed, and the isolated tree is used to detect a new output current signal; when the obtained abnormality score is greater than a preset abnormality threshold, it is determined that the vehicle-mounted starting power supply device has a fault; Based on the final sampling weight, an isolated tree of an isolated forest algorithm is constructed, including: when the isolated forest is used to sample the output current signal to construct the isolated tree, the mean value of the final sampling weights of the output current signals of all sliding windows is taken as a split threshold, and the isolated tree is constructed by randomly sampling the output current signals greater than or equal to the split threshold and the output current signals less than the split threshold, respectively.

2. The method of claim 1, wherein the method further comprises: The method for obtaining the discrete current signal is: using a time discretization method to discretely sample the output current signal to obtain the discrete current signal at each sampling time.

3. The method of claim 1, wherein the method further comprises: The method for obtaining the differential symbol is: ; in the formula ， is a differential symbol of the discrete current signal at each sampling moment, represents a first-order difference between the discrete current signal at each sampling moment and the discrete current signal at the immediately preceding adjacent sampling moment.

4. The method of claim 1, wherein the method further comprises: The method for calculating the difference inversion number is: for each sliding window, when the difference symbol at the sampling time is opposite to the difference symbol at the previous adjacent sampling time, it is regarded as one difference inversion.

5. The method of claim 1, wherein the method further comprises: The calculation formula of the waveform response factor is: calculating the mean value of the absolute values of the differences between the discrete current signals at all sampling times and the previous adjacent sampling times in each sliding window, and taking the product of the mean value and the difference inversion number as the waveform response factor.

6. The method of claim 1, wherein the method further comprises: The method for obtaining the harmonic energy is: using a window function to window the discrete current signal of each sliding window, using a fast Fourier transform to process the windowed discrete current signal to obtain the amplitudes of all harmonics of the discrete current signal, and calculating the harmonic energy based on the harmonic amplitudes.

7. The method of claim 1, wherein the method further comprises: The steps of calculation are all based on the output current signal within the sliding window.

8. An in-vehicle starting power supply device failure detection system characterized by comprising: The method comprises: A processor and a memory, the memory storing computer program instructions, when the computer program instructions are executed by the processor, a vehicle-mounted starting power supply device fault detection method according to any one of claims 1-7 is realized.

Images