Fault diagnosis method and system for special vehicle based on wide near infrared spectrum analysis

Abstract

The present application relates to the technical field of pattern recognition, and particularly relates to a special vehicle fault diagnosis method and system based on wide near-infrared spectrum analysis. The special vehicle fault diagnosis method based on wide near-infrared spectrum analysis comprises the following steps: S1: collecting ultraviolet-visible-near-infrared wide spectrum data of an oil sample; performing data layer fusion and spectrum preprocessing on the wide spectrum data to obtain final spectrum data; S2: based on the final spectrum data, using a continuous projection algorithm to screen characteristic wavelengths used for oil state analysis; S3: determining an initial characteristic wavelength based on the characteristic wavelengths; determining a set of strengthened characteristic wavelengths and a set of weakened characteristic wavelengths according to the initial characteristic wavelength. The present application fuses ultraviolet-visible-near-infrared wide spectrum and time sequence dynamic characteristics, screens key wavelengths through a continuous projection algorithm, combines clustering and change point detection, and realizes oil state evolution pattern recognition and early precise fault warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of pattern recognition, in particular to a special vehicle fault diagnosis method and system based on wide near-infrared spectrum analysis. BACKGROUND

[0002] Wide near-infrared spectrum analysis technology realizes rapid and non-destructive qualitative and quantitative analysis of oil chemical composition and physical state by detecting the absorption and scattering characteristics of substances in the ultraviolet to near-infrared band. This technology has gradually been applied to the field of oil monitoring of special vehicles, and has shown significant advantages in improving detection efficiency and realizing equipment portability by establishing a quantitative relationship between spectral characteristics and key parameters of oil, providing a new technical approach for equipment condition monitoring.

[0003] However, existing diagnosis methods based on wide near-infrared spectrum are mostly limited to static index analysis, lacking depth mining of dynamic evolution rules of oil state. Traditional methods are difficult to effectively capture the time sequence correlation and collaborative change pattern between multiple characteristic parameters, and are not sensitive to the weak characteristic response of early faults, resulting in the risk of state misjudgment under complex working conditions. In particular, when dealing with the dynamic coupling relationship between spectral characteristic wavelengths, the existing technology has not established an effective pattern recognition and evolution trend prediction mechanism.

[0004] Therefore, there is an urgent need to provide a spectrum analysis method capable of deeply analyzing the dynamic characteristics of oil state, to improve the accuracy and early warning ability of special vehicle fault diagnosis by constructing a multi-parameter collaborative evolution model. SUMMARY

[0005] In order to overcome the shortcomings of insufficient dynamic pattern recognition ability, the present application provides a special vehicle fault diagnosis method and system based on wide near-infrared spectrum analysis.

[0006] The technical implementation scheme of the present application is a special vehicle fault diagnosis method based on wide near-infrared spectrum analysis, comprising the following steps:

[0007] S1: Collecting ultraviolet-visible-near-infrared wide spectrum data of the oil sample; performing data layer fusion and spectrum pretreatment on the wide spectrum data to obtain final spectrum data;

[0008] S2: Based on the final spectrum data, using continuous projection algorithm to screen characteristic wavelengths for oil state analysis;

[0009] S3: Determining the initial characteristic wavelength based on the characteristic wavelength; determining the set of reinforced characteristic wavelengths and the set of weakened characteristic wavelengths according to the initial characteristic wavelength;

[0010] S4: determining a strengthening period according to the strengthening characteristic wavelength set; determining a weakening period according to the weakening characteristic wavelength set, and obtaining a mixed period according to the strengthening period and the weakening period;

[0011] S5: constructing an oil state evolution mode library based on the initial characteristic wavelength and the mixed period and determining an oil state evolution mode; determining a transient change rate and a statistically significant inflection point based on the oil state evolution mode, and generating a special vehicle fault diagnosis conclusion according to the transient change rate and the statistically significant inflection point.

[0012] Preferably, the ultraviolet-visible-near infrared wide spectrum data of the oil sample is collected; the wide spectrum data is subjected to data layer fusion and spectrum pretreatment to obtain final spectrum data, including:

[0013] The ultraviolet-visible-near infrared continuous wide spectrum data of the oil sample in the wavelength range of 200nm to 1700nm is collected;

[0014] The ultraviolet-visible continuous wide spectrum data and the near infrared continuous wide spectrum data are unified to the same wavelength coordinate point by using an interpolation algorithm;

[0015] The spectrum overlapping region of the ultraviolet-visible continuous wide spectrum data and the near infrared continuous wide spectrum data is subjected to smoothing processing using a transition function to obtain complete spectrum data;

[0016] The complete spectrum data is subjected to scattering correction using a standard normal transformation;

[0017] The complete spectrum data subjected to scattering correction is subjected to denoising using a Savitzky-Golay smoothing filter;

[0018] The complete spectrum data subjected to denoising is subjected to normalization processing to obtain final spectrum data.

[0019] Preferably, the characteristic wavelength used for oil state analysis is screened using a continuous projection algorithm based on the final spectrum data, including:

[0020] Step 1, setting a characteristic wavelength number parameter N to be screened by the continuous projection algorithm, randomly selecting an initial wavelength in the full wavelength range of the final spectrum data, and adding the initial wavelength to the characteristic wavelength set;

[0021] Step 2, taking the absorbance vector corresponding to the initial wavelength as an initial projection vector;

[0022] Step 3, calculating the projection of the absorbance vector corresponding to each remaining wavelength on the current projection vector in the remaining wavelengths not selected into the characteristic wavelength set;

[0023] Step 4, comparing the lengths of all the projections, selecting the wavelength corresponding to the projection with the maximum length as the characteristic wavelength of the next round, and adding the characteristic wavelength of the next round to the characteristic wavelength set;

[0024] Step 5, updating the projection vector, which is the orthogonal component of the absorbance vector corresponding to the characteristic wavelength of the next round and the current projection vector;

[0025] Step 6, if the number of wavelengths in the characteristic wavelength set does not reach the characteristic wavelength number parameter N, returning to step 3 for iteration;

[0026] Step 7, if the number of wavelengths in the characteristic wavelength set reaches the characteristic wavelength number parameter N, terminating the iteration and outputting the characteristic wavelength set as the screening result.

[0027] Preferably, the initial characteristic wavelength is determined based on the characteristic wavelength, comprising:

[0028] For each characteristic wavelength in the characteristic wavelength set, the absorbance value corresponding to each characteristic wavelength in each oil detection is recorded in chronological order to form the absorbance time series data of each characteristic wavelength;

[0029] The characteristic wavelength in the characteristic wavelength set that first deviates from the preset absorbance deviation threshold and is most relevant to the preset failure mode is selected as the initial characteristic wavelength.

[0030] Preferably, the initial characteristic wavelength is determined based on the characteristic wavelength, comprising:

[0031] If the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of another characteristic wavelength in the characteristic wavelength set exhibit a synergistic growth trend, the other characteristic wavelength is taken as the reinforced characteristic wavelength of the initial characteristic wavelength, and the reinforced characteristic wavelength set corresponding to the initial characteristic wavelength is constructed based on all the characteristic wavelengths satisfying this condition;

[0032] The synergistic growth trend refers to a Pearson correlation coefficient value greater than a preset positive correlation threshold, specifically, calculating the Pearson correlation coefficient between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of all other characteristic wavelengths in the characteristic wavelength set;

[0033] If the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of another characteristic wavelength in the characteristic wavelength set exhibit an antagonistic trend, the other characteristic wavelength is taken as the weakened characteristic wavelength of the initial characteristic wavelength, and the weakened characteristic wavelength set corresponding to the initial characteristic wavelength is constructed based on all the characteristic wavelengths satisfying this condition;

[0034] The antagonistic trend refers to the Pearson correlation coefficient value being less than a preset negative correlation threshold. Specifically, it involves calculating the Pearson correlation coefficient between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of all other characteristic wavelengths in the characteristic wavelength set.

[0035] Preferably, determining the enhancement period based on the set of enhancement characteristic wavelengths includes:

[0036] For each candidate characteristic wavelength in the enhanced characteristic wavelength set, calculate the cross-correlation function between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of the candidate characteristic wavelength;

[0037] In the sequence of cross-correlation functions, find the maximum value of the cross-correlation function and the time delay corresponding to the maximum value;

[0038] If the maximum value of the cross-correlation function is greater than the preset positive correlation threshold and the time delay is greater than zero, then the current candidate feature wavelength is determined as the effective enhancement feature wavelength, and the time delay is recorded as the enhancement period of the effective enhancement feature wavelength relative to the initial feature wavelength.

[0039] Preferably, the step of determining the weakening period based on the set of weakening characteristic wavelengths and obtaining a mixed period based on the strengthening period and the weakening period includes:

[0040] For each candidate characteristic wavelength in the weakened characteristic wavelength set, calculate the cross-correlation function between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of the candidate characteristic wavelength;

[0041] In the sequence of cross-correlation functions, find the minimum value of the cross-correlation function and the time delay corresponding to the minimum value;

[0042] If the minimum value of the cross-correlation function is less than the preset negative correlation threshold and the time delay is greater than zero, then the current candidate feature wavelength is determined as the effective weakened feature wavelength, and the time delay is recorded as the weakening period of the effective weakened feature wavelength relative to the initial feature wavelength.

[0043] The set consisting of the enhancement periods corresponding to all effective enhancement characteristic wavelengths and the weakening periods corresponding to all effective weakening characteristic wavelengths is defined as the mixing period corresponding to the initial characteristic wavelength.

[0044] Preferably, the step of constructing an oil state evolution model library based on the initial characteristic wavelength and the mixing period and determining the oil state evolution model includes:

[0045] Based on all initial characteristic wavelengths and corresponding mixing periods, a clustering algorithm is used to construct a library of oil state evolution models.

[0046] Each cluster in the oil state evolution model library represents a typical oil state evolution model, and each cluster is mapped to at least one preset fault type.

[0047] For real-time detection data, the initial feature wavelength and mixing period of the real-time detection data are extracted to generate a real-time feature vector;

[0048] By performing similarity matching between the real-time feature vector and the oil state evolution pattern library, the oil state evolution pattern to which the real-time feature vector belongs is determined, and a preliminary fault type inference is obtained.

[0049] Preferably, the step of determining the instantaneous change rate and statistically significant inflection point based on the oil state evolution model, and generating a special vehicle fault diagnosis conclusion based on the instantaneous change rate and statistically significant inflection point, includes:

[0050] Based on the established oil state evolution model, the sliding window linear fitting method is used to process the time series data of key oil state parameters and extract the instantaneous change rate of key oil state parameters.

[0051] Based on the established oil state evolution pattern, a change point detection algorithm is applied to identify statistically significant inflection points of key oil state parameters from the time series data of key oil state parameters.

[0052] If the instantaneous rate of change of a key oil condition parameter continues to exceed the historical baseline rate of the corresponding oil condition evolution mode, a fault acceleration warning will be triggered.

[0053] If a new statistically significant inflection point appears in the time series of key oil condition parameters, a fault mechanism mutation alarm will be triggered.

[0054] By integrating the oil state evolution model, instantaneous change rate status, and inflection point detection results, a final fault diagnosis method for key components of special vehicles is output.

[0055] A special vehicle fault diagnosis system based on wide near-infrared spectral analysis includes:

[0056] The wide near-infrared spectral acquisition module is used to acquire ultraviolet to near-infrared wide spectral data of oil samples, and to complete spectral data fusion and preprocessing to obtain the final spectral data;

[0057] The feature wavelength screening module uses a continuous projection algorithm to screen a set of feature wavelengths that characterize the state of the oil from the final spectral data.

[0058] The dynamic correlation analysis module identifies initial characteristic wavelengths through time-series data, constructs a set of enhanced and weakened characteristic wavelengths, and calculates the mixing period.

[0059] The fault diagnosis and decision-making module utilizes a hybrid cycle to construct a state evolution model library and combines the rate of change and inflection point detection to generate fault diagnosis conclusions.

[0060] Beneficial Effects: This invention acquires high-quality standard spectral data by collecting ultraviolet-visible-near-infrared broadband spectral data of oil samples and performing data layer fusion and spectral preprocessing. A continuous projection algorithm is used to autonomously screen the set of characteristic wavelengths representing oil state, effectively extracting key spectral information and eliminating multicollinearity interference between wavelengths. Initial characteristic wavelengths are identified based on time-series data, and through Pearson correlation coefficient and cross-correlation function analysis, sets of enhanced and weakened characteristic wavelengths and mixed periods are constructed, thereby deeply quantifying the dynamic correlation of oil state evolution and the temporal law of fault propagation. A clustering algorithm is used to construct an oil state evolution pattern library, achieving intelligent clustering identification and accurate mapping of typical fault modes. Furthermore, the instantaneous change rate of key state parameters is extracted using a sliding window linear fitting method, and a change point detection algorithm is applied to identify statistically significant inflection points, achieving early warning of fault acceleration trends and mechanism mutations. Ultimately, while improving the accuracy and stability of fault diagnosis for special vehicles, this invention significantly enhances their intelligent perception and prediction capabilities for weak fault characteristics and multi-stage evolution patterns under complex operating conditions. Attached Figure Description

[0061] Figure 1 This is a flowchart of the special vehicle fault diagnosis method based on wide near-infrared spectral analysis of the present invention;

[0062] Figure 2 This is a structural diagram of the special vehicle fault diagnosis system based on wide near-infrared spectral analysis according to the present invention. Detailed Implementation

[0063] The present invention will be further described below with reference to specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0064] Example 1: A fault diagnosis method for special vehicles based on wide near-infrared spectral analysis, such as... Figure 1 As shown, it includes the following steps:

[0065] S1: Collect UV-Vis-NIR broadband spectral data of the oil sample; perform data layer fusion and spectral preprocessing on the broadband spectral data to obtain the final spectral data, including:

[0066] The data layer fusion includes wavelength alignment and overlap region optimization; the spectral preprocessing includes scattering correction, denoising, and normalization.

[0067] Collect continuous broadband spectral data of oil samples in the ultraviolet-visible-near-infrared wavelength range of 200 nm to 1700 nm;

[0068] Interpolation algorithms were used to unify ultraviolet-visible continuous broadband data and near-infrared continuous broadband data to the same wavelength coordinate points;

[0069] The spectral overlap region between the ultraviolet-visible continuous broadband spectral data and the near-infrared continuous broadband spectral data is smoothed using a transition function to obtain complete spectral data.

[0070] The complete spectral data were scattering corrected using a standard normal transform.

[0071] The complete spectral data after scattering correction was denoised using Savitzky-Golay smoothing filtering;

[0072] The denoised complete spectral data is normalized to obtain the final spectral data.

[0073] It should be noted that in the monitoring of oil conditions in special vehicles, traditional spectral analysis methods often suffer from insufficient feature extraction stability due to inadequate fusion of multi-source spectral data and noise interference. To address this issue, this scheme first uses a UV-Vis-NIR composite optical sensing system to acquire the full-band transmission spectrum of oil samples from 200–1700 nm. Since the spectra obtained from different waveband sensors exhibit wavelength scaling differences and response shifts, interpolation algorithms (such as linear interpolation or cubic spline interpolation) are needed to unify the wavelength coordinates of the UV-Vis and NIR spectra. Furthermore, a transition function is used for weighted fusion in the overlapping spectral region to construct a continuous broad spectral curve. The transition function is a linear weighting function, i.e., in the overlapping region […]. , Within this range, the weight of the ultraviolet-visible spectrum decreases linearly from 1 to 0, while the weight of the near-infrared spectrum increases linearly from 0 to 1.

[0074] To further eliminate scattering effects and instrument random noise caused by suspended particles in the oil, a standard normal transformation was used to correct the spectral baseline, followed by Savitzky-Golay filtering to smooth high-frequency fluctuations. Finally, normalization (e.g., maximum normalization) was applied to unify the signal amplitude, forming standardized spectral data for subsequent feature learning. This preprocessing significantly improved spectral quality, laying a data foundation for subsequent robust oil condition analysis and fault diagnosis.

[0075] S2: Based on the final spectral data, a continuous projection algorithm is used to screen characteristic wavelengths for oil condition analysis, including:

[0076] Step 1: Set the number of feature wavelengths N to be screened by the continuous projection algorithm. Randomly select an initial wavelength within the entire band range of the final spectral data and add the initial wavelength to the feature wavelength set.

[0077] Step 2: Use the absorbance vector corresponding to the initial wavelength as the initial projection vector;

[0078] The 'absorbance vector' here refers to a vector composed of the absorbance values ​​of an oil sample measured at all wavelength points, and its dimension is equal to the total number of wavelength points.

[0079] Step 3: Among the remaining wavelengths that were not selected into the set of characteristic wavelengths, calculate the projection of the absorbance vector corresponding to each remaining wavelength onto the current projection vector;

[0080] Step 4: Compare the modulus of all the projections, select the wavelength corresponding to the projection with the largest modulus as the new round of characteristic wavelengths, and add the new round of characteristic wavelengths to the characteristic wavelength set.

[0081] Step 5, update the projection vector, wherein the updated projection vector is the orthogonal component of the absorbance vector corresponding to the characteristic wavelength in the new round and the current projection vector;

[0082] Step 6: If the number of wavelengths in the set of characteristic wavelengths does not reach the characteristic wavelength number parameter N, then return to step 3 for iteration;

[0083] Step 7: If the number of wavelengths in the set of characteristic wavelengths reaches the characteristic wavelength number parameter N, then terminate the iteration and output the set of characteristic wavelengths as the screening result.

[0084] It should be noted that in the field of spectral analysis, traditional oil detection methods often suffer from insufficient reliability due to the reliance on expert experience in selecting characteristic wavelengths and the existence of multicollinearity interference between wavelengths. This step achieves autonomous optimization and selection of characteristic wavelengths through a continuous projection algorithm. Specifically, the parameter N, representing the number of characteristic wavelengths, needs to be determined comprehensively based on the spectrometer resolution, the number of characteristic wavelengths to be selected, and the complexity of the final diagnostic rules. It is typically set to 5-20 key wavelengths. After the diagnostic system has been running for a period of time and accumulated sufficient new data, this cross-validation process is periodically (e.g., quarterly or annually) re-executed to verify or update the optimal N value, ensuring that the selected set of characteristic wavelengths can adapt to slow data changes. The full-band range refers to the complete data range acquired by the 200-1700nm spectrometer, covering the characteristic absorption bands of water, oxidation products, and wear particle components in the oil.

[0085] During algorithm initialization, an initial wavelength is randomly selected to eliminate human selection bias; the absorbance vector corresponding to this wavelength serves as the initial projection vector. In each iteration, the projection modulus of the remaining wavelength absorbance vectors and the current projection vector is calculated. The wavelength corresponding to the maximum modulus is selected into the feature wavelength set because it carries the most spectral information not represented by existing features. Subsequently, the projection vector is updated using the Gram-Schmidt orthogonalization method to ensure that subsequent selections are always performed in a space orthogonal to the selected feature wavelengths. The iteration terminates when the size of the feature wavelength set reaches the feature wavelength number parameter N, at which point the obtained feature wavelength set has both maximum information content and minimum redundancy.

[0086] For example, starting from the water characteristic peak at approximately 1450 nm, through six rounds of orthogonal projection iterations, an optimized set containing key wavelengths of approximately 1700 nm oxidation peak and approximately 1200 nm fuel oil peak was selected. Feature optimization was automatically completed through mathematical mechanisms, effectively overcoming the subjectivity and collinearity defects of traditional manual screening, laying the foundation for the subsequent construction of a high-precision oil state evolution model library and fault diagnosis system.

[0087] To illustrate the method for determining the key parameter N, this example uses engine oil from a certain type of special vehicle. 300 sets of oil sample spectral data and their moisture content reference values ​​were collected, covering different operating times, healthy and abnormal states. The dataset was divided into a training set (210 sets) and a validation set (90 sets) in a 7:3 ratio.

[0088] On the training set, the number of feature wavelengths, N, is iterated from 3 to 25. For each N value, a subset of feature wavelengths is selected using SPA (Special Spatial Analysis), and based on this subset, a partial least squares (PLSR) prediction function for moisture content is built on the training set to aid in evaluation. Subsequently, this PLSR function is applied to the validation set, and the root mean square error (RMSE) of its predictions is calculated to evaluate the predictive performance of the feature wavelength subsets under different N values.

[0089] The RMSE of the validation set was recorded for different values ​​of N, and the variation patterns are as follows: When N∈[3,10], the RMSE value decreases significantly and rapidly with the increase of N, indicating that increasing the number of feature wavelengths can effectively capture more key information. When N∈[11,15], the RMSE decreasing trend tends to level off. When N>15, the RMSE no longer decreases or even fluctuates, indicating that continuing to increase the number of wavelengths will introduce redundancy.

[0090] Based on the above patterns, on the RMSE-N relationship curve, N=12 is located at the inflection point where the curve transitions from a "rapid decline" to a "flattening out." At this point, the RMSE has dropped to a low level (0.076%), while redundancy is avoided. Therefore, in this application scenario, N=12 is chosen as the optimal parameter for the number of feature wavelengths. This method, by balancing prediction performance and complexity, ensures that the selected set of feature wavelengths retains core information while also possessing good robustness and practicality.

[0091] It is important to note that the moisture content PLSR prediction function used to determine parameter N in this example is a single-index, temporary auxiliary tool, whose purpose is solely to optimize the feature selection process through cross-validation. This auxiliary function itself is not the subject of this invention and is different from the oil state evolution model library for fault diagnosis ultimately constructed in this invention.

[0092] S3-1: Determining the initial characteristic wavelength based on the aforementioned characteristic wavelength includes:

[0093] For each characteristic wavelength in the set of characteristic wavelengths, the absorbance value of each characteristic wavelength in each oil detection is recorded in chronological order to form the absorbance time series data of each characteristic wavelength.

[0094] The earliest characteristic wavelength in the set of characteristic wavelengths whose absorbance value deviates from the preset absorbance deviation threshold and is most relevant to the preset fault mode is taken as the initial characteristic wavelength.

[0095] It should be noted that, to achieve early fault identification and tracing, this solution captures the dynamic evolution of oil state by constructing absorbance time-series data. The preset absorbance deviation threshold is determined based on the statistical distribution of historical data under normal equipment operation, typically within ±3 standard deviations of the mean absorbance of healthy samples. The wavelength that first exceeds this threshold and matches the spectral characteristics of the preset fault mode is set as the initial characteristic wavelength, as it both represents the temporal starting point of the fault and has clear fault indication characteristics. Taking coolant leakage faults as an example, when the absorbance of the water characteristic peak at approximately 1450nm exceeds the threshold for the first time in the third detection and has the highest matching degree with the spectral database of coolant contamination patterns, this wavelength is determined as the initial characteristic wavelength, serving as the benchmark anchor point for subsequent correlation feature mining and fault evolution analysis.

[0096] S3-2: Determine the set of enhanced characteristic wavelengths and the set of weakened characteristic wavelengths based on the initial characteristic wavelengths, including:

[0097] If the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of another characteristic wavelength in the set of characteristic wavelengths show a synergistic growth trend, then the other characteristic wavelength is used as the enhanced characteristic wavelength of the initial characteristic wavelength, and the set of enhanced characteristic wavelengths corresponding to the initial characteristic wavelength is constructed based on all characteristic wavelengths that meet this condition.

[0098] The synergistic growth trend refers to the Pearson correlation coefficient value being greater than a preset positive correlation threshold. Specifically, it involves calculating the Pearson correlation coefficient between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of all other characteristic wavelengths in the characteristic wavelength set.

[0099] If the absorbance time series data of the initial characteristic wavelength shows an antagonistic trend with the absorbance time series data of another characteristic wavelength in the set of characteristic wavelengths, then the other characteristic wavelength is taken as the weakened characteristic wavelength of the initial characteristic wavelength, and a set of weakened characteristic wavelengths corresponding to the initial characteristic wavelength is constructed based on all characteristic wavelengths that meet this condition.

[0100] The antagonistic trend refers to the Pearson correlation coefficient value being less than a preset negative correlation threshold. Specifically, it involves calculating the Pearson correlation coefficient between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of all other characteristic wavelengths in the characteristic wavelength set.

[0101] It should be noted that in the monitoring of fluids in special vehicles, fault evolution is essentially a dynamic process formed by the interaction of multiple chemical components. Traditional diagnostic methods are limited by focusing only on a single parameter threshold, failing to reveal the transmission mechanism of faults within the fluid system. This approach aims to establish a complete fault propagation path map by constructing a set of enhanced and weakened characteristic wavelengths.

[0102] The Pearson correlation coefficient was chosen as the quantification tool because of its sensitivity to linear relationships, which allows it to accurately capture the synchronicity of absorbance changes at different wavelengths. This method is used to characterize two fundamental interaction modes between oil parameters: when the absorbance of two characteristic wavelengths changes in the same direction, a synergistic effect is formed, such as the synchronous enhancement of water intrusion and metal corrosion; when they change in opposite directions, an antagonistic effect is formed, such as the trade-off between additive consumption and contaminant increase.

[0103] Threshold setting employs a statistical inference method based on historical data. The statistical significance boundary is determined by analyzing the distribution characteristics of correlation coefficients between characteristic wavelengths under normal equipment operating conditions, combined with hypothesis testing. For example: "By calculating the Pearson correlation coefficients between all characteristic wavelength pairs in a large amount of historical health data, a correlation coefficient distribution is formed. A specific quantile of this distribution (e.g., the 97.5th percentile as the positive threshold and the 2.5th percentile as the negative threshold) is taken as the threshold, with a statistical significance level corresponding to p < 0.05. The positive correlation threshold is used to identify synergistic growth patterns, and the negative correlation threshold is used to capture antagonistic trends. The determination of these two thresholds relies entirely on the statistical characteristics of a large amount of historical data to ensure effective differentiation between normal fluctuations and abnormal correlations."

[0104] This dual-set analysis mechanism elevates fault diagnosis from single-point judgment to the system level, enabling it to identify both directly related fault characteristics and indirect derivative effects, thus providing a more comprehensive analytical dimension for special vehicle fault early warning.

[0105] S4-1: Determine the enhancement period based on the set of enhancement characteristic wavelengths, including:

[0106] For each candidate characteristic wavelength in the enhanced characteristic wavelength set, calculate the cross-correlation function between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of the candidate characteristic wavelength;

[0107] In the sequence of cross-correlation functions, find the maximum value of the cross-correlation function and the time delay corresponding to the maximum value;

[0108] If the maximum value of the cross-correlation function is greater than the preset positive correlation threshold and the time delay is greater than zero, then the current candidate feature wavelength is determined as the effective enhancement feature wavelength, and the time delay is recorded as the enhancement period of the effective enhancement feature wavelength relative to the initial feature wavelength.

[0109] It should be noted that, addressing the deficiency of traditional oil diagnostic methods in adequately characterizing the temporal evolution of faults, this solution employs cross-correlation function analysis to establish dynamic correlations between characteristic wavelengths. This method constructs a complete cross-correlation function sequence by calculating the sliding correlation coefficient between the initial characteristic wavelength and candidate wavelengths on the absorbance time-series data point-by-point. This step is based on the principle of maximum correlation in signal processing theory, finding the maximum value and its corresponding time delay in the cross-correlation function sequence to capture the most significant unidirectional change pattern between two characteristic wavelengths. The method for setting the preset positive / negative correlation thresholds here is consistent with the correlation threshold setting method described in specific implementation S3-2.

[0110] Setting a constraint that the maximum value of the cross-correlation function must exceed a preset positive correlation threshold ensures the statistical significance of feature associations. Requiring a time delay strictly greater than zero follows the temporal logic of fault propagation, ensuring that changes in candidate features are indeed triggered by the initial feature. When both conditions are met, the candidate feature wavelength is assigned the identity of an effective enhancement feature wavelength, and its time delay is used as the enhancement period for subsequent construction of an oil state evolution model library. In engineering practice, this quantitative description of temporal relationships provides a crucial basis for predictive maintenance. For example, by observing the number of delay periods between metal corrosion features and moisture intrusion features, the development process of engine cooling system faults can be accurately predicted, thereby achieving a shift from passive inspection to proactive early warning.

[0111] S4-2: Determine the weakening period based on the set of weakening characteristic wavelengths, and obtain a mixed period based on the strengthening period and the weakening period, including:

[0112] For each candidate characteristic wavelength in the weakened characteristic wavelength set, calculate the cross-correlation function between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of the candidate characteristic wavelength;

[0113] In the sequence of cross-correlation functions, find the minimum value of the cross-correlation function and the time delay corresponding to the minimum value;

[0114] If the minimum value of the cross-correlation function is less than the preset negative correlation threshold and the time delay is greater than zero, then the current candidate feature wavelength is determined as the effective weakened feature wavelength, and the time delay is recorded as the weakening period of the effective weakened feature wavelength relative to the initial feature wavelength.

[0115] The set consisting of the enhancement periods corresponding to all effective enhancement characteristic wavelengths and the weakening periods corresponding to all effective weakening characteristic wavelengths is defined as the mixing period corresponding to the initial characteristic wavelength.

[0116] It should be noted that, addressing the shortcomings of traditional diagnostic methods in analyzing negative temporal relationships between features, this approach employs cross-correlation analysis to quantify antagonistic effects. This method constructs a complete cross-correlation function sequence by calculating the correlation coefficient distribution between the initial and candidate features within a sliding time window. The identification of the minimum value in the sequence is based on the principle of maximizing negative correlation, and its corresponding time delay characterizes the degree of reverse synchronization between the wavelength changes of the two features.

[0117] The minimum value must be set below a preset negative correlation threshold to ensure the statistical significance of the antagonistic relationship; simultaneously, the time delay must be greater than zero to conform to the causal temporal logic of fault propagation. The determination of the effective weakening characteristic wavelength and the recording of its weakening cycle provide key parameters for understanding the suppression mechanism triggered by the fault. Compared to the strengthening cycle, which reflects coordinated deterioration, the weakening cycle reveals the compensatory response mechanism within the system. By constructing a hybrid cycle set including strengthening and weakening cycles, a complete temporal description of the fault propagation process is formed. This dual-mode cycle integration strategy significantly improves the accuracy of fault evolution trend prediction and provides richer temporal characteristic basis for equipment maintenance decisions.

[0118] Based on the above 300 sets of healthy oil sample data, the following steps were performed to determine the key thresholds:

[0119] Absorbance deviation threshold: For each of the 12 characteristic wavelengths selected by SPA, calculate the mean (μ) and standard deviation (σ) of its absorbance in all healthy samples. The preset absorbance deviation threshold is set to μ ± 3σ. For example, for the 1450nm water characteristic peak, μ = 0.15 and σ = 0.02 are calculated, so its deviation threshold is [0.09, 0.21].

[0120] Correlation thresholds: Pearson correlation coefficients were calculated between all pairs of the 12 characteristic wavelengths in the healthy samples, resulting in 66 correlation coefficient values. A histogram of these correlation coefficients was plotted, and the 97.5th percentile (0.78) was used as the preset positive correlation threshold, and the 2.5th percentile (-0.75) was used as the preset negative correlation threshold. Statistically, this setting implies that the probability of a stronger correlation occurring in a healthy state is less than 5%.

[0121] Historical baseline rates: When constructing and applying a library of oil state evolution patterns for diagnostics, for each identified oil state evolution pattern (such as "coolant leakage pattern"), samples belonging to that pattern and in a "healthy" or "early" stage are extracted from the historical database (e.g., the first 50 test data). The instantaneous rate of change of key parameters (such as iron content) of these samples is calculated to form a "baseline rate set". The historical baseline rate warning threshold for this pattern is set to the mean of this set plus 2 standard deviations.

[0122] In this invention, "continuously exceeding" is defined as the instantaneous change rate exceeding the historical benchmark rate for three or more consecutive detection cycles.

[0123] S5-1: Construct an oil state evolution model library based on the initial characteristic wavelength and mixing period, and determine the oil state evolution model, including:

[0124] Based on all initial characteristic wavelengths and corresponding mixing periods, a clustering algorithm is used to construct a library of oil state evolution models.

[0125] Each cluster in the oil state evolution model library represents a typical oil state evolution model, and each cluster is mapped to at least one preset fault type.

[0126] For real-time detection data, the initial feature wavelength and mixing period of the real-time detection data are extracted to generate a real-time feature vector;

[0127] By performing similarity matching between the real-time feature vector and the oil state evolution pattern library, the oil state evolution pattern to which the real-time feature vector belongs is determined, and a preliminary fault type inference is obtained.

[0128] It should be noted that, to achieve intelligent diagnosis from oil spectral data to fault types, this solution constructs a complete oil state evolution pattern recognition system. Addressing the shortcomings of traditional methods that rely on expert experience and lack a systematic classification system, this step achieves automatic fault identification through pattern library construction and matching mechanisms.

[0129] The specific process for building the pattern library begins with constructing a feature matrix based on historical data. The initial feature wavelength value of each sample is used as the first dimension, and the period values ​​in the mixed period are arranged in a fixed order as subsequent dimensions, forming a standardized feature vector. Next, the K-means clustering algorithm is used, and the elbow rule is employed to determine the optimal number of clusters, k, to automatically group the feature matrix. Each cluster represents a fault mode with similar evolutionary patterns; for example, a cluster with a 1720nm water peak as the main feature and containing specific period combinations corresponds to a coolant leakage mode. Finally, through statistical analysis, the fault type with the highest frequency in each cluster is used as the preset label for that cluster, completing the construction of the fault knowledge graph.

[0130] In the real-time diagnostic execution process, for real-time detection data, the initial characteristic wavelength values ​​are first extracted, and the values ​​of each period in the mixing cycle are arranged in the same order to generate a feature vector with the same dimension as the historical data. Then, the Euclidean distance formula is used to calculate the distance between this real-time feature vector and the cluster centers in the pattern library, and the best-matching state evolution pattern is determined by the minimum distance criterion. For example, when a real-time sample with an initial characteristic wavelength of 1720nm and a mixing cycle of [8, 12, 5] is detected, its distance to each cluster center is calculated. It is found that the distance to the center of the "coolant leakage" pattern cluster is the smallest, thus inferring the fault type.

[0131] To illustrate the construction of the pattern library, this example is based on 500 sets of historical data containing clear fault labels (covering coolant leaks, fuel dilution, and normal wear). For each set of data, its "initial characteristic wavelength value" and "mixing cycle" sequence are extracted to form a feature vector.

[0132] Taking the typical failure mode of "coolant leakage accompanied by initial corrosion" as an example: Initial characteristic wavelength: Among the 12 wavelengths screened by SPA, 1450nm (water hydroxyl absorption peak) was identified as the initial characteristic wavelength of this mode. Mixed period: Through cross-correlation analysis, the following effective correlated wavelengths and periods were found: Strengthening characteristic wavelength: 500nm (iron absorption peak), with a strengthening period of 5 detection periods relative to the initial characteristic wavelength. Weakening characteristic wavelength: 1150nm (absorption peak of a certain antioxidant), with a weakening period of 8 detection periods relative to the initial characteristic wavelength. Feature vector construction: The feature vector of this mode is defined as [initial wavelength value, strengthening period 1, weakening period 1,...]. In this example, a representative feature vector is: [0.35, 5, 8]. Where: 0.35 is the normalized absorbance value of the initial characteristic wavelength (1450nm). 5 is the strengthening period of the strengthening characteristic wavelength (500nm). 8 is the weakening period of the weakening characteristic wavelength (1150nm). The K-means clustering algorithm was used to cluster the feature vectors of all 500 samples. The optimal number of clusters, k=6, was determined using the elbow rule. The center of one cluster was calculated to be [0.32, 5.2, 7.8]. 85% of the samples in this cluster were labeled as "coolant leakage". Therefore, a mapping relationship was established between this cluster and the fault type "coolant leakage accompanied by initial corrosion", and its cluster center [0.32, 5.2, 7.8] was stored in the pattern library.

[0133] Real-time diagnostic example: In a real-time detection, the extracted feature vector is... =[0.34,5,8]. The Euclidean distance between this sample and the cluster center [0.32,5.2,7.8] of the "coolant leakage" pattern in the pattern library is calculated to be 0.28, and the distances to the cluster centers of other fault mode patterns are all greater than 1.5. Therefore, this real-time sample is determined to belong to the "coolant leakage accompanied by initial corrosion" pattern.

[0134] This pattern recognition-based diagnostic mechanism automatically determines the type of fault by quantifying feature similarity, which not only ensures the repeatability of the diagnostic process but also significantly improves diagnostic efficiency, providing a reliable intelligent diagnostic solution for the hydraulic systems of special vehicles.

[0135] S5-2: Based on the oil state evolution model, determine the instantaneous change rate and statistically significant inflection point, and generate a special vehicle fault diagnosis conclusion based on the instantaneous change rate and statistically significant inflection point, including:

[0136] Based on the established oil state evolution model, the sliding window linear fitting method is used to process the time series data of key oil state parameters and extract the instantaneous change rate of key oil state parameters.

[0137] The key oil state parameters are based on the oil state evolution model and are related to the time series data of absorbance at characteristic wavelengths or the oil state indices (such as water content and iron concentration) calculated from it.

[0138] Based on the established oil state evolution pattern, a change point detection algorithm is applied to identify statistically significant inflection points of key oil state parameters from the time series data of key oil state parameters.

[0139] If the instantaneous rate of change of a key oil condition parameter continues to exceed the historical baseline rate of the corresponding oil condition evolution mode, a fault acceleration warning will be triggered.

[0140] If a new statistically significant inflection point appears in the time series of key oil condition parameters, a fault mechanism mutation alarm will be triggered.

[0141] By integrating the oil state evolution model, instantaneous change rate status, and inflection point detection results, a final fault diagnosis method for key components of special vehicles is output.

[0142] It should be noted that in intelligent diagnosis of oil condition, traditional methods, lacking quantitative tracking of the dynamic evolution of faults, often only confirm faults after they occur, failing to provide early warnings. This step introduces two dynamic indicators—instantaneous rate of change and statistically significant inflection points—aiming to capture the critical process of a fault's transformation from quantitative to qualitative change, thereby achieving predictive diagnosis.

[0143] Instantaneous rate of change is used to quantify the speed of fault development. Specifically, it operates by using a sliding window linear fitting method to process time-series data of key parameters (such as iron content) based on a defined oil state evolution model (e.g., a "coolant leakage model"). This method slides a fixed-length data window along the time axis, performs linear regression on the data points within the window, and uses the slope of the resulting fitted line as the instantaneous rate of change at the center of the window. This rate reflects the average rate of change of the parameter per unit time.

[0144] Statistically significant inflection points are used to identify turning points in the development of a fault. The specific operation involves applying a change point detection algorithm (such as a likelihood ratio test-based method) to scan time-series data for the same key parameter. When the algorithm identifies a significant abrupt change in the statistical characteristics (such as mean and variance) of the data sequence, it is marked as a statistically significant inflection point. This inflection point signifies that the fault has entered a new and more severe stage.

[0145] The triggering logic for warnings and alerts is entirely based on comparisons with historical benchmarks:

[0146] The triggering logic for the accelerated fault warning is as follows: First, the instantaneous rate of change of key parameters is calculated. When this rate is consistently higher than the historical benchmark rate of the corresponding oil state evolution mode, it indicates that wear is accelerating abnormally, and a warning is triggered.

[0147] It should be noted that the accelerated fault warning mechanism relies on establishing a quantitative and comparative rate of change benchmark for each "oil state evolution mode". The method for determining this historical benchmark rate is as follows: Defining a subset of benchmark data: From the constructed "oil state evolution mode library", for a specific mode (e.g., "coolant leakage mode"), all sample data that were historically identified as belonging to this mode but were in the early stages of its fault development are retrospectively extracted. These early-stage samples constitute the "intra-mode health / early benchmark dataset" for calculating the benchmark. Calculating the instantaneous rate of change: For each sample in this benchmark dataset, based on its position in the time series, a sliding window linear fitting method is used to calculate the instantaneous rate of change of its key oil state parameters (e.g., iron content). This yields a "sample distribution of instantaneous rate of change values" representing the mode at a normal evolution rate. Setting a statistical warning threshold: Statistical analysis is performed on all the instantaneous rate of change values ​​obtained above, calculating their mean and standard deviation. The historical benchmark rate warning threshold is ultimately set as the mean + 2 × standard deviation. Statistically, this threshold means that if the rate of change of a new sample exceeds this value, the probability that it does not belong to the early normal evolution rate distribution of the pattern is over 95%, thus triggering an early warning. Furthermore, "consistently exceeding" is defined in this invention as: within the most recent M consecutive detection periods (e.g., M=3), the calculated instantaneous rate of change of the key parameters is higher than the historical baseline rate warning threshold of its corresponding pattern.

[0148] Fault mechanism mutation alarm: When a new statistically significant inflection point is identified in the time series of key parameters (such as viscosity), it indicates that the physicochemical state of the oil has undergone a fundamental change, thereby triggering a mechanism mutation alarm.

[0149] The final diagnostic method is a decision based on the fusion of the above multi-source information. For example:

[0150] Pattern recognition: Determines the current state as "coolant leakage mode".

[0151] Dynamic analysis: The system detected that "the instantaneous rate of change of iron content continuously exceeds the historical baseline rate of this mode", triggering a fault acceleration warning; at the same time, a new statistically significant inflection point was detected in the "oil viscosity" data, triggering a fault mechanism mutation alarm.

[0152] Comprehensive Diagnosis: Based on the basic "coolant leakage" pattern, the abnormal wear trend indicated by the fault acceleration warning, and the lubrication performance quality change signal indicated by the fault mechanism sudden change alarm, the final accurate diagnostic conclusion is: "The engine has abnormal wear caused by coolant leakage, and the lubrication performance has begun to deteriorate. Immediate inspection is recommended."

[0153] Example 2: Based on Example 1, a special vehicle fault diagnosis system based on wide near-infrared spectral analysis, such as... Figure 2 As shown, it includes:

[0154] The wide near-infrared spectral acquisition module is used to acquire ultraviolet to near-infrared wide spectral data of oil samples, and to complete spectral data fusion and preprocessing to obtain the final spectral data;

[0155] The feature wavelength screening module uses a continuous projection algorithm to screen a set of feature wavelengths that characterize the state of the oil from the final spectral data.

[0156] The dynamic correlation analysis module identifies initial characteristic wavelengths through time-series data, constructs a set of enhanced and weakened characteristic wavelengths, and calculates the mixing period.

[0157] The fault diagnosis and decision-making module utilizes a hybrid cycle to construct a state evolution model library and combines the rate of change and inflection point detection to generate fault diagnosis conclusions.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing a fault of a special vehicle based on wide near infrared spectrum analysis, characterized in that, The method comprises the following steps: S1: collecting ultraviolet-visible-near infrared wide spectrum data of the oil sample; performing data layer fusion and spectrum pretreatment on the wide spectrum data to obtain final spectrum data; S2: screening characteristic wavelengths for oil liquid state analysis based on the final spectrum data by using a continuous projection algorithm; S3: determining an initial characteristic wavelength based on the characteristic wavelengths, comprising: recording absorbance values corresponding to each characteristic wavelength in previous oil liquid detection in chronological order to form absorbance time series data of each characteristic wavelength for each characteristic wavelength in the characteristic wavelength set; taking the characteristic wavelength that first appears with absorbance values deviating from a preset absorbance deviation threshold and being most relevant to a preset failure mode as the initial characteristic wavelength; determining a set of reinforced characteristic wavelengths and a set of weakened characteristic wavelengths according to the initial characteristic wavelength, comprising: if the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of another characteristic wavelength in the characteristic wavelength set present a synergistic growth trend, taking the another characteristic wavelength as a reinforced characteristic wavelength of the initial characteristic wavelength, and constructing a set of reinforced characteristic wavelengths corresponding to the initial characteristic wavelength based on all characteristic wavelengths satisfying the condition; the synergistic growth trend refers to a Pearson correlation coefficient value greater than a preset positive correlation threshold, specifically, calculating the Pearson correlation coefficient between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of all other characteristic wavelengths in the characteristic wavelength set; if the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of another characteristic wavelength in the characteristic wavelength set present an antagonistic trend, taking the another characteristic wavelength as a weakened characteristic wavelength of the initial characteristic wavelength, and constructing a set of weakened characteristic wavelengths corresponding to the initial characteristic wavelength based on all characteristic wavelengths satisfying the condition; the antagonistic trend refers to a Pearson correlation coefficient value less than a preset negative correlation threshold, specifically, calculating the Pearson correlation coefficient between the absorbance time series data of the initial characteristic wavelength and the absorbance time series data of all other characteristic wavelengths in the characteristic wavelength set; S4: determining a reinforced period according to the set of reinforced characteristic wavelengths; determining a weakened period according to the set of weakened characteristic wavelengths, and obtaining a mixed period according to the reinforced period and the weakened period; S5: constructing an oil liquid state evolution mode library and determining an oil liquid state evolution mode based on the initial characteristic wavelength and the mixed period; determining an instantaneous change rate and a statistically significant inflection point based on the oil liquid state evolution mode, and generating a special vehicle failure diagnosis conclusion according to the instantaneous change rate and the statistically significant inflection point.

2. The method for diagnosing a fault of a special vehicle based on wide near infrared spectrum analysis according to claim 1, characterized in that, The method comprises the following steps: S1: collecting ultraviolet-visible-near infrared wide spectrum data of the oil sample; performing data layer fusion and spectrum pretreatment on the wide spectrum data to obtain final spectrum data, comprising: collecting ultraviolet-visible-near infrared continuous wide spectrum data of the oil sample in a wavelength range of 200nm to 1700nm; unifying the ultraviolet-visible continuous wide spectrum data and the near infrared continuous wide spectrum data to the same wavelength coordinate point by using an interpolation algorithm; Smooth the spectral overlap region of the ultraviolet-visible continuous wide spectrum data and the near-infrared continuous wide spectrum data using a transition function to obtain complete spectrum data; Perform scattering correction on the complete spectrum data using a standard normal transform; Perform denoising on the scattering-corrected complete spectrum data using a Savitzky-Golay smoothing filter; Perform normalization processing on the denoised complete spectrum data to obtain final spectrum data.

3. The method for diagnosing a fault of a special vehicle based on wide near infrared spectrum analysis according to claim 1, characterized in that, The feature wavelengths for oil condition analysis are screened from the final spectrum data using a continuous projection algorithm, including: Step 1, setting a feature wavelength number parameter N to be screened by the continuous projection algorithm, randomly selecting an initial wavelength in the full wavelength range of the final spectrum data, and adding the initial wavelength to a feature wavelength set; Step 2, taking the absorbance vector corresponding to the initial wavelength as an initial projection vector; Step 3, calculating the projection of the absorbance vector corresponding to each remaining wavelength on the current projection vector among the remaining wavelengths not selected into the feature wavelength set; Step 4, comparing the lengths of all the projections, and selecting the wavelength corresponding to the maximum length of the projection as a new round of feature wavelength, and adding the new round of feature wavelength to the feature wavelength set; Step 5, updating the projection vector, which is the orthogonal component of the absorbance vector corresponding to the new round of feature wavelength and the current projection vector; Step 6, if the number of wavelengths in the feature wavelength set does not reach the feature wavelength number parameter N, returning to step 3 for iteration; Step 7, if the number of wavelengths in the feature wavelength set reaches the feature wavelength number parameter N, terminating the iteration and outputting the feature wavelength set as the screening result.

4. The method for diagnosing a fault of a special vehicle based on wide near infrared spectrum analysis according to claim 1, characterized in that, The determination of the strengthening period according to the strengthening feature wavelength set includes: For each candidate feature wavelength in the strengthening feature wavelength set, calculate the cross-correlation function between the absorbance time series data of the initial feature wavelength and the absorbance time series data of the candidate feature wavelength; In the cross-correlation function sequence, find the maximum value of the cross-correlation function and the time delay corresponding to the maximum value; If the maximum value of the cross-correlation function is greater than a preset positive correlation threshold and the time delay is greater than zero, the current candidate feature wavelength is determined as an effective strengthening feature wavelength, and the time delay is recorded as the strengthening period of the effective strengthening feature wavelength relative to the initial feature wavelength.

5. The method for diagnosing a fault of a special vehicle based on wide near infrared spectrum analysis according to claim 1, characterized in that, The determination of the weakening period according to the weakening feature wavelength set, and the determination of the mixing period according to the strengthening period and the weakening period include: For each candidate feature wavelength in the weakening feature wavelength set, calculate the cross-correlation function between the absorbance time series data of the initial feature wavelength and the absorbance time series data of the candidate feature wavelength; In the cross-correlation function sequence, find the minimum value of the cross-correlation function and the time delay corresponding to the minimum value; If the minimum value of the cross-correlation function is less than a preset negative correlation threshold and the time delay is greater than zero, the current candidate feature wavelength is determined as an effective weakening feature wavelength, and the time delay is recorded as the weakening period of the effective weakening feature wavelength relative to the initial feature wavelength. The set composed of the strengthening period corresponding to all the effective strengthening characteristic wavelengths and the weakening period corresponding to all the effective weakening characteristic wavelengths is defined as a mixed period corresponding to the initial characteristic wavelength.

6. The method for diagnosing a fault of a special vehicle based on wide near infrared spectrum analysis according to claim 1, characterized in that, The method comprises the following steps of: constructing an oil state evolution mode library based on the initial characteristic wavelength and the mixed period, and determining an oil state evolution mode; constructing an oil state evolution mode library based on all the initial characteristic wavelengths and the corresponding mixed periods by using a clustering algorithm; each clustering cluster in the oil state evolution mode library represents a typical oil state evolution mode, and each clustering cluster is mapped to at least one preset fault type; for real-time detection data, extracting the initial characteristic wavelength and the mixed period of the real-time detection data to generate a real-time feature vector; 7. The method for diagnosing a fault of a special vehicle based on wide near infrared spectrum analysis according to claim 1, characterized in that, performing similarity matching between the real-time feature vector and the oil state evolution mode library to determine the oil state evolution mode to which the real-time feature vector belongs, and obtaining a preliminary fault type inference. The method comprises the following steps of: based on the determined oil state evolution mode, using a sliding window linear fitting method to process time series data of a key oil state parameter, and extracting an instantaneous change rate of the key oil state parameter; based on the determined oil state evolution mode, applying a change point detection algorithm to identify a statistically significant inflection point of the key oil state parameter from the time series data of the key oil state parameter; if the instantaneous change rate of the key oil state parameter continuously exceeds a historical reference rate of the corresponding oil state evolution mode, a fault acceleration warning is triggered; if a new statistically significant inflection point appears in the time series sequence of the key oil state parameter, a fault mechanism mutation alarm is triggered; 8. A system for diagnosing faults in a special vehicle based on wide near infrared spectroscopy analysis, for implementing the method for diagnosing faults in a special vehicle based on wide near infrared spectroscopy analysis according to any one of claims 1 to 7, characterized in that, comprehensively considering the oil state evolution mode, the instantaneous change rate state and the inflection point detection result, an ultimate fault diagnosis method of a key component of a special vehicle is output. The method comprises the following steps of: a wide near-infrared spectrum acquisition module is configured to acquire oil sample ultraviolet to near-infrared wide spectrum data, complete spectrum data fusion and preprocessing to obtain final spectrum data; a characteristic wavelength screening module is configured to screen a characteristic wavelength set representing an oil state from the final spectrum data based on a continuous projection algorithm; a dynamic correlation analysis module is configured to identify an initial characteristic wavelength through time series data, construct a strengthening and weakening characteristic wavelength set and calculate a mixed period; and a fault diagnosis decision module is configured to construct a state evolution mode library by using the mixed period, and generate a fault diagnosis conclusion in combination with a change rate and an inflection point detection.

Images