Self-adaptive sand supply device fault diagnosis method and system

Abstract

The invention relates to the technical field of chips of the Internet of Things, and discloses a fault diagnosis method and system for a self-adaptive sand supply device. The method comprises the following steps: acquiring multi-channel original data of vibration, temperature and pressure of a sand supply device, and recombining a signal sequence after time synchronization and frequency band division processing; performing multi-scale decomposition denoising to obtain a smooth data set, and extracting time domain and frequency domain features to construct a feature vector set; comparing with historical feature data to analyze a long-term trend, and positioning an abnormal point and calculating coupling influence intensity and abnormal fluctuation probability if the long-term trend is abnormal; deducing a fault evolution path and defining a signal change range, and calculating a fault probability progressive increase dynamic index; and risk threshold grading evaluation is compared, an evaluation result and an index change trend are fused, and an accurate fault early warning signal is generated. According to the method, accurate fault identification and prospective early warning of the sand supply device can be realized.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) chip technology, and in particular to a fault diagnosis method and system for an adaptive sand supply device. Background Technology

[0002] Currently, the rapid development of IoT chip technology provides core support for intelligent monitoring of industrial equipment. Its low power consumption and high integration characteristics enable efficient acquisition and real-time processing of multi-dimensional data. As a key piece of equipment in industrial production, the operational stability of sand supply devices directly affects production continuity and economic benefits. Building a precise and efficient fault diagnosis system based on IoT chips has become an urgent need for the industry.

[0003] In existing technologies, fault diagnosis of sand supply devices often relies on traditional fixed-rule judgment or single-threshold monitoring methods. These methods collect single-dimensional data such as vibration, temperature, or pressure from independent sensors and directly compare them with preset thresholds to determine abnormalities. These methods do not leverage the technological advantages of IoT chips, making it impossible to achieve synchronous integration and in-depth analysis of multi-channel data. They can only passively respond to obvious faults.

[0004] In summary, existing technologies are unable to distinguish between effective normal fluctuations and real fault signals in sand supply devices, and cannot meet the needs of industrial production for intelligent operation and maintenance of sand supply devices under the empowerment of IoT chips. Summary of the Invention

[0005] This invention provides an adaptive sand supply device fault diagnosis method and system to distinguish between effective normal fluctuations and real fault signals of the sand supply device, thereby meeting the needs of industrial production for intelligent operation and maintenance of sand supply devices under the empowerment of IoT chips.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a fault diagnosis method and system for an adaptive sand supply device, comprising: Acquire multi-channel raw data of the sand supply device operation, including vibration data, temperature data and pressure data; The multi-channel raw data is time-synchronized and frequency-segmented, and time-domain signal segments exceeding a preset reference energy threshold are extracted. The time-domain signal segments are then reassembled to obtain a reassembled signal sequence. The recombined signal sequence is subjected to multi-scale decomposition and denoising to obtain a smoothed dataset; Extract time-domain and frequency-domain features from the smoothed dataset to construct a multi-dimensional feature vector set; The feature vector set is compared with the pre-acquired historical feature data in the time dimension to analyze the long-term trend of feature changes. If the long-term trend shows abnormal fluctuations, then potential anomalies are located and the coupling influence strength between multiple variables is analyzed, and the coupling influence strength is converted into the probability of abnormal fluctuations. Based on the abnormal fluctuation probability, the fault evolution path is deduced, and the signal change range is defined. Combining the abnormal fluctuation probability, the signal change range, and the fluctuation amplitude of the long-term change trend, a dynamic index of increasing fault probability is calculated. The dynamic indicators are compared with preset risk assessment thresholds. The fault risk is graded and assessed based on the extent to which the dynamic indicators exceed the risk assessment thresholds. The graded assessment results and the changing trends of the dynamic indicators are then integrated to generate accurate fault early warning signals.

[0007] Secondly, the present invention provides an adaptive sand supply device fault diagnosis system, comprising: The data acquisition module is used to acquire multi-channel raw data of the sand supply device operation, including vibration data, temperature data and pressure data. The signal processing module is used to perform time synchronization and frequency band processing on the multi-channel raw data, extract time-domain signal segments that exceed a preset reference energy threshold, and reassemble the time-domain signal segments to obtain a reassembled signal sequence. The denoising module is used to perform multi-scale decomposition and denoising on the recombined signal sequence to obtain a smooth dataset. The feature extraction module is used to extract time-domain and frequency-domain features from the smoothed dataset and construct a multi-dimensional feature vector set. The trend judgment module is used to compare the feature vector set with the pre-acquired historical feature data in the time dimension to analyze the long-term trend of feature changes. An anomaly analysis module is used to locate potential anomaly points and analyze the coupling influence strength between multiple variables if the long-term trend shows abnormal fluctuations, and to convert the coupling influence strength into anomaly fluctuation probability. The indicator calculation module is used to deduce the fault evolution path based on the abnormal fluctuation probability, define the signal change range, and calculate the dynamic indicator of increasing fault probability by combining the abnormal fluctuation probability, the signal change range, and the fluctuation amplitude of the long-term change trend. The early warning generation module is used to compare the dynamic indicators with preset risk judgment thresholds, classify and assess the fault risk according to the extent to which the dynamic indicators exceed the risk judgment thresholds, and integrate the classification assessment results with the changing trend of the dynamic indicators to generate accurate fault early warning signals.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention acquires multi-channel raw data of the sand supply device and performs time synchronization and frequency segmentation processing on the multi-channel raw data to obtain a recombined signal sequence, so that vibration data, temperature data and pressure data can form a consistent data expression under the same time reference, avoiding the information fragmentation and time misalignment problems caused by single-dimensional acquisition in the prior art; at the same time, frequency segmentation processing can effectively separate the key frequency band features of the equipment operation from the complex background, so that subsequent diagnosis no longer relies on coarse-grained judgment with fixed thresholds, but can form a more accurate state characterization based on the comprehensive information of multiple variables and multiple frequency bands, thereby significantly improving the reliability of basic data for fault diagnosis and the ability to capture anomalies, and solving the problem of difficulty in distinguishing normal fluctuations from real fault signals and frequent false alarms and missed alarms in the prior art.

[0009] (2) The present invention obtains a smooth dataset by performing multi-scale decomposition and denoising on the signal sequence, and further extracts time-domain features and frequency-domain features from the smooth dataset to construct a multi-dimensional feature vector set, so that the weak anomalies of the sand supply device under multi-variable interference environment can be effectively highlighted; wherein, multi-scale decomposition and denoising weaken the influence of high-frequency random noise on the diagnostic results, and the joint modeling of time-domain features and frequency-domain features can characterize the equipment operating status from the two levels of waveform change and energy distribution, so that the fault judgment is transformed from the traditional experience threshold comparison to a quantifiable multi-feature comprehensive analysis, thereby improving the identification accuracy of early wear and progressive faults, and solving the problems of poor adaptability to complex working conditions and inability to maintain diagnostic stability under multi-variable coupled interference in the prior art.

[0010] (3) This invention combines pre-acquired historical feature data, compares the feature vector set with the historical feature data in the time dimension to determine the long-term trend of feature changes, and locates potential anomalies when the long-term trend shows abnormal fluctuations and analyzes the coupling effect between multiple variables to obtain the probability of abnormal fluctuations. Furthermore, it infers the fault evolution path and defines the signal change range based on the probability of abnormal fluctuations, calculates the dynamic index of increasing fault probability and generates accurate fault warning signals, so that fault diagnosis is upgraded from post-event alarm to an active warning mechanism of trend prediction and risk classification. This mechanism can identify trend deviation and probability increase characteristics before the fault becomes apparent, and achieves accurate output of warning signals through risk judgment threshold comparison and classification assessment, thereby significantly reducing the probability of unplanned equipment downtime and improving the pertinence and timeliness of maintenance decisions. It solves the problem that the existing technology lacks predictive ability in the face of dynamic working conditions and cannot conduct forward assessment of fault development. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of a fault diagnosis method for an adaptive sand supply device provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an adaptive sand supply device fault diagnosis system provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a fault diagnosis method for an adaptive sand supply device, comprising the following steps: S11, acquire multi-channel raw data of the sand supply device operation, the multi-channel raw data including vibration data, temperature data and pressure data; S12, perform time synchronization and frequency segmentation processing on the multi-channel raw data, extract time-domain signal segments that exceed the preset reference energy threshold, and reassemble the time-domain signal segments to obtain a reassembled signal sequence; S13, Perform multi-scale decomposition and denoising on the recombined signal sequence to obtain a smoothed dataset; S14, extract time-domain features and frequency-domain features from the smoothed dataset, and construct a multi-dimensional feature vector set; S15, compare the set of feature vectors with the pre-acquired historical feature data in the time dimension to analyze the long-term trend of feature changes; S16. If the long-term trend shows abnormal fluctuations, then locate potential anomalies and analyze the coupling influence strength between multiple variables, and convert the coupling influence strength into the probability of abnormal fluctuations. S17. Based on the abnormal fluctuation probability, the fault evolution path is deduced, and the signal change range is defined. Combining the abnormal fluctuation probability, the signal change range, and the fluctuation amplitude of the long-term change trend, a dynamic index of increasing fault probability is calculated. S18, compare the dynamic indicator with the preset risk judgment threshold, classify and assess the fault risk according to the extent to which the dynamic indicator exceeds the risk judgment threshold, and integrate the classification assessment result with the changing trend of the dynamic indicator to generate an accurate fault early warning signal.

[0014] In step S11, multi-channel raw data of the sand supply device operation is acquired. This multi-channel raw data includes vibration data, temperature data, and pressure data, including: A sensor array is deployed to collect vibration, temperature, and pressure data during the operation of the sand supply device, resulting in multi-channel raw data.

[0015] It should be noted that specific sensor models were selected when deploying the sensor array to ensure data accuracy and reliability. Vibration data acquisition used a triaxial piezoelectric accelerometer, installed near the motor bearing housing of the sand supply device and rigidly secured with bolts to reduce installation resonance. The sampling frequency was set to 2000Hz to capture high-frequency vibration components. Temperature data acquisition used a platinum resistance temperature sensor, attached to the outer ring of the bearing or the surface of the motor stator housing, and thermally conductive silicone grease was used to enhance thermal coupling. The sampling frequency was set to 50Hz to match the thermal inertial response speed. Pressure data acquisition used a strain gauge pressure sensor, installed in the pressure stabilizing section of the sand supply pipeline or at the front end of the throttling component, also with a sampling frequency of 50Hz.

[0016] In step S12, the process of performing time synchronization and frequency segmentation on the multi-channel raw data, extracting time-domain signal segments exceeding a preset reference energy threshold, and reassembling the time-domain signal segments to obtain a reassembled signal sequence includes: The vibration data, temperature data, and pressure data are timestamped to generate a time-synchronized data stream. The time-synchronized data stream is subjected to spectral analysis to decompose it into a set of sub-signals in multiple frequency bands; If the instantaneous energy value of any frequency band in the sub-signal set exceeds the preset reference energy threshold, then the corresponding time point is locked and a time domain signal segment is extracted. The time-domain signal segments are vectorized and recombined in chronological order to obtain a recombined signal sequence.

[0017] It should be noted that, firstly, timestamp alignment uses the hardware clock of the data acquisition unit as a unified reference. A global time index is generated based on the sampling point sequence number of the vibration channel. The vibration sampling frequency is 2000Hz, and the temperature and pressure sampling frequencies are 50Hz. Temperature and pressure data are aligned to 2000Hz using linear interpolation and synchronized via timestamps to generate a time-synchronized data stream. After alignment, the units for each channel remain unchanged: vibration is recorded in acceleration (g), temperature in degrees Celsius (°C), and pressure in megapascals (MPa), ensuring consistent physical meaning in subsequent processing. For example, if the vibration signal experiences an impact peak at 120.250 seconds, the temperature and pressure data can be matched at the same time point after interpolation, forming synchronized multivariate data.

[0018] In this embodiment, the spectrum analysis is implemented using a wavelet packet decomposition algorithm. The wavelet basis is Daubechies4, and the decomposition level is set to three, dividing the signal into eight sub-band components, covering the typical operating frequency band of the sand supply device. Each sub-band component is reconstructed into a time-domain sub-signal, forming a multi-band sub-signal set for instantaneous energy calculation. This decomposition method is suitable for non-stationary impact and friction noise conditions, allowing the impact component to be concentrated in the high-frequency sub-bands. For example, after decomposition, the high-frequency sub-bands show energy increases during friction, while the low-frequency sub-bands show less change, which can be combined with the pressure sub-signal pulsation enhancement to pinpoint abnormal events.

[0019] It should be noted that the baseline energy threshold is set based on statistical data from the past 12 months of healthy operation, and is configured independently for the data characteristics of different channels such as vibration, temperature, and pressure. The specific setting method is as follows: First, for vibration data, the focus is on the high-frequency energy generated by mechanical impact. Since mechanical vibration energy typically follows a skewed distribution, probability density analysis of historical health datasets shows that its median energy is approximately 0.04 g² (representing the normal energy level during stable equipment operation). To cover the vast majority of occasional load fluctuations under normal operating conditions, the 95th quantile of the cumulative probability distribution function is calculated, yielding a value of 0.09 g². Therefore, 0.09 g² is set as the baseline energy threshold for vibration data. This threshold effectively filters out 95% of background noise, thus accurately capturing subtle impacts caused by bearing wear or abnormal gear meshing.

[0020] Secondly, for pressure data, the main focus is on monitoring abnormal fluctuations in fluid pulsation. Based on fluid dynamics characteristics, the pressure pulsation energy in the sand supply pipeline exhibits a wide distribution range. Statistical results show that the median pressure energy under healthy conditions is 0.25 MPa²; however, considering the normal pulsation peaks present during fluid transport, its 95th percentile reaches 0.40 MPa². Therefore, 0.40 MPa² is set as the baseline energy threshold for pressure data to effectively distinguish normal sand supply fluid pulsation from pressure surges caused by pipeline blockage or leakage.

[0021] Finally, for temperature data, the focus is on energy fluctuations caused by changes in thermal inertia. Energy statistics are performed on the sub-bands (usually low-frequency bands) after temperature signal decomposition. Although the median energy of thermal fluctuations under healthy conditions is only 0.5℃² (reflecting the stability of temperature changes), to prevent false alarms caused by sudden changes in ambient temperature, the result calculated based on the 95th percentile of the statistical distribution is 1.5℃². Therefore, setting 1.5℃² as the baseline energy threshold for temperature data allows for the rapid identification of abnormal temperature rise rates caused by motor overheating or frictional heat generation, while maintaining a sufficient safety margin.

[0022] It should be noted that all the above thresholds have dynamic adjustment capabilities. By reviewing the median energy of the latest health data weekly, if the median energy of a certain channel drifts by more than 10%, the corresponding threshold of that channel will be automatically corrected proportionally to ensure that the system's sensitivity to capturing various faults remains stable when the equipment ages or the operating conditions change.

[0023] The instantaneous energy value is calculated within a sliding window with a window length of 0.5 seconds. The calculation method involves summing the squares of the amplitudes at each sampling point within the window and dividing by the number of sampling points to obtain the average energy value, thus eliminating the influence of the window length. The energy value is measured in squares of the original data unit of the corresponding channel. Specifically, vibration energy is measured in g², temperature energy in ℃², and pressure energy in MPa². Threshold determination is strictly performed by comparing the energy value with the respective set thresholds within the same frequency band of the same channel. If the instantaneous energy of any frequency band in the sub-signal set exceeds the reference energy threshold for that channel, the corresponding time point is locked and a time-domain signal segment is extracted. For example, if the background noise energy of the high-frequency sub-signal in the vibration channel is 0.04 g², and an impact peak causes the instantaneous energy to rise to 0.16 g², exceeding the preset vibration energy threshold of 0.09 g², then this time point is locked and the extraction process begins. Similarly, if the pressure signal energy fluctuates to 0.50 MPa², exceeding its threshold of 0.40 MPa², extraction is also triggered.

[0024] It is worth noting that the time-domain signal segment truncation adopts a bidirectional extension method, truncating a pre-truncation duration of 0.2 seconds before the locked time point and a post-truncation duration of 0.8 seconds after the anomaly, which can cover the baseline segment before the anomaly occurs and the attenuation segment after the anomaly occurs. The post-truncation duration is adaptively extended according to the time required for the energy to fall below the threshold, avoiding the anomaly process being truncated. During segment truncation, segments of vibration, temperature, and pressure within the same time range are output simultaneously to ensure that the multivariate coupling relationship is preserved. For example, if an impact-type anomaly forms a 1-second segment, and the energy falls back after 1.6 seconds, the post-truncation is extended to 1.6 seconds to include the complete tail process.

[0025] Finally, vectorized recombination uses fixed-length segments as basic units. For each segment, feature vectors of a uniform dimension are extracted from each frequency band and concatenated. These features include time-domain statistical features such as root mean square (RMS), peak value, kurtosis, and peak factor, and frequency-domain features such as main peak frequency, spectral centroid, and band energy percentage. To eliminate the influence of different physical dimensions on the subsequent model weights, uniform min-max normalization is performed on all dimensional features. Specifically, for both vibration RMS and pressure pulsation values, based on their statistical distribution in the healthy sample set, the 1st quantile is selected as the lower bound of normalization, and the 99th quantile as the upper bound, linearly mapping their original values ​​to the 0-1 interval. If real-time data exceeds these statistical boundaries, it is clamped to 0 or 1, thus ensuring a high degree of consistency in the numerical scale of all feature vectors. Multiple feature vectors are then concatenated in chronological order to obtain the reconstructed signal sequence.

[0026] In step S13, the multi-scale decomposition and denoising process of the reconstructed signal sequence to obtain a smoothed dataset includes: The recombined signal sequence is decomposed into approximation coefficients and detail coefficients by multi-scale decomposition. By combining the approximation coefficients and the detail coefficients, signal denoising and reconstruction are performed to generate a denoised signal sequence; Calculate the first-order difference variance of the denoised signal sequence, select the continuous sequence segments with the smallest variance, and form a smoothed dataset.

[0027] It should be noted that the discrete wavelet transform algorithm is used when performing multi-scale decomposition on the reconstructed signal sequence. The wavelet basis is Daubechies4, and the decomposition level is set to 4 levels. The signal is decomposed into approximation coefficients and detail coefficients to capture features at different scales. The approximation coefficients preserve the low-frequency trend components of the signal, while the detail coefficients contain high-frequency fluctuations and noise information.

[0028] In this embodiment, the threshold is set based on statistical analysis of health operation data over the past 12 months. The median distribution of the absolute values ​​of detail coefficients at each level is statistically analyzed. The median of the higher-level detail coefficients is multiplied by a scaling factor to obtain the basic threshold. The scaling factor is determined by the noise statistical distribution characteristics of the health data, and its specific value follows the robust standard deviation estimation principle. The noise standard deviation is estimated using the median absolute deviation (MAD). (0.6745), setting the scale factor to 2.5 is equivalent to setting it to approximately 3.7. The filtering boundary is set at 2.5 / 0.6745 ≈ 3.7. This value aims to ensure coverage of over 99.9% of random Gaussian white noise fluctuations, thereby suppressing background noise while preserving non-Gaussian fault impact characteristics to the greatest extent. Threshold adjustability is achieved through weekly health data review. If the median of detail coefficients in the latest week's health data drifts by more than 10%, the threshold is adjusted synchronously according to the new median to ensure that the noise reduction effect adapts to changes in equipment status. For example, after the vibration signal of the sand supply device is decomposed into four layers, the median of detail coefficients in the fourth layer is 0.03, and the base threshold is set to 0.075. When the median of the health data drops to 0.025, the threshold is lowered to 0.0625 to maintain noise suppression stability.

[0029] Next, when performing signal denoising and reconstruction using approximation coefficients and detail coefficients, a soft thresholding function is applied to the detail coefficients. The thresholding function is defined as an amplitude compression mode: coefficients with absolute values ​​below the threshold are set to zero, while those above the threshold retain the compressed amplitude by sign. During reconstruction, an inverse wavelet transform is performed using the approximation coefficients and the processed detail coefficients to generate a denoised signal sequence. It is worth noting that since the input reconstructed signal sequence has already undergone normalization, the generated denoised signal sequence also maintains a dimensionless normalized state, with its numerical distribution remaining within the 0 to 1 range. It does not possess physical units such as vibration (g), temperature (°C), or pressure (MPa), thus ensuring the consistency of numerical scales for each channel in subsequent feature extraction steps.

[0030] For example, the detail coefficient of a normalized vibration signal has a noise peak with an amplitude of 0.12 in the 1500Hz frequency band. After exceeding the threshold of 0.075, it is compressed to 0.04. The dimensionless background noise level of the reconstructed signal is reduced from 0.05 to 0.02, while the normalized impact peak value of 0.5, which represents the fault characteristics, is still completely preserved.

[0031] It's worth noting that when calculating the first-order difference variance of the denoised signal sequence, a sliding window is used to first calculate one segment of the denoised sequence's difference sequence, and then calculate the variance of that segment. The window length is set based on the shortest duration of the stable segment in the health data. The duration distribution of stable segments in the health data over the past 12 months is statistically analyzed, with a common duration of 5 seconds used as the default window length. After variance calculation, it is normalized to the 0-1 interval. The upper and lower bounds of normalization are determined by the 1st and 99th percentiles of all variance values ​​in the first-order difference sequence of the health data. When selecting the continuous sequence segment with the smallest variance, the segment length must be no less than the minimum length required for model processing (10 seconds). If multiple candidate segments have similar variance values, a relative difference threshold is used for judgment. The relative difference ratio between the variance of each candidate segment and the global minimum variance is calculated. If the variance of a segment does not exceed 5% of the global minimum variance (i.e., the relative difference threshold), the two are considered to have comparable stability. In this case, following the time priority principle, the segment with the most recent timestamp is directly selected to ensure that the data used for diagnosis reflects the latest operating status of the equipment.

[0032] Finally, these optimal segments with the smallest variance are extracted and concatenated to form a smoothed dataset, ensuring that subsequent feature extraction is always based on the highest quality data samples within the current time period. For example, the system calculates the variances of three consecutive candidate segments to be 0.005, 0.002, and 0.008, respectively. After comparison, the segment with a variance of 0.002 is identified as the current minimum and is directly selected as a component of the smoothed dataset, while other segments with relatively large variances are automatically discarded, thus eliminating interference from unstable data.

[0033] In step S14, extracting time-domain and frequency-domain features from the smoothed dataset to construct a multi-dimensional feature vector set includes: Calculate the mean, variance, peak value, and kurtosis of the smoothed dataset to form a time-domain feature set; The smoothed dataset is subjected to spectral transformation to extract the main frequency, power spectrum peak value and frequency band energy, forming a frequency domain feature set; The time-domain feature set and the frequency-domain feature set are concatenated according to their dimensions to construct a multi-dimensional feature vector set.

[0034] In this embodiment, time-domain feature extraction includes calculating the mean, variance, peak value, and kurtosis of the smoothed dataset. The mean reflects the overall signal offset level, the variance characterizes the fluctuation intensity, the peak value captures instantaneous amplitude anomalies, and the kurtosis quantifies the steepness of the signal distribution to identify impact components. It is calculated by taking the average of the fourth power of the difference between each point in the smoothed dataset and the mean, then dividing by the square of the variance (i.e., the fourth-order normalized moment). This index is extremely sensitive to non-Gaussian pulse signals generated by early faults (such as microcracks), approaching 3 under normal conditions and significantly increasing under fault conditions. Frequency-domain feature extraction involves performing a Fast Fourier Transform on the smoothed dataset to achieve spectrum conversion. A Hamming window is used to reduce spectral leakage, extracting the dominant frequency (the frequency value where the power spectral density is maximum), the power spectral peak value (the amplitude corresponding to that frequency), and the band energy (the power spectral integral value within a preset band).

[0035] To eliminate the impact of dimensional differences on model weights, all extracted time-domain and frequency-domain features (specifically including dominant frequency and kurtosis) undergo uniform min-max normalization. Specifically, each feature is normalized based on its statistical distribution in the past 12 months of health data, using the 1 quantile as the lower bound and the 99 quantile as the upper bound, linearly mapping the original feature values ​​to the 0-1 range. If the real-time calculated feature value exceeds this statistical boundary, its value is clamped to 0 or 1, ensuring all feature data are on a uniform numerical scale before being input into the model. Before feature concatenation, dimensionality consistency is checked. The time-domain feature set contains 4-dimensional vectors, and the frequency-domain feature set contains 3-dimensional vectors, concatenated in a fixed order to form a 7-dimensional feature vector set, ensuring that each feature vector represents the comprehensive state of the same time segment.

[0036] For example, in the vibration signal analysis of a sand supply device, the original eigenvalues ​​of a segment of the smoothed dataset before normalization are as follows: mean = 0.5, variance = 0.3, peak value = 0.8, kurtosis = 4.2, dominant frequency = 35Hz, peak power spectrum = 0.5, and frequency band energy = 0.6. After normalization based on the statistical boundaries of the health data, the original kurtosis of 4.2 is mapped to 0.75, and the original dominant frequency of 35Hz is mapped to 0.35. The final concatenated eigenvector is [0.5, 0.3, 0.8, 0.75, 0.35, 0.5, 0.6], where all elements are in the range of 0 to 1, and are used for subsequent trend analysis.

[0037] It should be noted that the frequency band range in the frequency band energy calculation is set based on statistical data from the past 12 months of healthy operation. By analyzing the energy distribution of healthy vibration signals, the main energy concentration band of 0 to 200 Hz is set as the basic frequency band, and the secondary frequency band of 200 to 500 Hz is set as the extended frequency band. The basic frequency band covers the dominant components of normal operation, and the extended frequency band captures abnormal high-frequency harmonics. The adjustability of the frequency band boundary is achieved through weekly health data review. If the center of gravity of the energy distribution in the latest health data shifts by more than 10%, the frequency band boundary is adjusted proportionally according to the new distribution. This setting has been verified in multiple scenarios and can effectively distinguish between background noise and fault harmonics when the sand supply device is operating stably.

[0038] In step S15, comparing the feature vector set with pre-acquired historical feature data over a time dimension to analyze the long-term trend of feature changes includes: Construct a historical data window matrix based on pre-acquired historical feature data; Calculate the difference between the feature vector set and the corresponding dimension in the historical data window matrix to generate a dimension difference sequence; The dimensional difference sequence is smoothed to filter out short-term random fluctuations, resulting in a stationary difference sequence. A trend change curve is generated by fitting the stationary difference sequence; The fluctuation amplitude of the trend change curve is calculated. If the fluctuation amplitude exceeds a preset trend fluctuation threshold, the long-term trend of the feature is determined to exhibit abnormal fluctuation.

[0039] It should be noted that a sliding window approach is used to process historical feature data when constructing the historical data window matrix. The window length W is set based on the statistically significant duration of stable operation data over the past 12 months. The most common duration of continuous stable segments in the statistically analyzed health data is used as the default window length. For example, if the stable segment of the sand supply device under standard load lasts for more than 24 hours, then W is set to the number of time points corresponding to 24 hours. The step size G is one-tenth of the window length to ensure trend continuity. When equipment operating conditions change frequently, G is adjusted synchronously according to the shortening ratio of stable segments. Each row of the matrix represents the feature evolution within a time window, facilitating alignment and comparison with the current feature vector set by time point.

[0040] For example, historical feature data is generated into a vector every 10 minutes. W is set to 24 hours, corresponding to 144 vectors, and G is set to 1 hour, corresponding to 6 vectors. Then the number of matrix rows covers the entire historical period, and the features at the current moment can be directly compared with those of the same historical window period.

[0041] Next, when calculating the dimensional difference sequence, a periodic time alignment method is used to compare the current feature vector set with the historical data window matrix. This periodic time alignment refers to identifying the intraday time period (e.g., 10:00-10:10 AM) or process cycle stage (e.g., the second hour after sand supply starts) of the current real-time data, and retrieving all historical windows in the same time period or process stage from the historical data window matrix. For example, if the current time is the "high-pressure conveying stage" of the sand supply operation, only windows in the historical data that are also in the "high-pressure conveying stage" are selected as the benchmark. Subsequently, the absolute difference in each dimension between the current feature vector and the median of the features within these selected historical windows is calculated, i.e., the absolute value of the current value minus the absolute value of the historical benchmark median. After the difference sequence is generated, it is arranged in chronological order to form the difference evolution curve for each dimension.

[0042] When smoothing the dimensional difference sequence, a moving average algorithm is used. The coefficient is set based on the autocorrelation statistical analysis of the difference sequence of health data over the past 12 months: First, the autocorrelation function (ACF) of the health difference sequence is calculated to identify the number of lag steps required for the autocorrelation coefficient to decay to 0.5 (correlation length). If the lag step is short (e.g., less than 3 steps), it indicates strong random fluctuations in the sequence, requiring a smaller smoothing coefficient (0.1-0.15) to enhance smoothing. If the lag step is long (e.g., more than 10 steps), it indicates strong trend in the sequence, and the smoothing coefficient can be appropriately increased (0.2-0.3). Statistically, the average number of autocorrelation decay steps for the sand supply device health data is 5 steps, corresponding to a smoothing coefficient of 0.2, which best balances noise reduction and response speed. Smoothing is performed independently on each dimension. The new smoothed value is calculated as the current difference value multiplied by the coefficient A, plus the previous smoothed value multiplied by 1 minus A, resulting in a stationary difference sequence. The coefficient A is adjustable. When the equipment noise level increases, causing increased sequence jitter, A is lowered to 0.1 to enhance the smoothing effect; when a rapid response to deterioration trends is required, A is raised to 0.3 to reduce hysteresis. For example, if a differential sequence in a certain dimension experiences a jump from 0.05 to 0.20 within 10 minutes, the smoothed sequence fluctuates stably around 0.14, and the spikes are effectively suppressed, resulting in a stable differential sequence.

[0043] In this embodiment, a robust locally weighted scatter smoothing algorithm is used to fit the trend change curve. The regression span is set to 0.3, which determines the proportion of data points in the neighborhood participating in the local regression. First, power spectral density analysis is performed on the stationary difference sequence to identify the dominant cycle length under healthy conditions (e.g., the 24-hour diurnal cycle of a sand supply device affected by ambient temperature). To effectively filter out periodic fluctuations and retain long-term trends, the local regression window must cover at least three complete dominant cycles. Therefore, the regression span value is equal to the local regression window length divided by the total analysis window length. The value of this parameter was determined through comparative experiments: historical datasets containing early fault characteristics were selected, and trend fitting was performed with regression spans of 0.1, 0.3, and 0.5, respectively. Experimental results show that when the span is 0.1, the fitted curve retains a large amount of periodic fluctuations, leading to a false alarm rate of 15%; when the span is 0.5, the fitted curve is too smooth, resulting in a delay of more than 2 hours in identifying the inflection point of the fault trend; while when the span is 0.3, it completely filters out the interference of the diurnal cycle and controls the delay in identifying the fault inflection point to within 10 minutes. Therefore, a value of 0.3 achieves the best balance between eliminating periodic noise and preserving the inflection point of fault evolution.

[0044] It should be noted that the Huber loss function is used in conjunction with iterative reweighted least squares to reduce the impact of outliers. In each iteration, the residuals are calculated. For data points with residuals less than a threshold (normal points), a squared error loss (weight 1) is used; for data points with residuals greater than the threshold (outliers, such as instantaneous spikes in sensor data), a linear error loss is used, with its weight decreasing inversely with increasing residuals. The threshold is set to 1.345 times the absolute median of the residuals. Through 3 to 5 iterations, the weights of sudden outliers are automatically reduced to near zero, thus ensuring that the fitted curve closely follows the main trend of the data rather than being skewed by noise.

[0045] For example, a steady-state difference in a certain dimension shows a slow upward trend over two weeks, increasing by an average of 0.05 units per day, while also being superimposed with a daily diurnal cycle fluctuation with an amplitude of 0.2. Without robust fitting, the curve would be pulled by the daily peaks and troughs; however, by using a regression span of 0.3 and introducing Huber loss, the fitted curve can completely smooth out the 0.2 cycle fluctuation and automatically ignore occasional instantaneous impact noise above 0.5, outputting a highly smooth linear evolution trajectory with a slope stable at 0.05, clearly characterizing the monotonic degradation direction of equipment performance.

[0046] Next, when calculating the fluctuation range of the trend change curve, the range method is used for quantification, that is, the maximum value of the trend change curve within the current evaluation window is selected and the minimum value is subtracted to obtain the fluctuation range value.

[0047] It's worth noting that the trend fluctuation threshold is set based on the statistical analysis of the fluctuation range of health data over the past 12 months. The peak-to-trough difference distribution of the trend curve under health conditions is calculated for each dimension, and the 95th percentile is used as the base threshold. Threshold adjustability is achieved through weekly health data review. If the median fluctuation range of the latest week's health data drifts by more than 10%, the threshold is adjusted proportionally. For example, if the health fluctuation range threshold for a certain dimension is 0.18, and the current peak-to-trough difference of the trend curve is 0.22, exceeding the threshold is considered abnormal fluctuation. If increased equipment load causes the median health fluctuation range to rise, the threshold is simultaneously adjusted to 0.20 to maintain judgment stability.

[0048] In step S16, if the long-term trend exhibits abnormal fluctuations, potential anomalies are located and the coupling influence strength between multiple variables is analyzed. This coupling influence strength is then converted into an abnormal fluctuation probability, including: Perform a second-order difference operation on the trend change curve to generate a trend change rate sequence; Identify the time coordinate points where the signs flip in the trend change rate sequence and mark them as a set of candidate anomalies; A multivariate state matrix is ​​constructed based on the set of candidate anomalies, and the cross-covariance values ​​between variables in the multivariate state matrix are calculated to quantify the intensity of coupling influence. If the coupling influence strength exceeds a preset coupling strength threshold, the coupling influence strength is mapped to a probability value through probability modeling to obtain the abnormal fluctuation probability.

[0049] It should be noted that, firstly, when performing second-order difference operations on the trend change curve, the difference is calculated point by point according to the time index, which is used to characterize the degree of curvature and acceleration change of the trend curve, and to generate a trend change rate sequence.

[0050] For example, after a slow rise for three consecutive days, a significant inflection point appears in a certain trend curve. The second difference jumps from a level close to 0 to 0.06 near the inflection point and remains there for several time steps. This phenomenon corresponds to the change in the state of the sand supply device from a stable evolution to a stage of accelerated deterioration.

[0051] Next, when identifying the time coordinates of sign reversal in the trend rate of change sequence, sign reversal is defined as the product of the second differences of two adjacent points being less than 0 and both absolute values ​​exceeding the minimum change threshold R, thus eliminating false reversals caused by numerical fluctuations close to 0. The minimum change threshold R is set based on the distribution of the absolute values ​​of the second differences of health data over the past 12 months. High quantiles (such as the 95th percentile) of the absolute values ​​of the second differences under statistically healthy conditions are used as the basic threshold, ensuring that candidate points mainly come from obvious bends rather than random fluctuations. The adjustability of the threshold R is achieved through weekly health data review. If the median of the second differences of the latest week's health data drifts by more than 10 percent, the R value is adjusted proportionally to the new median.

[0052] Subsequently, when constructing the multivariate state matrix based on the candidate outlier set, a long time window H is determined with the candidate outlier as the center. The multidimensional feature vectors within the time window are arranged into sub-matrices in chronological order, and then the sub-matrices of each candidate point are summarized to form the multivariate state matrix.

[0053] In this embodiment, the time window H is set based on the statistical average duration of the impact of confirmed abnormal events. The duration distribution of impact duration is extracted from the abnormal time periods corresponding to maintenance records over the past year, and the duration covering the majority of events is taken as the default window length. The adjustability of the window length H is adjusted in conjunction with the energy fallback time of the most recent month's abnormal segment and the recovery time of the differential sequence. For example, 30 minutes are taken before and after a candidate abnormal point to form a submatrix with a window length H of 60 minutes, containing multi-dimensional features such as vibration dominant frequency, peak pressure, and average temperature.

[0054] It is worth noting that when calculating the cross-covariance values ​​between variables in a multivariable state matrix, each variable is first normalized to zero, and then the covariance is calculated by combining the variables pairwise to obtain the coupling effect strength.

[0055] The coupling strength threshold is set based on the cross-covariance statistics of the health window over the past 12 months. The absolute values ​​of the covariance of each variable combination in the health window are summarized, and the 95th percentile is taken as the coupling strength threshold.

[0056] Finally, when the coupling influence strength exceeds the coupling strength threshold, a logistic regression model is used to convert the coupling influence strength into the probability of abnormal fluctuations through probabilistic modeling. The model training uses historical data from the past 12 months to construct a training set. Training samples are generated in windows: positive samples come from maintenance records and manually confirmed abnormal windows, while negative samples come from healthy windows confirmed by inspections. Stratified sampling maintains a near-balanced ratio of positive to negative samples. Input features include coupling strength indices and their multidimensional statistics. Specifically, the coupling strength index refers to the normalized cross-covariance sequence between pairs of variables (e.g., vibration and pressure, temperature and vibration) calculated within a time window; while the statistics include the peak value (maximum instantaneous interaction force between variables), mean (average level of coupling over time), variance (stability of the coupling relationship), and the first-order slope of the sequence (rate of increase in coupling strength). These features are combined into a high-dimensional vector input to the model, first undergoing min-max normalization to scale to the 0-1 interval. The upper and lower bounds of the normalization are determined by the 1st and 99th quantiles of the training set features. Logistic regression uses the Sigmoid activation function to output probabilities, employs binary cross-entropy as the loss function, and selects the Adam optimization algorithm. The initial learning rate is set to 0.01 and decays to 0.8 every 10 epochs, with a maximum training duration of 200 epochs. Iteration stops when the loss on the validation set decreases by less than 0.0005 for 20 consecutive epochs. The probability threshold is obtained by maximizing the F1 score on the validation set and can be adjusted within the range of 0.55 to 0.75 based on the false positive tolerance. After model pre-training is complete, real-time data is input into the model to output the corresponding probability of abnormal fluctuations.

[0057] For example, after normalizing the coupling strength within a candidate outlier window and inputting it into the model, the output probability of abnormal fluctuation is 0.81.

[0058] In step S17, the step of deducing the fault evolution path based on the abnormal fluctuation probability and defining the signal change range, and calculating the dynamic index of increasing fault probability by combining the abnormal fluctuation probability, the signal change range, and the fluctuation amplitude of the long-term trend, includes: Based on the abnormal fluctuation probability, integrate the equipment operating parameters to construct the initial state vector of fault evolution; Simulate multiple sets of future state transition trajectories of the initial state vector, analyze the statistical distribution of the trajectories, calculate the confidence interval based on the analysis results, and define the range of signal changes. Risk intervals are divided based on the range of signal changes, and the weight of each risk interval is calculated by combining the fluctuation amplitude of the long-term trend. Based on the abnormal fluctuation probability, the risk interval weight, and the distribution density of the trajectory statistical distribution, a dynamic index of increasing failure probability is calculated.

[0059] It should be noted that when deducing the fault evolution path and calculating the dynamic index of increasing fault probability based on the abnormal fluctuation probability, the initial state vector of fault evolution is first constructed by integrating equipment operating parameters based on the abnormal fluctuation probability. The abnormal fluctuation probability is taken from the output of step S16. The equipment operating parameters include multi-dimensional measurements such as the vibration acceleration of the sand supply device, bearing temperature, and sand supply pressure. These parameters are obtained through the preprocessing process of steps S11 to S13 and have been normalized to a dimensionless state.

[0060] The initial state vector is constructed using a concatenation method. First, the normalized values ​​of vibration acceleration, temperature, and pressure are obtained to form a basic state feature group. Next, the abnormal fluctuation probability is multiplied and interactively operated with each normalized value in the basic state feature group to generate a risk enhancement feature group representing the risk intensity. Finally, the basic state feature group and the risk enhancement feature group are concatenated dimensionally to form a composite initial state vector. At this point, the normalization interval for each dimension of the vector is 0 to 1, and the upper and lower bounds of the normalization are set based on statistical data from the past 12 months of healthy operation.

[0061] For example, if the probability of abnormal fluctuation at a certain moment is 0.8, the normalized value of vibration acceleration is 0.6, the normalized value of temperature is 0.5, and the normalized value of pressure is 0.7, then the basic state characteristics are first obtained as [0.6, 0.5, 0.7]. The risk enhancement characteristics are calculated as [0.6×0.8, 0.5×0.8, 0.7×0.8], which is [0.48, 0.40, 0.56]. The final constructed composite initial state vector is [0.6, 0.5, 0.7, 0.48, 0.40, 0.56].

[0062] In this embodiment, a Hidden Markov Model (HMM) is used to simulate multiple future state transition trajectories of the initial state vector. The model training is based on a training set constructed from nearly one year of historical health and fault data. Using the fault occurrence time in the equipment maintenance record as a benchmark, multi-channel data within a 24-hour period is extracted backward as the fault precursor data. A sliding window method (24 steps in length, 1 step size) is used to slice and generate positive sample sequences, labeled as high-risk or medium-risk. Simultaneously, data from periods more than 7 days after the last maintenance and with stable operating conditions are selected and sliced ​​in the same way to generate negative sample sequences, labeled as low-risk. The training set contains 5000 state sequence samples. Stratified sampling maintains a near 1:1 ratio of positive to negative samples to eliminate the impact of class imbalance on model training.

[0063] In this model, the number of hidden states in the Hidden Model (HMM) is set to 3, corresponding to three discrete states: low risk, medium risk, and high risk. The observation layer consists of a continuous multi-dimensional state vector. The Baum-Welch algorithm is used for model training, with a maximum of 100 iterations. The learning rate is initially 0.01, decreasing by 0.1 every 10 iterations. The loss function is negative log-likelihood, and training stops when the loss decreases by less than 0.001 for five consecutive iterations. For trajectory generation, the mapping from continuous observations to discrete hidden states is addressed first. The constructed composite initial state vector is input into the trained model, and the emission probability density of this vector belonging to each hidden state is calculated using a Gaussian mixture emission probability distribution function. The posterior probability is then calculated based on the initial state probability distribution. The hidden state with the highest posterior probability is selected as the starting point of the true hidden state at the current moment. Subsequently, based on this starting point, a sequence of hidden states for the next preset number of steps is generated by sampling according to the state transition probability matrix. The corresponding observation state sequence is then generated using the emission distribution. The step size H is set to 24 steps based on the average fault development time, with each step corresponding to 10 minutes. The number of trajectories K is set to 1000 to cover the main evolution path. For example, after the initial state vector is input into the HMM, 1000 state sequences for the next 4 hours are generated. Each sequence contains 24 state points, and most trajectories show a migration from low risk to high risk.

[0064] Subsequently, when analyzing the statistical distribution of the trajectories and calculating the confidence interval to define the range of signal changes, kernel density estimation was performed on the state values ​​of the K trajectories at each time point. A Gaussian kernel was selected as the kernel function, and the bandwidth was adaptively adjusted according to the standard deviation of the trajectory values. The standard deviation of the state values ​​of all trajectories at the current time point was calculated, and the bandwidth was set to 1.06 times the standard deviation multiplied by the negative fifth power of the number of trajectories (i.e., following Silverman's rule of thumb). This rule dynamically optimizes the smoothing effect based on the dispersion of the predicted data: when the trajectory distribution is compact (small standard deviation), it automatically reduces the bandwidth to retain details; when the trajectory is divergent (large standard deviation), it automatically increases the bandwidth to extract the overall trend. The confidence interval is set at the 95% level, i.e., the 2.5 and 97.5 quantiles of the state values ​​at each time point are used as the upper and lower bounds, forming a corridor for the signal variation range over time. The upper and lower bounds of the interval are set based on historical fault data statistics. The distribution of the state values ​​before the fault window over the past year is statistically analyzed, and the corresponding quantiles are used as the basic boundaries. When the equipment load changes, the boundaries fluctuate proportionally to the median of recent health data. For example, if the statistical analysis of the vibration acceleration variation range in the 24 hours before the fault shows that the 95% interval is 0.3 to 0.9, then the basic boundaries of the signal variation range for this parameter are set as a lower limit of 0.3 and an upper limit of 0.9. If the current healthy median drifts by 10%, the boundaries are adjusted by 10% accordingly.

[0065] It should be noted that when dividing risk intervals based on signal variation range, the signal value range is evenly divided into three risk intervals: low, medium, and high. The interval boundaries are determined by the deviation from the healthy baseline. The low-risk interval is calculated as the range from the lower boundary of the signal variation range to the lower boundary plus 0.3 times the range width; the medium-risk interval is calculated as the range from the lower boundary plus 0.3 to the lower boundary plus 0.7 times the range width; and the high-risk interval is the remaining portion. The weight of each risk interval is calculated based on the fluctuation amplitude of the long-term trend. The fluctuation amplitude is the peak-to-trough difference of the trend curve within the assessment period, normalized to the 0-1 range. The weight is calculated as a weighted sum of the fluctuation amplitude and the interval level. The base weight for the low-risk interval is 0.2, for the medium-risk interval 0.5, and for the high-risk interval 0.8. The fluctuation amplitude weight coefficient is 0.4, and the interval level weight coefficient is 0.6. This combination is based on historical fault backtesting.

[0066] Specifically, the backtracking process involves selecting 100 typical fault samples and 100 healthy samples labeled in the historical database to construct a parameter grid search experiment. The interval level weights are set to vary from 0.1 to 0.9, and the fault warning accuracy and lead time under different weight combinations are calculated. Experimental statistics show that when the interval level weight is 0.6 and the fluctuation amplitude weight is 0.4, the system achieves the lowest false alarm and false negative rates for progressive faults. This result is also explained at the physical level: the interval level directly reflects the current risk level and is the primary concern, thus receiving a higher weight (0.6); while the fluctuation amplitude reflects trend instability and serves as an auxiliary judgment indicator, receiving a lower weight (0.4). The fluctuation amplitude reflects trend instability and receives a lower weight, while the interval level directly relates to the risk level and receives a higher weight.

[0067] For example, if a signal changes from 0.3 to 0.9, then the low-risk range is 0.3 to 0.48, the medium-risk range is 0.48 to 0.72, and the high-risk range is 0.72 to 0.9. If the fluctuation range is 0.6, then the weight of the low-risk range is 0.2×0.6+0.6×0.4=0.36, the medium-risk range is 0.62, and the high-risk range is 0.88.

[0068] Next, when calculating the dynamic indicator of increasing failure probability using weighted averages, the dynamic indicator is a weighted sum of abnormal fluctuation probability, risk interval weight, and trajectory distribution density. The weight for abnormal fluctuation probability is set to 0.5, the risk interval weight to 0.3, and the trajectory distribution density weight to 0.2. This weighting combination is based on statistical analysis of failure warning effectiveness over the past year. Abnormal fluctuation probability, directly reflecting the current probability of anomalies, is given the highest weight; the risk interval weight reflects the importance of spatial distribution; and the trajectory distribution density, representing future uncertainty, is given a lower weight. The trajectory distribution density is obtained by kernel density estimation of the frequency of trajectory occurrence within the risk interval, normalized to the 0-1 range. After weighted calculation, the dynamic indicator ranges from 0 to 1, with a larger value indicating a faster increase in failure probability. For example, with an abnormal fluctuation probability of 0.8, a risk interval weight of 0.62, and a trajectory distribution density of 0.7, the dynamic indicator is 0.8 × 0.5 + 0.62 × 0.3 + 0.7 × 0.2 = 0.726.

[0069] In step S18, comparing the dynamic indicator with a preset risk assessment threshold, classifying the fault risk according to the extent to which the dynamic indicator exceeds the risk assessment threshold, and fusing the classification assessment result with the changing trend of the dynamic indicator to generate a precise fault early warning signal includes: The dynamic indicators are compared with preset risk assessment thresholds, and the risk levels are divided according to the extent of exceedance to obtain the graded assessment results. Calculate the slope of the dynamic index over time to determine the rate of increase in the failure probability; By integrating the graded assessment results with the failure probability growth rate, a comprehensive early warning vector is constructed. The comprehensive early warning vector is analyzed to identify the fault-related components and their development trends, thereby generating accurate fault early warning signals.

[0070] It should be noted that when comparing dynamic indicators with preset risk assessment thresholds, the risk assessment thresholds are statistically set based on the past 12 months of healthy operation data and confirmed fault data. First, the distribution of dynamic indicator values ​​under healthy operation conditions is collected, and its 95th percentile is calculated as the base threshold. Simultaneously, the lowest value of the dynamic indicator within the window before the fault occurred is statistically analyzed to ensure that the threshold can effectively distinguish between normal fluctuations and abnormal states. Threshold adjustability is achieved through a weekly health data review mechanism. If the median of the dynamic indicator in the latest week's health data drifts by more than 5%, the threshold is adjusted proportionally to maintain the stability of risk assessment.

[0071] It should be noted that the risk level is determined by calculating the relative proportion of dynamic indicators exceeding the risk assessment threshold. If the relative proportion is between 0% and 20%, it is considered low risk, indicating that the fault is in its early stages; if it is between 20% and 50%, it is considered medium risk, indicating that the fault characteristics are significant but have not yet disrupted operation; if it exceeds 50%, it is considered high risk.

[0072] In this embodiment, a first-order difference algorithm is used to calculate the slope of the dynamic index over time. Based on the time points of the dynamic index sequence, the difference between index values ​​at adjacent time points is calculated and divided by the time interval to obtain the slope sequence. To suppress the impact of short-term fluctuations, a sliding window median smoothing process is applied to the slope sequence. The window length is set according to the data update frequency. If the dynamic index is updated every 10 minutes, the window length is set to 5 points. The median of the slope within the window is taken as the final slope value. The slope calculation result is normalized to the interval between -1 and +1, where negative values ​​indicate a decreasing probability of failure and positive values ​​indicate an increasing probability.

[0073] Specifically, the distribution of the absolute values ​​of the slope of this dynamic indicator's change over the past 12 months of historical data is analyzed, and the 99th percentile is selected as the maximum historical rate of change. For example, the maximum rate is calculated to be 0.1 / min. The normalized slope is the real-time slope divided by the maximum historical rate of change. If the calculated result exceeds the range of [-1, 1], it is truncated. This method ensures that the normalized value objectively reflects the severity of the current rate of change relative to the historical limit. Negative values ​​indicate a decreasing (mitigation) probability of failure, while positive values ​​indicate an increasing (deterioration) probability. For example, if a dynamic indicator rises from 0.60 to 0.72 within one hour (10-minute interval), the slope is calculated as (0.72 - 0.60) ÷ 6 = 0.02, and the normalized slope is 0.2.

[0074] It should be noted that when integrating the grading assessment results with the failure probability growth rate, the grading results are mapped to numerical form: 0.3 for low risk, 0.6 for medium risk, and 1.0 for high risk. The failure probability growth rate, i.e., the absolute value of the slope, is normalized to the range of 0 to 1. A weighted summation method is used to construct a comprehensive early warning vector. The weight of the grading result is set to 0.6, and the weight of the slope is set to 0.4. This weight combination is determined based on historical early warning effect statistics. The grading result, which directly reflects the current risk level, is given a higher weight, while the slope, which provides trend change information, is given a lower weight. For example, if the grading result is medium risk (0.6) and the normalized slope value is 0.7, then the comprehensive early warning vector is calculated as 0.6 × 0.6 + 0.7 × 0.4 = 0.64.

[0075] Finally, when analyzing the comprehensive early warning vector, the mapping rules are based on the historical fault case library. The rule library is constructed from the correlation between faulty components and dynamic indicators in the maintenance records of the past year.

[0076] Specifically, this rule base performs statistical distribution analysis on the comprehensive warning vector values ​​at the time of occurrence of different fault types in historical data. First, all fault samples confirmed as bearing wear are extracted from historical records, and their comprehensive warning vector values ​​within the 24 hours prior to the fault occurrence are traced back to form a bearing wear characteristic distribution cluster. Statistical analysis reveals that the vector values ​​of this cluster are mainly concentrated between 0.62 and 0.78 (corresponding to medium risk with a continuously increasing trend). Second, to achieve early warning and cover the vast majority of fault samples, the quintile (lower boundary) of this distribution cluster is selected as the judgment threshold. In this example, the calculated lower boundary value is approximately 0.60. Therefore, 0.60 is set as the trigger threshold for bearing wear. Similarly, for sudden and severe faults such as motor burnout, the comprehensive warning vector values ​​of historical samples are usually extremely high (concentrated above 0.85), so their threshold is set to 0.80.

[0077] For example, if the comprehensive warning vector is 0.64 and the slope is positive, exceeding the bearing wear threshold of 0.6, a warning signal will be generated to indicate that the bearing wear is accelerating and it is recommended to arrange an inspection within 24 hours. If the vector exceeds 0.8, it will be upgraded to a motor failure risk and it is recommended to stop the machine immediately.

[0078] In summary, this invention discloses an adaptive sand supply device fault diagnosis method, comprising: acquiring multi-channel raw data of vibration, temperature, and pressure from the sand supply device; reconstructing the signal sequence after time synchronization and frequency band processing; obtaining a smoothed dataset through multi-scale decomposition and denoising; extracting time-domain and frequency-domain features to construct a feature vector set; comparing and analyzing long-term trends with historical feature data; locating anomalies and calculating the coupling influence strength and abnormal fluctuation probability; deducing the fault evolution path and defining the signal change range; calculating a dynamic index of increasing fault probability; comparing and classifying risk thresholds; fusing the assessment results with the index change trend; and generating an accurate fault early warning signal. This method can distinguish between effective normal fluctuations and real fault signals in the sand supply device, meeting the needs of intelligent operation and maintenance of sand supply devices in industrial production empowered by IoT chips.

[0079] Reference Figure 2 The second embodiment of the present invention provides an adaptive sand supply device fault diagnosis system, comprising: The data acquisition module is used to acquire multi-channel raw data of the sand supply device operation, including vibration data, temperature data and pressure data. The signal processing module is used to perform time synchronization and frequency band processing on the multi-channel raw data, extract time-domain signal segments that exceed a preset reference energy threshold, and reassemble the time-domain signal segments to obtain a reassembled signal sequence. The denoising module is used to perform multi-scale decomposition and denoising on the recombined signal sequence to obtain a smooth dataset. The feature extraction module is used to extract time-domain and frequency-domain features from the smoothed dataset and construct a multi-dimensional feature vector set. The trend judgment module is used to compare the feature vector set with the pre-acquired historical feature data in the time dimension to analyze the long-term trend of feature changes. An anomaly analysis module is used to locate potential anomaly points and analyze the coupling influence strength between multiple variables if the long-term trend shows abnormal fluctuations, and to convert the coupling influence strength into anomaly fluctuation probability. The indicator calculation module is used to deduce the fault evolution path based on the abnormal fluctuation probability, define the signal change range, and calculate the dynamic indicator of increasing fault probability by combining the abnormal fluctuation probability, the signal change range, and the fluctuation amplitude of the long-term change trend. The early warning generation module is used to compare the dynamic indicators with preset risk judgment thresholds, classify and assess the fault risk according to the extent to which the dynamic indicators exceed the risk judgment thresholds, and integrate the classification assessment results with the changing trend of the dynamic indicators to generate accurate fault early warning signals.

[0080] It should be noted that the adaptive sand supply device fault diagnosis system provided in this embodiment of the invention is used to execute all the process steps of the adaptive sand supply device fault diagnosis method in the above embodiment. The working principle and beneficial effects of the two are one-to-one, so they will not be described again.

[0081] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0082] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A fault diagnosis method for an adaptive sand supply device, characterized in that, include: Acquire multi-channel raw data of the sand supply device operation, including vibration data, temperature data and pressure data; The multi-channel raw data is time-synchronized and frequency-segmented, and time-domain signal segments exceeding a preset reference energy threshold are extracted. The time-domain signal segments are then reassembled to obtain a reassembled signal sequence. The recombined signal sequence is subjected to multi-scale decomposition and denoising to obtain a smoothed dataset; Extract time-domain and frequency-domain features from the smoothed dataset to construct a multi-dimensional feature vector set; The feature vector set is compared with the pre-acquired historical feature data in the time dimension to analyze the long-term trend of feature changes. If the long-term trend shows abnormal fluctuations, then potential anomalies are located and the coupling influence strength between multiple variables is analyzed, and the coupling influence strength is converted into the probability of abnormal fluctuations. Based on the abnormal fluctuation probability, the fault evolution path is deduced, and the signal change range is defined. Combining the abnormal fluctuation probability, the signal change range, and the fluctuation amplitude of the long-term change trend, a dynamic index of increasing fault probability is calculated. The dynamic indicators are compared with preset risk assessment thresholds. The fault risk is graded and assessed based on the extent to which the dynamic indicators exceed the risk assessment thresholds. The graded assessment results and the changing trends of the dynamic indicators are then integrated to generate accurate fault early warning signals.

2. The fault diagnosis method for the adaptive sand supply device according to claim 1, characterized in that, Acquire multi-channel raw data of the sand supply device operation, including vibration data, temperature data, and pressure data, including: A sensor array is deployed to collect vibration, temperature, and pressure data during the operation of the sand supply device, resulting in multi-channel raw data.

3. The fault diagnosis method for the adaptive sand supply device according to claim 1, characterized in that, The process involves time synchronization and frequency segmentation of the multi-channel raw data, extracting time-domain signal segments exceeding a preset reference energy threshold, and reassembling these time-domain signal segments to obtain a reassembled signal sequence, including: The vibration data, temperature data, and pressure data are timestamped to generate a time-synchronized data stream. The time-synchronized data stream is subjected to spectral analysis to decompose it into a set of sub-signals in multiple frequency bands; If the instantaneous energy value of any frequency band in the sub-signal set exceeds the preset reference energy threshold, then the corresponding time point is locked and a time domain signal segment is extracted. The time-domain signal segments are vectorized and recombined in chronological order to obtain a recombined signal sequence.

4. The fault diagnosis method for the adaptive sand supply device according to claim 1, characterized in that, The process of performing multi-scale decomposition and denoising on the reconstructed signal sequence to obtain a smoothed dataset includes: The recombined signal sequence is decomposed into approximation coefficients and detail coefficients by multi-scale decomposition. By combining the approximation coefficients and the detail coefficients, signal denoising and reconstruction are performed to generate a denoised signal sequence; Calculate the first-order difference variance of the denoised signal sequence, select the continuous sequence segments with the smallest variance, and form a smoothed dataset.

5. The fault diagnosis method for the adaptive sand supply device according to claim 1, characterized in that, The step of extracting time-domain and frequency-domain features from the smoothed dataset to construct a multi-dimensional feature vector set includes: Calculate the mean, variance, peak value, and kurtosis of the smoothed dataset to form a time-domain feature set; The smoothed dataset is subjected to spectral transformation to extract the main frequency, power spectrum peak value and frequency band energy, forming a frequency domain feature set; The time-domain feature set and the frequency-domain feature set are concatenated according to their dimensions to construct a multi-dimensional feature vector set.

6. The fault diagnosis method for the adaptive sand supply device according to claim 1, characterized in that, The step of comparing the feature vector set with pre-acquired historical feature data over a time dimension to analyze the long-term trend of feature changes includes: Construct a historical data window matrix based on pre-acquired historical feature data; Calculate the difference between the feature vector set and the corresponding dimension in the historical data window matrix to generate a dimension difference sequence; The dimensional difference sequence is smoothed to filter out short-term random fluctuations, resulting in a stationary difference sequence. A trend change curve is generated by fitting the stationary difference sequence; The fluctuation amplitude of the trend change curve is calculated. If the fluctuation amplitude exceeds a preset trend fluctuation threshold, the long-term trend of the feature is determined to exhibit abnormal fluctuation.

7. The fault diagnosis method for the adaptive sand supply device according to claim 6, characterized in that, If the long-term trend exhibits abnormal fluctuations, then potential anomalies are located and the coupling influence strength between multiple variables is analyzed. This coupling influence strength is then converted into an abnormal fluctuation probability, including: Perform a second-order difference operation on the trend change curve to generate a trend change rate sequence; Identify the time coordinate points where the signs flip in the trend change rate sequence and mark them as a set of candidate anomalies; A multivariate state matrix is ​​constructed based on the set of candidate anomalies, and the cross-covariance values ​​between variables in the multivariate state matrix are calculated to quantify the intensity of coupling influence. If the coupling influence strength exceeds a preset coupling strength threshold, the coupling influence strength is mapped to a probability value through probability modeling to obtain the abnormal fluctuation probability.

8. The fault diagnosis method for the adaptive sand supply device according to claim 1, characterized in that, The step of deducing the fault evolution path based on the abnormal fluctuation probability, defining the signal change range, and calculating a dynamic index of increasing fault probability by combining the abnormal fluctuation probability, the signal change range, and the fluctuation amplitude of the long-term trend, includes: Based on the abnormal fluctuation probability, integrate the equipment operating parameters to construct the initial state vector of fault evolution; Simulate multiple sets of future state transition trajectories of the initial state vector, analyze the statistical distribution of the trajectories, calculate the confidence interval based on the analysis results, and define the range of signal changes. Risk intervals are divided based on the range of signal changes, and the weight of each risk interval is calculated by combining the fluctuation amplitude of the long-term trend. Based on the abnormal fluctuation probability, the risk interval weight, and the distribution density of the trajectory statistical distribution, a dynamic index of increasing failure probability is calculated.

9. The fault diagnosis method for the adaptive sand supply device according to claim 1, characterized in that, The step of comparing the dynamic indicator with a preset risk assessment threshold, classifying the fault risk according to the extent to which the dynamic indicator exceeds the risk assessment threshold, and fusing the classification assessment result with the changing trend of the dynamic indicator to generate a precise fault early warning signal includes: The dynamic indicators are compared with preset risk assessment thresholds, and the risk levels are divided according to the extent of exceedance to obtain the graded assessment results. Calculate the slope of the dynamic index over time to determine the rate of increase in the failure probability; By integrating the graded assessment results with the failure probability growth rate, a comprehensive early warning vector is constructed. The comprehensive early warning vector is analyzed to identify the fault-related components and their development trends, thereby generating accurate fault early warning signals.

10. A fault diagnosis system for an adaptive sand supply device, characterized in that, include: The data acquisition module is used to acquire multi-channel raw data of the sand supply device operation, including vibration data, temperature data and pressure data. The signal processing module is used to perform time synchronization and frequency band processing on the multi-channel raw data, extract time-domain signal segments that exceed a preset reference energy threshold, and reassemble the time-domain signal segments to obtain a reassembled signal sequence. The denoising module is used to perform multi-scale decomposition and denoising on the recombined signal sequence to obtain a smooth dataset. The feature extraction module is used to extract time-domain and frequency-domain features from the smoothed dataset and construct a multi-dimensional feature vector set. The trend judgment module is used to compare the feature vector set with the pre-acquired historical feature data in the time dimension to analyze the long-term trend of feature changes. An anomaly analysis module is used to locate potential anomaly points and analyze the coupling influence strength between multiple variables if the long-term trend shows abnormal fluctuations, and to convert the coupling influence strength into anomaly fluctuation probability. The indicator calculation module is used to deduce the fault evolution path based on the abnormal fluctuation probability, define the signal change range, and calculate the dynamic indicator of increasing fault probability by combining the abnormal fluctuation probability, the signal change range, and the fluctuation amplitude of the long-term change trend. The early warning generation module is used to compare the dynamic indicators with preset risk judgment thresholds, classify and assess the fault risk according to the extent to which the dynamic indicators exceed the risk judgment thresholds, and integrate the classification assessment results with the changing trend of the dynamic indicators to generate accurate fault early warning signals.

Images