Flexible temperature and pressure sensing array detection method for wear of roller skin of belt conveyor

Abstract

The invention relates to a flexible temperature and pressure sensing array detection method for belt conveyor roller skin wear in the cross technical field of mechanical engineering, sensor technology and intelligent algorithm combination. The method comprises the following steps: synchronously calibrating parameters of a sensing array and a conveying system; dynamic temperature and pressure data are collected in real time; carrying out multi-source interference adaptive processing; reconstructing the surface temperature and pressure cloud picture of the carrier roller; wear fault differentiation feature extraction; fault type identification based on feature matching; carrying out fault positioning and real-time early warning output; and dynamically optimizing and updating model parameters. According to the invention, a special algorithm model is constructed based on the existing flexible temperature and pressure sensing array hardware basis, accurate detection of the wear fault of the roller skin is realized, and key technical support is provided for intelligent maintenance of mining equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical engineering, sensor technology and intelligent algorithms, in particular to a flexible temperature and pressure sensing array detection method for belt conveyor roller cylinder skin wear. BACKGROUND

[0002] Roller wear is mainly divided into two types: circumferential wear and axial wear. Circumferential wear usually manifests as the complete loss of the outer skin of the roller, which is caused by long-term friction between the roller and the conveyor belt. The outer side bears the maximum frictional force, and under the combined action of high load and harsh environment, the skin gradually wears out until it is lost. Axial wear is the abnormal wear of a certain area along the axis, which may be caused by uneven material distribution, resulting in excessive local stress on the roller. Both types of wear change the contact state between the roller and the conveyor belt, breaking the original uniform friction force distribution and causing abnormal friction. Abnormal friction can rapidly increase the temperature of the roller and the conveyor belt, accelerating the damage of the equipment. When the wear is severe, the conveyor belt may tear due to the inability to withstand the abnormal tension, or the roller may be stuck, causing the equipment to stop. These situations not only cause production interruptions and increase maintenance costs, but also may cause coal accumulation, affecting the production progress of the entire mine and causing significant economic losses to the mine.

[0003] In terms of hardware detection methods, existing temperature and pressure sensors mostly use single-point measurement. This method can only obtain the temperature or pressure data at a single point on the roller surface, and cannot comprehensively reflect the temperature and pressure distribution on the roller surface, making it difficult to build a complete wear fault feature model. Because roller wear is a complex process, the wear degree and characteristics of different parts may be different, and single-point measurement cannot cover these information. Machine vision-based detection methods also face many challenges in coal mines. The light conditions are poor underground, and additional lighting equipment is needed, but even so, uneven light and dust obstruction can still affect image acquisition quality. Moreover, machine vision has insufficient accuracy in quantitative analysis of cylinder skin wear, making it difficult to accurately measure key parameters such as wear depth and area, and cannot meet the demand for accurate detection of roller wear in actual production. SUMMARY

[0004] The present application aims to solve the core problems in existing coal mine belt conveyor roller wear fault detection technology, specifically including the following aspects:

[0005] Accurate extraction of wear failure features: The circumferential wear and axial wear of the roller show different features on the temperature and pressure cloud map, but these features are intertwined in the actual dynamic change cloud map, and it is difficult to clearly separate them. Accurate extraction of the "outside-inside area difference" feature of the circumferential wear from the complex cloud map requires accurate analysis of the change rule of temperature and pressure values in the circumferential direction of the roller to distinguish the boundary between the normal area and the wear area; for axial wear, accurate identification of the "axial local area anomaly" feature requires detailed gradient analysis and pattern matching of temperature and pressure data along the axial direction to determine the location and range of the abnormal area. However, existing algorithms often fail to accurately extract features when dealing with these complex feature extraction tasks due to the dynamic nature of the cloud data and noise interference, making it impossible to accurately identify the two types of wear.

[0006] Noise interference problem under complex working conditions: The coal mine underground is a complex industrial environment with many interference factors. Strong vibration can cause additional stress on the sensor array, causing fluctuations in the collected temperature and pressure data; electromagnetic interference can affect the signal transmission of the sensor, causing data distortion; material impact can instantaneously change the stress and temperature distribution on the surface of the roller, producing abnormal sensor data. These noise interferences can seriously affect the quality of the sensor data and reduce the accuracy of fault detection. Existing signal processing algorithms lack robustness when dealing with such complex noise interference, and cannot effectively suppress noise, so in actual application, noise signals are often misjudged as fault signals, or the real fault signals are missed, resulting in a significant reduction in the reliability of fault detection.

[0007] Real-time online detection engineering needs: With the continuous expansion of coal mining scale, the transportation distance of belt conveyors is getting longer and longer, and the transportation volume is also getting larger and larger. In this case, real-time online detection of roller wear has higher requirements. Most existing detection methods cannot meet the real-time monitoring needs of long-distance and large-volume operation, and cannot timely detect early wear failure of the roller. When the roller shows early wear, existing methods cannot capture these subtle changes and cannot timely issue a warning, leading to gradual development of the fault and eventually causing serious equipment damage and production accidents. Moreover, accurately and quickly locating the position of the faulty roller on a long-distance conveying line is also a difficult problem, and existing methods cannot accurately locate the faulty roller, causing great difficulty in equipment maintenance.

[0008] To solve the above problems, the present application provides a belt conveyor roller cylinder skin wear flexible temperature and pressure sensing array detection method, comprising the following steps:

[0009] S1 Synchronization calibration of sensor array and conveying system parameters: By measuring the data of the sensor array at different positions of the roller multiple times, the optimal value is calculated using the least squares method, thereby realizing accurate coordinate mapping;

[0010] S2 Real-time collection of dynamic temperature and pressure data: According to the continuous contact scanning of the flexible sensor array on the surface of the roller, the temperature and pressure changes on the surface of the roller are sensed, and the specific position of each data point on the surface of the roller is determined through coordinate mapping;

[0011] S3 Adaptive processing of multi-source interference: A two-stage adaptive noise reduction model is constructed to eliminate different types of interference and retain key wear characteristics;

[0012] S4 Reconstruction of roller surface temperature and pressure cloud map: A two-dimensional grid matrix is constructed, and the missing data in the gaps between the sensing units is processed using a bilinear interpolation algorithm. At the same time, the temperature cloud map and the pressure cloud map are fused to generate a fusion cloud map that combines temperature and pressure information;

[0013] S5 Extraction of differentiated features of wear faults: By reasonably dividing the fusion cloud map into different regions, the temperature, pressure statistical features and gradient change rate of different regions are calculated, thereby extracting quantified feature parameters that can effectively represent wear faults;

[0014] S6 Fault type recognition based on feature matching: An improved K-nearest neighbor-based fault recognition model is introduced, a rich and accurate feature sample library is constructed, and a scientific and reasonable distance measurement and classification decision mechanism is adopted to realize accurate identification of the type of roller wear faults;

[0015] S7 Fault positioning and real-time warning output: Combined with the roller number and coordinate mapping relationship marked during data collection, accurate positioning of the fault roller on the conveying line is realized, the wear condition of the roller is accurately and comprehensively evaluated through the wear severity grading based on fuzzy logic algorithm, and real-time warning is realized;

[0016] S8 Dynamic optimization and updating of model parameters: A model parameter adaptive updating mechanism is established, and combined with regional special calibration, the fault recognition model continuously adapts to the complex and variable underground environment.

[0017] In order to ensure the accurate synchronization of the flexible temperature and pressure sensor array with the key parameters of the conveying machine and realize accurate monitoring of the roller state, the step S1 specifically includes the following steps:

[0018] The linear speed v of the belt is obtained in real time by the speed sensor installed on the driving drum, and combined with the known shaft length L and circumference C of the roller, a mathematical model of the data collection frequency f is established:

[0019]

[0020] wherein D is the diameter of the roller;

[0021] The time period for the sensor array to scan a full roller is recorded Ensure the time interval of single group data collection Synchronized with the circumferential displacement of the roller surface;

[0022] Define the mapping relationship between the physical coordinates of the sensor array (x p ,y p ) and the axial coordinate x∈[0,L] and the circumferential coordinate y∈0,C] as follows:

[0023] x=x p +δ x ,y=v·t+δ y

[0024] wherein δ x , δ y are installation offset errors, which are fitted and calibrated by the least squares method to eliminate the position mapping deviation.

[0025] In order to ensure that the temperature and pressure changes on the surface of the roller can be accurately perceived, the step S2 specifically comprises the following steps:

[0026] When the sensor array is driven by the belt to pass through the roller area, each sensor unit in the array synchronously collects the temperature value T i (t) and the pressure value P i (t) of the corresponding contact point at a frequency f, wherein i represents the i-th sensor unit, thereby forming a time series:

[0027]

[0028] wherein M is the number of axial sensor units, N is the number of circumferential sensor units, x j ,y k are calibrated spatial coordinates, and through accurate coordinate mapping, the data traceability accuracy is ensured to reach the millimeter level.

[0029] In order to effectively improve the data quality, a two-level adaptive noise reduction model is constructed to eliminate different types of interference and retain key wear characteristics, and the step S3 specifically comprises the following steps:

[0030] S3.1 The first level adopts an outlier rejection algorithm based on a sliding window, and through statistical analysis of a large amount of temperature and pressure data collected by the standard roller under normal operating conditions, the normal fluctuation range of the data is determined;

[0031] Assuming that the mean of the temperature data is μ T , the standard deviation is σ T , and the mean of the pressure data is μP , standard deviation is σ P , according to the 3σ rule in statistics, set the normal fluctuation range of temperature as μ T -3σ T , μ T +3σ T ], the normal fluctuation range of pressure is [μ P -3σ P , μ P +3σ P ];

[0032] When the temperature value T(n) or the pressure value P(n) in the collected real-time data exceeds the corresponding range, it is determined as an abnormal point; for the abnormal point T(n), the neighborhood mean filling method is used for processing, and the formula is as follows:

[0033]

[0034] Wherein, is the filled temperature value, K is the sliding window half-width, and the formula indicates that the average value of the K data points before and after the abnormal point T(n) is taken as the center is calculated as the filling value;

[0035] S3.2 The second stage uses an adaptive wavelet filtering algorithm to perform multi-layer wavelet decomposition on the signal S(n) after the first stage of noise reduction, and decomposes it into high-frequency noise coefficients W H and low-frequency signal coefficients W L ; By analyzing the frequency domain characteristics of the data, the adaptive threshold function λ(f) = α·std(W H (f)) is determined, where α is a noise sensitivity coefficient, and std(W H (f)) is the standard deviation of the high-frequency noise coefficients at frequency f, and its optimal value is determined through training and analysis of a large amount of historical data;

[0036] According to the adaptive threshold function λ(f), the high-frequency noise coefficients W H are processed to filter out high-frequency interference components; for each coefficient in the high-frequency noise coefficients W H , if its absolute value is less than λ(f), it is set to 0, considering that the coefficient is caused by noise; if its absolute value is greater than λ(f), the coefficient is retained, considering that it contains useful signal information; After threshold processing, the high-frequency noise coefficients and the low-frequency signal coefficients are reconstructed through wavelet inverse transform to obtain the filtered signal The formula is as follows:

[0037]

[0038] Wherein, Thresh(W H, λ(f)) represents the high-frequency noise coefficient W H the result after threshold processing.

[0039] In order to more accurately identify the fault feature and improve the accuracy and reliability of fault diagnosis, the step S4 specifically comprises the following steps:

[0040] S4.1 constructs a MxN grid matrix with the roller axial X∈[0, L] and Y∈[0, C], and each node coordinate is Wherein m=0, 1, …, M-1, n=0, 1, …, N-1, wherein L is the roller shaft length, C is the roller circumference, M and N respectively represent the number of grids divided in the axial and circumferential directions, in this way, the roller surface is discretized into a series of grid nodes, each node corresponds to a specific position on the roller surface; the sensor unit data (T j,k , P j,k ) is mapped to the grid node, wherein T j,k represents the temperature value collected by the jth axial sensor unit and the kth circumferential sensor unit, and P j,k represents the corresponding pressure value;

[0041] For the gap missing data, a bilinear interpolation method is adopted, assuming that the four vertex coordinates around the interpolation point P(x, y) are Q 11 (x1, y1), Q 12 (x1, y2), Q 21 (x2, y1), and Q 22 (x2, y2), and the corresponding temperature values are T 11 , T 12 , T 21 , and T 22 , then the temperature cloud map T(x, y) of the interpolation point is calculated by the following formula:

[0042] Similarly, the same method is used to obtain the pressure cloud map P(x, y);

[0043] S4.2 After obtaining the temperature cloud map T(x, y) and the pressure cloud map P(x, y), a channel fusion algorithm is used to fuse them into a three-dimensional matrix F(x, y)=[T(x, y); P(x, y)], generating a fusion cloud map with temperature and pressure information.

[0044] In order to accurately identify the wear failure type of the roller, it is necessary to design a partition feature extraction strategy according to the different characteristics of the circumferential wear and the axial wear, and to provide accurate data support for subsequent fault identification, and the step S5 specifically comprises the following steps:

[0045] The fused cloud image is divided into an outer side y∈[0, 0.2C]∪[0.8C, C], a middle y∈[0.4C, 0.6C], and an inner side y∈[0.2C, 0.4C]∪[0.6C, 0.8C] in the circumferential direction, wherein C is the length of the circumference of the roller;

[0046] The fused cloud image is divided into a left x∈[0, 0.3L], a middle x∈[0.3L, 0.7L], and a right x∈[0.7L, L] in the axial direction, wherein L is the length of the axis of the roller;

[0047] For circumferential wear feature extraction, the mean and variance differences of the temperature and pressure of the outer side region and the inner side region are calculated; the mean μ of the region is calculated according to the following formula:

[0048]

[0049] wherein z i is the temperature or pressure value of the i-th grid node in the region, and N is the total number of grid nodes in the region; through the formula, the mean μ 外侧 of the outer side region and the mean μ 内侧 of the inner side region can be obtained, and then the mean difference Δμ = |μ 外侧 - μ 内侧 | is calculated.

[0050] The variance σ 2 is calculated according to the following formula:

[0051]

[0052] The variance σ of the outer side region and the variance σ of the inner side region are calculated through the formula, and then the variance ratio σ

[0053] By setting a difference threshold, it is determined whether there is a feature that the values of the outer side region and the inner side region are significantly different.

[0054] For axial wear feature extraction, the gradient change rate of each strip-shaped region in the axial direction is calculated; the gradient change rate G is used to measure the change rate of the temperature T and the pressure P in the axial direction x, and the calculation formula is as follows:

[0055]

[0056] By calculating the gradient change rate of each strip-shaped region, the region of gradient mutation is located, and it is determined whether there is a feature that the values of the local region in the axial direction are abnormal.

[0057] At the same time, three quantitative feature parameters of the wear region, including the area A, the mean difference Δμ, and the gradient peak G max are extracted, and a feature vector f = [A, Δμ, G max ] is formed.

[0058] In order to realize the accurate identification of the carrier roller wear fault type, and provide strong guarantee for timely maintenance and safety production of the equipment, the step S6 specifically includes the following steps:

[0059] A large number of labeled normal and two types of wear fault carrier roller cloud image feature vectors are used as training samples to construct a feature sample library Where f i is the feature vector of the i-th sample, containing quantitative feature parameters such as wear area, mean difference degree, gradient peak value, l i is the corresponding fault label;

[0060] In the identification process, for the feature vector f q of the to-be-detected carrier roller, the distance weighted KNN algorithm is used to calculate the distance d(f q ,f i ) between it and each feature vector f i in the sample library, and the formula is:

[0061]

[0062] Where w k is the weight of the k-th feature dimension, which is determined through analysis and normalization processing of the training data to reflect the importance of different feature dimensions to fault identification, and D is the dimension of the feature vector;

[0063] By calculating the distance, the K nearest samples are selected, and a weighted voting mechanism is used to determine the fault type of the to-be-detected carrier roller The formula is as follows:

[0064]

[0065] Where δ(l i , l) is the Kronecker function.

[0066] In order to realize the accurate positioning of the fault carrier roller on the conveying line, in the step S7, the fault positioning specifically includes the following steps:

[0067] The carrier roller number is the unique identification of each carrier roller in the belt conveyor system, which is closely related to the conveying line coordinate mapping table Map(ID) = (x pos ,y pos ) ID(x pos ,y pos )x pos y pos ,

[0068] Where ID is the carrier roller number, (x pos ,y pos) is the two-dimensional coordinate of the roller on the conveying line, x pos is the position coordinate along the direction of the conveying belt, y pos is the position coordinate perpendicular to the direction of the conveying belt.

[0069] To ensure that the fault identification model continuously maintains high identification accuracy in the complex and changeable downhole environment, the step 8 specifically comprises the following steps:

[0070] A sample feedback closed loop is established, and the fault data confirmed by the ground monitoring center is fed back to the edge computing node, and model parameter updating is triggered every time a certain group of new samples is accumulated;

[0071] In the parameter updating process, the gradient descent algorithm is used to adjust the feature weight vector w, and the formula is:

[0072]

[0073] Where w is the feature weight vector, η is the learning rate, is the gradient of the loss function J(w) with respect to the old weight vector wold.

[0074] The loss function J(w) adopts the cross-entropy function, and its definition is:

[0075]

[0076] Where n is the number of samples, y i is the true label of sample i, is the label probability of the model predicting sample i;

[0077] When the area recognition accuracy Acc(r) is less than 85%, special calibration is started, wherein Acc(r) represents the recognition accuracy of the model in the area r;

[0078] The special calibration adopts the K-means algorithm to re-cluster the feature vectors in the area, and the formula is:

[0079] Calibrate(r) = K-means(f i | Region(i) = r)

[0080] Where f i | Region(i) = r represents the set of all feature vectors in the area r.

[0081] To sum up, compared with the prior art, the present application has the following technical advantages and beneficial effects: based on the existing flexible temperature and pressure sensing array hardware basis, a special algorithm model is constructed to realize accurate detection of the roller cylinder skin wear fault, and key technical support is provided for intelligent maintenance of mine equipment. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 is a flow chart of the belt conveyor idler drum skin wear flexible temperature and pressure sensing array detection method of the present application;

[0083] Figure 2 is a performance comparison table of the present application and prior art;

[0084] Figure 3 is a sample size change curve of fault recognition accuracy of the present application;

[0085] Figure 4 is a wear severity grading confusion matrix of the present application. DETAILED DESCRIPTION

[0086] The preferred embodiments of the present application are described in detail below with reference to the accompanying drawings, so that the advantages and features of the present application can be more easily understood by those skilled in the art, and the scope of protection of the present application is more clearly defined.

[0087] As shown in a belt conveyor idler drum skin wear flexible temperature and pressure sensing array detection method, comprising the following steps: Figure 1

[0088] S1 Sensing array and conveying system parameter synchronous calibration: through multiple measurements of the data of the sensing array at different positions of the idler, the optimal value is calculated by using the least square method, so as to realize accurate coordinate mapping;

[0089] S2 Real-time acquisition of dynamic temperature and pressure data: according to the continuous contact scanning of the flexible sensing array on the surface of the idler, the temperature and pressure changes of the surface of the idler are sensed, and the specific position of each data point on the surface of the idler is determined through coordinate mapping;

[0090] S3 Multi-source interference adaptive processing: a two-stage adaptive noise reduction model is constructed, different types of interference are eliminated, and key wear features are retained;

[0091] S4 Reconstruction of idler surface temperature and pressure cloud map: a two-dimensional grid matrix is constructed, and the missing data of the sensing unit gap is processed by using the bilinear interpolation algorithm, and the temperature cloud map and the pressure cloud map are channel fused to generate a fusion cloud map with temperature and pressure information;

[0092] S5 Wear fault differential feature extraction: through reasonable regional division of the fusion cloud map, the temperature, pressure statistical features and gradient change rate of different regions are calculated, so as to extract the quantitative feature parameters which can effectively represent the wear fault;

[0093] ​S6 Fault type identification based on feature matching: An improved K-Nearest Neighbor based fault identification model is introduced. By constructing a rich and accurate feature sample library, and using a scientific and reasonable distance measurement and classification decision mechanism, accurate identification of the fault type of the roller is realized.

[0094] S7 Fault location and real-time warning output: Combined with the marked roller number and coordinate mapping relationship during data collection, accurate positioning of the fault roller on the conveying line is realized. Through the wear severity grading based on fuzzy logic algorithm, the wear condition of the roller is accurately and comprehensively evaluated, and real-time warning is realized.

[0095] S8 Dynamic optimization and updating of model parameters: An adaptive updating mechanism of model parameters is established, combined with regional special calibration, so that the fault identification model can continuously adapt to the complex underground environment.

[0096] The present application will be described in detail below.

[0097] S1 Synchronization calibration of sensor array and conveying system parameters

[0098] In the complex operating environment of the coal mine belt conveyor, ensuring the accurate synchronization matching of the flexible temperature and pressure sensor array and the key parameters of the conveying system is the basis for accurate monitoring of the roller state. This process not only concerns the accuracy of data collection, but also directly affects the reliability of subsequent fault diagnosis.

[0099] Data acquisition frequency calibration and spatial coordinate mapping: The linear speed of the belt v (unit: m / s) is obtained in real time through the speed sensor installed on the driving drum, combined with the known shaft length L (unit: m) and circumference C = πD (D is the diameter of the roller, unit: m) of the roller, a mathematical model of the data acquisition frequency f is established:

[0100]

[0101] The formula shows that the data acquisition frequency is proportional to the linear speed of the belt and inversely proportional to the circumference of the roller. For example, when the linear speed of the belt v = 2 m / s and the diameter of the roller D = 0.1 m, the data acquisition frequency

[0102] Take the standard non-wearing roller as a reference, record the time period of the sensor array scanning a complete roller Ensure the time interval of single data acquisition Synchronize with the circumferential displacement of the roller surface. This means that every time Δt, the data collected by the sensor array corresponds to a specific position on the circumferential surface of the roller, thereby ensuring the continuity and accuracy of the data in the circumferential direction.

[0103] Define the physical coordinates (x p ,yp ) and the axial coordinate x e [0, L] and the circumferential coordinate y e

[0104] The mapping relationship between the axial coordinate x e [0, L] and the circumferential coordinate y e

[0105] x = x p + δ x , y = v · t + δ y

[0106] wherein δ x , δ y are installation offset errors, which are fitted and calibrated by the least squares method to eliminate the position mapping deviation. The principle of the least squares method is to determine the best fitting parameters by minimizing the sum of squares of errors between the observed values and the model predicted values. In this scenario, the optimal values are calculated by the least squares method through multiple measurements of the data of the sensor array at different positions of the idler, thereby realizing accurate coordinate mapping.

[0107] S2 Real-time acquisition of dynamic temperature and pressure data

[0108] Contact scanning and data stream fragmentation storage: relying on the flexible sensor array built into the bottom of the belt, continuous contact scanning of the idler surface is realized with the belt running. The flexible sensor array is driven by the belt and closely contacts the idler surface, ensuring that the temperature and pressure changes of the idler surface can be accurately perceived. When the belt drives the sensor array to pass through the idler area, each sensor unit in the array synchronously collects the temperature value T i (t) and the pressure value P i (t) of the corresponding contact point at a frequency f. Wherein i represents the i-th sensor unit, thereby forming a time series:

[0109]

[0110] wherein M is the number of axial sensor units, N is the number of circumferential sensor units, x j , y k are the calibrated spatial coordinates, and through accurate coordinate mapping, the data traceability accuracy reaches millimeter level. Therefore, the specific position of each data point on the idler surface can be accurately determined, providing high-precision position information for subsequent analysis.

[0111] S3 Adaptive noise reduction processing of multiple sources

[0112] The coal mine environment is complex, and the belt conveyor idler monitoring data is easily affected by multiple sources of interference such as vibration, electromagnetic and material impact, which seriously affects the data accuracy and fault diagnosis reliability. In order to effectively improve the data quality, a two-level adaptive noise reduction model is constructed to eliminate different types of interference and retain key wear characteristics.

[0113] S3.1 Sliding window outlier rejection algorithm

[0114] The first stage adopts a sliding window-based outlier rejection algorithm, mainly to deal with sudden disturbances such as material impact. Through statistical analysis of a large number of temperature and pressure data collected by the standard roller under normal operating conditions, the normal fluctuation range of the data is determined. Assuming that the mean of the temperature data is μ T , the standard deviation is σ T , the mean of the pressure data is μ P , and the standard deviation is σ P , according to the 3σ rule in statistics, the normal fluctuation range of temperature is set as [μ T -3σ T , μ T +3σ T ], and the normal fluctuation range of pressure is set as [μ P -3σ P , μ P +3σ P ]. This rule believes that under normal circumstances, the probability of data falling within the range of mean plus or minus 3 times the standard deviation is about 99.7%, so data points outside this range are likely to be outliers caused by sudden disturbances.

[0115] When the temperature value T(n) or pressure value P(n) in the collected real-time data exceeds the corresponding range above, it is determined as an abnormal point. For the abnormal point T(n), the neighborhood mean filling method is used for processing, the formula is as follows:

[0116]

[0117] Where T is the filled temperature value, K is the half-width of the sliding window, and the formula represents taking the previous and subsequent K data points (including itself) centered on the abnormal point T(n), calculating the average of these data points as the filling value. In this way, the statistical characteristics of the neighborhood data are used to replace the outliers, effectively eliminating the impact of sudden disturbances such as material impact. For example, when K = 2, for the abnormal point T(n), its filling value is

[0118]

[0119] The selection of the sliding window half-width K has a significant impact on the denoising effect. To determine the optimal window size, Fourier transform is used to analyze the original signal. Fourier transform can convert time-domain signals into frequency-domain signals. By observing the frequency-domain characteristics, the main frequency components of the signal and the frequency distribution of the noise can be found. When selecting K, ensure that the window size can effectively eliminate outliers while maximizing the preservation of low-frequency characteristics of the signal, as low-frequency characteristics often contain key information about the roller's operating state, such as normal wear trends. If the window is too small, it may not fully utilize the statistical properties of the neighborhood data, resulting in incomplete outlier removal. If the window is too large, it may smooth out some useful signal details, affecting subsequent extraction of roller wear characteristics. Through multiple experiments and data analysis, the optimal K value for the current monitoring environment is determined.

[0120] S3.2 Adaptive Wavelet Filtering Algorithm

[0121] The second stage uses an adaptive wavelet filtering algorithm, mainly targeting electromagnetic and vibration interference. Wavelet transform is a time-frequency analysis method that can decompose signals into subbands of different frequencies, allowing for analysis and processing of signals at different scales. It is well suited for processing non-stationary signals, such as roller monitoring signals affected by electromagnetic and vibration interference.

[0122] The signal S(n) after the first stage of denoising is subjected to multi-level wavelet decomposition, resulting in high-frequency noise coefficients W H and low-frequency signal coefficients W L . High-frequency noise coefficients mainly contain high-frequency interference components such as electromagnetic and vibration, while low-frequency signal coefficients retain the main characteristic information of the roller's operation. By analyzing the frequency-domain characteristics of the data, an adaptive threshold function λ(f) = α · std(W H (f)) is determined, where α is a noise sensitivity coefficient determined through training and analysis of a large amount of historical data. std(W H (f)) represents the standard deviation of the high-frequency noise coefficients at frequency f. This threshold function can automatically adjust the threshold based on the statistical characteristics of the high-frequency noise coefficients to adapt to different levels of interference.

[0123] The high-frequency noise coefficients W H are processed according to the adaptive threshold function λ(f), filtering out high-frequency interference components. Specifically, for each coefficient in the high-frequency noise coefficients W H , if its absolute value is less than λ(f), it is set to 0, indicating that the coefficient is caused by noise. If its absolute value is greater than λ(f), the coefficient is retained, indicating that it contains useful signal information. After threshold processing, the high-frequency noise coefficients and the low-frequency signal coefficients are reconstructed through inverse wavelet transform (IDWT) to obtain the filtered signal The formula is as follows:

[0124]

[0125] where Thresh(W H ,λ(f)) represents the result of threshold processing on the high-frequency noise coefficient W H . In the noise reduction process, the local gradient feature of the signal is crucial. The local gradient feature reflects the rate of change of the signal at adjacent time points, which is of great significance for identifying the wear characteristics of the roller. If excessive filtering leads to distortion of the local gradient feature, it may cause deviation in the subsequent wear fault feature extraction, thereby affecting the accuracy of fault diagnosis. Therefore, when designing the adaptive wavelet filtering algorithm, by reasonably selecting the wavelet basis function, the number of decomposition layers, and the threshold function parameters, etc., the local gradient feature of the signal is ensured to be retained to the greatest extent while effectively filtering out high-frequency interference.

[0126] S4 Roller surface temperature and pressure cloud map reconstruction

[0127] After completing the noise reduction processing of the roller monitoring data, in order to more intuitively and comprehensively show the temperature and pressure distribution of the roller surface, so as to facilitate the subsequent wear fault feature extraction and analysis, it is necessary to reconstruct the cloud map of the data. This process constructs a two-dimensional grid matrix, and uses a bilinear interpolation algorithm to process the missing data between the sensing unit gaps, and simultaneously fuses the temperature cloud map and the pressure cloud map to generate a fusion cloud map with both temperature and pressure information, providing multi-dimensional data support for subsequent feature extraction and fault diagnosis.

[0128] S4.1 Two-dimensional grid matrix construction and bilinear interpolation

[0129] Two-dimensional grid matrix construction and bilinear interpolation: a two-dimensional grid matrix is constructed with the roller axial direction X ∈ [0, L] and Y ∈ [0, C], and each node coordinate is where m = 0, 1, …, M-1, n = 0, 1, …, N-1. Here L is the roller shaft length (unit: m), C is the roller circumference (unit: m), and M and N represent the number of grids divided in the axial and circumferential directions, respectively. In this way, the roller surface is discretized into a series of grid nodes, each of which corresponds to a specific position on the roller surface. The sensing unit data (T j,k , P j,k ) is mapped to the grid nodes, where T j,k represents the temperature value (unit: ℃) collected by the jth axial sensing unit and the kth circumferential sensing unit, and P j,k ​represents the corresponding pressure value (unit: Pa). However, since the distribution of the sensing unit does not completely cover the surface of the roll, there are gaps, so it is necessary to process the missing data at these gaps.

[0130] For the gap missing data, a bilinear interpolation method is adopted. Assuming that the coordinates of the four vertices around the point P(x, y) to be interpolated are Q 11 (x1, y1), Q 12 (x1, y2), Q 21 (x2, y1), and Q 22 (x2, y2), and the corresponding temperature values are T 11 , T 12 , T 21 , and T 22 , respectively, then the temperature cloud map T(x, y) of the point to be interpolated is calculated by the following formula:

[0131]

[0132] The derivation of this formula is based on the principle of bilinear interpolation. First, perform linear interpolation twice in the x direction to obtain R1 and R2:

[0133]

[0134]

[0135] Then perform linear interpolation on R1 and R2 in the y direction to obtain the final interpolation result T(xy):

[0136]

[0137] Expanding and combining the above two equations, the previous bilinear interpolation formula is obtained.

[0138] For example, in actual application, if the temperature values of the four vertices Q 11 (1, 1), Q 12 (1, 2), Q 21 (2, 1), and Q 22 (2, 2) are 30°C, 32°C, 35°C, and 38°C, respectively, and the point to be interpolated P(1.5, 1.5), then Substituting the formula gives:

[0139] T(1.5, 1.5) = (1-0.5)(1-0.5) x 30 + (1-0.5) x 0.5 x 32 + 0.5 x (1-0.5)

[0140] x 35 + 0.5 x 0.5 x 38 = 33.75°C

[0141] In the same way, the interpolation calculation of the pressure cloud map can smoothly fill in the missing data of the sensor unit gap, ensuring the continuity and integrity of the cloud map, and providing more accurate data basis for subsequent analysis.

[0142] S4.2 Channel fusion to generate multi-dimensional cloud map

[0143] After obtaining the temperature cloud map T(x, y) and the pressure cloud map P(x, y), in order to comprehensively utilize the temperature and pressure information, a channel fusion algorithm is adopted to fuse the two into a three-dimensional matrix F(x, y) = [T(x, y); P(x, y)]. This fusion method enables the cloud map to simultaneously display the temperature and pressure distribution on the surface of the roller, providing multi-dimensional data support for subsequent wear fault feature extraction. Through analysis of the fused cloud map, the running state of the roller can be more comprehensively understood. For example, in some wear fault conditions, the temperature and pressure distribution on the surface of the roller will exhibit specific change patterns. Through observation and analysis of the fused cloud map, these fault features can be more accurately identified, improving the accuracy and reliability of fault diagnosis.

[0144] S5 Wear fault differential feature extraction

[0145] After completing the reconstruction of the temperature and pressure cloud maps of the roller surface, in order to accurately identify the wear fault types of the roller, a partition feature extraction strategy needs to be designed for the different features of circumferential wear and axial wear. By reasonably dividing the fused cloud map into different regions, the temperature, pressure statistical features and gradient change rates of different regions are calculated, thereby extracting quantified feature parameters that can effectively represent wear faults, providing accurate data support for subsequent fault identification.

[0146] S5.1 Region division and feature parameter calculation

[0147] The cloud map is divided into outer side y ∈ [0, 0.2C] ∪ [0.8C, C], middle y ∈ [0.4C, 0.6C], and inner side y ∈ [0.2C, 0.4C] ∪ [0.6C, 0.8C] along the circumferential direction, where C is the circumference of the roller (unit: m). This division method can better capture the wear features at different positions along the circumference of the roller. For example, when the roller exhibits uneven wear along the circumference, the temperature and pressure distribution in the outer and inner regions may differ significantly.

[0148] The axial direction is divided into left x ∈ [0, 0.3L], middle x ∈ [0.3L, 0.7L], and right x ∈ [0.7L, L] three bar regions, where L is the axial length of the roller (unit: m). This axial region division helps to locate the local wear areas in the axial direction, as in actual operation, the roller may exhibit abnormal wear in certain local areas due to installation deviation, uneven material distribution, etc.

[0149] For circumferential wear feature extraction, the temperature and pressure mean values and variance differences of the outer region and the inner region are calculated. The region mean value μ is calculated as follows:

[0150]

[0151] where z i is the temperature or pressure value of the i-th grid node in the region, and N is the total number of grid nodes in the region. Through this formula, the mean value μ 外侧 of the outer region and the mean value μ 内侧 of the inner region can be obtained, and the mean difference Δμ = | μ 外侧 - μ 内侧 | is calculated. The mean difference can intuitively reflect the average difference of temperature or pressure in different circumferential regions. When the roller has circumferential wear, this value will increase significantly.

[0152] The variance σ 2 is calculated as follows:

[0153]

[0154] Through this formula, the variance σ of the outer region and the variance σ of the inner region are calculated, and the variance ratio σ is obtained. The variance ratio can measure the difference in the dispersion degree of data in different circumferential regions. When the roller has circumferential wear, the dispersion degree of data in the wear region will usually change, resulting in a variance ratio deviating from the normal range.

[0155] By setting a difference threshold (based on the statistical value of the difference of normal roller regions), it is determined whether there is a feature of "significant difference between the outer and inner region values". For example, after analyzing a large amount of data of normal rollers, it is determined that the normal threshold of temperature mean difference is Δμ th = 5℃. When the calculated Δμ > Δμ th , it is considered that there is a circumferential wear feature.

[0156] For axial wear feature extraction, the gradient change rate of each axial strip region is calculated. The gradient change rate G is used to measure the change rate of temperature T and pressure P in the axial x direction, and the calculation formula is as follows:

[0157]

[0158] In actual calculation, the difference approximation method is used to approximate the derivative for discrete grid node data. Taking a node (x n , y m ) as an example, the axial temperature gradient is approximately:

[0159]

[0160] The same principle applies to pressure gradients. By calculating the gradient change rate of each strip region, areas of abrupt gradient changes can be located to determine if there are any "abnormal axial local area values." When the idler roller experiences wear in a localized axial region, the rate of temperature and pressure change in that region will differ from that in the normal region, manifesting as an abnormal increase or decrease in the gradient change rate.

[0161] Simultaneously, the area A of the wear region, the mean difference Δμ, and the gradient peak G are extracted. max Three quantized feature parameters form the feature vector f = [A, Δμ, G] max The wear area A is obtained by statistically analyzing the pixels in the wear area of ​​the fused cloud map, and reflects the severity of the wear. Gradient peak value G. max The maximum value among the gradient change rates of each strip region along the axial direction highlights the location and degree of most severe axial wear. These characteristic parameters describe the wear state of the idler roller from different perspectives, providing a comprehensive and accurate data foundation for subsequent fault type identification based on feature matching.

[0162] S6 Fault Type Identification Based on Feature Matching

[0163] After extracting the wear fault features of the idler rollers, accurately and efficiently utilizing these features to identify fault types becomes a crucial step in achieving intelligent operation and maintenance of belt conveyor idler rollers. Traditional fault identification methods often suffer from insufficient accuracy and poor adaptability when facing the complex and ever-changing working conditions in coal mines. Therefore, a fault identification model based on improved K-Nearest Neighbors (KNN) is introduced. By constructing a rich and accurate feature sample library and employing a scientifically sound distance metric and classification decision mechanism, accurate identification of idler roller wear fault types can be achieved, providing strong support for timely equipment maintenance and safe production.

[0164] Improved K-Nearest Neighbor Fault Recognition Model: To achieve accurate identification of idler roller wear fault types, an improved K-Nearest Neighbor (KNN) fault recognition model is constructed. First, a feature sample library is built using a large number of labeled idler roller feature vectors of normal wear and two types of wear faults (circumferential wear and axial wear). Where f i Let l be the feature vector of the i-th sample, containing quantized feature parameters such as the area of ​​the wear region, the mean difference, and the gradient peak. i The corresponding fault labels are (0 represents normal, 1 represents circumferential wear, and 2 represents axial wear).

[0165] During the identification process, for the feature vector f of the idler roller to be detected q The distance-weighted KNN algorithm is used to calculate its relationship with each feature vector f in the sample database.i the distance d(f q ,f i ), the formula is:

[0166]

[0167] where w k is the weight of the kth feature dimension, which is determined by analyzing and normalizing the training data to reflect the importance of different feature dimensions to fault recognition. For example, through analysis of a large number of training samples, it is found that the mean difference degree of the wear area has high sensitivity in distinguishing circumferential wear faults, so a higher weight can be given to this feature dimension. D is the dimension of the feature vector, and in this model D = 3, corresponding to the three quantitative feature parameters mentioned above.

[0168] By calculating the distance, the K nearest neighbors (i.e. K-nearest neighbors) are selected, and a weighted voting mechanism is used to determine the fault type of the to-be-detected roller The formula is as follows:

[0169]

[0170] where δ(l i , l) is the Kronecker function, δ(l i , l) = 1 when l i = l; otherwise, δ(l i , l) = 0. The meaning of this formula is that for each possible fault category l, the weighted voting sum of the samples belonging to this category in the K-nearest neighbors is calculated, and the category with the highest voting sum is the predicted fault type of the to-be-detected roller. Here, a distance weighting strategy is adopted, i.e. samples closer to the to-be-detected feature vector are given higher weights Because the closer the sample, the more similar the features of the sample to the to-be-detected sample, the higher the reference value for fault type judgment. For example, if K = 5, among the 5 nearest neighbor samples calculated, there are 3 samples belonging to the circumferential wear fault (l i = 1), and the distances between these 3 samples and the to-be-detected sample are relatively close, and the corresponding weighted voting sum is the highest among all categories, then according to the above formula, the to-be-detected roller will be judged as a circumferential wear fault.

[0171] To improve the reliability of fault identification, a confidence threshold θ = 0.9 is set. Confidence represents the degree of certainty of the model on the identification result, which is measured by calculating the voting proportion of the predicted category in K nearest neighbors. When the highest voting proportion is lower than θ, it means that the model has insufficient confidence in the current identification result, which is marked as suspected fault. At this time, the continuous monitoring mode is started, and the scanning frequency of the roller is increased to obtain more data samples for analysis until the fault state is clear. This can effectively avoid misjudgment caused by insufficient data or unobvious features, and improve the accuracy and stability of fault identification.

[0172] S7. Fault location and real-time warning output

[0173] S7.1 Positioning and severity classification

[0174] When the improved K nearest neighbor-based fault identification model successfully identifies the wear fault of the roller, the roller number and coordinate mapping relationship marked by the collected data are quickly combined to realize accurate positioning of the faulty roller on the conveying line. The roller number serves as the unique identifier of each roller in the belt conveyor system, and the conveying line coordinate mapping table Map(ID) = (x pos ,y pos )ID(x pos ,y pos )x pos y pos establishes a close relationship.

[0175] Where ID represents the roller number, (x pos ,y pos ) represents the two-dimensional coordinates of the roller on the conveying line, x pos is the position coordinate along the direction of the conveyor belt (unit: m), and y pos is the position coordinate perpendicular to the direction of the conveyor belt (unit: m). Through this mapping relationship, when a roller with number ID is detected to have a fault, its corresponding accurate position can be immediately queried from the mapping table, with an accuracy of ±0.5 m, which is of great significance for quickly locating the fault point and timely maintenance. For example, when the roller with number ID = 1005 is identified to have a wear fault, by querying the mapping table Map(1005) = (25.5, 1.2), it can be determined that the faulty roller is located at a distance of 25.5 from the starting point along the conveyor belt, and at a distance of 1.2 from a reference line perpendicular to the direction of the conveyor belt.

[0176] For the assessment of the severity of wear, the fuzzy logic algorithm comprehensively considers two key factors of the wear area A and the mean difference degree Δμ. The wear area A reflects the size of the wear range. The actual area of the wear area (unit: m 2 ) can be accurately calculated by counting the pixel points of the wear area in the fused cloud image and combining the actual physical size of the cloud image (determined according to the shaft length L and the circumference C of the roller). 外侧 The mean difference degree Δμ reflects the difference degree of the wear area and the normal area in the mean values of temperature and pressure, which is determined by the mean difference degree formula Δμ = μ 内侧 | calculated in the circumferential wear feature extraction process mentioned above.

[0177] The fuzzy logic algorithm maps the wear area A and the mean difference degree Δμ to different severity levels by establishing a fuzzy rule base. The specific classification standards are as follows:

[0178] Mild wear: When the wear area A < 0.1L·0.1C and the mean difference degree Δμ < 0.3σ, it is determined as mild wear. Here, 0.1L·0.1C represents 1% of the total area of the roller surface, and σ is the standard deviation of the mean difference degree of the normal roller area, which is obtained by analyzing and statistically processing a large number of normal roller data. In this case, the wear degree of the roller is relatively light, and the impact on the operation of the conveyor belt is small, but it still needs to be closely monitored for its development trend. For example, if the shaft length L of a certain roller is 1 m, the circumference C is 0.5 m, and the standard deviation σ of the mean difference degree of the normal roller area is 2, when the calculated wear area A of the roller is 0.004 m 2 < 0.1 × 1 × 0.1 × 0.5 = 0.005 m 2 , and the mean difference degree Δμ = 0.5 < 0.3 × 2 = 0.6, it is determined that the roller is in a state of mild wear.

[0179] Moderate wear: When 0.1L·0.1C ≤ A < 0.3L·0.3C or 0.3σ ≤ Δμ < 0.6σ, it is determined as moderate wear. At this time, the wear of the roller is already relatively obvious, and it may have some impact on the running stability of the conveyor belt, so maintenance personnel need to be arranged for inspection and evaluation, and a corresponding maintenance plan needs to be developed. For example, when the wear area A of a certain roller is 0.015 m 2 , it satisfies 0.005 m 2 ≤ 0.015 m 2 < 0.3 × 1 × 0.3 × 0.5 = 0.045 m 2 , or the mean difference degree Δμ = 0.4, which satisfies 0.3 × 2 = 0.6 ≤ 0.4 < 0.6 × 2 = 1.2, it is determined that the roller is in a state of moderate wear.

[0180] Severe wear: If the wear area A ≥ 0.3L·0.3C and the mean difference Δμ ≥ 0.6σ, it is determined as severe wear. In this case, the wear of the roller is already very serious, which may cause the conveyor belt to fail at any time, and immediate shutdown for replacement or repair is needed to avoid production accidents. For example, when the wear area A = 0.05m 2 ≥ 0.045m 2 and the mean difference Δμ = 1.5 ≥ 1.2, it can be determined that the roller is in a state of severe wear.

[0181] Through this wear severity classification method based on fuzzy logic algorithm, the wear condition of the roller can be more accurately and comprehensively evaluated, providing a scientific basis for subsequent early warning and maintenance decision-making.

[0182] S7.2 Early warning information transmission and historical data storage

[0183] Once the location and wear severity of the faulty roller are determined, the early warning information is quickly transmitted to the ground monitoring center through the underground industrial Ethernet. As an important infrastructure for communication in coal mine underground, the industrial Ethernet has the characteristics of high speed, reliability, strong anti-interference ability, etc., which can ensure the rapid and accurate transmission of early warning information in complex underground environment. The early warning information is output in the form of sound and light alarm and text prompt, with rich and detailed content, including key information such as fault roller number, location, wear type and severity, collection time, etc. For example, when the roller with ID = 1005 is detected to have circumferential wear at location (25.5, 1.2) and the wear severity is moderate, and the collection time is 2025-11-05 14:30:00, the ground monitoring center will attract the attention of the staff with a prominent sound and light alarm, and display detailed text prompt on the monitoring interface: “Fault roller number: 1005, location: (25.5, 1.2), wear type: circumferential wear, severity: moderate, collection time: 2025-11-05 14:30:00”, so that the staff can quickly understand the fault condition and take appropriate measures in time.

[0184] At the same time, fault data and cloud images are stored in the historical database, providing valuable data support for the whole life cycle management of the equipment. The historical database uses a high-performance database management system, with characteristics of large-capacity storage, efficient data retrieval, and safety and reliability. When storing fault data, in addition to recording the basic information of the fault roller, such as the number, position, wear type, severity, and collection time, the device operating condition parameters are also attached, such as the belt running speed, the type and weight of the conveyed material, the environmental temperature and humidity, etc. These operating condition parameters have important reference value for analyzing the fault cause, evaluating the equipment performance, and formulating maintenance strategies. For example, by analyzing the wear condition of the roller under different operating conditions, the main factors causing the roller wear can be found, such as the hardness, particle size, conveying capacity of the material, etc., so as to optimize the equipment operating parameters and prolong the service life of the roller. At the same time, the fault data and cloud images in the historical database also provide a data basis for the whole life cycle management of the equipment. Through the analysis of the historical data of the equipment operation, the remaining life of the equipment can be predicted, the maintenance plan can be made in advance, the equipment failure rate can be reduced, and the production efficiency can be improved.

[0185] S8 model parameter dynamic optimization update

[0186] Adaptive updating mechanism and regional special calibration: To ensure that the fault identification model maintains high identification accuracy in the complex and variable downhole environment, it is essential to establish a model parameter adaptive updating mechanism. This mechanism can enable the model to continuously adjust its parameters according to the actual operating conditions, so as to better adapt to the feature drift caused by performance degradation, environmental changes, etc. during the roller operation.

[0187] A sample feedback loop is established, and the fault data (fnew) confirmed by the ground monitoring center is returned to the edge computing node. Every 50 new samples accumulated trigger model parameter update. Here, fnew is a vector containing quantitative feature parameters such as wear area, mean difference, and gradient peak value, representing the feature data of the newly confirmed fault roller. By feeding these new data back to the edge computing node, the model can use these latest information to optimize its parameters.

[0188] In the parameter updating process, the gradient descent algorithm is used to adjust the feature weight vector w, the formula is:

[0189]

[0190] where w is the feature weight vector, which determines the importance of each feature parameter in the fault identification model. For example, when judging the circumferential wear fault, if the mean difference degree of the wear area has a greater impact on the identification result, then the weight value corresponding to this feature in the weight vector will be relatively high. η is the learning rate, which controls the step size of each parameter update. The selection of the learning rate is very critical. If the learning rate is too large, the model may skip the optimal solution during training, resulting in failure to converge; if the learning rate is too small, the training speed of the model will be very slow, requiring more training time and data. represents the gradient of the loss function J(w) with respect to the old weight vector wold, which reflects the direction and rate of change of the loss function under the current weight vector. By updating the weight vector in the opposite direction of the gradient, the loss function can be gradually reduced, thereby improving the identification accuracy of the model.

[0191] The loss function J(w) uses the cross-entropy function, which is defined as:

[0192]

[0193] where n is the number of samples, which is the total number of fault samples used to train and update the model in the fault identification model. i is the true label of sample i, taking values of 0 (normal), 1 (circumferential wear), or 2 (axial wear), which represents the actual operating state of the roller. is the label probability predicted by the model for sample i, which is the prediction result of the model based on the input feature vector f. The cross-entropy function evaluates the performance of the model by measuring the difference between the model's prediction result and the true label f. When the model's prediction result is closer to the true label, the value of the cross-entropy is smaller, indicating that the performance of the model is better; otherwise, the value of the cross-entropy is larger, indicating that the performance of the model is worse.

[0194] When the area recognition accuracy Acc(r) < 85%, start special calibration. Here, Acc(r) represents the recognition accuracy of the model in area r, which is obtained by calculating the ratio of the number of correctly identified samples to the total number of samples in that area. For example, if there are 100 fault recognitions in a certain area, of which 80 are correctly identified, then the recognition accuracy of that area is When the recognition accuracy of a certain area is continuously lower than 85%, it indicates that the performance of the model in that area has a problem and needs to be calibrated.

[0195] Special calibration uses the K-means algorithm to re-cluster the feature vectors in that area, with the formula:

[0196] Calibrate(r) = K-means(f i|Region(i)=r)

[0197] Among them, f i |Region(i) = r represents the set of all feature vectors within region r. The K-means algorithm is a commonly used clustering algorithm. Its basic idea is to divide the data into K clusters, minimizing the distance between data points within each cluster and maximizing the distance between clusters. In specialized calibration, re-clustering the feature vectors within region r using the K-means algorithm can find more reasonable feature cluster centers, thereby optimizing the threshold parameter and improving the model's recognition accuracy in that region. For example, when performing specialized calibration on a certain region, the K-means algorithm may reclassify feature vectors that were previously misclassified into the correct clusters, enabling the model to more accurately identify the type of idler roller fault within that region.

[0198] By combining this adaptive update mechanism with regional specialized calibration, the fault identification model can continuously adapt to the complex and ever-changing downhole environment, maintain high identification accuracy, and provide reliable technical support for fault diagnosis of belt conveyor idlers.

[0199] The following section verifies the flexible temperature and pressure sensor array detection method for wear of belt conveyor idler roller skin according to the present invention.

[0200] like Figure 2 As shown in the table, the data was obtained through comparative experiments under 100 sets of identical operating conditions (belt speed 2m / s, idler diameter 0.1m, downhole electromagnetic interference + vibration interference). This system, utilizing "sensor synchronous calibration + dual-stage noise reduction + improved KNN recognition" technology, improves the recognition accuracy by 23.8% compared to a single sensor, reduces the positioning error to 0.5m, compresses the detection response time to the second level, and lowers the false negative rate to below 2%, completely solving the problems of "slow, coarse, and false negatives" in existing technologies in complex downhole environments.

[0201] like Figure 3As shown, the effectiveness of the "model parameter dynamic optimization update" mechanism is verified, highlighting the system's ability to adapt to feature drift over a long period of operation. 500 groups of fault samples (including 100 groups of normal rollers, 200 groups of circumferential wear, and 200 groups of axial wear) are selected, and the samples are accumulated in stages and tested for recognition accuracy. When the sample size increases from 100 groups to 500 groups, the accuracy of the traditional KNN model increases from 82% to 89% (an increase of 8.5%), while the accuracy of the improved KNN of the system (including the adaptive update mechanism) increases from 88% to 96.8% (an increase of 10%). The key difference is that when the sample size is <300 groups, the system uses regional special calibration (K-means clustering optimization threshold) to always lead the traditional KNN by ≥6% in accuracy; when the sample size is ≥300 groups, dynamic weight adjustment stabilizes the accuracy at >95%, solving the problem of "strong sample dependence and long-term precision decay" of existing models.

[0202] As shown in Figure 4 To quantify the system's differentiated recognition ability for "mild / moderate / severe" wear, highlighting the relevance of feature extraction, a confusion matrix is used to verify the wear severity grading ability.

[0203] The rows of the confusion matrix represent the true labels, and the columns represent the predicted labels. The diagonal elements are the number of correct recognitions. The test set contains 150 groups of graded samples (50 groups of mild wear, 50 groups of moderate wear, and 50 groups of severe wear). The system's recognition accuracy for mild wear is 92%, for moderate wear is 98%, and for severe wear is 100%, with an overall grading accuracy of 96.7%. Compared to traditional methods (grading accuracy of 81%), the core advantage is that the "area + mean difference" two-dimensional fuzzy logic grading solves the confusion problem between mild wear and normal rollers, and between moderate and severe wear (traditional method mild wear misjudgment rate 18%, system only 8%).

[0204] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Therefore, any equivalent changes made in terms of structure, shape, principle, and application direction of the present application should be covered within the protection scope of the present application.

Claims

1. A method for detecting a belt conveyor idler drum skin wear flexible warm pressure sensing array, characterized by, Comprise the following steps: S1, the sensor array is calibrated with the conveying system parameter synchronization: through measuring the data of the sensor array at different positions of the roller for many times, the optimal value is calculated by using the least square method, so as to realize accurate coordinate mapping; S2, real-time acquisition of dynamic temperature and pressure data: according to the continuous contact scanning of the flexible sensor array on the surface of the roller, the temperature and pressure changes of the roller surface are sensed, and the specific position of each data point on the surface of the roller is determined through coordinate mapping; S3, adaptive processing of multi-source interference: a two-stage adaptive noise reduction model is constructed, different types of interference are eliminated, and key wear features are retained; S4, reconstruction of roller surface temperature and pressure cloud map: a two-dimensional grid matrix is constructed, and the missing data of the sensing unit gap is processed by using the bilinear interpolation algorithm, and the temperature cloud map and the pressure cloud map are fused into a three-dimensional matrix F(x, y) = [T(x, y); P(x, y)], and a fusion cloud map with temperature and pressure information is generated; S5, wear fault differential feature extraction: through reasonable regional division of the fusion cloud map, the temperature, pressure statistical features and gradient change rate of different regions are calculated, so as to extract the quantitative feature parameters which can effectively represent the wear fault; S6, fault type recognition based on feature matching: an improved K nearest neighbor based fault recognition model is introduced, a rich and accurate feature sample library is constructed, a scientific and reasonable distance measurement and classification decision mechanism is adopted, and the accurate identification of the roller wear fault type is realized; S7, fault positioning and real-time early warning output: combined with the marked roller number and coordinate mapping relationship when collecting data, the accurate positioning of the fault roller on the conveying line is realized, the wear severity grading based on fuzzy logic algorithm is adopted, the wear condition of the roller is accurately and comprehensively evaluated, and real-time early warning is realized; S8, dynamic optimization and update of model parameters: a model parameter adaptive update mechanism is established, combined with regional special calibration, the fault recognition model is continuously adapted to the complex underground environment.

2. The method of claim 1, wherein the method further comprises: The step S1 specifically comprises the following steps: The line speed v of the belt is obtained in real time through the speed sensor installed on the driving drum, and a mathematical model of the data acquisition frequency f is established combined with the known shaft length L and circumference C of the roller: Wherein, D is the diameter of the roller; With reference to a standard non-abradable carrier roller, the time period for the sensor array to scan the entire carrier roller in one revolution is recorded Ensuring a single set of data collection time intervals Synchronized with the circumferential displacement of the carrier roller surface; Definition of the mapping between the physical coordinates of the sensor array (x p ,y p ) and the axial coordinate x e [0, L] and the circumferential coordinate y e [0, C] of the roller is: x = x p + δ x y = v · t + δ y where δ x , δ y is the installation offset error, which is fitted by least square calibration to eliminate the position mapping bias.

3. The method of claim 2, wherein the flexible warm pressure sensor array is attached to the belt conveyor idler drum skin. The step S2 specifically comprises the following steps: ​ When the belt drives the sensing array through the area of the idler roller, each sensing cell in the array synchronously acquires the temperature value T of the corresponding contact point at the frequency f i (t) with the pressure value P i (t), where i denotes the i-th sensing cell, thereby forming a time series: Wherein, M is the number of axial sensing units, N is the number of circumferential sensing units, x j ,y k The calibrated spatial coordinates are obtained by accurate coordinate mapping, and the data traceability accuracy reaches millimeter level.

4. The method of claim 3, wherein the method further comprises: The step S3 specifically comprises the following steps: S3.1, the first stage adopts an outlier elimination algorithm based on sliding window, a large amount of temperature and pressure data collected under normal operating state of the standard roller is statistically analyzed, and the normal fluctuation range of the data is determined; Assume the mean of temperature data is μ T , the standard deviation is σ T , the mean of pressure data is μ P , the standard deviation is σ P , according to the 3σ criterion in statistics, set the normal fluctuation range of temperature as [μ T -3σ T , μ T +3σ T ], the normal fluctuation range of pressure as [μ P -3σ P , μ P +3σ P ] When the temperature value T(n) or the pressure value P(n) in the collected real-time data exceeds the above corresponding range, it is determined as an abnormal point; for the abnormal point T(n), the neighborhood mean filling method is adopted for processing, and the formula is as follows: wherein, K is the half width of the sliding window, the formula indicates that the average value of the K data points before and after the abnormal point T(n) is taken as the filling value. S3.2 The second stage adopts an adaptive wavelet filtering algorithm to perform multi-layer wavelet decomposition on the signal S(n) after the first stage of noise reduction, and decomposes it into high-frequency noise coefficients W H and low-frequency signal coefficients W L ; By analyzing the frequency domain characteristics of the data, an adaptive threshold function λ(f) = α·std(W H (f)) is determined, where α is a noise sensitivity coefficient, and std(W H (f)) is the standard deviation of the high-frequency noise coefficients at frequency f, and its optimal value is determined through training and analysis of a large amount of historical data; According to the adaptive threshold function λ(f) to high frequency noise coefficient W H is processed to filter out high frequency interference components; for each coefficient in high frequency noise coefficient W H , if its absolute value is less than λ(f), it is set to 0, considering that the coefficient is caused by noise; if its absolute value is greater than λ(f), the coefficient is retained, considering that it contains useful signal information; after threshold processing, the high frequency noise coefficient is reconstructed with the low frequency signal coefficient through wavelet inverse transform to obtain the filtered signal The formula is as follows: where Thresh(W H , λ(f)) represents the result of threshold processing on the high-frequency noise coefficient W H .

5. The method for detecting wear of belt conveyor idler roller skin using a flexible temperature and pressure sensor array according to claim 4, characterized in that, The step S4 specifically comprises the following steps: S4.1 Construct a MxN grid matrix with the axial direction of the roller X ∈ [0, L] and the circumferential direction Y ∈ [0, C], and each node coordinate is wherein m = 0, 1, …, M-1, n = 0, 1, …, N-1, wherein L is the axial length of the roller, C is the circumferential length of the roller, M and N respectively represent the number of grids divided in the axial and circumferential directions, in this way, the surface of the roller is discretized into a series of grid nodes, each node corresponds to a specific position on the surface of the roller; map the sensor unit data (T j,k , P j,k ) to the grid nodes, wherein T j,k represents the temperature value collected by the jth axial sensor unit and the kth circumferential sensor unit, and P j,k represents the corresponding pressure value; For the gap missing data, a bilinear interpolation method is adopted, assuming that the coordinates of four vertices around the interpolation point P(x, y) are Q 11 (x1, y1), Q 12 (x1, y2), Q 21 (x2, y1), and Q 22 (x2, y2), and the corresponding temperature values are T 11 , T 12 , T 21 , and T 22 , respectively. The temperature image T(x, y) of the interpolation point is calculated by the following formula: Similarly, the pressure cloud map P(x, y) is obtained by the same method; S4.2, after obtaining the temperature cloud map T(x, y) and the pressure cloud map P(x, y), the two are fused into a three-dimensional matrix F(x, y) = [T(x, y); P(x, y)] by using channel fusion algorithm, and a fusion cloud map with temperature and pressure information is generated.

6. The belt conveyor idler shell wear flexible warm press sensor array detection method of claim 5, wherein, The step S5 specifically comprises the following steps: The fused cloud image is divided into an outer side y∈[0, 0.2C]∪[0.8C, C], a middle y∈[0.4C, 0.6C], and an inner side y∈[0.2C, 0.4C]∪[0.6C, 0.8C] in the circumferential direction, wherein C is the length of the circumference of the roller; The fused cloud image is divided into a left x∈[0, 0.3L], a middle x∈[0.3L, 0.7L], and a right x∈[0.7L, L] in the axial direction, wherein L is the length of the axis of the roller; For circumferential wear feature extraction, the mean and variance differences of the temperature and pressure of the outer side and inner side regions are calculated; the mean μ of the region is calculated according to the following formula: wherein z i is the temperature or pressure value of the i-th grid node in the region, and N is the total number of grid nodes in the region; through the formula, the mean μ 外侧 of the outer region and the mean μ 内侧 of the inner region can be obtained, and then the mean difference Δμ = |μ 外侧 - μ 内侧 | is calculated. Variance σ 2 The calculation formula is: The variance of the outer region is calculated by the formula and the variance of the inner region and the variance ratio is obtained By setting a difference threshold, it is determined whether there is a feature with a significantly different value between the outer side and inner side regions; For axial wear feature extraction, the gradient change rate of each strip-shaped region in the axial direction is calculated; the gradient change rate G is used to measure the change rate of the temperature T and the pressure P in the axial direction x, and is calculated according to the following formula: By calculating the gradient change rate of each strip-shaped region, the region with a gradient mutation is located, and it is determined whether there is a feature with an abnormal value in the local region in the axial direction; At the same time, the area A of the wear region, the mean difference degree Δμ, and the gradient peak value G are extracted max Three quantified characteristic parameters form a characteristic vector f = [A, Δμ, G max ].

7. The method of claim 6, wherein the method further comprises: The step S6 specifically includes the following steps: A large number of labeled normal and two types of worn-out fault roller cloud feature vectors are taken as training samples to construct a feature sample library wherein f i is the feature vector of the i th sample, containing quantitative characteristic parameters such as the area of the wear area, the mean difference degree, and the gradient peak value, and l i is the corresponding fault label; In the identification process, for the feature vector f q of the to-be-detected roller i , the distance d(f q , f i ) between f q and each feature vector f i in the sample library is calculated by using the distance-weighted KNN algorithm, and the formula is as follows: wherein w k is the weight of the kth feature dimension, determined by analyzing and normalizing the training data to reflect the importance of different feature dimensions to fault recognition, and D is the dimension of the feature vector; By calculating the distance, the nearest K samples are selected, and the weighted voting mechanism is used to determine the fault type of the detected roller The formula is as follows: where δ(l i ,l) is the Kronecker delta.

8. The method of claim 7, wherein the method further comprises: In the step S7, the fault positioning specifically includes the following steps: The idler number is the unique identification of each idler in the belt conveyor system, and the conveying line coordinate mapping table Map(ID) = (x pos ,y pos ) pos ID(x pos ,y pos )x pos Establishes a close relationship, wherein ID is the id of the roller, (x pos ,y pos ) is the two-dimensional coordinate of the roller on the conveying line, x pos is the position coordinate along the direction of the conveying belt, and y pos is the position coordinate perpendicular to the direction of the conveying belt.

9. The method of claim 8, wherein the method further comprises: The step 8 specifically includes the following steps: A sample feedback closed loop is established, and the fault data confirmed by the ground monitoring center is fed back to the edge computing node; each time a certain group of new samples is accumulated, the model parameter update is triggered; In the parameter update process, the gradient descent algorithm is used to adjust the feature weight vector w, and the formula is as follows: where w is a feature weight vector, η is a learning rate, is the gradient of the loss function J(w) with respect to the old weight vector wold; The loss function J(w) adopts a cross-entropy function, and its definition is as follows: where n is the number of samples, y i is the true label of sample i, is the label probability predicted by the model for sample i; When the region recognition accuracy Acc(r) is less than 85%, special calibration is started, wherein Acc(r) represents the recognition accuracy of the model in the region r; The special calibration adopts a K-means algorithm to re-cluster the feature vectors in the region, and the formula is as follows: Calibrate(r) = K-means(f i | Region(i) = r) where f i | Region(i) = r denotes the set of all feature vectors within region r.

Images