Carrier roller skin damage fault diagnosis method based on distributed sound and temperature fusion monitoring

Abstract

The invention relates to a carrier roller skin damage fault diagnosis method based on distributed sound and temperature fusion monitoring in the technical field of coal mine underground conveying equipment fault monitoring. The carrier roller skin damage fault diagnosis method sequentially comprises the following steps of S1 sound and temperature data acquisition, S2 data preprocessing, S3 wear feature extraction, S4 sound and temperature fusion diagnosis model training and S5 fault diagnosis and mathematical model verification. Aiming at a core pain point of an existing carrier roller wear fault diagnosis technology, through synchronous data acquisition, targeted preprocessing, fault sensitive feature extraction, attention weighted fusion and full-process optimization of a robust classification model, a technical breakthrough is realized, the problems of time alignment and quality of multi-source data are solved, the distinguishing capability of different wear types is improved, and the fault diagnosis accuracy is improved. The diagnosis accuracy and the anti-interference generalization ability are improved, the fault detection rate is guaranteed, and key fault omission is avoided.

Description

Technical Field

[0001] This invention belongs to the field of fault monitoring technology for underground conveying equipment in coal mines, specifically relating to diagnostic technology for belt conveyor roller skin damage faults, and particularly to a diagnostic method for accurately determining roller skin damage faults based on sound and temperature fusion monitoring. Background Technology

[0002] In underground coal mine production, belt conveyors are critical material transport equipment, and the stable operation of their idlers directly affects production efficiency and operational safety. Idler shell damage is one of the most common types of failure. Currently, the industry primarily relies on manual inspections to diagnose idler shell damage. This method is not only resource-intensive but also prone to errors due to subjective human factors, leading to missed or incorrect detections, thus increasing equipment maintenance costs and safety risks. Although existing technologies have developed solutions for monitoring idler shell damage using single or combined methods such as sound, temperature, vibration, and images, these solutions mostly focus on monitoring idler bearing failures. However, idler shell damage and bearing failures differ significantly in their physical characteristics, such as sound features, temperature changes, and vibration patterns. This makes it difficult for existing monitoring technologies to effectively address the diagnostic needs of idler shell damage. Therefore, a dedicated monitoring technology for idler shell damage is urgently needed to fill this industry gap. Summary of the Invention

[0003] This invention aims to solve the technical problems existing in the monitoring of idler roller skin damage in underground belt conveyors in coal mines: on the one hand, traditional personnel inspection methods are inefficient, costly, and prone to missed or false detections due to subjective factors, making it difficult to meet the requirements for safe and stable operation of equipment; on the other hand, existing idler roller fault monitoring technologies based on sound, temperature, vibration, etc., are mostly designed for bearing faults and do not fully consider the special physical characteristics of idler roller skin damage, resulting in insufficient accuracy and poor adaptability in identifying roller skin damage faults, making it impossible to achieve accurate and efficient diagnosis of such faults, thereby affecting the timeliness of belt conveyor maintenance and production continuity.

[0004] To address the aforementioned problems, this invention provides a method for diagnosing roller skin damage based on distributed sound and temperature fusion monitoring, comprising the following steps:

[0005] S1 Sound and Temperature Data Acquisition: Sound and temperature sensors are reasonably arranged around the belt conveyor to collect sound and temperature signals from the idler rollers;

[0006] S2 Data Preprocessing: Preprocesses the collected sound and temperature signals to achieve wavelet transform noise reduction, temperature data cleaning, and data standardization.

[0007] S3 Wear Feature Extraction: Axial and circumferential wear features are extracted from the preprocessed sound and temperature signals;

[0008] S4 Sound and Temperature Fusion Diagnostic Model Training: By performing feature layer fusion and diagnostic model training on data after wear feature extraction, the advantages of multi-source information can be complemented.

[0009] S5 Fault Diagnosis and Verification Mathematical Model: Construct a complete fault diagnosis process and verify the performance of the mathematical model.

[0010] To improve the accuracy of data acquisition and lay a foundation for time consistency in subsequent fusion analysis, the steps for acquiring the sound and temperature signals of the idler rollers in step S1 are as follows:

[0011] Let the frequency of the sound signal be f. s The temperature signal acquisition frequency is f t To fully capture the periodic sound changes and dynamic temperature trends of axial wear, the following must be satisfied:

[0012] f s ≥44100,Hz,f t ≥1 Hz

[0013] Let the acoustic signal acquisition time series be... Temperature signal acquisition time series is The timestamps of both are made to satisfy the interpolation algorithm. Ensure that sound and temperature data at the same time point can be matched.

[0014] To improve the quality of the preprocessed data, step S2 involves the following preprocessing steps for the acquired sound and temperature signals:

[0015] S2.1 Wavelet Transform Denoising

[0016] Let the original acoustic signal be x(t), which consists of effective signal and noise:

[0017] x(t) = s(t) + n(t)

[0018] Where s(t) is the effective signal related to roller wear, and n(t) is the noise component, which is denoised using wavelet transform, defined as follows:

[0019]

[0020] Where a is the scaling factor, b is the translation factor, and ψ(t) is the mother wavelet function. * Its conjugate function, W x (a,b) are wavelet coefficients;

[0021] The noise reduction process is achieved through threshold processing: for the wavelet coefficient W x (a,b), a reasonable threshold λ is set, the coefficients with amplitudes greater than or equal to the threshold are retained, and the low-amplitude noise coefficients are removed. Then, the signal is reconstructed through wavelet inverse transform to achieve noise suppression;

[0022] S2.2 Temperature data cleaning

[0023] S2.2.1 Outlier identification: Let the temperature sequence be T = [T1, T2,..., T n , calculate the sequence mean μ T and the standard deviation σ T . The 3σ criterion is used to determine outliers. If |T i -μ T |>3σ T , then T i is determined as an outlier and removed;

[0024] S2.2.2 Missing value interpolation: For the data missing position t, linear interpolation is used to complete it. The formula is:

[0025]

[0026] where t0 < t < t1 are the valid sampling times adjacent to the missing point, and the continuity of temperature data is ensured through linear fitting;

[0027] S2.3 Data standardization

[0028] Min-max standardization is used to map the data to the [0, 1] interval:

[0029]

[0030] where,

[0031] x is the original feature value;

[0032] x min 、x max are the minimum and maximum values of the feature samples;

[0033] x norm is the value after standardization.

[0034] In order to accurately extract the feature parameters related to the acoustic and temperature signals after preprocessing, in step S3,

[0035] S3.1 Axial wear feature extraction includes the following steps:

[0036] Acoustic feature extraction

[0037] Periodic Amplitude Difference: Axial wear exhibits periodic sound changes with roller rotation. Let the roller speed be n (r / min), then the rotation period T0 = 60 / n (s). In the k-th period, the sound signal amplitude in the wear area is... Normal region amplitude is The difference between the two values ​​reflects the degree of wear:

[0038]

[0039] Fourier transform: Wear-related characteristic frequencies can be identified through frequency domain analysis. The frequency domain representation of the acoustic signal s(t) is as follows:

[0040]

[0041] Where f is the frequency, S(f) is the spectral amplitude, and the characteristic frequency f is... c Defined as the frequency corresponding to the maximum amplitude of the spectrum, reflecting the dominant frequency of wear and friction; the spectral energy distribution is characterized by the energy proportion within the frequency band [f1, f2].

[0042]

[0043] Temperature feature extraction

[0044] Regional temperature mean: Axial wear leads to localized temperature increases. The axial direction is divided into M regions, and the temperature sequence of the i-th region is T. i =[T i1 ,T i2 ,...,T iN The average value can reflect the regional temperature level.

[0045]

[0046] Temperature difference change rate: During the wear development process, the temperature difference between regions will change with time. The temperature difference between adjacent regions i and i+1 is ΔT. i (t)=T i (t)-T i+1 (t), the rate of change of which can characterize the dynamic development trend of wear:

[0047]

[0048] Where Δt is the time interval, r i (t) is in °C / s;

[0049] S3.2 Circumferential wear feature extraction includes the following steps:

[0050] Acoustic feature extraction

[0051] Sound pressure level: The frictional sound pressure level in the circumferential wear region is significantly higher than that in the normal region. The formula for calculating the sound pressure level in the j-th circumferential region is:

[0052]

[0053] Where, p j For the effective sound pressure level of the region, p0 = 2 × 10 -5 Pa is the reference sound pressure level, SPL j The unit is dB. The location of wear can be identified by comparing the sound pressure levels in different areas.

[0054] Frequency band energy distribution: The sound energy from circumferential wear is concentrated in a specific frequency band. Let the total frequency band energy be... Main frequency band [f a ,f b Energy is The energy percentage is:

[0055]

[0056] Temperature feature extraction

[0057] High-temperature zone identification: The K-means clustering algorithm is used to cluster the circumferential temperature data, and the sample point T is defined. j With cluster center c k The distance is:

[0058] d(T j ,c k )=|T j -c k |

[0059] By iteratively optimizing the cluster centers, the temperature data is divided into high-temperature clusters and low-temperature clusters. After clustering, the high-temperature cluster centers are c. high The corresponding area is the high-temperature wear zone;

[0060] High and low temperature difference: The temperature difference between the high-temperature zone and the low-temperature zone can directly reflect the degree of circumferential wear. The average temperature of the high-temperature zone is μ high With the average temperature μ in the low temperature region low The difference:

[0061] ΔT HL =μ high -μ low .

[0062] To further improve diagnostic accuracy, in step S4,

[0063] S4.1 Feature layer fusion includes the following steps:

[0064] Feature splicing: Integrating axial and circumferential acoustic and temperature features into a unified feature vector, let the axial acoustic feature vector be... Axial temperature characteristics are Fa t =[μT i ,r i ]; Peripheral acoustic characteristics are Circumferential temperature characteristics are The fused feature vector is:

[0065]

[0066] Attention Mechanism Weighted Approach: This approach uses an attention mechanism to highlight key features. Let the feature dimension be D, and the attention weight be α. d Calculated using the softmax function:

[0067]

[0068] Among them, w d For learnable weight parameters, F d Let d be the eigenvalues ​​of dimension d. The weighted eigenvector is:

[0069]

[0070] The training of the S4.2 diagnostic model includes the following steps:

[0071] Random Forest Model: This model uses a random forest as the core classification model. The forest contains K decision trees, and the input features F... att The output class prediction h of the Kth tree k (F att The final classification result is determined through a majority voting mechanism.

[0072]

[0073] Where I(·) is the indicator function, {normal, axial, circumferential} represents the three categories of the model classification, and arg max represents the category c that makes the following expression take the maximum value;

[0074] Model optimization metrics: The number of decision trees K in the model parameters is optimized using the five-fold cross-validation method, and the validation set accuracy is used as the core evaluation metric.

[0075]

[0076] TP, TN, FP, and FN are confusion matrix elements, with the goal of achieving a validation set accuracy of Acc ≥ 95%.

[0077] To ensure the effectiveness of this invention in real-world scenarios, in step S5,

[0078] The steps of the S5.1 fault diagnosis process are as follows:

[0079] In real-time diagnosis, the collected sound and temperature signals are first preprocessed according to step S2 to remove noise and outliers; then, the fused feature vector to be diagnosed is obtained through the feature extraction method in step S3; finally, the features are input into the trained fusion diagnosis model, and the model outputs the fault diagnosis result and the corresponding confidence score. The fault confidence score is defined as the voting percentage of the category in the random forest.

[0080]

[0081] The steps for verifying the mathematical model in S5.2 are as follows:

[0082] Precision and recall are used as the core validation metrics:

[0083] Accuracy: Similar to model optimization metrics, it reflects the overall diagnostic accuracy.

[0084] Recall rate: For a specific fault category c, the recall rate is defined as:

[0085]

[0086] Where TP(c) is the number of true positives in category c, and FN(c) is the number of false negatives in category c.

[0087] The present invention provides a fault diagnosis method for idler roller skin damage based on distributed sound and temperature fusion monitoring. This method addresses the core pain points of existing idler roller wear fault diagnosis technologies (such as incomplete information from a single sensor, time misalignment of multi-source data, low diagnostic accuracy due to noise interference, and weak ability to distinguish different wear types). Through full-process optimization including synchronous data acquisition, targeted preprocessing, fault-sensitive feature extraction, attention-weighted fusion, and a robust classification model, a technological breakthrough is achieved. Specific advantages are as follows:

[0088] (1) Solving the problem of time alignment and quality of multi-source data

[0089] Existing technologies often employ single sound or temperature sensors. Furthermore, the use of multiple sensors frequently results in data mismatch due to timestamp discrepancies. Additionally, environmental noise (such as motor noise) and sensor anomalies (such as temperature fluctuations) severely interfere with signal quality. This invention achieves strict timestamp alignment by setting sampling frequencies of ≥44100Hz for sound signals and ≥1Hz for temperature signals, combined with interpolation. Then, through wavelet transform noise reduction (preserving high-frequency sound signals related to wear) and 3σ outlier removal plus linear interpolation (repairing temperature data), the signal-to-noise ratio of the preprocessed signal is improved by ≥30%, laying a high-quality data foundation for subsequent diagnostics.

[0090] (2) Improve the ability to distinguish between different wear types

[0091] Existing technologies often rely on single features (such as sound pressure level or average temperature alone), which cannot effectively distinguish between "axial wear (periodic sound changes + local temperature rise)" and "circumferential wear (regional sound pressure differences + high temperature concentration)". This invention specifically extracts fault-sensitive features such as axial periodic amplitude differences, circumferential sound pressure level (SPL), and high and low temperature differences, and combines them with attention mechanism weighting (increasing the weight of features with high wear contribution by 2-3 times), which improves the accuracy of distinguishing axial / circumferential wear by ≥25% compared with single feature methods.

[0092] (3) Improve diagnostic accuracy and anti-interference generalization ability

[0093] Existing technologies commonly use classifiers such as SVM and single decision trees, but their accuracy tends to drop below 80% when faced with complex operating conditions (such as noise fluctuations and imbalanced samples). This invention employs random forest (anti-overfitting) + five-fold cross-validation, fusing information from both sound and temperature sources, resulting in a stable overall diagnostic accuracy of ≥95%. Furthermore, the accuracy remains ≥92% even when the noise level is ≤20dB. This significantly improves the anti-interference capability compared to single sound sensor diagnosis (accuracy ≤65% at 20dB noise) and single temperature sensor diagnosis (accuracy ≤60% at 20dB noise).

[0094] (4) Ensure the fault detection rate and avoid missing critical faults.

[0095] Existing technologies often have a recall rate of less than 80% for minor axial / circumferential wear, which can easily lead to the expansion of faults. This invention captures early wear signals by targeting specific features (such as temperature difference change rate and frequency band energy ratio), so that the axial wear recall rate is ≥96% and the circumferential wear recall rate is ≥97%, ensuring that no critical faults are missed and reducing equipment maintenance costs. Attached Figure Description

[0096] Figure 1 This is a flowchart of the fault diagnosis method for idler roller skin damage based on distributed sound and temperature fusion monitoring according to the present invention. Detailed Implementation

[0097] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0098] like Figure 1 The method for diagnosing roller skin damage based on distributed acoustic and temperature fusion monitoring includes the following steps in sequence: S1 Acquisition of acoustic and temperature data, S2 Data preprocessing, S3 Wear feature extraction, S4 Training of acoustic and temperature fusion diagnostic model, and S5 Fault diagnosis and verification mathematical model.

[0099] The following is a detailed explanation.

[0100] S1 sound and temperature data acquisition

[0101] Data acquisition is a fundamental step in achieving accurate diagnosis of idler roller wear faults. Its core objective is to acquire high-quality, time-aligned acoustic and temperature signals to provide reliable data support for subsequent feature extraction and model training.

[0102] The acquisition frequencies of the sound and temperature signals need to be reasonably set according to the time scale of the fault characteristics. Let the sound signal acquisition frequency be f. s (Unit: Hz), temperature signal acquisition frequency is f t (Unit: Hz) To fully capture the periodic sound changes (typically including high-frequency components) and dynamic temperature trends of axial wear, the following must be satisfied:

[0103] f s ≥44100,Hz,f t ≥1 Hz

[0104] Time synchronization is a prerequisite for multi-sensor data fusion, achieved through timestamp alignment technology. Let the acoustic signal acquisition time series be... Temperature signal acquisition time series is The timestamps of both are made to satisfy the interpolation algorithm. This ensures that sound and temperature data at the same point in time can be matched, laying a foundation for temporal consistency in subsequent fusion analysis.

[0105] S2 Data Preprocessing

[0106] Raw acquired signals often contain problems such as environmental interference, sensor noise, and missing data, which require preprocessing to improve data quality.

[0107] S2.1 Wavelet Transform Denoising

[0108] Sound signals are easily interfered with by background noise such as motor operation and belt friction. Let the original sound signal be x(t) (unit: Pa), which consists of effective signal and noise:

[0109] x(t) = s(t) + n(t)

[0110] Where s(t) is the effective signal related to roller wear, and n(t) is the noise component. Wavelet transform is used for noise reduction, defined as follows:

[0111]

[0112] Where a is the scaling factor (dimensionless, controlling frequency scaling), b is the translation factor (unit: s, controlling time translation); ψ(t) is the mother wavelet function (dimensionless), ψ * Its conjugate function; Wx (a, b) is the wavelet coefficient (unit: Pa·s).

[0113] The noise reduction process is achieved through threshold processing: for the wavelet coefficient W x (a, b), a reasonable threshold λ is set, and the coefficients with amplitudes greater than or equal to the threshold (mainly corresponding to the effective signal) are retained, while the low-amplitude noise coefficients are removed, and then the signal is reconstructed through wavelet inverse transform to achieve noise suppression.

[0114] S2.2 Temperature data cleaning

[0115] The temperature signal may have abnormal jumps or data missing due to sensor failures, and cleaning processing is required:

[0116] S2.2.1 Outlier identification: Let the temperature sequence be T = [T1, T2,..., T n , (unit: °C), calculate the sequence mean μ T and the standard deviation σ T , and use the 3σ criterion to determine outliers. If |T i - μ T | > 3σ T , then T i is determined as an outlier and removed to avoid abnormal data interfering with trend analysis.

[0117] S2.2.2 Missing value interpolation: For the data missing position t, linear interpolation method is used to complete it, and the formula is:

[0118]

[0119] where t0 < t < t1 are the effective sampling times adjacent to the missing point (unit: s), and the continuity of temperature data is ensured through linear fitting.

[0120] S2.3 Data standardization

[0121] The dimensions and amplitude ranges of the acoustic signal and the temperature signal vary greatly, and standardization processing is required to eliminate the influence of dimensions. Min-max standardization is used to map the data to the [0, 1] interval:

[0122]

[0123] where,

[0124] x is the original feature value (such as sound pressure level, temperature);

[0125] x min 、x max are the minimum and maximum values of the feature samples;

[0126] x normThis is the standardized value (dimensionless).

[0127] Standardized data ensures that different types of features have equal weight in model training, improving the effectiveness of fusion analysis.

[0128] S3 Wear Feature Extraction Method

[0129] The pre-processed sound and temperature signals need to be further extracted to extract characteristic parameters related to wear faults. These characteristics need to be able to effectively distinguish the normal state, axial wear and circumferential wear of the idler roller.

[0130] S3.1 Axial Wear Feature Extraction

[0131] Axial wear manifests as abnormal friction in localized areas along the axial direction of the idler roller. Its acoustic signal exhibits periodic variations, and the temperature shows regional differences, requiring targeted feature extraction.

[0132] Acoustic feature extraction

[0133] Periodic Amplitude Difference: Axial wear exhibits periodic sound changes with roller rotation. Let the roller speed be n (r / min), then the rotation period T0 = 60 / n (s). In the k-th period, the acoustic signal amplitude in the wear area is... Normal region amplitude is The difference between the two values ​​reflects the degree of wear:

[0134]

[0135] Fourier transform: Wear-related characteristic frequencies can be identified through frequency domain analysis. The frequency domain representation of the acoustic signal s(t) is as follows:

[0136]

[0137] Where f is the frequency (unit: Hz), and S(f) is the spectral amplitude (unit: Pa / Hz). Characteristic frequency f c Defined as the frequency corresponding to the maximum amplitude of the spectrum, reflecting the dominant frequency of wear and friction; the spectral energy distribution is characterized by the energy proportion within the frequency band [f1, f2].

[0138]

[0139] Temperature feature extraction

[0140] Regional temperature mean: Axial wear leads to localized temperature increases. The axial direction is divided into M regions, and the temperature sequence of the i-th region is T. i =[T i1 ,T i2 ,...,T iN The average value can reflect the regional temperature level.

[0141]

[0142] Temperature difference change rate: During the wear development process, the temperature difference between regions will change with time. The temperature difference between adjacent regions i and i+1 is ΔT. i (t)=T i (t)-T i+1 (t), the rate of change of which can characterize the dynamic development trend of wear:

[0143]

[0144] Where Δt is the time interval (unit: s), r i (t) is in °C / s.

[0145] S3.2 Circumferential Wear Feature Extraction

[0146] Circumferential wear manifests as continuous friction in a specific area of ​​the roller's circumference, with regional differences in sound pressure levels in the acoustic signal and a significant high-temperature zone in the temperature field. Feature extraction needs to focus on these differences.

[0147] Acoustic feature extraction

[0148] Sound Pressure Level (SPL): The frictional sound pressure level in the circumferential wear region is significantly higher than that in the normal region. The formula for calculating the sound pressure level in the j-th circumferential region is:

[0149]

[0150] Where, p j The effective sound pressure level (in Pa) is p0 = 2 × 10⁻⁶. -5 Pa is the reference sound pressure level, SPL j The unit is dB. The location of wear can be identified by comparing the sound pressure levels in different areas.

[0151] Frequency band energy distribution: The sound energy from circumferential wear is concentrated in a specific frequency band. Let the total frequency band energy be... Main frequency band [f a ,f b Energy is The energy percentage is:

[0152]

[0153] This indicator can quantify the energy distribution characteristics of wear and tear sound.

[0154] Temperature feature extraction

[0155] High-temperature zone identification: The K-means clustering algorithm is used to cluster the circumferential temperature data, and the sample point T is defined. j With cluster center c kThe distance is:

[0156] d(T j ,c k )=|T j -c k |

[0157] By iteratively optimizing the cluster centers, the temperature data is divided into high-temperature clusters and low-temperature clusters. After clustering, the high-temperature cluster centers are c. high The corresponding area is the high-temperature wear zone.

[0158] High and low temperature difference: The temperature difference between the high-temperature zone and the low-temperature zone can directly reflect the degree of circumferential wear. The average temperature of the high-temperature zone is μ high With the average temperature μ in the low temperature region low The difference:

[0159] ΔT HL =μ high -μ low

[0160] S4 sound and temperature fusion diagnostic model training

[0161] Single sensor features are insufficient to fully characterize wear faults. By fusing features and training models, the advantages of multi-source information can be complemented, thereby improving diagnostic accuracy.

[0162] S4.1 Feature Layer Fusion

[0163] Feature splicing: Integrating axial and circumferential acoustic and temperature features into a unified feature vector, let the axial acoustic feature vector be... Axial temperature characteristics are Fa t =[μT i ,r i ]; Peripheral acoustic characteristics are Circumferential temperature characteristics are The fused feature vector is:

[0164]

[0165] Attention-based weighting: Different features contribute differently to fault diagnosis; an attention mechanism is used to highlight key features. Let the feature dimension be D, and the attention weight be α. d Calculated using the softmax function:

[0166]

[0167] Among them, w d For learnable weight parameters (dimensionless), F d Let be the d-th eigenvalue. The weighted eigenvector is:

[0168]

[0169] The purpose of this process is to increase attention to fault-sensitive features (such as axial periodic acoustic differences and circumferential high and low temperature differences) and suppress irrelevant interference.

[0170] S4.2 Diagnostic Model Training

[0171] Random Forest Model: This model uses random forests as the core classification model, which has the advantages of strong resistance to overfitting and stability in handling high-dimensional features. Let the forest contain K decision trees, and for the input feature F... att The output class prediction h of the Kth tree k (F att The final classification result is determined through a majority voting mechanism.

[0172]

[0173] Where I(·) is an indicator function (1 represents true, 0 represents false), {normal, axial, circumferential} represents the three categories to be classified by the model, namely "normal", "axial" and "circumferential", and arg max represents the category c that makes the following expression take the maximum value.

[0174] Model optimization metrics: The number of decision trees K in the model parameters is optimized using the five-fold cross-validation method, and the validation set accuracy is used as the core evaluation metric.

[0175]

[0176] TP (true positive), TN (true negative), FP (false positive), and FN (false negative) are confusion matrix elements, with the goal of achieving a validation set accuracy of Acc ≥ 95% to ensure the model's generalization ability.

[0177] S5 Fault Diagnosis and Verification Mathematical Model

[0178] By building a complete diagnostic process and validating model performance, the effectiveness of the method in real-world scenarios can be ensured.

[0179] S5.1 Fault Diagnosis Process

[0180] During real-time diagnosis, the acquired sound and temperature signals are first preprocessed according to step S2 to remove noise and outliers. Then, the feature extraction method in step S3 is used to obtain the fused feature vector to be diagnosed. Finally, the features are input into the trained fusion diagnostic model, and the model outputs the fault diagnosis result (normal state, axial wear fault, circumferential wear fault) and the corresponding confidence score. The fault confidence score is defined as the percentage of votes cast for that category in the random forest.

[0181]

[0182] Confidence level can quantify the reliability of diagnostic results and assist in on-site decision-making.

[0183] S5.2 Verification of Mathematical Model

[0184] To comprehensively evaluate the performance of diagnostic methods, accuracy and recall were used as the core validation metrics:

[0185] Accuracy: Similar to model optimization metrics, it reflects the overall diagnostic accuracy. The calculation formula is shown in step S4.

[0186] Recall: For a specific fault category c, recall is defined as follows:

[0187]

[0188] TP(c) represents the number of true positives in category c, and FN(c) represents the number of false negatives in category c. This indicator reflects the fault detection capability and ensures that critical faults are not missed.

[0189] After adopting the above method, the accuracy of the present invention is significantly better than that of the prior art, as shown in Table 1 below:

[0190] Table 1. Comparison of Accuracy of Different Diagnostic Methods

[0191]

[0192] This table compares the diagnostic accuracy of "existing technologies (single acoustic sensor + SVM, single temperature sensor + decision tree)" with "this invention (acoustic-temperature fusion + attention + random forest)". It is based on 1000 samples (300 normal, 350 axial, 350 circumferential), with each sample tested five times and the average value taken. The results show that the accuracy of this invention reaches 97.5%, an improvement of 15.5 percentage points compared to a single acoustic sensor (82.0%) and 19.5 percentage points compared to a single temperature sensor (78.0%), validating the advantages of dual-source fusion.

[0193] Meanwhile, the noise intensity and diagnostic accuracy of the present invention are significantly better than those of the prior art, as shown in Table 2 below:

[0194] Table 2. Relationship between noise intensity and diagnostic accuracy

[0195]

[0196] This table simulates the accuracy changes of three diagnostic methods under different environmental noise intensities (0-25dB). The results show that, due to the use of wavelet denoising, the accuracy of this invention remains ≥92% when the noise intensity is ≤20dB, while the accuracy of a single acoustic sensor (without denoising) drops to 75% at 15dB, and the accuracy of a single temperature sensor drops to 60% at 20dB, verifying the strong anti-interference capability of this invention.

[0197] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape, principle and application direction of this application should be covered within the scope of protection of this application.

Claims

1. A roller shell breakage fault diagnosis method based on distributed sound and temperature fusion monitoring, characterized in that, Comprising the following steps: S1 sound, temperature data acquisition: reasonably arrange sound sensors and temperature sensors around the belt conveyor, collect the sound signal and temperature signal of the carrier roller; S2 data preprocessing: preprocessing the collected sound signal and temperature signal, realizing wavelet transform denoising, temperature data cleaning and data standardization; S3 wear feature extraction: axial wear feature extraction and circumferential wear feature extraction are performed on the preprocessed sound signal and temperature signal; S4 sound, temperature fusion diagnosis model training: through feature layer fusion and diagnosis model training on the data after wear feature extraction, the advantages of multi-source information are realized; S5 fault diagnosis and verification mathematical model: build a complete fault diagnosis process and verify the performance of the mathematical model.

2. The method according to claim 1, wherein the method is characterized by, In step S1, the sound signal and temperature signal of the carrier roller are collected as follows: Let the sound signal collection frequency be f s , and the temperature signal collection frequency be f t , in order to completely capture the periodic sound changes and dynamic trends of temperature of the axial wear, the following conditions must be met: f s ≥44100,Hz,f t ≥1,Hz The time sequence of sound signal collection is The time sequence of temperature signal collection is Both time stamps satisfy Ensure that the sound and temperature data at the same time point can be matched.

3. The method according to claim 2, wherein the method is characterized by, In step S2, the collected sound signal and temperature signal are preprocessed as follows: S2.1 wavelet transform denoising Let the original sound signal be x(t), which is composed of effective signal and noise: x(t)=s(t)+n(t) Where s(t) is the effective signal related to the wear of the carrier roller, n(t) is the noise component, and wavelet transform is used for denoising: where a is a scale factor, b is a translation factor, ψ(t) is a mother wavelet function, ψ * is its conjugate function, W x (a,b) is a wavelet coefficient; The noise reduction process is achieved by threshold processing: the wavelet coefficients W x (a,b) Set a reasonable threshold λ, keep the coefficients whose amplitude is greater than or equal to the threshold, eliminate the low-amplitude noise coefficients, and then reconstruct the signal through inverse wavelet transform Noise suppression is achieved; S2.2 temperature data cleaning S2.2.1 Outlier Identification: Let the temperature sequence be T = [T1, T2, ..., T...]. n ] Calculate the sequence mean μ T and standard deviation σ T Outliers are identified using the 3σ criterion; if |T i -μ T |>3σ T Then determine T i Outliers were identified and removed. S2.2.2 interpolation of missing values: for the missing position t, linear interpolation method is used to complete, the formula is: Where t0<t<t1 is the adjacent effective sampling time of the missing point, and the continuity of the temperature data is ensured by linear fitting; S2.3 data standardization The min-max standardization is used to map the data to the [0,1] interval: Where x is the original feature value; In step S3, x min , x max is the minimum and maximum value of the feature sample; x norm is the standardized value.

4. The method according to claim 3, wherein the method is characterized by, S3.1 axial wear feature extraction includes the following steps: Sound feature extraction Fourier transform: through frequency domain analysis, the characteristic frequency related to wear can be identified, and the frequency domain representation of the sound signal s(t) is: Periodic amplitude difference: the axial wear presents periodic sound changes with the rotation of the roller. If the rotation speed of the roller is n, r / min, then the rotation period T0=60 / n. In the kth period, the amplitude of the sound signal of the wear area is The amplitude of the normal area is The difference between the two can reflect the degree of wear: Temperature feature extraction where f is the frequency, S(f) is the spectral amplitude, and f c The characteristic frequency f is defined as the frequency corresponding to the maximum spectral amplitude, which reflects the dominant frequency of the wear friction. The spectral energy distribution is characterized by the proportion of energy within the frequency band [f1, f2]: S3.2 circumferential wear feature extraction includes the following steps: The average temperature of each region: The local temperature will increase due to the axial wear, and the shaft is divided into M regions, and the temperature sequence of the ith region is T i = [T i1 , T i2 , ..., T iN ] , and the average value of the temperature of each region can reflect the temperature level of the region: Rate of change of temperature difference: During the wear development, the temperature difference between regions will change over time, the temperature difference between adjacent regions i and i+1 is ΔT i (t) = T i (t) - T i+1 (t), the rate of change of which can characterize the dynamic development trend of wear: where Δt is the time interval, r i (t) in units of °C / s; Sound feature extraction Sound pressure level: the friction sound pressure level of the circumferential wear area is significantly higher than that of the normal area, and the sound pressure level calculation formula of the circumferential j area is: Temperature feature extraction wherein p j is the area sound pressure effective value, p0=2x10 -5 Pa is the reference sound pressure, SPL j in dB, and the wear position can be identified by comparing the different area sound pressure levels; Band energy proportion: the sound energy of the circumferential wear is concentrated in a certain frequency band, and the total frequency band energy is Main frequency band [f a ,f b ] energy is Then the energy proportion is: In step S4, High temperature zone identification: K-means clustering algorithm is used to cluster the circumferential temperature data, and the distance between the sample point T j and the cluster center c k is defined as: d(T j ,c k ) = |T j -c k | By iteratively optimizing the cluster centers, the temperature data is divided into a high-temperature cluster and a low-temperature cluster, and the high-temperature cluster center c high The corresponding area is the high-temperature wear area. High-low temperature difference: the temperature difference between the high temperature zone and the low temperature zone can directly reflect the degree of circumferential wear, and the difference between the average temperature μ high of the high temperature zone and the average temperature μ low of the low temperature zone: ΔT HL = μ high - μ low .

5. The method according to claim 4, wherein the method is characterized by, S4.1 feature layer fusion includes the following steps: S4.2 diagnosis model training includes the following steps: Feature concatenation: integrate axial and circumferential acoustic and temperature features into a unified feature vector. Let the axial acoustic feature vector be The axial temperature feature is Fa t = [μT i , r i ]; the circumferential acoustic feature is The circumferential temperature feature is The fused feature vector is: Attention mechanism weighting: use attention mechanism to highlight key features, set feature dimension D, attention weight α d Calculate by softmax function: where w d is a learnable weight parameter, F d is the dthdimensional feature value, and the weighted feature vector is Where I(·) is the indicator function, {normal, axial, circumferential} represents the three categories of model classification, and arg max represents the category c corresponding to the maximum value of the expression behind it; Random forest model: Random forest is used as the core model of classification. The forest contains K decision trees. For input features F att , the Kth tree outputs the category prediction h k (F att ). The final classification result is determined by the majority voting mechanism. Model optimization index: five-fold cross-validation method is used to optimize the number of decision trees K in the model parameters, and the verification set accuracy is used as the core evaluation index: Where TP, TN, FP, FN are the elements of the confusion matrix, and the target is to make the verification set accuracy Acc≥95%. In step S5, 6. The method according to claim 5, wherein the method is characterized by, S5.1 fault diagnosis process steps are as follows: ​ In real-time diagnosis, firstly, the collected acoustic and temperature signals are preprocessed to remove noise and outliers according to the method of step S2; then the fusion feature vector to be diagnosed is obtained through the feature extraction method of step S3; finally, the features are input into the trained fusion diagnosis model, and the model outputs the fault diagnosis result and the corresponding confidence. The fault confidence is defined as the proportion of votes of this category in the random forest: S5.2 Verification of mathematical model The steps are as follows: The accuracy and recall rate are used as the core verification indicators: Accuracy: Same as the model optimization indicator, reflecting the overall diagnosis accuracy; Recall rate: For a specific fault category c, the recall rate is defined as: Where TP(c) is the true positive number of category c, and FN(c) is the false negative number of category c.

Images