Belt machine auscultation anomaly management method based on contrast learning and time series modeling

Abstract

The present application relates to the technical field of belt conveyor auscultation, and discloses a belt conveyor auscultation abnormality management method based on contrast learning and time sequence modeling, comprising the following steps: step 1, obtaining running sound and equipment parameters, extracting frame features and obtaining a voiceprint baseline; step 2, calculating a synchronous frequency of a carrier roller, constructing a cycle and a sideband fingerprint prototype vector; step 3, inputting the frame features and the fingerprint prototype vector into a contrast learning network to obtain a contrast embedding vector and a time sequence feature vector; step 4, a time sequence modeling network outputs three types of probabilities and a classification original output value; step 5, calculating an abnormality score, generating an abnormal event and determining a level; step 6, calculating a priority score and dispatching a work order; and step 7, backfilling disposal information, calculating a total cost and storing in a database. The present application realizes accurate detection, hierarchical management and operation and maintenance closed-loop processing of belt conveyor operation abnormalities.

Description

Technical Field

[0001] This invention belongs to the field of belt conveyor auscultation technology, specifically relating to a belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling. Background Technology

[0002] Belt conveyors are widely used in the transportation of bulk materials in industries such as coal mining, metallurgy, and power. Their operational stability directly affects the continuity and safety of production. During long-term operation, belt conveyors are prone to malfunctions such as idler wear, bearing damage, or belt misalignment. Failure to detect and address these issues promptly often leads to equipment downtime and significant economic losses. Current condition monitoring primarily relies on manual inspections or vibration signal-based methods. Manual inspections are limited by personnel experience, resulting in low efficiency, strong subjectivity, and delayed detection. Vibration signal-based methods require the installation of numerous sensors, leading to high deployment and maintenance costs, and suffer from low signal-to-noise ratios in noisy environments, making it difficult to identify early anomalies promptly and accurately.

[0003] In recent years, sound signals have been increasingly used for belt conveyor condition monitoring. However, existing acoustic methods mostly rely on simple feature threshold judgments or traditional classification models, failing to fully utilize the temporal correlation features and spectral modulation information in the operating sounds. Furthermore, due to the complex operating environment of belt conveyors, with frequent changes in load and belt speed, existing methods lack robustness under different operating conditions, easily leading to false alarms or missed alarms. In addition, existing technologies typically only focus on anomaly detection, lacking hierarchical management of abnormal events and effective linkage with maintenance work orders and cost assessments, making it difficult to meet the actual needs of intelligent and refined management in production sites. Summary of the Invention

[0004] This invention provides a belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling, which solves the technical problems of insufficient anomaly detection accuracy, poor robustness across working conditions, and lack of anomaly classification and closed-loop operation and maintenance management in related technologies.

[0005] This invention provides a method for managing auscultatory abnormalities in belt conveyors based on contrastive learning and temporal modeling, comprising the following steps:

[0006] Step 1: Obtain the operating sound data and equipment parameters of the belt conveyor; extract frame features based on the operating sound data; and obtain the acoustic baseline under normal operating conditions. The equipment parameters include the idler diameter and belt speed.

[0007] Step 2: Calculate the roller synchronization frequency based on the equipment parameters, and construct the period and side-band fingerprint prototype vector, where the period and side-band fingerprint prototype vector includes the modulation ratio and kurtosis ratio;

[0008] Step 3: Input the frame features and periodicity with the sideband fingerprint prototype vector into the contrast learning network, perform contrast learning according to the set data augmentation and pairing rules, and generate contrast embedding vectors and temporal feature vectors;

[0009] Step 4: Input the temporal feature vector into the temporal modeling network, and output the probabilities of three classes: normal, mildly abnormal, and severely abnormal, as well as the original classification output values.

[0010] Step 5: Calculate the anomaly score based on the three probabilities, the original output value of the classification, the period and the side-band fingerprint prototype and the voiceprint baseline, generate anomaly events based on the anomaly scores, calculate the event intensity, and determine the anomaly event level according to the set threshold.

[0011] Step 6: Calculate the priority score based on the abnormal event level, event intensity, equipment criticality, and real-time load, and generate and dispatch work orders according to the set service level agreement.

[0012] Step 7: The maintenance personnel complete the on-site handling according to the work order and fill in the handling information. They calculate the total cost based on the handling information and downtime, and store the handling information and total cost in the database.

[0013] Furthermore, frame features are extracted based on the running sound data, and a voiceprint baseline is obtained under normal operating conditions, including:

[0014] Step 11: Divide the running sound data into frames according to the first preset frame length and the second preset frame shift. Perform a fast Fourier transform on each frame and calculate the Mel frequency cepstral coefficients through the Mel filter bank. At the same time, extract the instantaneous energy in each frame to obtain the combination of Mel frequency cepstral coefficients and instantaneous energy.

[0015] Step 12: Calculate the first-order difference and the second-order difference based on the Mel frequency cepstral coefficients, and combine the Mel frequency cepstral coefficients, the first-order difference, the second-order difference, and the instantaneous energy into a frame feature set; wherein, the first-order difference is obtained by the difference between adjacent frame features, and the second-order difference is obtained by the difference between adjacent first-order differences.

[0016] Step 13: Under normal operating conditions, collect a set of frame features for multiple consecutive frames. The set of frame features includes energy values ​​for each frequency dimension obtained by the Mel filter bank. Calculate the median of the energy value sequence for each frequency dimension as the reference energy level for that frequency, and calculate the difference between the 95th percentile of the energy value sequence for that frequency and the median as the upper limit of the energy deviation for that frequency.

[0017] Step 14: Form an ordered pair between the reference energy level and the upper limit of energy deviation for each frequency, and then form the acoustic baseline by combining all ordered pairs.

[0018] Furthermore, based on the equipment parameters, the idler synchronization frequency is calculated, and a prototype of the cycle and side-band fingerprint is constructed, including:

[0019] Step 21: Calculate the circumference based on the roller diameter and pi, and use the ratio of belt speed to circumference as the roller synchronization frequency.

[0020] Step 22: Divide the key frequency bands with the idler synchronization frequency as the center, and select the frequency corresponding to the energy peak as the modulation frequency; the positive band and the negative band are the idler synchronization frequency plus or minus the modulation frequency, respectively.

[0021] Step 23: Calculate the ratio of the sum of frequency component energy in the positive and negative band ranges to the sum of frequency component energy in the main frequency band range to obtain the modulation ratio. The main frequency band is the range of preset percentages above and below the idler roller synchronization frequency.

[0022] Step 24: Calculate the ratio of the kurtosis of the main frequency band signal to the kurtosis of the full frequency range signal to obtain the kurtosis ratio; where kurtosis is the ratio of the fourth moment of the signal deviating from the mean to the square of the variance.

[0023] Step 25: Combine the modulation ratio and kurtosis ratio to form the periodic and sideband fingerprint prototype vectors.

[0024] Furthermore, a comparison is made between the generation of embedding vectors and temporal feature vectors, including:

[0025] Step 31: Perform data augmentation processing on the frame feature set, the data augmentation including random time axis shifting and band-limited noise mixing;

[0026] Step 32: Construct paired samples based on the enhanced frame feature set and the period and side-band fingerprint prototype vector. Positive sample pairs are feature sequences of the same device under the same operating conditions, while negative sample pairs are feature sequences of different devices or under different operating conditions. The operating conditions are determined based on the classification results of the device's load and belt speed.

[0027] Step 33: Input the paired samples into the encoder and the projector. The encoder includes a one-dimensional convolutional layer and a gated recurrent unit for extracting temporal feature vectors. The projector includes a multi-layer fully connected network for compressing the temporal feature vectors to obtain contrast embedding vectors.

[0028] Step 34: Construct a loss function and optimize the parameters of the encoder and projector. The loss function includes contrast loss and periodic regularization loss, which are weighted and combined by preset weights. The contrast loss is calculated based on the similarity between the contrast embedding vectors of positive sample pairs and the contrast embedding vectors of negative sample pairs. The periodic regularization loss is calculated based on the temporal feature vectors of the same signal before and after phase shift.

[0029] Step 35: The encoder and projector processed in step 34 are used for forward computation to output the contrast embedding vector and temporal feature vector, which are used as inputs to the temporal modeling network.

[0030] Furthermore, the data enhancement also includes frequency interval masking and time frame masking. Frequency interval masking refers to randomly selecting a continuous frequency interval in the frequency dimension of the frame feature set and setting the instantaneous energy of the interval to zero. Time frame masking refers to randomly selecting a continuous time frame in the time dimension of the frame feature set and setting the Mel frequency cepstral coefficients, first-order differences, second-order differences, and instantaneous energy of the selected time frame in each frequency dimension to zero.

[0031] Furthermore, the generation of the three types of probabilities includes:

[0032] Step 41: The optimized temporal feature vector is segmented, and the temporal feature vectors of consecutive frames are divided into segment-level feature sequences according to a preset length.

[0033] Step 42: Input the segment-level feature sequence into the branch network for decomposition. The low-frequency trend branch extracts long-term dependency features through gated recurrent units, and the high-frequency impulse branch extracts local impulse features through convolution operations and gating mechanisms. The long-term dependency features and local impulse features are concatenated on the same feature dimension to form an enhanced feature vector.

[0034] Step 43: Input the enhanced feature vector into the classifier, and output the original output value vector containing normal, mildly abnormal and severely abnormal classification values. Then, normalize the vector using the softmax function to obtain the probabilities of the three classes.

[0035] Furthermore, the calculation of anomaly scores and the determination of anomaly event levels include:

[0036] Step 51: The difference between 1 and the normal probability is taken as the classification contribution, the part of the difference between the sum of squares of the original output value vectors of two adjacent classification segments that is greater than zero is taken as the spectral flux contribution, the modulation ratio is taken as the sideband contribution, and the three parts are weighted and summed according to the preset weights to obtain the anomaly score.

[0037] Step 52: Detect abnormal score sequences based on continuous segments. Calculate the centered abnormal score by subtracting it from the upper limit of the abnormal score, and perform a one-sided cumulative sum on the centered abnormal score. When the one-sided cumulative sum exceeds the preset deviation threshold and continues for more than three consecutive segments, it is determined to be an abnormal event.

[0038] Step 53: In the window corresponding to the abnormal event, first perform an exponential transformation on the abnormal score, then take the mean and natural logarithm of the transformed score to obtain the event intensity.

[0039] Step 54: When the event intensity is below the first threshold, it is determined to be normal; when the event intensity is between the first threshold and the second threshold, it is determined to be slightly abnormal; when the event intensity is not lower than the second threshold, it is determined to be severely abnormal.

[0040] Furthermore, step 6 specifically includes:

[0041] Step 61: Obtain the abnormal event level, event intensity, device criticality, and real-time load, and obtain the priority score by weighted summation;

[0042] Step 62: Perform service level agreement mapping based on priority scores and abnormal event levels;

[0043] Step 63: Under the premise of satisfying skill matching, the generated work order is dispatched to the maintenance personnel. The skill matching is based on the correspondence between the skill tags required by the work order and the skill tags of the maintenance personnel, and the shortest path distance between the work order location and the current location of the maintenance personnel is used as the dispatch priority condition.

[0044] Furthermore, the disposal information includes disposal time, material consumption, replacement part codes, and disposal results;

[0045] The equipment recovery time is determined based on the processing time of the handling information, and the downtime is obtained by combining the downtime start time;

[0046] The total cost includes: fixed costs, material costs, and downtime costs corresponding to the work order type. Downtime costs are obtained by multiplying the downtime loss rate per unit time by the downtime duration.

[0047] The beneficial effects of this invention are as follows: At the feature level, the invention introduces the construction of frame features and acoustic print baselines, which can accurately reflect the acoustic characteristics of normal equipment operation; by establishing periodic and side-band fingerprint prototype vectors, the device operation mechanism is combined with spectral features, improving the physical rationality and interpretability of feature representation; at the modeling level, robust representation across operating conditions is achieved using contrastive learning, and the model's dual sensitivity to trend anomalies and sudden anomalies is enhanced through the decoupling and fusion of low-frequency trend branches and high-frequency impact branches; at the decision-making level, hierarchical calculation of anomaly scores, anomaly events, and event intensity enables quantitative identification and hierarchical judgment of anomalies; further, combining equipment criticality, real-time load, and production schedule to calculate priority scores ensures the rationality of work order dispatch and the timeliness of response. Overall, this invention achieves closed-loop management from belt conveyor sound signal acquisition, anomaly detection, operation and maintenance handling, and cost archiving, significantly improving the intelligence and refinement level of belt conveyor operation and maintenance, and reducing downtime risks and maintenance costs. Attached Figure Description

[0048] Figure 1This is a flowchart of the belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling of the present invention. Detailed Implementation

[0049] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0050] like Figure 1 As shown, the belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling includes the following steps:

[0051] Step 1: Obtain the operating sound data and equipment parameters of the belt conveyor; extract frame features based on the operating sound data; and obtain the acoustic baseline under normal operating conditions. The equipment parameters include the idler diameter and belt speed.

[0052] Step 2: Calculate the roller synchronization frequency based on the equipment parameters, and construct the period and side-band fingerprint prototype vector, where the period and side-band fingerprint prototype vector includes the modulation ratio and kurtosis ratio;

[0053] Step 3: Input the frame features and periodicity with the sideband fingerprint prototype vector into the contrast learning network, perform contrast learning according to the set data augmentation and pairing rules, and generate contrast embedding vectors and temporal feature vectors;

[0054] Step 4: Input the temporal feature vector into the temporal modeling network, and output the probabilities of three classes: normal, mildly abnormal, and severely abnormal, as well as the original classification output values.

[0055] Step 5: Calculate the anomaly score based on the three probabilities, the original output value of the classification, the period and the side-band fingerprint prototype and the voiceprint baseline, generate anomaly events based on the anomaly scores, calculate the event intensity, and determine the anomaly event level according to the set threshold.

[0056] Step 6: Calculate the priority score based on the abnormal event level, event intensity, equipment criticality, and real-time load, and generate and dispatch work orders according to the set service level agreement.

[0057] Step 7: The maintenance personnel complete the on-site handling according to the work order and fill in the handling information. They calculate the total cost based on the handling information and downtime, and store the handling information and total cost in the database.

[0058] In one embodiment of the present invention, the operating sound data refers to the original sound signal collected by a microphone during the normal operation of the belt conveyor;

[0059] Frame features are extracted from the operating sound data, and a voiceprint baseline is obtained under normal operating conditions, including:

[0060] Step 11: Divide the running sound data into frames according to the first preset frame length and the second preset frame shift. Perform a fast Fourier transform on each frame and calculate the Mel frequency cepstral coefficients using the Mel filter bank. At the same time, extract the instantaneous energy in each frame to obtain the combination of Mel frequency cepstral coefficients and instantaneous energy. Instantaneous energy refers to the sum of squares of the amplitude of the signal in each frame, which can reflect the intensity of the signal in that time segment.

[0061] Step 12: Calculate the first-order difference and the second-order difference based on the Mel frequency cepstral coefficients, and combine the Mel frequency cepstral coefficients, the first-order difference, the second-order difference, and the instantaneous energy into a frame feature set; wherein, the first-order difference is obtained by the difference between adjacent frame features, and the second-order difference is obtained by the difference between adjacent first-order differences.

[0062] Step 13: Under normal operating conditions, collect a set of frame features for multiple consecutive frames. The set of frame features includes energy values ​​for each frequency dimension obtained by the Mel filter bank. Calculate the median of the energy value sequence for each frequency dimension as the reference energy level for that frequency, and calculate the difference between the 95th percentile of the energy value sequence for that frequency and the median as the upper limit of the energy deviation for that frequency.

[0063] Step 14: The reference energy level and the upper limit of energy deviation for each frequency are arranged into ordered pairs, and all ordered pairs are used to form the acoustic waveform baseline, which is used to characterize the set of statistical features of the sound signal of the belt conveyor under normal operating conditions, and to provide a reference for subsequent anomaly detection.

[0064] In one embodiment of the present invention, the process of calculating the idler synchronization frequency based on equipment parameters and constructing a cycle and side-band fingerprint prototype includes:

[0065] Step 21: Calculate the circumference based on the roller diameter and pi, and use the ratio of belt speed to circumference as the roller synchronization frequency.

[0066] Step 22: Divide the key frequency band with the idler synchronization frequency as the center, and select the frequency corresponding to the energy peak as the modulation frequency. When there are multiple candidate peaks, search in order of energy from high to low and select the frequency that forms the most significant side band pair with the idler synchronization frequency as the modulation frequency. If it cannot be determined, take the one with the largest energy. The positive side band and the negative side band are the idler synchronization frequency plus or minus the modulation frequency, respectively.

[0067] Step 23: Calculate the ratio of the sum of frequency component energy in the positive and negative side bands to the sum of frequency component energy in the main frequency band to obtain the modulation ratio. The main frequency band is the range of preset percentages above and below the idler synchronous frequency. The modulation ratio reflects the degree of distribution of main frequency band energy to side bands and is used to reflect the modulation effect during idler operation.

[0068] Step 24: Calculate the ratio of the kurtosis of the main frequency band signal to the kurtosis of the full frequency range signal to obtain the kurtosis ratio; where kurtosis is the ratio of the fourth moment of the signal deviating from the mean to the square of the variance.

[0069] Step 25: Combine the modulation ratio and kurtosis ratio to form a prototype vector of period and sideband fingerprint. This vector not only contains information on energy distribution and modulation effect, but also comprehensively characterizes the pulse characteristics of the signal.

[0070] In one embodiment of the present invention, the generation of the embedded vector and the temporal feature vector is compared, including:

[0071] Step 31 involves data augmentation of the frame feature set. This data augmentation includes random time axis shifting and band-limited noise mixing. Specifically, random time axis shifting refers to randomly shifting the entire feature sequence along the time dimension to eliminate phase differences caused by different sampling start points. Band-limited noise mixing involves superimposing random noise within a preset frequency range to enhance the model's robustness in noisy environments. The augmented frame feature set can more comprehensively cover the sound feature distribution under actual operating conditions.

[0072] Step 32: Construct paired samples based on the enhanced frame feature set and the period and side-band fingerprint prototype vector. Positive sample pairs are feature sequences of the same device under the same operating conditions, while negative sample pairs are feature sequences of different devices or under different operating conditions. The operating conditions are determined based on the classification results of the device's load and belt speed. This method can ensure that the contrastive learning process captures both the consistency of the device's internal mechanism and the differences across devices.

[0073] Step 33: Input the paired samples into the encoder and the projector. The encoder includes a one-dimensional convolutional layer and a gated recurrent unit for extracting temporal feature vectors. The projector includes a multi-layer fully connected network for compressing the temporal feature vectors to obtain contrast embedding vectors.

[0074] Step 34: Construct a loss function and optimize the parameters of the encoder and projector. The loss function includes contrastive loss and periodic regularization loss, which are weighted and combined using preset weights. The contrastive loss is calculated based on the similarity between the contrastive embedding vectors of positive sample pairs and the contrastive embedding vectors of negative sample pairs. The periodic regularization loss is calculated based on the temporal feature vectors of the same signal before and after phase shift. Specifically, the formula for calculating the loss function is:

[0075] ;

[0076] in, This represents the value of the loss function. and Let represent the first and second weighting coefficients, respectively, used to balance the contributions of the contrastive loss and the periodicity regularization loss. P represents the positive sample pair. This represents the set containing the corresponding positive sample j and all negative samples, with sample i as the anchor. and Let represent the contrast embedding vectors of the i-th and j-th positive samples, respectively. The coefficient represents the temperature coefficient, used to adjust the smoothness of the distribution. i and j represent the indices of the positive samples, and k represents the index of the negative samples. Let N represent the contrastive embedding vector of the k-th negative sample, and N represent the number of samples. Represents the cosine similarity function. express and cosine similarity, This represents the m-th time-series feature vector. This represents the time-series feature vector obtained after phase shifting the same signal. express The square of the resulting Euclidean distance;

[0077] Step 35: The encoder and projector processed in step 34 are used for forward computation to output the contrast embedding vector and temporal feature vector, which are used as inputs to the temporal modeling network.

[0078] Through the above optimization process, this embodiment can obtain the optimized contrast embedding vector and temporal feature vector, and use them as the input of the subsequent temporal modeling network. By introducing paired samples and loss functions, it can simultaneously ensure the consistent modeling of the equipment mechanism features in the embedding space and the ability to distinguish differences in different working conditions. Through periodic regularization, the model becomes insensitive to phase changes caused by roller rotation, thereby enhancing its robustness across working conditions.

[0079] In one embodiment of the present invention, the data augmentation further includes frequency interval masking and time frame masking. Frequency interval masking refers to randomly selecting consecutive frequency intervals in the frequency dimension of the frame feature set and setting the instantaneous energy of those intervals to zero. Mel frequency cepstral coefficients, first-order differences, and second-order differences are not organized according to the frequency dimension and do not need to be set to zero. The frequency dimension of the frame feature set is composed of multiple frequency channels output by the Mel filter bank, each channel corresponding to an energy component within a specific frequency range. By masking consecutive frequency channels, the loss of some frequency band signals can be simulated, allowing the model to learn independently without relying on features of a specific frequency band, thereby improving robustness under different noise environments. Time frame masking refers to randomly selecting consecutive time frames in the time dimension of the frame feature set and setting the Mel frequency cepstral coefficients, first-order differences, second-order differences, and instantaneous energy of the selected time frames to zero in each frequency dimension. The time dimension is composed of a sequence of frames formed by dividing the running sound data according to a preset frame length and frame shift, with each frame corresponding to a sound feature of a time segment. By masking consecutive time frames, we can simulate situations where signals are missing or interfered with in certain time segments, enabling the model to learn to maintain its ability to recognize the overall signal pattern even under conditions of missing data during training.

[0080] By introducing frequency range masking and temporal frame masking, the model can avoid over-reliance on specific frequency features or local temporal features, thereby improving the robustness of contrastive learning networks in the face of signal loss, noise interference, or operating condition fluctuations. This masking mechanism complements the aforementioned random shift of the time axis and band-limited noise mixing, and can simultaneously improve the model's generalization ability and stability in both the frequency and time domains.

[0081] In one embodiment of the present invention, the generation of the three types of probabilities includes:

[0082] Step 41: The optimized temporal feature vector is segmented, and the temporal feature vectors of consecutive frames are divided into segment-level feature sequences according to a preset length.

[0083] Step 42: The segment-level feature sequence is input into the branch network for decomposition. The low-frequency trend branch extracts long-term dependency features through gated recurrent units, while the high-frequency impulse branch extracts local impulse features through convolution operations and gating mechanisms. The long-term dependency features and local impulse features are concatenated on the same feature dimension to form an enhanced feature vector. Through this decomposition and fusion structure, not only can long-term and short-term features be taken into account, but feature complementarity and information enhancement can also be achieved in a unified feature space.

[0084] Step 43: Input the enhanced feature vector into the classifier, and output the original output value vector containing normal, mildly abnormal and severely abnormal classification values. Then, normalize the vector using the softmax function to obtain the probabilities of the three classes.

[0085] This embodiment, through the synergistic effect of low-frequency trend branches and high-frequency impact branches, can simultaneously capture long-term evolutionary features and instantaneous pulse features, and achieve fusion in the enhanced feature vector; by combining a classifier with softmax normalization, the enhanced features can be accurately mapped into three probabilities: normal, mildly abnormal, and severely abnormal, thus providing a reliable basis for subsequent abnormal event identification and hierarchical processing.

[0086] It should be noted that step 33 focuses on jointly representing the frame feature set and the periodic fingerprint prototype vector. Specifically, it extracts a general temporal feature vector through convolution and recurrent structures, and further compresses it in the projector to form a contrastive embedding vector. The core of this process is to establish a unified embedding representation space so as to achieve cross-device and cross-operating condition discrimination capability through contrastive loss. Step 42 performs branch decoupling processing based on the optimized temporal feature vector; the low-frequency trend branch is used to extract long-term dependent features, and the high-frequency impulse branch is used to capture local impulse features, and they are spliced ​​and fused on the same feature dimension to form an enhanced feature vector. The core of this process is feature enhancement for classification tasks, enabling the model to utilize both long-term trend information and short-term impulse information simultaneously.

[0087] In one embodiment of the present invention, the calculation of the anomaly score and the determination of the anomaly event level include:

[0088] Step 51: The difference between 1 and the normal probability is taken as the classification contribution. The part of the difference between the sum of squares of the original output value vectors of two adjacent segments is taken as the spectral flux contribution, which is used to measure the cumulative change of abnormal energy in the signal during the time-series evolution. The modulation ratio is taken as the sideband contribution, which is used to reflect the abnormal enhancement of the sideband components of the signal spectrum relative to the main frequency band. The three parts are weighted and summed according to the preset weights to obtain the abnormal score.

[0089] Step 52: Detect anomaly scores based on a continuous sequence of anomaly scores. A centered anomaly score is calculated by subtracting the upper limit of the anomaly score from the upper limit. A one-sided cumulative sum is then performed on the centered anomaly score. An anomaly event is identified when the one-sided cumulative sum exceeds a preset deviation threshold and persists for more than three consecutive segments. The upper limit of the anomaly score is obtained by adding the median to the percentile. The centered anomaly score is obtained by comparing the anomaly score of each segment with this upper limit and calculating the difference between the two. The one-sided cumulative sum means that when a segment's centered anomaly score is positive, it is added to the cumulative sum; when it is negative, the cumulative sum is set to zero. This method can amplify anomaly trends that continuously deviate from normal levels without causing misjudgment due to a single noise point.

[0090] Step 53: Within the window corresponding to the abnormal event, first perform an exponential transformation on the abnormal score, then take the mean and natural logarithm of the transformed score to obtain the event intensity; specifically, the formula for calculating the event intensity is: ,in, The event intensity is represented by Q, the number of consecutive segments contained within the abnormal time window is represented by e, and the natural logarithm is represented by e. This represents the anomalous score of the q-th segment within the anomalous time window. In this way, numerical scale compression can be achieved while performing nonlinear amplification, making the event intensity index more suitable for comparison with the threshold.

[0091] Step 54: When the event intensity is below the first threshold, it is determined to be normal; when the event intensity is between the first threshold and the second threshold, it is determined to be slightly abnormal; when the event intensity is not lower than the second threshold, it is determined to be severely abnormal.

[0092] This embodiment integrates classification contribution, spectral flux contribution, and sideband contribution from multiple dimensions. The anomaly score can not only reflect the degree of deviation of the classification result, but also characterize the trend of signal energy change and spectral modulation characteristics. By combining the constraints of cumulative deviation and threshold conditions, the false alarm rate can be effectively reduced. Through the nonlinear transformation and hierarchical judgment of event intensity, the quantitative characterization and hierarchical identification of abnormal events can be realized, thereby significantly improving the accuracy and robustness of belt conveyor anomaly detection and diagnosis.

[0093] In one embodiment of the present invention, step 6 specifically includes:

[0094] Step 61: Obtain the abnormal event level, event intensity, equipment criticality, and real-time load, and obtain the priority score by weighted summation; where, equipment criticality indicates the importance of the belt conveyor in the overall production system, which can be set according to the dependence of production links and the degree of downtime loss; real-time load refers to the operating load level of the belt conveyor at the current moment, reflecting its operating intensity and potential risks.

[0095] Step 62: Perform service level agreement mapping based on priority scores and abnormal event levels. Specifically, through preset service level agreement rules, different levels of abnormalities and priority scores are mapped to corresponding service timeliness and response requirements to ensure that high-priority and high-risk events are handled quickly, while low-priority events are allocated resources reasonably to improve overall maintenance efficiency and resource utilization.

[0096] Step 63: Under the premise of skill matching, the generated work order is dispatched to the maintenance personnel. The skill matching is based on the correspondence between the skill tags required for the work order and the skill tags of the maintenance personnel, and the shortest path distance between the work order location and the maintenance personnel's current location is used as the dispatch priority condition. The skill tags are feature markers used to describe the maintenance personnel's professional capabilities and the work order task requirements; skill tags include, but are not limited to, categories such as electrical repair, mechanical repair, vibration analysis, acoustic diagnosis, and safe operation.

[0097] This embodiment achieves intelligent and efficient handling of belt conveyor abnormal events by integrating priority scoring, service level agreement mapping, and skill matching and path optimization dispatch strategies. It can not only ensure priority response to abnormalities of critical equipment, but also improve the rationality of maintenance resource allocation, and significantly improve the response speed and processing efficiency of the overall operation and maintenance management of belt conveyors.

[0098] In one embodiment of the present invention, the handling information includes handling time, material consumption, replacement part code, and handling result; handling time is used to record the time from the arrival of maintenance personnel on site to the restoration of equipment operation; material consumption is used to record the types and quantities of materials used in the maintenance process; replacement part code is used to identify the specific part model that was replaced; and handling result is used to indicate whether the equipment has returned to normal operation or whether there are any residual hidden dangers.

[0099] The equipment recovery time is determined based on the processing time of the handling information, and the downtime is obtained by combining the downtime start time;

[0100] The total cost includes: fixed costs corresponding to the work order type, material costs, and downtime costs. Downtime costs are obtained by multiplying the downtime loss rate per unit time by the downtime duration. This method allows for the quantification of the direct and indirect economic impact of a single conveyor belt maintenance and repair process on production and operations.

[0101] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0102] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of the present embodiments, all of which are within the protection scope of the present embodiments.

Claims

1. A method for managing auscultatory abnormalities in belt conveyors based on contrastive learning and temporal modeling, characterized in that, Includes the following steps: Step 1: Obtain the operating sound data and equipment parameters of the belt conveyor; extract frame features based on the operating sound data; and obtain the acoustic baseline under normal operating conditions. The equipment parameters include the idler diameter and belt speed. Step 2: Calculate the roller synchronization frequency based on the equipment parameters, and construct the period and side-band fingerprint prototype vector, where the period and side-band fingerprint prototype vector includes the modulation ratio and kurtosis ratio; Step 3: Input the frame features and periodicity with the sideband fingerprint prototype vector into the contrast learning network, perform contrast learning according to the set data augmentation and pairing rules, and generate contrast embedding vectors and temporal feature vectors; Step 4: Input the temporal feature vector into the temporal modeling network, and output the probabilities of three classes: normal, mildly abnormal, and severely abnormal, as well as the original classification output values. Step 5: Calculate the anomaly score based on the three probabilities, the original output value of the classification, the period and the side-band fingerprint prototype and the voiceprint baseline, generate anomaly events based on the anomaly scores, calculate the event intensity, and determine the anomaly event level according to the set threshold. Step 6: Calculate the priority score based on the abnormal event level, event intensity, equipment criticality, and real-time load, and generate and dispatch work orders according to the set service level agreement. Step 7: The maintenance personnel complete the on-site handling according to the work order and fill in the handling information. They calculate the total cost based on the handling information and downtime, and store the handling information and total cost in the database.

2. The belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling according to claim 1, characterized in that, Frame features are extracted from the operating sound data, and a voiceprint baseline is obtained under normal operating conditions, including: Step 11: Divide the running sound data into frames according to the first preset frame length and the second preset frame shift. Perform a fast Fourier transform on each frame and calculate the Mel frequency cepstral coefficients through the Mel filter bank. At the same time, extract the instantaneous energy in each frame to obtain the combination of Mel frequency cepstral coefficients and instantaneous energy. Step 12: Calculate the first-order difference and the second-order difference based on the Mel frequency cepstral coefficients, and combine the Mel frequency cepstral coefficients, the first-order difference, the second-order difference, and the instantaneous energy into a frame feature set; wherein, the first-order difference is obtained by the difference between adjacent frame features, and the second-order difference is obtained by the difference between adjacent first-order differences. Step 13: Under normal operating conditions, collect a set of frame features for multiple consecutive frames. The set of frame features includes energy values ​​for each frequency dimension obtained by the Mel filter bank. Calculate the median of the energy value sequence for each frequency dimension as the reference energy level for that frequency, and calculate the difference between the 95th percentile of the energy value sequence for that frequency and the median as the upper limit of the energy deviation for that frequency. Step 14: Form an ordered pair between the reference energy level and the upper limit of energy deviation for each frequency, and then form the acoustic baseline by combining all ordered pairs.

3. The belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling according to claim 1, characterized in that, Calculate the idler synchronization frequency based on equipment parameters, and construct a cycle and side-band fingerprint prototype, including: Step 21: Calculate the circumference based on the roller diameter and pi, and use the ratio of belt speed to circumference as the roller synchronization frequency. Step 22: Divide the key frequency bands with the idler synchronization frequency as the center, and select the frequency corresponding to the energy peak as the modulation frequency; the positive band and the negative band are the idler synchronization frequency plus or minus the modulation frequency, respectively. Step 23: Calculate the ratio of the sum of frequency component energy in the positive and negative band ranges to the sum of frequency component energy in the main frequency band range to obtain the modulation ratio. The main frequency band is the range of preset percentages above and below the idler roller synchronization frequency. Step 24: Calculate the ratio of the kurtosis of the main frequency band signal to the kurtosis of the full frequency range signal to obtain the kurtosis ratio; where kurtosis is the ratio of the fourth moment of the signal deviating from the mean to the square of the variance. Step 25: Combine the modulation ratio and kurtosis ratio to form the periodic and sideband fingerprint prototype vectors.

4. The belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling according to claim 2, characterized in that, The generation of embedding vectors and temporal feature vectors is compared, including: Step 31: Perform data augmentation processing on the frame feature set, the data augmentation including random time axis shifting and band-limited noise mixing; Step 32: Construct paired samples based on the enhanced frame feature set and the period and side-band fingerprint prototype vector. Positive sample pairs are feature sequences of the same device under the same operating conditions, while negative sample pairs are feature sequences of different devices or under different operating conditions. The operating conditions are determined based on the classification results of the device's load and belt speed. Step 33: Input the paired samples into the encoder and the projector. The encoder includes a one-dimensional convolutional layer and a gated recurrent unit for extracting temporal feature vectors. The projector includes a multi-layer fully connected network for compressing the temporal feature vectors to obtain contrast embedding vectors. Step 34: Construct a loss function and optimize the parameters of the encoder and projector. The loss function includes contrast loss and periodic regularization loss, which are weighted and combined by preset weights. The contrast loss is calculated based on the similarity between the contrast embedding vectors of positive sample pairs and the contrast embedding vectors of negative sample pairs. The periodic regularization loss is calculated based on the temporal feature vectors of the same signal before and after phase shift. Step 35: The encoder and projector processed in step 34 are used for forward computation to output the contrast embedding vector and temporal feature vector, which are used as inputs to the temporal modeling network.

5. The belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling according to claim 4, characterized in that, The data augmentation also includes frequency interval masking and time frame masking. Frequency interval masking refers to randomly selecting a continuous frequency interval in the frequency dimension of the frame feature set and setting the instantaneous energy of the interval to zero. Time frame masking refers to randomly selecting a continuous time frame in the time dimension of the frame feature set and setting the Mel frequency cepstral coefficients, first-order differences, second-order differences and instantaneous energy of the selected time frame in each frequency dimension to zero.

6. The belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling according to claim 1, characterized in that, The generation of the three types of probabilities includes: Step 41: The optimized temporal feature vector is segmented, and the temporal feature vectors of consecutive frames are divided into segment-level feature sequences according to a preset length. Step 42: Input the segment-level feature sequence into the branch network for decomposition. The low-frequency trend branch extracts long-term dependency features through gated recurrent units, and the high-frequency impulse branch extracts local impulse features through convolution operations and gating mechanisms. The long-term dependency features and local impulse features are concatenated on the same feature dimension to form an enhanced feature vector. Step 43: Input the enhanced feature vector into the classifier, and output the original output value vector containing normal, mildly abnormal and severely abnormal classification values. Then, normalize the vector using the softmax function to obtain the probabilities of the three classes.

7. The belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling according to claim 1, characterized in that, The calculation of anomaly scores and the determination of anomaly event levels include: Step 51: The difference between 1 and the normal probability is taken as the classification contribution, the part of the difference between the sum of squares of the original output value vectors of two adjacent classification segments that is greater than zero is taken as the spectral flux contribution, the modulation ratio is taken as the sideband contribution, and the three parts are weighted and summed according to the preset weights to obtain the anomaly score. Step 52: Detect abnormal score sequences based on continuous segments. Calculate the centered abnormal score by subtracting it from the upper limit of the abnormal score, and perform a one-sided cumulative sum on the centered abnormal score. When the one-sided cumulative sum exceeds the preset deviation threshold and continues for more than three consecutive segments, it is determined to be an abnormal event. Step 53: In the window corresponding to the abnormal event, first perform an exponential transformation on the abnormal score, then take the mean and natural logarithm of the transformed score to obtain the event intensity. Step 54: When the event intensity is below the first threshold, it is determined to be normal; when the event intensity is between the first threshold and the second threshold, it is determined to be slightly abnormal; when the event intensity is not lower than the second threshold, it is determined to be severely abnormal.

8. The belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling according to claim 1, characterized in that, Step 6 specifically includes: Step 61: Obtain the abnormal event level, event intensity, device criticality, and real-time load, and obtain the priority score by weighted summation; Step 62: Perform service level agreement mapping based on priority scores and abnormal event levels; Step 63: Under the premise of satisfying skill matching, the generated work order is dispatched to the maintenance personnel. The skill matching is based on the correspondence between the skill tags required by the work order and the skill tags of the maintenance personnel, and the shortest path distance between the work order location and the current location of the maintenance personnel is used as the dispatch priority condition.

9. The belt conveyor auscultation anomaly management method based on contrastive learning and temporal modeling according to claim 1, characterized in that, The disposal information includes disposal time, material consumption, replacement part codes, and disposal results; The equipment recovery time is determined based on the processing time of the handling information, and the downtime is obtained by combining the downtime start time; The total cost includes: fixed costs, material costs, and downtime costs corresponding to the work order type. Downtime costs are obtained by multiplying the downtime loss rate per unit time by the downtime duration.

Images