A cloud-edge collaboration-based remote fault diagnosis method and system for a rice mill

Abstract

The application discloses a cloud-edge cooperation-based remote fault diagnosis method and system for rice millers, relates to the intelligent operation and maintenance technical field of agricultural machinery, and discloses the cloud-edge cooperation-based remote fault diagnosis method and system for rice millers. The disclosed method is characterized in that the original sensing signals are subjected to feature refining and pre-judgment processing through an edge device, and a low-dimensional feature vector is uploaded to a cloud platform only in an abnormal state, thereby effectively reducing data transmission volume and improving diagnosis efficiency, reducing data transmission bandwidth occupation and improving fault diagnosis real-time performance. Meanwhile, the application can also reduce redundant information transmission and improve cloud computing efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance of agricultural machinery, and particularly relates to a rice mill remote fault diagnosis method and system based on cloud-edge collaboration. BACKGROUND

[0002] In the field of remote monitoring of agricultural machinery, rice mill fault diagnosis is a key link to realize intelligent operation and maintenance of equipment. The traditional technical solution usually adopts a method of uploading the full amount of original vibration signals collected by sensors to the cloud server for processing. This method has significant limitations: first, the high-frequency vibration signals generated during the operation of the rice mill have the characteristics of large data volume and high sampling rate. Direct transmission of raw data will occupy a large amount of network bandwidth resources. Second, due to the unstable network environment in the agricultural operation site, large-scale data transmission is easy to cause communication delay, so that the fault diagnosis response time often reaches seconds or even tens of seconds, which cannot meet the real-time requirement. Third, the original sensor signals contain a large amount of redundant information. The transmission of invalid data not only wastes cloud computing resources, but also affects the diagnosis efficiency. In addition, the existing solution lacks effective edge preprocessing mechanism, which cannot realize feature refinement at the data source, resulting in that the cloud diagnosis model needs to process high-dimensional complex data, further increasing the computing burden and diagnosis delay.

[0003] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0004] The main purpose of the present application is to provide a rice mill remote fault diagnosis method and system based on cloud-edge collaboration, aiming to reduce data transmission bandwidth occupation and improve fault diagnosis real-time performance.

[0005] To achieve the above purpose, the present application provides a rice mill remote fault diagnosis method based on cloud-edge collaboration, which comprises:

[0006] Collecting original sensor signals of the rice mill based on the edge device;

[0007] Performing edge feature refinement processing on the original sensor signals based on the edge device to generate corresponding low-dimensional feature vectors;

[0008] Performing state classification on the low-dimensional feature vectors based on the edge device to generate a pre-judgment state label; the pre-judgment state label includes a normal label and an abnormal label;

[0009] When the pre-judgment state label is an abnormal label, the pre-judgment state label is attached to the low-dimensional feature vector and uploaded to the cloud platform;

[0010] Performing incremental migration diagnosis on the low-dimensional feature vector based on the cloud platform, generating a fault type and a confidence result, and delivering to an edge device and a user terminal.

[0011] In an embodiment, the step of performing edge feature refinement processing on the original sensing signal based on the edge device to generate a corresponding low-dimensional feature vector comprises:

[0012] Selecting an optimal wavelet basis from a set of preset candidate wavelet bases;

[0013] Performing wavelet packet decomposition on the original sensing signal using the selected optimal wavelet basis to generate a set of node signals;

[0014] Calculating the kurtosis and energy entropy values of each node signal, and screening a target node with the largest product of kurtosis and energy entropy;

[0015] Extracting the normalized energy entropy, spectral kurtosis, and root mean square value of the target node signal, and combining them to form the low-dimensional feature vector.

[0016] In an embodiment, the step of selecting an optimal wavelet basis from a set of preset candidate wavelet bases comprises:

[0017] For each candidate wavelet basis in the set of preset candidate wavelet bases, calculating the average kurtosis value and the average energy entropy value of each sub-band node at a preset decomposition level;

[0018] Weighted fusion calculation is performed on the average kurtosis value and the average energy entropy value of each candidate wavelet basis according to a preset weight ratio to obtain a joint criterion score;

[0019] Selecting the candidate wavelet basis with the highest joint criterion score as the optimal wavelet basis.

[0020] In an embodiment, the step of performing state classification on the low-dimensional feature vector based on the edge device to generate a pre-judgment state label comprises:

[0021] Extracting the root mean square value of the low-dimensional feature vector;

[0022] When the root mean square value is greater than a preset device baseline value, directly outputting an abnormal label;

[0023] When the root mean square value is less than the preset device baseline value, calculating the product between the normalized energy entropy and the spectral kurtosis in the feature vector to obtain a discriminant factor;

[0024] If the discriminant factor is greater than a dynamic label threshold, outputting an abnormal label; if the discriminant factor is not greater than the dynamic label threshold, outputting a normal label.

[0025] In an embodiment, the method further comprises:

[0026] Collecting a vibration signal with a preset time length of stable running with load when the device is first debugged, and extracting a root mean square value of the vibration signal, and taking an average value as the baseline value of the device;

[0027] Collecting a low-dimensional feature vector set during a normal working condition period on the same day, and calculating a discriminant factor of each feature vector in the low-dimensional feature vector set, and taking a preset quantile of the discriminant factor of each feature vector as a new dynamic threshold.

[0028] In an embodiment, the cloud platform-based low-dimensional feature vector performs incremental migration diagnosis, and the step of generating a fault type and a confidence result includes:

[0029] Inputting the low-dimensional feature vector into a pre-trained base model to output a primary feature vector;

[0030] Inputting the primary feature vector into an incremental adaptation layer, and calculating a same-class feature aggregation loss, a different-class feature separation loss, and a new and old data distribution alignment loss in combination with a historical feature library;

[0031] Weighted sum of the same-class feature aggregation loss, the different-class feature separation loss, and the new and old data distribution alignment loss is performed to update the weight of the incremental adaptation layer, so that the incremental adaptation layer outputs a reinforced feature vector;

[0032] Performing full connection classification on the reinforced feature vector to generate a fault type, and generating a confidence result based on the maximum value of the Softmax probability.

[0033] In an embodiment, the method further includes:

[0034] Generating a pseudo feature vector set by sampling a device historical feature library through a cloud platform;

[0035] Training a lightweight convolutional network with the pseudo feature vector set as input;

[0036] Weight pruning is performed on the trained lightweight convolutional network to remove channels with a contribution degree lower than a preset contribution degree threshold, generating the lightweight diagnostic model;

[0037] The lightweight diagnostic model is issued to an edge device to update the state classification standard of the low-dimensional feature vector by the edge device.

[0038] In an embodiment, the method further includes:

[0039] When the pre-judgment state label is a normal label, a low-dimensional feature vector in a random time window is retained in a preset polling period.

[0040] The sampled and retained low-dimensional feature vector is bound to the pre-judgment state label and uploaded to the cloud.

[0041] In an embodiment, the method further comprises:

[0042] If the uploading to the cloud platform fails, the edge device re-uploads the low-dimensional feature vector and the pre-judgment state label according to a preset retry interval and a maximum number of retries;

[0043] If the number of retries exceeds the preset number of times, the edge device stores the low-dimensional feature vector and triggers a local alarm signal, and simultaneously reports a failure log to a user terminal.

[0044] In addition, to achieve the above-mentioned purpose, the application further provides a rice mill remote fault diagnosis system based on cloud-edge collaboration, which comprises a memory, a processor, and a rice mill remote fault diagnosis program based on cloud-edge collaboration stored on the memory and capable of running on the processor. The rice mill remote fault diagnosis program based on cloud-edge collaboration is configured to implement the steps of the rice mill remote fault diagnosis method based on cloud-edge collaboration.

[0045] The rice mill remote fault diagnosis method and system based on cloud-edge collaboration provided by the application can effectively reduce data transmission volume and improve diagnosis efficiency by performing feature refining and pre-judgment processing on original sensing signals through an edge device, and uploading low-dimensional feature vectors to a cloud platform only in an abnormal state, thereby reducing data transmission bandwidth occupation and improving fault diagnosis real-time performance. Meanwhile, the application can also reduce redundant information transmission and improve cloud computing efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0046] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the application and, together with the specification, serve to explain the principles of the application.

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0048] Figure 1 A flowchart is provided for an embodiment of the rice mill remote fault diagnosis method based on cloud-edge collaboration of the application;

[0049] Figure 2 A structural diagram is provided for an embodiment of the rice mill remote fault diagnosis system based on cloud-edge collaboration of the application.

[0050] Explanation of the reference signs:

[0051] 10, memory; 20, processor.

[0052] The object, functional characteristics and advantages of the present application will be further explained in conjunction with the embodiments, with reference to the accompanying drawings. DETAILED DESCRIPTION

[0053] It should be understood that the specific embodiments described herein are merely intended to explain the technical solutions of the present application, and are not intended to limit the present application.

[0054] In order to better understand the technical solutions of the present application, the specific embodiments will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0055] The main solution of the embodiments of the present application is: based on the edge device to collect the original sensing signal of the rice mill; and based on the edge device to perform edge feature refining processing on the original sensing signal to generate a corresponding low-dimensional feature vector; then based on the edge device to perform state classification on the low-dimensional feature vector to generate a pre-judgment state label; the pre-judgment state label includes a normal label and an abnormal label; when the pre-judgment state label is an abnormal label, the pre-judgment state label is attached to the low-dimensional feature vector and uploaded to the cloud platform; finally, based on the cloud platform to perform incremental migration diagnosis on the low-dimensional feature vector to generate a fault type and a confidence result, and to issue to the edge device and the user terminal.

[0056] In the present embodiment, in order to facilitate description, the following will be described with the recognition of the cloud edge collaborative rice mill remote fault diagnosis system as the execution subject.

[0057] Since the prior art usually adopts the method of uploading the original vibration signal collected by the sensor to the cloud server for processing, this method has significant limitations: first, the high-frequency vibration signal generated during the operation of the rice mill has the characteristics of large data volume and high sampling rate, and direct transmission of the original data will occupy a large amount of network bandwidth resources; second, due to the unstable network environment at the agricultural operation site, large-scale data transmission is easy to cause communication delay, so that the fault diagnosis response time often reaches seconds or even tens of seconds, which cannot meet the real-time requirement; third, the original sensing signal contains a large amount of redundant information, and invalid data transmission not only wastes cloud computing resources, but also affects the diagnosis efficiency. In addition, the existing scheme lacks effective edge preprocessing mechanism, which cannot realize feature refining at the data source, resulting in that the cloud diagnosis model needs to process high-dimensional complex data, further increasing the computing burden and diagnosis delay.

[0058] The solution provided by the present application can effectively reduce the data transmission amount and improve the diagnosis efficiency by performing feature refining and pre-judgment processing on the original sensing signal by the edge device, and can reduce the data transmission bandwidth occupation and improve the fault diagnosis real-time performance. At the same time, the present application can also reduce the transmission of redundant information and improve the cloud computing efficiency.

[0059] It should be noted that the execution subject of the embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, a cloud-edge collaborative rice mill remote fault diagnosis system, etc. The following takes the cloud-edge collaborative rice mill remote fault diagnosis system as an example to illustrate the embodiment and the following embodiments.

[0060] In the prior art, the way of uploading sensor raw data in full to the cloud for artificial intelligence analysis is generally used in the field of agricultural machinery remote monitoring. Due to the large amount of high-frequency vibration signal data generated during the operation of the rice mill, in the environment of limited network bandwidth in remote areas, this data transmission method is easy to cause communication channel congestion, leading to significant delay in fault diagnosis, and a large amount of invalid data occupies transmission resources, affecting the overall operation efficiency of the system.

[0061] In order to solve the above problems, the researchers noticed that the edge computing device has local data processing capability, and tried to optimize the data transmission mechanism to improve the system response speed. Through analysis, it is found that the data generated during the normal operation of the rice mill accounts for the main part of the total transmission amount, and the abnormal data that really needs to be analyzed in depth by the cloud accounts for a limited proportion. Based on this, a preliminary diagnosis mechanism is established on the edge side, and only suspicious data triggers cloud collaborative analysis, and a feature dimension compression method is designed to reduce the transmission load.

[0062] Therefore, the embodiment of the present application provides a cloud-edge collaborative rice mill remote fault diagnosis method, which refers to Figure 1 In the embodiment, the cloud-edge collaborative rice mill remote fault diagnosis method includes steps S100-S500, wherein:

[0063] Step S100, collecting original sensing signals of the rice mill based on the edge device;

[0064] Step S200, performing edge feature refining processing on the original sensing signals based on the edge device to generate corresponding low-dimensional feature vectors;

[0065] Step S300, performing state classification on the low-dimensional feature vectors based on the edge device to generate a pre-judgment state label; the pre-judgment state label includes a normal label and an abnormal label;

[0066] Step S400, when the pre-judgment state label is an abnormal label, the pre-judgment state label is attached to the low-dimensional feature vector and uploaded to the cloud platform;

[0067] Step S500, performing incremental transfer diagnosis on the low-dimensional feature vector based on the cloud platform to generate fault type and confidence results, and delivering to the edge device and the user terminal.

[0068] In this embodiment, the edge feature refinement process is a process of converting original high-dimensional data into a simplified feature vector through signal decomposition and feature screening technology, which can be realized by wavelet packet decomposition combined with feature parameter calculation, effectively removing redundant information. The low-dimensional feature vector refers to a feature set that retains key state information after dimension reduction processing, which can be formed by combining standardized energy entropy, spectral kurtosis and root mean square value parameters, to ensure the feature representation ability while reducing the data dimension. The pre-judgment state label refers to the preliminary judgment result of the device running state based on local computing resources, which can be generated by setting a threshold comparison mechanism, used to trigger the transmission conditions of cloud diagnosis. Incremental migration diagnosis refers to the process of continuously optimizing the diagnosis model by the cloud platform using historical data, which can be realized by feature space transfer learning and model parameter updating method, to improve the adaptability to new fault modes.

[0069] In this embodiment, after the edge device continuously collects the rice mill vibration sensor signal, it first performs multi-scale signal decomposition and feature parameter calculation to generate a low-dimensional feature vector containing the core state information of the device. Then, through a threshold comparison mechanism, the feature vector is preliminarily classified, and when it is judged as an abnormal state, the feature data and classification label are transmitted to the cloud platform together. After receiving the data, the cloud platform maps the feature space and performs incremental training of the model, finally outputs the accurate fault type and diagnosis confidence, forming a closed-loop diagnosis process from edge rapid response to cloud deep analysis.

[0070] It can be understood that the existing scheme needs to continuously upload all original data to the cloud, while the present scheme effectively reduces the amount of data to be transmitted by completing data preprocessing and preliminary diagnosis on the edge side. The existing technology uses a fixed diagnosis model to process all data, while the present scheme can dynamically adapt to changes in device running state and the emergence of new fault modes through cloud incremental learning mechanism. In this way, the present application significantly reduces the data transmission amount while ensuring the accuracy of diagnosis, so that the diagnosis response speed is effectively improved under the same network bandwidth conditions. At the same time, through the cloud-edge collaborative mechanism, the rational allocation of computing resources is realized, the edge device undertakes basic feature processing tasks, and the cloud focuses on complex pattern recognition, forming an efficient hierarchical diagnosis system. This collaborative working mode is particularly suitable for agricultural operation environments with unstable network conditions, maintaining real-time monitoring capability of the device while avoiding excessive consumption of communication resources.

[0071] In a feasible implementation, a three-axis accelerometer and a temperature sensor can be deployed synchronously at the bearing position of key components such as the motor, main shaft and rubber roller of the rice mill, and a current sensor can be installed in the motor control cabinet. The edge computing device collects raw waveform data of vibration and current at a high frequency (such as 25.6 kHz) and collects scalar data of temperature and rotating speed (if any) at a lower frequency (such as 1 Hz). These signals from different physical dimensions together constitute an original and high-fidelity data set for judging the health status of the rice mill, providing a solid data foundation for subsequent edge feature refinement and cloud intelligent diagnosis.

[0072] In a feasible implementation, the step of performing edge feature refinement processing on the original sensing signals by the edge device to generate a corresponding low-dimensional feature vector includes: selecting an optimal wavelet basis from a pre-set candidate wavelet basis set; performing wavelet packet decomposition on the original sensing signals using the selected optimal wavelet basis to generate a set of node signals; calculating the kurtosis and energy entropy values of each node signal, and screening a target node with the maximum product of kurtosis and energy entropy; extracting the standardized energy entropy, spectral kurtosis and root mean square value of the target node signal, and combining them to form the low-dimensional feature vector.

[0073] In this embodiment, the candidate wavelet basis set refers to a predefined set containing a plurality of wavelet basis functions, and specifically, db4, sym5, coif3 and the like can be selected as the candidate set to match the optimal decomposition mode of different signal characteristics. The optimal wavelet basis refers to the wavelet basis function that is obtained by filtering through a preset evaluation index and is most suitable for the current signal decomposition, and specifically, the adaptability of signal decomposition can be enhanced by calculating the energy distribution difference of the frequency band after decomposition of different wavelet bases. The wavelet packet decomposition is a process of multi-layer decomposition of signals in the time-frequency domain, and specifically, the decomposition depth of 3-5 layers can be adopted to divide the original signal into a plurality of sub-band node signals, so as to realize fine analysis of signal characteristics. The kurtosis is a statistical quantity for measuring the degree of deviation of signal amplitude distribution from normal distribution, and specifically, it can be calculated by the fourth moment, and is used to represent the strength of the impact component in the signal. The energy entropy value is an index reflecting the chaotic degree of signal energy distribution, and specifically, it can be calculated by the entropy value of the proportion of each node energy, and is used to reflect the concentration characteristics of signal energy distribution. The target node refers to the node signal with high kurtosis and high energy entropy, and specifically, the node containing impact characteristics and energy concentration characteristics can be selected by product operation to ensure that the key fault information can be effectively captured. The standardized energy entropy is an energy entropy index after normalization processing, and specifically, the maximum and minimum value normalization method can be adopted to eliminate the dimension difference under different working conditions. The spectral kurtosis is a parameter for representing the impact characteristics of the signal in the frequency domain, and specifically, it can be realized by short-time Fourier transform combined with kurtosis calculation, and is used to improve the recognition degree of the abnormal characteristics in the frequency domain. The root mean square value is a statistical index for measuring the time energy level of the signal, and specifically, it can be obtained by square mean value square root operation, and is used to represent the overall vibration strength of the signal. This feature is extremely sensitive to the capture of instantaneous impact events unique to the rice mill. For example, the spectral kurtosis can sensitively capture the high-frequency instantaneous impact signal generated by the abnormal collision between the grain particles and the metal sieve frame due to the damage of the sieve.

[0074] In this embodiment, at the edge device end, first, the optimal decomposition basis function is selected from the pre-set wavelet basis set. For example, the candidate set can cover 5-8 commonly used wavelet bases, and by calculating the weighted scores of the average kurtosis and energy entropy of each wavelet basis under a preset decomposition level, the wavelet basis with the highest score is selected as the decomposition tool. Then, the selected wavelet basis is used to carry out 3-layer wavelet packet decomposition on the original vibration signal to generate a signal set containing 8 sub-band nodes. The kurtosis and energy entropy value of each node signal are calculated respectively, and the product of the two is taken as the criterion for the importance of the node, and the target node with the largest product is selected. Finally, the three feature parameters of the standardized energy entropy, the spectral kurtosis and the time domain root mean square value are extracted from the target node to form a low-dimensional feature vector with a dimension of 3, and the compression conversion of the signal from the original high-dimensional data to the key feature vector is completed.

[0075] In this embodiment, the decomposition precision is improved by dynamically selecting the optimal wavelet basis, and the key nodes are screened by combining the joint criterion of kurtosis and energy entropy, which can reduce the data dimension while retaining the core fault information. At the same time, the amount of data required for cloud diagnosis is greatly reduced to the original signal, significantly reducing network transmission delay while ensuring the completeness of the features required for subsequent fault classification. For example, in the bearing wear fault scenario of a rice mill, this method can accurately extract the spectral kurtosis features and energy distribution anomaly features reflecting high-frequency impact, avoiding the problem of feature loss caused by insufficient sampling rate of the original signal or mismatch of the fixed wavelet basis.

[0076] In a feasible implementation, the step of selecting the optimal wavelet basis from the preset candidate wavelet basis set comprises: calculating the average kurtosis value and the average energy entropy value of each sub-band node under the preset decomposition level for each candidate wavelet basis in the preset candidate wavelet basis set; calculating the weighted fusion of the average kurtosis value and the average energy entropy value of each candidate wavelet basis according to the preset weight ratio to obtain a joint criterion score; and selecting the candidate wavelet basis with the highest joint criterion score as the optimal wavelet basis.

[0077] In the embodiment, referring to the above embodiment, the candidate wavelet basis set refers to a predefined set containing multiple wavelet basis functions, which serves as a basis for the selection of different signal decomposition characteristics. The average kurtosis value refers to the average value of the kurtosis calculation results of the decomposed sub-band node signals, which can be obtained by calculating the ratio of the fourth-order central moment to the square of the second-order central moment of each node signal. Its function is to measure the strength of the impact component of the signal. The average energy entropy value refers to the average value of the energy entropy calculation results of the decomposed sub-band node signals, which can be achieved by calculating the entropy value of the energy distribution of each node signal. Its function is to represent the complexity of the signal energy distribution. The preset weight ratio refers to the weighting coefficient set for the kurtosis and energy entropy indicators, which can be determined by empirical values or by means of optimization algorithms, for example, setting the kurtosis weight to 0.6 and the energy entropy weight to 0.4. The purpose is to balance the contribution of different indicators to signal characteristics. The joint criterion score refers to the comprehensive score obtained by weighting and fusing the kurtosis and energy entropy indicators, which can be realized by a linear weighted sum formula. Its significance lies in quantifying the representation ability of the candidate wavelet basis for signal characteristics. In the embodiment, when the edge device processes the rice mill sensing signal, the wavelet basis that best fits the current signal characteristics needs to be selected from the candidate wavelet basis set. For each candidate wavelet basis, a wavelet packet decomposition operation of a preset number of layers is carried out to obtain multiple sub-band node signals. Then, the kurtosis value and the energy entropy value of each node signal are calculated, and the average value of all nodes is taken as the evaluation indicator of the wavelet basis. Further, according to the preset weight ratio, the average kurtosis value and the average energy entropy value are weighted and calculated to obtain the joint criterion score. Finally, the wavelet basis with the highest score is selected as the optimal decomposition tool, thereby ensuring the accuracy and stability of signal decomposition in the subsequent feature extraction process.

[0078] In the embodiment, by fusing the dual indicators of kurtosis and energy entropy and introducing a weight adjustment mechanism, the non-stationary characteristics and energy distribution characteristics of the signal can be more comprehensively reflected, thereby improving the adaptability of wavelet basis selection. In this way, the present application solves the problem of inaccurate feature extraction caused by improper selection of wavelet basis in the prior art, optimizes the selection process of wavelet basis through a multi-index weighted evaluation mechanism, enhances the ability of the edge device to capture fault features in the vibration signal, provides a more reliable low-dimensional feature vector for subsequent state classification, and at the same time avoids the risk of misjudgment or omission caused by insufficient signal decomposition.

[0079] In an implementable embodiment, the step of performing state classification on the low-dimensional feature vector by the edge device to generate a pre-judgment state label comprises: extracting a root mean square value of the low-dimensional feature vector; directly outputting an abnormal label when the root mean square value is greater than a preset device baseline value; calculating a product between a normalized energy entropy and a spectral kurtosis in the feature vector to obtain a discriminant factor when the root mean square value is less than the preset device baseline value; outputting an abnormal label if the discriminant factor is greater than a dynamic label threshold; and outputting a normal label if the discriminant factor is not greater than the dynamic label threshold.

[0080] In the embodiment, the root mean square value is an effective amplitude representation of a signal in a time domain, and can be obtained by square root mean calculation on the vibration acceleration signal. The root mean square value reflects the overall vibration intensity of the device. The preset device baseline value is a vibration intensity baseline of the device in a stable running state with load, that is, the root mean square average of the vibration signal of the rice mill in the stable running state with load is set as a threshold for judging whether the device is in abnormal vibration. The discriminant factor is a composite index composed of the normalized energy entropy and the spectral kurtosis, and can be obtained by product operation on the two. The discriminant factor is used to quantify the non-stationarity of the signal and the intensity of the impact component. The dynamic label threshold is a classification boundary automatically adjusted according to the normal working condition data of the day. The dynamic label threshold can be determined by selecting a preset quantile after the discriminant factor distribution of the normal feature vector is counted, so as to adapt to the dynamic change of the device running state.

[0081] In the embodiment, when the edge device performs state classification, the root mean square value of the low-dimensional feature vector is first extracted and compared with the preset device baseline value. If the root mean square value exceeds the baseline value, it is directly determined that the device is in an abnormal state, and subsequent calculation is not needed. If the root mean square value does not exceed the baseline value, the product of the normalized energy entropy and the spectral kurtosis is further calculated to generate a discriminant factor. The discriminant factor can comprehensively reflect the non-stationary characteristics and impact characteristics of the signal. By comparing the discriminant factor with the dynamic label threshold, the feature fluctuations caused by slight abnormalities and the random fluctuations in the normal working condition can be effectively distinguished. This hierarchical determination mechanism can quickly identify significant abnormalities and capture early signs of potential faults through the composite index.

[0082] In this embodiment, by introducing a two-stage decision mechanism, combining time domain vibration intensity and frequency domain composite indicators, the robustness of the classification logic is enhanced while preserving fast response. At the same time, the dynamic threshold updating strategy avoids the problem of adaptability decline caused by fixed threshold due to equipment aging or environmental changes. In this way, the application can effectively reduce the misjudgment rate of edge devices in the state classification process, avoid false alarms of random fluctuations in normal working conditions as fault states, and reduce the uploading frequency of invalid data. At the same time, by dynamically adjusting the classification threshold, the state changes of the device in different operating stages can be adapted to ensure the consistency of the classification standard and the actual working condition.

[0083] In a feasible implementation, the method further comprises collecting vibration signals in a steady running state with load for a preset time length when the device is first commissioned, and extracting the root mean square value of the vibration signals, and taking the average value as the device baseline value; collecting a low-dimensional feature vector set during the normal working condition period of the day, and calculating the discriminant factor of all feature vectors in the low-dimensional feature vector set respectively, and taking the preset quantile of the discriminant factor of all feature vectors as a new dynamic threshold.

[0084] In this embodiment, the preset device baseline value refers to the vibration intensity reference of the device in a steady running state with load, that is, the setting for the average root mean square value of the vibration signals of the rice mill in a steady running state with load. Specifically, the sliding average of the root mean square value of the vibration signals collected continuously for 30 seconds, for example, can be calculated to achieve this value, which is used to construct the reference baseline of the device in a normal running state. The discriminant factor refers to the product of the normalized energy entropy and the spectral kurtosis, which can be obtained by multiplying the numerical values of the normalized energy entropy value and the spectral kurtosis coefficient. This factor is used to quantify the degree of abnormality of the device running state. The dynamic label threshold refers to the classification boundary value dynamically adjusted based on normal working condition data, which can be determined by statistical methods such as 75% quantile or 90% quantile. This threshold is used to adapt to the state fluctuations of the device under different working conditions.

[0085] In this embodiment, during the first installation and commissioning stage of the device, vibration signals in a running state with load for 5 minutes, for example, are collected, the root mean square value is extracted and the arithmetic mean is calculated to form the device baseline value, which is used as the baseline for subsequent state judgment. During daily operation, low-dimensional feature vectors generated during normal working hours are continuously collected, for example, 100 feature vectors are extracted every hour, and the discriminant factor value of each feature vector is calculated. By statistically analyzing the distribution of all discriminant factor values for the day, the value corresponding to the 85% quantile, for example, is selected as the new dynamic threshold, which will replace the initial set value for state classification judgment the next day.

[0086] In this embodiment, by dynamically updating the threshold mechanism, the natural changes of the device running state can be automatically tracked, such as the adjustment of the bearing gap caused by seasonal temperature changes, or the change of the motor vibration mode caused by long-term operation, thereby avoiding the problem of rising false positive rate caused by fixed threshold. In this way, the application realizes the adaptive adjustment of the device state classification threshold, effectively solving the technical problem of the decline of classification accuracy of the traditional fixed threshold in the long-term running scene. By using the quantile statistical method to determine the dynamic threshold, the accurate identification of abnormal states is ensured, and the threshold fluctuation caused by individual noise data interference is avoided, such as filtering the abnormally high discriminant factor value caused by instantaneous impact vibration, to ensure the rationality of the threshold setting.

[0087] In an available implementation, the step of performing incremental migration diagnosis on the low-dimensional feature vector based on the cloud platform to generate a fault type and a confidence result includes: inputting the low-dimensional feature vector into a pre-trained base model to output a primary feature vector; inputting the primary feature vector into an incremental adaptation layer and combining historical feature library to calculate same-class feature aggregation loss, different-class feature separation loss, and new and old data distribution alignment loss; weighting and summing the same-class feature aggregation loss, different-class feature separation loss, and new and old data distribution alignment loss to update the weights of the incremental adaptation layer, so that the incremental adaptation layer outputs a strengthened feature vector; performing full connection classification on the strengthened feature vector to generate a fault type, and generating a confidence result based on the maximum value of the Softmax probability.

[0088] In this embodiment, the pre-trained base model refers to a deep neural network model pre-trained by historical fault data, which can be implemented by using a multi-layer convolutional neural network or a residual network. Its function is to extract primary features with distinguishing degrees from the low-dimensional feature vector. The incremental adaptation layer refers to a network layer that can dynamically adjust the weights, which can be achieved by a full connection layer combined with an adaptive activation function, and is used to fuse the feature distribution information in the historical feature library and optimize the feature representation. The same-class feature aggregation loss refers to the distance measurement of the same-class fault features in the feature space, which can be calculated by using cosine similarity or Euclidean distance to promote the aggregation of the same-class fault features. The different-class feature separation loss refers to the separation measurement of different-class fault features in the feature space, which can be achieved by using a margin loss function, aiming to force the features of different fault types to maintain the minimum interval. The new and old data distribution alignment loss refers to the difference measurement of the data distribution between the new input features and the historical feature library, which can be calculated by using the maximum mean difference algorithm to reduce the impact of data distribution deviation on diagnosis.

[0089] In this embodiment, when the cloud platform receives the low-dimensional feature vector uploaded by the edge device, it first extracts a primary feature vector with the help of a pre-trained base model, which has learned the general feature patterns of historical fault data. Then, the primary feature vector is input into the incremental adaptation layer, which combines the fault feature data stored in the historical feature library to synchronously calculate the same-class feature aggregation loss, different-class feature separation loss, and new-old data distribution alignment loss. By weighting and summing the three loss functions and updating the network weights of the incremental adaptation layer in a backpropagation manner, the reinforced feature vector can meet the requirements of same-class feature aggregation, different-class feature separation, and new-old data distribution alignment. Finally, the reinforced feature vector is classified through the full connection layer, and the fault type and confidence are determined according to the maximum value of the Softmax output probability.

[0090] In this embodiment, the incremental adaptation layer is used to realize online updating of model parameters, and the historical feature library is used for feature distribution optimization, which can quickly adapt to new fault types without retraining the entire model, while ensuring diagnostic accuracy. In this way, the application effectively solves the problem of decreased diagnostic accuracy caused by changes in data distribution. Through the incremental migration diagnosis mechanism, the compatibility of new fault features and historical data is ensured while reducing the computational amount of model updating. For example, when a new type of bearing wear fault occurs in a rice mill, the cloud platform can quickly adjust the feature extraction rules based on the incremental adaptation layer to avoid misjudgment due to the model not learning the fault pattern, significantly improving the reliability of the diagnosis result.

[0091] In a feasible implementation, the method further includes generating a set of pseudo feature vectors by sampling the device historical feature library through the cloud platform; training a lightweight convolutional network with the set of pseudo feature vectors as input; performing weight pruning on the trained lightweight convolutional network to remove channels with a contribution degree lower than a preset contribution degree threshold, to generate the lightweight diagnosis model; and distributing the lightweight diagnosis model to the edge device to enable the edge device to update the state classification standard for the low-dimensional feature vector.

[0092] In this embodiment, the set of pseudo feature vectors refers to simulated feature data generated by random sampling or cluster sampling from the device historical feature library, which can be implemented using the Monte Carlo method or K-means clustering algorithm to simulate the feature distribution under real working conditions. The lightweight convolutional network refers to a convolutional neural network structure with small parameter quantity and low computational complexity, which can be implemented using depth separable convolution or channel pruning technology to adapt to the limited computing resources of the edge device. Weight pruning refers to pruning operation based on the contribution of channels in the classification task, which can evaluate the importance of channels using the L1 norm and remove redundant channels that have a lower impact on the classification result, thereby compressing the model size.

[0093] In the embodiment, the cloud platform extracts part of samples from the historical feature library, generates a pseudo-feature vector set, and inputs the pseudo-feature vector set as training data into the lightweight convolutional network to carry out model training. After the training is completed, the activation strength of each channel in the network is analyzed by means of a weight pruning algorithm, for example, the channels with an activation strength lower than 0.05 are identified as redundant structures and removed, and finally a compressed lightweight diagnostic model is generated. The model is issued to the edge device under the cloud-edge collaborative mechanism, replaces the original state classification module, and enables the edge device to make more accurate state judgments on the real-time generated low-dimensional feature vectors based on the updated classification standard.

[0094] In the embodiment, after the continuous optimization of the diagnostic model is completed in the cloud through the pseudo-feature generation and model pruning technology, only the lightweight model parameters need to be issued to realize the dynamic improvement of the edge-side classification capability, avoiding the network load caused by frequent transmission of a large amount of original data. Through the above technical solutions, the online iterative optimization of the state classification model of the edge device is realized, and the problem that the traditional cloud-trained model is too large to adapt to the edge device is solved. Through the pseudo-feature training and weight pruning operation, the model calculation complexity is reduced while the classification accuracy is ensured, so that the edge device can quickly complete the state judgment based on the lightweight diagnostic model, and the dependence on the cloud computing resources is reduced.

[0095] In a feasible implementation manner, the method further includes when the pre-judgment state label is a normal label, reserving a low-dimensional feature vector of only one random time window in a preset polling period; and uploading the low-dimensional feature vector reserved by sampling and the pre-judgment state label to the cloud.

[0096] In the embodiment, the preset polling period refers to a fixed time interval of data transmission from the edge device to the cloud, and can be implemented by using a time length of, for example, 5 minutes or 10 minutes, for balancing the data real-time performance and transmission efficiency. The random time window refers to a continuous time period randomly selected in each polling period, for example, the window length is randomly determined in the range of 10 seconds to 30 seconds, and the sampling interval is dynamically adjusted to avoid the monitoring blind area caused by the fixed sampling mode. The low-dimensional feature vector refers to a compressed data set generated after feature refining, and can be generated by wavelet packet decomposition and feature screening steps, and has a dimension of, for example, 3 to 5, for reducing the data volume while retaining the key state information.

[0097] In the embodiment, when the device operation state is determined to be normal, the edge device will no longer continuously upload all the monitoring data, but will retain only the low-dimensional feature vector corresponding to a randomly selected time window in each preset polling period. The start point and duration of the time window are randomly generated within the period, for example, in a 5-minute period, a window starting at 2 minutes and 10 seconds and lasting 15 seconds is randomly selected. The sampled and retained low-dimensional feature vector is bound with a normal label and uploaded to the cloud, so that the cloud can periodically obtain device operation state samples while avoiding transmission of redundant data.

[0098] In the embodiment, through the random sampling mechanism within the polling period, the transmission of invalid data under normal conditions is reduced while ensuring that representative state data is obtained by the cloud, significantly reducing the amount of data uploaded under normal conditions, reducing the storage and computing load of the cloud, and effectively alleviating network bandwidth pressure. At the same time, the random time window design maintains the ability to continuously track the device state, avoiding the risk of missing potential abnormal events due to fixed sampling intervals.

[0099] In a feasible implementation, the method further includes, if uploading to the cloud platform fails, the edge device re-uploads the low-dimensional feature vector and the pre-judgment state label according to a preset retry interval and a maximum number of retries; if the number of retries exceeds the preset number of times, the edge device stores the low-dimensional feature vector and triggers a local alarm signal, and reports the failure log to the user terminal.

[0100] In the embodiment, the preset retry interval and the maximum number of retries refer to a pre-set data retransmission time interval and a maximum number of allowed attempts, which can be implemented in a fixed time interval or a dynamically adjusted interval, for example, the retry interval can be set to a stepwise increasing strategy of 5 seconds, 10 seconds, and 30 seconds, and the maximum number of retries can be set to 3 times. This mechanism is used to avoid frequent invalid transmission when the network fluctuates, while ensuring the ultimate accessibility of critical data. Storing the low-dimensional feature vector means persistently saving the feature data that fails to be uploaded in the local storage medium, which can be implemented by using an embedded database or a file system, and is used to prevent data loss and provide original information for subsequent fault tracing. Triggering the local alarm signal means notifying the on-site operator through a sound and light device or a message push, which can be implemented by using a buzzer, an LED indicator light, or a local area network message broadcast, and is used to maintain basic alarm capability when the cloud communication is interrupted. Reporting the failure log to the user terminal means transmitting key parameters of the transmission failure event to the maintenance personnel device through a backup communication link, which can be implemented by using a short message, an email, or a low-bandwidth communication protocol, and is used to assist the operation and maintenance personnel in locating the network anomaly cause.

[0101] In this embodiment, when the edge device detects a data upload failure, a retry process is first started according to a preset strategy. For example, after the first failure, a first retry is performed after waiting for 5 seconds, if it fails again, a second retry is performed after waiting for 10 seconds, and the third retry interval is extended to 30 seconds. If the three retries are not successful, the current low-dimensional feature vector and the associated time stamp are automatically stored in the local flash memory, and the red warning light and intermittent buzzer of the device body are activated. At this time, the device continues to record network state parameters, and every 10 minutes attempts to send a short message containing the failure time, the number of retries and the network error code to the bound maintenance personnel's mobile phone through the short message gateway.

[0102] In this embodiment, the hierarchical retry strategy balances the transmission success rate and resource consumption, the local storage ensures data integrity, the dual alarm mechanism takes into account the on-site warning and remote notification needs, and the detailed recording of failure logs provides data support for network optimization. In this way, the application effectively solves the problem of missing diagnostic data caused by network anomalies, and can still ensure the complete retention of key fault information when the communication link is unstable. At the same time, it can prompt the on-site personnel to intervene and handle in time through local alarm, and provide accurate network fault analysis basis for the operation and maintenance team, forming a complete abnormal handling closed loop.

[0103] In the embodiments of the application, the remote fault diagnosis method for rice mill based on cloud-edge collaboration refines and predicts the original sensing signals through the edge device, uploads the low-dimensional feature vector to the cloud platform only in abnormal state, effectively reduces the data transmission amount and improves the diagnosis efficiency, can reduce the data transmission bandwidth occupation and improve the fault diagnosis real-time performance, at the same time, the application can also reduce the transmission of redundant information and improve the cloud computing efficiency.

[0104] It should be noted that the above examples are only used to understand the application and do not constitute a limitation on the remote fault diagnosis method for rice mill based on cloud-edge collaboration of the application. More forms of simple transformation based on this technical concept are within the protection scope of the application.

[0105] The application also provides a remote fault diagnosis system for rice mill based on cloud-edge collaboration, which refers to Figure 2 The system comprises a memory 10, a processor 20, and a remote fault diagnosis program for rice mill based on cloud-edge collaboration stored on the memory 10 and executable on the processor 20, wherein the remote fault diagnosis program for rice mill based on cloud-edge collaboration is configured to implement the steps of the remote fault diagnosis method for rice mill based on cloud-edge collaboration.

[0106] The remote fault diagnosis system of the rice mill based on cloud edge cooperation provided in the application adopts the remote fault diagnosis method of the rice mill based on cloud edge cooperation in the above embodiment, and can reduce data transmission bandwidth occupation and improve fault diagnosis real-time performance. Compared with the prior art, the beneficial effects of the remote fault diagnosis system of the rice mill based on cloud edge cooperation provided in the application are the same as those of the remote fault diagnosis method of the rice mill based on cloud edge cooperation provided in the above embodiment, and other technical features in the remote fault diagnosis system of the rice mill based on cloud edge cooperation are the same as those disclosed in the above embodiment method, and will not be repeated here.

[0107] The above is only part of the embodiments of the application, and does not limit the patent scope of the application. Any equivalent structural transformation made by using the content of the specification and drawings, or direct / indirect application in other related technical fields under the technical concept of the application is included in the patent protection scope of the application.

Claims

1. A remote fault diagnosis method for rice milling machines based on cloud-edge collaboration, characterized in that, The method includes: Raw sensor signals from the rice milling machine are acquired using edge devices; The original sensing signal is refined based on edge features using edge devices to generate corresponding low-dimensional feature vectors. Based on edge devices, low-dimensional feature vectors are classified into states to generate predicted state labels; the predicted state labels include normal labels and abnormal labels. When the predicted state label is an abnormal label, the predicted state label is appended to the low-dimensional feature vector and uploaded to the cloud platform; Incremental migration diagnostics are performed on the low-dimensional feature vectors based on the cloud platform to generate fault type and confidence results, which are then sent to edge devices and user terminals. The step of refining the edge features of the original sensing signal based on the edge device to generate the corresponding low-dimensional feature vector includes: Select the optimal wavelet basis from a pre-defined set of candidate wavelet basis bases; The original sensor signal is decomposed into a set of nodal signals using the selected optimal wavelet basis. Calculate the kurtosis and energy entropy values ​​of the signals at each node, and select the target node with the largest product of kurtosis and energy entropy; Extract the standardized energy entropy, spectral kurtosis, and root mean square value of the target node signal, and combine them to form the low-dimensional feature vector; The steps of performing incremental migration diagnostics on the low-dimensional feature vector based on the cloud platform to generate fault type and confidence results include: Input the low-dimensional feature vector into the pre-trained base model and output the primary feature vector; The primary feature vector is input into the incremental adaptation layer, and the loss of aggregation of similar features, the loss of separation of dissimilar features, and the loss of alignment of distribution of new and old data are calculated in combination with the historical feature library. The weighted sum of the loss for aggregation of similar features, the loss for separation of dissimilar features, and the loss for alignment of distribution of new and old data is used to update the weights of the incremental adaptation layer, so that the incremental adaptation layer outputs an enhanced feature vector. Perform fully connected classification on the enhanced feature vectors to generate fault types, and generate confidence results based on the maximum probability of Softmax.

2. The remote fault diagnosis method for rice milling machines based on cloud-edge collaboration as described in claim 1, characterized in that, The step of selecting the optimal wavelet basis from the preset set of candidate wavelet basis bases includes: For each candidate wavelet basis in the preset set of candidate wavelet basis bases, calculate the average kurtosis and average energy entropy of each sub-band node under the preset number of decomposition layers; The average kurtosis and average energy entropy of each candidate wavelet basis are weighted and fused according to a preset weight ratio to obtain a joint criterion score. The candidate wavelet basis with the highest joint criterion score is selected as the optimal wavelet basis.

3. The remote fault diagnosis method for rice milling machines based on cloud-edge collaboration as described in claim 1, characterized in that, The step of classifying low-dimensional feature vectors based on edge devices to generate predicted state labels includes: Extract the root mean square value of the low-dimensional feature vector; When the root mean square value is greater than the preset device baseline value, an abnormal label is output directly; When the root mean square value is less than the preset device baseline value, the product between the standardized energy entropy and the spectral kurtosis in the feature vector is calculated to obtain the discriminant factor. If the discrimination factor is greater than the dynamic label threshold, an abnormal label is output; if the discrimination factor is not greater than the dynamic label threshold, a normal label is output.

4. The remote fault diagnosis method for rice milling machines based on cloud-edge collaboration as described in claim 3, characterized in that, The method further includes: During the initial commissioning of the equipment, vibration signals were collected for a preset period of stable operation under load, and the root mean square value of the vibration signals was extracted. The average value was then used as the baseline value of the equipment. Collect a set of low-dimensional feature vectors during normal working conditions on the same day, and calculate the discriminant factor of each feature vector in the set. Take the preset quantile of the discriminant factor of all feature vectors as the new dynamic threshold.

5. The remote fault diagnosis method for rice milling machines based on cloud-edge collaboration as described in claim 1, characterized in that, The method further includes: A set of pseudo-feature vectors is generated by sampling from the device's historical feature database through a cloud platform; A lightweight convolutional network is trained using a set of pseudo-feature vectors as input. The trained lightweight convolutional network is weighted and channels with contributions below a preset contribution threshold are removed to generate a lightweight diagnostic model. Lightweight diagnostic models are deployed to edge devices to enable them to update their state classification criteria for low-dimensional feature vectors.

6. The remote fault diagnosis method for rice milling machines based on cloud-edge collaboration as described in claim 1, characterized in that, The method further includes: When the predicted state label is a normal label, a low-dimensional feature vector of only one random time window is retained within the preset polling period. The sampled low-dimensional feature vectors are bound to the predicted state labels and uploaded to the cloud.

7. The remote fault diagnosis method for rice milling machines based on cloud-edge collaboration as described in claim 1, characterized in that, The method further includes: If uploading to the cloud platform fails, the edge device will re-upload the low-dimensional feature vector and the predicted status label according to the preset retry interval and the maximum number of retries. If the number of retries exceeds the preset number, the edge device stores the low-dimensional feature vector and triggers a local alarm signal, while simultaneously reporting the failure log to the user terminal.

8. A remote fault diagnosis system for rice milling machines based on cloud-edge collaboration, characterized in that, The system includes: a memory, a processor, and a cloud-edge collaborative remote fault diagnosis program for rice milling machines stored in the memory and executable on the processor, wherein the cloud-edge collaborative remote fault diagnosis program for rice milling machines is configured to implement the steps of the cloud-edge collaborative remote fault diagnosis method for rice milling machines as described in any one of claims 1 to 7.

Images