Intelligent management system for energy consumption optimization and fault self-diagnosis of cleaning equipment

Abstract

The invention discloses an intelligent management system for energy consumption optimization and fault self-diagnosis of cleaning equipment, and relates to the field of intelligent maintenance of the cleaning equipment, and the system comprises the steps: obtaining three groups of core parameters, i.e., a historical vibration spectrum, a motor current harmonic component and a bearing temperature gradient, constructing a dynamic failure mode knowledge graph, performing time sequence correlation analysis on the historical fault data to obtain failure mode analysis data; establishing a multi-dimensional analysis platform, identifying a high-risk component, and updating a fault threshold value; introducing a service time attenuation factor and a working condition correction coefficient, establishing an aging degree quantitative model, and calculating an aging coefficient; and constructing and developing an energy consumption-reliability joint optimization module, and adjusting equipment operation parameters. The method has the advantages that the dynamic knowledge graph and the time sequence analysis model are constructed by integrating multi-source sensor data, precise diagnosis and self-adaptive threshold adjustment of the coupling fault of the cleaning equipment are achieved, aging evaluation and task scheduling optimization are combined, the energy consumption efficiency is improved, and the maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent maintenance of cleaning equipment, and specifically relates to an intelligent management system for energy consumption optimization and fault self-diagnosis of cleaning equipment. Background Art

[0002] Energy consumption optimization of cleaning equipment refers to reducing the energy consumption (such as electric energy, fuel, etc.) of cleaning equipment during operation through technical means and management strategies while ensuring the cleaning effect, finding the best balance among performance, efficiency, cost, and environmental protection, thereby improving energy utilization efficiency, extending the equipment's battery life, reducing carbon emissions, and operating costs. However, during the use of existing intelligent management systems for energy consumption optimization and fault self-diagnosis of cleaning equipment, their fault diagnosis relies on single-parameter threshold alarms, unable to identify the coupled failure modes of bearing wear and motor abnormalities. At the same time, the threshold settings are fixed, making it difficult to adapt to the performance degradation of components under different environmental conditions, resulting in false alarms in their detection. During continuous use, energy consumption optimization is disconnected from the equipment's health status, lacking a dynamic power adjustment mechanism based on the degree of component aging, leading to lag in equipment maintenance, an increase in the average fault repair time, and an increase in the annual energy consumption waste rate. Summary of the Invention

[0003] To solve the above technical problems, an intelligent management system for energy consumption optimization and fault self-diagnosis of cleaning equipment is provided. This technical solution solves the problems in the above background art that the single diagnosis method relying on fixed thresholds is difficult to adapt to the performance degradation of components under different environmental conditions, and at the same time, energy consumption optimization is disconnected from the equipment's health status, lacking a dynamic power adjustment mechanism based on the degree of component aging.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An intelligent management system for energy consumption optimization and fault self-diagnosis of cleaning equipment, comprising: Obtain three groups of core parameters of the cleaning equipment, namely the historical vibration spectrum, motor current harmonic components, and bearing temperature gradient, construct a dynamic failure mode knowledge graph, and perform time-series correlation analysis on historical fault data to obtain failure mode analysis data; Obtain the standard data released by the cleaning equipment manufacturer, establish a multi-dimensional analysis platform. The multi-dimensional analysis platform includes a failure mode analysis module, an aging assessment model, and an energy consumption optimization model. Import the failure mode analysis data into the failure mode analysis module to identify high-risk components and update the fault threshold; Import the historical equipment operation time, number of work cycles, and environmental data into the aging assessment model, introduce a service time decay factor and a working condition correction coefficient, generate a risk heat map, combine the cumulative number of work cycles of the equipment and the environmental temperature and humidity parameters, establish an aging degree quantification model, and calculate the aging coefficient; Based on the sensor network monitoring, the operating parameters of the load adjustment device are monitored in real time. Combining the fault threshold and aging coefficient of high-risk components, an energy consumption-reliability joint optimization module is constructed and developed to adjust the operating parameters of the device to reduce energy consumption.

[0005] Preferably, the specific steps of obtaining the historical vibration spectrum, motor current harmonic components, and bearing temperature gradient of the cleaning device and constructing the dynamic failure mode knowledge graph are as follows: Obtain the historical vibration spectrum, motor current harmonic components, and bearing temperature gradient of the cleaning device. After aligning the timestamps of the three groups of data, perform a joint time-frequency domain analysis on the historical vibration spectrum, extract the fundamental frequency, harmonic frequency, and sideband energy distribution feature vectors, calculate the harmonic distortion rate of the motor current harmonic components and quantify the spectral sparsity to generate an abnormal current index; Perform a sliding window difference processing on the bearing temperature gradient to construct a temperature gradient tensor with two rows and two columns. The first row represents the number of bearing surface temperature sensors, the second row represents the number of time segments divided by the sliding window, and the two columns respectively store the difference values of the axial and radial temperature change rates of the corresponding sensors within a specific time window; Input the feature vectors, abnormal current index, and temperature gradient matrix into a graph neural network to establish a dynamic association relationship between the component state nodes and the failure mode edges, denoted as the knowledge graph topological structure; Based on the failure mode labels in the historical fault case database, embed the same nodes with the same timestamp into the knowledge graph topological structure to form a dynamic failure mode knowledge graph containing the fault propagation path.

[0006] Preferably, the specific steps of performing a temporal correlation analysis on the historical fault data to obtain the failure mode analysis data are as follows: Align the timestamps of the fault event sequences in the historical fault data, extract the fundamental frequency offset of the vibration spectrum, the mutation value of the current harmonic distortion rate, and the extreme value of the bearing temperature gradient corresponding to each fault event, and construct a multi-dimensional time series database; Use a sliding time window to segment the multi-dimensional time series database, calculate the variance of the fundamental frequency offset, the change slope of the harmonic distortion rate, and the occurrence frequency of the extreme value of the temperature gradient within the window, summarize to obtain the data of each index, perform normalization processing on the data of each index, and then obtain the fault correlation intensity index of each time segment through weighted fusion; Use a temporal graph convolutional network to extract spatio-temporal features from the multi-dimensional time series matrix, capture the temporal causal relationship between cross-sensor parameters, and generate the probability weight of the fault propagation link; Input the fault correlation intensity index and the fault propagation link probability weight into the Bayesian network, calculate the conditional probability distribution of different failure modes in the time dimension, and output the failure mode analysis data including temporal dependence relationships. The failure mode analysis data includes the fault precursor event sequence, the cross-component failure coupling probability, and the dynamic failure threshold drift trend.

[0007] Preferably, the importing the failure mode analysis data into the failure mode analysis module, identifying high-risk components, and updating the fault threshold specifically includes: Extract the fault propagation link probability weight as the failure mode occurrence probability benchmark value, obtain the downtime duration and repair cost caused by the same type of faults in the historical fault case library, and calculate the impact coefficient of each failure mode; Label the role of each component in the functional chain, read the number of downstream functions and the total number of functions of each component, divide the number of downstream functions of each component by the total number of functions to obtain the component criticality factor, multiply the occurrence probability benchmark value, the impact coefficient, and the component criticality factor to obtain the risk priority number. The number of downstream functions is the number of downstream functional modules directly affected when the component fails, and the total number of functions is the total number of all independent functional modules in the overall functional chain of the cleaning equipment; Screen the components of the failure modes whose risk priority numbers exceed the dynamic threshold, record them as high-risk components, and update the fault threshold.

[0008] Preferably, the calculating the impact coefficient of each failure mode specifically includes: Obtain the downtime duration data set and the repair cost data set corresponding to the same type of fault cases in the historical fault case library; Perform logarithmic normalization on the downtime duration data, take the natural logarithm of each downtime duration data, and select the maximum value among them. Calculate the ratio of the natural logarithm of each downtime duration to the maximum value of all natural logarithms, and record it as the standardized downtime impact factor; Perform range normalization on the repair cost data, select the highest and lowest repair costs after normalization, calculate the difference between the highest and lowest repair costs, and record it as the standard range. Divide the difference between each repair cost and the lowest repair cost by the standard range to obtain the standardized cost impact factor; The user inputs the preset downtime duration weight and repair cost weight, perform weighted summation on the standardized downtime impact factor and the standardized cost impact factor to generate the impact factor of each fault case, take the arithmetic mean of the impact factors of the same type of fault cases, and output the impact coefficient of this failure mode.

[0009] Preferably, the screening the components of the failure modes whose risk priority numbers exceed the dynamic threshold and recording them as high-risk components specifically includes: Obtain the risk priority numbers of all components in the current system, calculate the distribution characteristics of historical risk priority numbers based on the sliding time window, calculate the mean and standard deviation of the risk priority numbers in the window, and use the sum of the mean and three times the standard deviation of the risk priority number as the dynamic threshold baseline value; Collect equipment operating status parameters in real time, and obtain a threshold correction model by training the mapping relationship between historical operating status and abnormal fluctuations in risk priority numbers through neural network training. Input the equipment operating status parameters into the threshold correction model and output a dynamic threshold correction coefficient. The equipment operating status parameters include ambient temperature and humidity, current load rate, and continuous operation time of the equipment. Multiply the dynamic threshold baseline value by the dynamic threshold correction coefficient to generate a real-time dynamic threshold. Summarize the ratio of the risk priority number to the real-time dynamic threshold to form ratio data. Perform spatial cluster analysis on components with ratio data greater than 1 to eliminate isolated abnormal points caused by sensor noise. Based on the component groups formed after clustering, the function correlation parameters of all components in the group are extracted, and at least two components in the group belong to the same function chain, and the function correlation parameters exceed the preset threshold, and the group as a whole is marked as a high-risk component set; Reversely trace the failure mode of each component in the high-risk component set, verify its risk priority number, calculate the data integrity of the fault propagation link probability weight, influence coefficient and component criticality factor involved in the calculation process, and generate a verified high-risk component list.

[0010] Preferably, the update failure threshold specifically includes: Obtain the attenuation factor associated with the aging coefficient, add 1 to the product of the component criticality factor and the attenuation factor associated with the aging coefficient, take the logarithm, obtain the intermediate value, calculate the product of the intermediate value and the real-time dynamic threshold, and obtain the initial threshold adjustment amount; Through the real-time collection of environmental temperature and humidity parameters and load fluctuation rate, an environment-load correction matrix is ​​constructed, and the initial threshold adjustment amount is multiplied with the correction matrix by tensor product operation to generate the threshold offset after environmental condition compensation. The behavioral environmental parameters of the environment-load correction matrix are listed as the period of load fluctuation rate; The sliding window length is set to the preset number of days, and the exponentially weighted moving average model is used to perform trend fitting on the threshold drift in the window. The model residuals are extracted for normality test. If the residuals obey the normal distribution, the confidence interval is constructed according to the sliding window distribution. If the residuals are non-normal, the quantiles of the residuals are directly taken as the upper and lower bounds. When the threshold offset exceeds the confidence interval, the exponential smoothing method is used to dynamically damp the offset. Adding the processed threshold offset to the original fault threshold to generate an updated dynamic fault threshold, which is synchronized to the alarm threshold database of the fault self-diagnosis module; Cross-device group verification is performed on the updated dynamic fault threshold. The median value of the thresholds of non-high-risk components in the same model device group is selected as the reference value. When the deviation of the updated threshold from the reference value exceeds 20%, a threshold feedback is triggered and a calibration alarm signal is generated.

[0011] Preferably, the establishment of the aging degree quantification model and the calculation of the aging coefficient specifically include: Obtain the cumulative working cycle times of the device, and combine the temperature and humidity parameters collected by the environmental temperature and humidity sensor to obtain a temperature and humidity - cycle times composite attenuation factor through weighting. Based on the maximum eigenvalue of the bearing temperature gradient matrix, calculate the friction loss index of the rolling element and the cage, introduce the service time attenuation factor, and construct an exponential decay relationship between friction loss and service time. Divide the environmental temperature and humidity parameters into temperature gradient intervals and humidity gradient intervals. In the coordinate system, use the temperature and humidity gradient intervals and the working cycle times as the coordinate axes respectively, calculate the variance value of the bearing temperature gradient within each cubic unit, and generate a risk heat map. Extract the local texture features of the risk heat map through a convolutional neural network, and combine the working condition correction coefficient to smooth and correct the heat distribution, and output the aging feature vector. Weight the service time attenuation factor, the friction loss index and the aging feature vector to obtain the aging coefficient.

[0012] Preferably, the real-time monitoring of the load adjustment device operation parameters based on the sensor network, combined with the fault threshold and the aging coefficient of high-risk components, to construct a development energy consumption - reliability joint optimization module specifically includes: Real-time monitor the motor load rate through the current sensor, use the Kalman filter to predict the load fluctuation, and generate the load prediction curve for the next three working cycles. Extract the load characteristic parameters based on the historical task data, classify the tasks to be executed by the cleaning device, and the user inputs the task threshold. When the motor power demand and the bearing friction torque volatility are lower than the task threshold, this task is defined as a low-load task, and the remaining tasks are divided into medium and high load levels according to the power interval. Adopt a sliding time window to analyze the time distribution curve of the aging coefficient of high-risk components, calculate the ratio of the standard deviation to the mean value of the aging coefficient within each time window. When the ratio exceeds the preset threshold, mark this time period as the aging coefficient peak window. During the same time period, low-load tasks preferentially occupy the aging coefficient peak window, and medium and high load tasks are not allowed to continuously occupy more than 2 adjacent time windows. The task execution must meet the preset operation cycle deadline. The user inputs an error threshold. The real-time load prediction error is obtained by subtracting the real-time monitoring of the motor load by the current sensor from the load prediction curve. When the real-time load prediction error exceeds the error threshold, a sliding window re-planning is triggered, and the task allocation for the next three cycles is recalculated. For the medium and high load tasks that have been scheduled to the peak window, the power output of the peak window is reduced by adjusting the duty cycle of the motor; The optimized task time sequence queue is output, and the low load tasks are concentrated and deployed during the peak period of the aging coefficient. At the same time, it is ensured that the medium and high load tasks are evenly distributed during the non-peak period, and finally an energy consumption-reliability joint optimization module is generated.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention proposes an intelligent management system for energy consumption optimization and fault self-diagnosis of a cleaning device. First, a heterogeneous data acquisition layer is constructed by integrating vibration spectra, current harmonics, and temperature gradient tensors, and a heterogeneous graph structure including vibration nodes, current nodes, and temperature nodes is established. The GraphSAGE model is used to achieve dynamic association. Finally, cross-sensor spatio-temporal features are extracted through a temporal graph convolutional network, and the probability of the fault propagation chain is calculated by combining a Bayesian network to form a spatio-temporal feature fusion module. Thus, the fundamental frequency / sideband energy features can be extracted through joint time-frequency domain analysis, the temperature gradient tensor is constructed to capture the spatial thermal distribution of the bearing, the graph neural network introduces an attention mechanism to calculate the edge weights, encodes the physical connection relationship and the fault propagation path into the topological structure of the knowledge graph, and combines historical fault labels to achieve dynamic node classification, thereby improving the accuracy of fault diagnosis, enhancing the ability to identify coupled faults, detecting the collaborative failure of bearing wear and motor insulation aging, and at the same time eliminating false alarms of single sensors through spatio-temporal feature fusion, reducing the false alarm rate; The present invention proposes an intelligent management system for energy consumption optimization and fault self-diagnosis of cleaning equipment. A load prediction engine is constructed by predicting the future load curve through a Kalman filter. Inside the health status evaluation layer, a three-dimensional risk heat map is analyzed through a convolutional neural network to output an aging feature vector. Finally, through dynamic scheduling, high-risk time periods can be dynamically divided based on the coefficient of variation of the aging coefficient during use, and a dual-constraint optimization model is constructed, so as to form a hard constraint that medium and high-load tasks are prohibited from continuously occupying two adjacent time windows and a soft constraint that the task deadline deviation ≤ 8 minutes, realizing the dynamic planning of the task queue. When the prediction error exceeds the limit, a sliding window re-planning is triggered. The bearing friction torque is calculated through the Hertz contact theory, and the motor output power is dynamically adjusted, so as to reduce energy consumption, improve the task deadline achievement rate, shorten the peak load period, and reduce the risk of sudden failures. At the same time, the dynamic optimization of the scheduling algorithm and the accurate risk assessment provided by the knowledge graph form a closed loop. The failure propagation probability weight output by the knowledge graph provides a risk quantification basis for task scheduling, while the real-time task load data feeds back the time-series correlation analysis of the knowledge graph, realizing the co-evolution of diagnosis-optimization, so that the comprehensive energy efficiency ratio under high-temperature and high-humidity working conditions is improved and the mean time between failures is extended. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flowchart of the present invention; Figure 2 is a flowchart of obtaining three core parameters of the historical vibration spectrum, motor current harmonic components, and bearing temperature gradient of the cleaning equipment in the present invention and constructing a dynamic failure mode knowledge graph; Figure 3 is a flowchart of performing time-series correlation analysis on historical fault data to obtain failure mode analysis data in the present invention; Figure 4 is a flowchart of importing the failure mode analysis data into the failure mode analysis module, identifying high-risk components, and updating the fault threshold in the present invention; Figure 5 is a flowchart of calculating the influence coefficient of each failure mode in the present invention; Figure 6 is a flowchart of screening components of failure modes whose risk priority numbers exceed the dynamic threshold and recording them as high-risk components in the present invention; Figure 7 is a flowchart of updating the fault threshold in the present invention; Figure 8 is a flowchart of establishing an aging degree quantification model and calculating the aging coefficient in the present invention; Figure 9It is a schematic flow chart of constructing and developing an energy consumption - reliability joint optimization module based on real - time monitoring of the operating parameters of load adjustment equipment by a sensor network in the present invention, in combination with the fault threshold and aging coefficient of high - risk components. Detailed implementation manners

[0015] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments in the following description are only examples, and those skilled in the art can think of other obvious variations.

[0016] Refer to Figure 1 As shown, an intelligent management system for energy consumption optimization and fault self - diagnosis of a cleaning device includes: Obtain three groups of core parameters of the cleaning device, namely the historical vibration spectrum, the harmonic component of the motor current, and the bearing temperature gradient, construct a dynamic failure mode knowledge graph, and perform time - series correlation analysis on the historical fault data to obtain failure mode analysis data; Obtain the standard data released by the cleaning device manufacturer, establish a multi - dimensional analysis platform. The multi - dimensional analysis platform includes a failure mode analysis module, an aging assessment model, and an energy consumption optimization model. Import the failure mode analysis data into the failure mode analysis module to identify high - risk components and update the fault threshold; Import the historical equipment operation time, the number of working cycles, and environmental data into the aging assessment model, introduce a service - time decay factor and a working - condition correction coefficient, generate a risk heat map, combine the cumulative number of working cycles of the equipment and the environmental temperature and humidity parameters, establish an aging degree quantification model, and calculate the aging coefficient; Based on the sensor network monitoring, real - time monitor the operating parameters of the load adjustment equipment, combine the fault threshold and aging coefficient of high - risk components, construct and develop an energy consumption - reliability joint optimization module, and adjust the equipment operating parameters to reduce energy consumption.

[0017] Integrate three groups of heterogeneous data, namely the historical vibration spectrum, the harmonic component of the motor current, and the bearing temperature gradient, through timestamp alignment technology, adopt a time - frequency domain joint analysis method to extract the fundamental frequency feature vector, combine the calculation of the harmonic distortion rate and the construction of the temperature gradient tensor, and use a graph neural network to establish a dynamic correlation relationship between component state nodes and failure modes. Compared with the traditional single - parameter analysis method, this technology realizes the deep fusion of multi - physical quantity spatio - temporal characteristics and can accurately capture the coupled failure modes of bearing wear and motor winding abnormalities.

[0018] The multi-dimensional analysis platform constructs a collaborative calculation framework for failure mode analysis, aging assessment, and energy consumption optimization by importing standard operating condition parameters provided by equipment manufacturers. A calculation model for the criticality factor of components based on functional chain analysis is proposed. The importance of components is quantified by the ratio of the number of downstream functional modules to the total number of functional modules. Combining the probability of fault propagation and the maintenance cost impact coefficient, a three-dimensional risk priority assessment matrix is constructed. An adaptive threshold dynamic update mechanism based on the equipment operating state and environmental parameters is realized, improving the accuracy of high-risk component identification.

[0019] For equipment aging assessment, the service time decay factor and operating condition correction coefficient are introduced into the quantification model. By constructing a temperature-humidity - cycle number composite decay factor matrix and combining the spatial variance calculation of the bearing temperature gradient, a three-dimensional risk heat map is generated. A convolutional neural network is used to extract local texture features of the heat map, and an aging coefficient calculation model is constructed by fusing the friction loss index. Thus, the effect of predicting bearing failure in advance and accurately quantifying the performance decay rate under different environmental operating conditions is achieved.

[0020] In the energy consumption optimization module, the motor load rate and bearing torque fluctuation data are collected in real time through a high-density sensor network, and the load prediction for the next three working cycles is realized using a Kalman filter. Combining the time distribution curve of the aging coefficient, an energy consumption - reliability joint optimization algorithm based on task classification is designed. By dynamically adjusting the distribution strategy of low / high load tasks in the peak window, energy consumption is reduced while ensuring the operation cycle, and task scheduling is deeply coupled with the component health state.

[0021] Refer to Figure 2 As shown, the construction of the dynamic failure mode knowledge graph by obtaining three groups of core parameters, namely the historical vibration spectrum, motor current harmonic components, and bearing temperature gradient of the cleaning equipment, specifically includes: Obtain three groups of core parameters of the cleaning equipment, namely the historical vibration spectrum, motor current harmonic components, and bearing temperature gradient. After aligning the timestamps of the three groups of data, perform a joint time-frequency domain analysis on the historical vibration spectrum, extract the fundamental frequency, harmonic frequency, and sideband energy distribution eigenvectors, calculate the harmonic distortion rate of the motor current harmonic components and quantify the spectral sparsity, and generate the current anomaly index; Perform a sliding window difference process on the bearing temperature gradient to construct a temperature gradient tensor with two rows and two columns. The first row is the number of bearing surface temperature sensors, the second row is the number of time segments divided by the sliding window, and the two columns are the difference values of the axial and radial temperature change rates of the corresponding sensors within a specific time window respectively; Input the eigenvectors, current anomaly index, and temperature gradient matrix into a graph neural network to establish a dynamic association relationship between the failure mode edges with component state nodes, denoted as the knowledge graph topological structure; Based on the failure mode tags in the historical failure case library, the same node with the same timestamp is embedded into the knowledge graph topology to form a dynamic failure mode knowledge graph containing the fault propagation path.

[0022] Connect the 8 equally spaced temperature sensor nodes on the bearing housing surface according to their physical positions. The axial / radial temperature gradient difference value is used as the node attribute vector to ensure that the graph neural network can capture the spatial correlation features. At the same time, the IEEE 1588 Precision Time Protocol is used for timestamp alignment to achieve microsecond-level synchronization of vibration, current, and temperature data. First, a time synchronization mechanism based on the NTP protocol is used to align the millisecond-level timestamps of three groups of heterogeneous data: vibration spectrum, current harmonics, and temperature gradient. For the vibration spectrum data, short-time Fourier transform is used for joint time-frequency domain analysis. A Hanning window function is set for 128-point windowing. The fundamental frequency component is extracted through the spectral peak search algorithm, and the amplitude ratio of each harmonic component to the fundamental frequency is calculated as the harmonic frequency feature. For the sideband energy distribution, the frequency band division method is used to divide the spectrum into a frequency band range of fundamental frequency ±5%. The ratio of the energy integral in each sub-band to the total energy is calculated to form a 7-dimensional feature vector.

[0023] In the current harmonic analysis, the total harmonic distortion rate is calculated according to the IEC 61000-4-7 standard , where LH is the effective value of the hth harmonic current and LI is the effective value of the fundamental wave current. The Gini coefficient method is used for spectrum sparsity quantization, and the calculation method is , where x{k} is the spectrum amplitude arranged in ascending order and μ is the mean value of the spectrum amplitude. The product of THD and the Gini coefficient is used as the current anomaly index.

[0024] The sliding window mechanism is used for bearing temperature gradient processing, with the window length set to 30 seconds and the step size to 5 seconds. For the temperature data within each window, the axial temperature change rate is calculated, which is the ratio of the difference between the end temperature value and the start temperature value to the time difference. The radial temperature change rate is the difference between the maximum and minimum values in the sensor data installed at different radial positions on the bearing housing. In the constructed 2×2 temperature gradient tensor, the first dimension stores the mean value of ΔT in the axial and radial directions, and the second dimension stores the standard deviation of the two.

[0025] When the above features are input into the graph neural network, the GraphSAGE architecture is adopted to construct a heterogeneous graph structure. Three types of nodes are defined: vibration nodes (including fundamental frequency, 3rd harmonic frequency, 5th harmonic frequency features), current nodes (THD, Gini coefficient), and temperature nodes (axial / radial gradient). The edge attributes are jointly determined by the timestamp synchronization relationship and the physical connection relationship, and the attention mechanism is used to calculate the edge weights. During the training process, the fault type labels of historical fault cases are used for node classification, and the node embedding vectors are iteratively updated through graph convolution operations, finally forming a dynamic knowledge graph containing the fault propagation path. Thus, through the deep feature fusion of multi-source sensing data and the dynamic relationship modeling of the graph neural network, the association between time-frequency domain vibration features and electro-thermal parameters in a unified graph structure is realized, the problem of difficult cross-modal feature fusion is solved, and the Gini coefficient is effectively introduced to quantify the sparsity of the current spectrum, achieving the effect of distinguishing normal operating conditions from abnormal harmonic distributions caused by insulation aging. At the same time, the temperature gradient tensor processed by sliding window difference can capture both the spatial gradient and the time variation trend of the bearing temperature field, improving the detection sensitivity of early wear and the recognition accuracy of the knowledge graph for bearing-motor coupling faults.

[0026] Referring to Figure 3 as shown, the historical fault data is subjected to temporal correlation analysis to obtain the failure mode analysis data, which specifically includes: Align the timestamps of the fault event sequences in the historical fault data, extract the fundamental frequency offset of the vibration spectrum, the mutation value of the current harmonic distortion rate, and the extreme value of the bearing temperature gradient corresponding to each fault event, and construct a multi-dimensional time series database; Use a sliding time window to segment the multi-dimensional time series database, calculate the variance of the fundamental frequency offset, the change slope of the harmonic distortion rate, and the occurrence frequency of the extreme value of the temperature gradient within the window, summarize to obtain the index data of each index, and after normalizing the index data of each index, obtain the fault correlation intensity index of each time segment through weighted fusion; Use a temporal graph convolutional network to extract spatio-temporal features from the multi-dimensional time series matrix, capture the temporal causal relationship between cross-sensor parameters, and generate the probability weights of the fault propagation link; Input the fault correlation intensity index and the probability weights of the fault propagation link into a Bayesian network, calculate the conditional probability distribution of different failure modes in the time dimension, and output the failure mode analysis data containing temporal dependence relationships. The failure mode analysis data includes the fault precursor event sequence, the cross-component failure coupling probability, and the dynamic failure threshold drift trend.

[0027] First, align the timestamps of historical fault events at the millisecond level, and use the dynamic time warping algorithm to eliminate the differences in sensor sampling rates. For vibration signals, after extracting the fundamental frequency component through fast Fourier transform, calculate the fundamental frequency offset Δf between adjacent fault events, and its variance calculation formula is , where μ is the mean value within the window. For the current harmonic distortion rate THD sequence, the least squares method is used to fit the change slope within the time window . For the temperature gradient extreme value frequency statistics, the PeakOverThreshold method is used to set the dynamic threshold is the third quartile, and IQR is the interquartile range). The number of extreme values exceeding θ is counted, so that it can be calculated through the sliding window statistic to effectively capture the gradual change characteristics of early faults.

[0028] The hierarchical normalization framework performs logarithmic normalization on the fundamental frequency variance , the harmonic slope is Z-score normalized, and the extreme value frequency is Min-Max normalized. When performing weighted fusion, the weights of each index are dynamically calculated by the entropy weight method: calculate the information entropy of each index , weight , where p ij is the proportion of the i-th sample in the j-th index, and the final fault correlation strength index .

[0029] The time series graph convolutional network adopts a three-layer ST-GCN structure. The convolutional kernel in the time dimension is set to 3×1, and the adjacency matrix based on the physical connection relationship of the device is constructed in the space dimension. The network input is a three-dimensional tensor of [time step × feature dimension × number of sensor nodes]. The time propagation mode across nodes is extracted through spatio-temporal convolution, and the transition probability weight P(e|t) of each fault propagation path is output. The Bayesian network construction adopts a dynamic structure learning algorithm. The nodes include three categories: fault type, correlation strength interval, and component status. The conditional probability calculation formula is P(Failure|F,P)=α·P(F|Failure)·P(P|Failure)·P(Failure), where is the normalization factor, and the prior probability P(Failure) comes from historical fault statistics. The network parameters are updated in real time through variational inference, and the final output is the conditional probability distribution including the time decay effect, which can quantify the occurrence probability curve of different failure modes within 72 hours. The detection sensitivity for coupled faults is improved, the false alarm rate is reduced to, and the trend of the prediction threshold drift can be predicted.

[0030] Referring to Figure 4 as shown, the specific steps of importing the failure mode analysis data into the failure mode analysis module, identifying high-risk components, and updating the fault threshold include: Extract the probability weight of the fault propagation link as the benchmark value of the failure mode occurrence probability, obtain the downtime and repair cost caused by the same type of fault in the historical fault case library, and calculate the influence coefficient of each failure mode; Label the roles of each component in the functional chain, read the number of downstream functions and the total number of functions of each component, divide the number of downstream functions of each component by the total number of functions to obtain the component criticality factor, multiply the occurrence probability baseline value, the impact coefficient, and the component criticality factor to obtain the risk priority number. The number of downstream functions is the number of downstream functional modules directly affected when the component fails, and the total number of functions is the total number of all independent functional modules in the overall functional chain of the cleaning device; Screen the components of the failure modes whose risk priority numbers exceed the dynamic threshold, mark them as high-risk components, and update the failure threshold.

[0031] Extract the probability weight of the fault propagation link as the occurrence probability baseline value through the fault propagation path transition probability generated by the self-temporal graph convolutional network. Then, combined with the dataset of the downtime and repair cost of similar faults in the historical fault case library, adopt double processing of logarithmic normalization (downtime) and range normalization (repair cost) to construct a standardized impact factor. Through the user-preset downtime weight (0.6) and cost weight (0.4), generate the impact coefficient of each failure mode by weighting, and realize the quantitative integration of economic indicators and operation indicators.

[0032] During its actual use, determine the component criticality using the functional chain topology analysis method. By analyzing the control logic diagram of the cleaning device, construct a functional dependency directed graph structure. The number of downstream functions of each component is obtained by traversing through the depth-first search algorithm. For example, the failure of the motor component will directly affect 5 downstream functions such as the transmission module (level 1) and the brush disk drive (level 2). If the total functional chain contains 12 independent modules, then its criticality factor is 5 / 12 = 0.416. Multiply the occurrence probability baseline value, the impact coefficient, and the criticality factor to calculate the risk priority number (RPN). Among them, the criticality factor is used as an adjustment item to avoid misjudgment of components with high probability and low impact. That is, through the probability-impact-criticality three-dimensional model, components such as the bearing component (RPN = 158) and the motor component (RPN = 145) can be accurately distinguished due to the difference in criticality under the same fault probability.

[0033] The dynamic threshold generation mechanism adopts a dual-track system of sliding window statistics and neural network correction. With a 30-day window period, the mean value μ = 85 and standard deviation σ = 12 of the historical RPN are calculated, and the dynamic threshold baseline value is set to μ + 3σ = 121. At the same time, an LSTM neural network model is constructed, which inputs real-time parameters such as ambient temperature and humidity (25°C / 60%RH), current load rate (78%), and continuous operation duration (120 hours), and outputs a threshold correction coefficient β = 1.15. The final real-time dynamic threshold is 121×1.15 = 139. Compared with the traditional fixed threshold method, this mechanism enables the threshold to adapt to environmental mutations (such as the threshold automatically rising by 12% under high-temperature and high-humidity conditions), reduces the false alarm rate, and can effectively identify the dominant failure components in coupled faults. The service time decay factor is introduced in the threshold update link, so that the thresholds of equipment that has been in operation for many years are automatically lowered, which is more in line with the actual aging state and prevents potential undetected faults.

[0034] Refer to Figure 5 As shown, the specific calculation of the influence coefficient of each failure mode includes: Obtain the dataset of downtime duration and the dataset of maintenance cost corresponding to the same type of fault cases in the historical fault case library; Perform logarithmic normalization on the downtime duration data, take the natural logarithm of each downtime duration data, select the maximum value among them, and calculate the ratio of the natural logarithm of each downtime duration to the maximum value of all natural logarithms, which is recorded as the standardized downtime impact factor; Perform range normalization on the maintenance cost data, select the highest and lowest maintenance costs after normalization, calculate the difference between the highest and lowest maintenance costs, which is recorded as the standard range, and divide the difference between each maintenance cost and the lowest maintenance cost by the standard range to obtain the standardized cost impact factor; The user inputs the preset weights of downtime duration and maintenance cost, performs weighted summation on the standardized downtime impact factor and the standardized cost impact factor to generate the impact factor of each fault case, takes the arithmetic mean of the impact factors of the same type of fault cases, and outputs the impact coefficient of this failure mode.

[0035] Extract the dataset of downtime duration and maintenance cost of the same type of fault cases from the fault repair database of the equipment management system. For the downtime duration data, logarithmic normalization is used to eliminate the influence of the long-tailed distribution. The specific method is to take the natural logarithm of each downtime duration value, and the calculation formula is , (add 1 to avoid zero values). Through double normalization, it can not only eliminate the interference of the skewed distribution of downtime duration data, but also solve the problem of inconsistent dimensions of maintenance costs. Then, through the extreme value normalization formula Generate the standardized downtime impact factor, where the denominator takes the maximum natural logarithm value in the dataset. For the maintenance cost data, the range normalization method is used to eliminate the dimension difference, and the calculation formula is Among which C max and C min are the highest and lowest maintenance costs in the dataset respectively. Through a preset weight allocation mechanism, the user can input the downtime weight w1 and the maintenance cost weight w2 (which need to satisfy w1 + w2 = 1) to perform weighted fusion on the standardized two-dimensional impact factors and generate the impact factor of a single sample. Finally, by taking the arithmetic mean of all ϕi of similar failure cases, the impact coefficient of this failure mode is output. Introducing a user-defined weight mechanism enables equipment managers to flexibly adjust the importance of evaluation dimensions according to actual operation requirements (such as production continuity priority or cost control priority). At the same time, calculating the impact coefficient by arithmetic mean effectively reduces the interference of individual abnormal cases and improves the robustness of the evaluation results. The two cooperate with each other to accurately reflect the true severity of the failure impact in different scenarios and provide reliable input for subsequent risk priority calculation.

[0036] Refer to Figure 6 As shown, the components of the failure modes whose screened risk priority numbers exceed the dynamic threshold are recorded as high-risk components, specifically including: Obtain the risk priority numbers of all components in the current system, statistically analyze the distribution characteristics of historical risk priority numbers based on a sliding time window, calculate the mean and standard deviation of the risk priority numbers within the window, and take the sum of the mean of the risk priority numbers and three times the standard deviation as the dynamic threshold baseline value; Real-time collect the equipment operation status parameters, obtain the threshold correction model by training the mapping relationship between historical operation status and abnormal fluctuations of risk priority numbers through a neural network, input the equipment operation status parameters into the threshold correction model, and output the dynamic threshold correction coefficient. The equipment operation status parameters include environmental temperature and humidity, current load rate, and equipment continuous operation duration; Multiply the dynamic threshold baseline value by the dynamic threshold correction coefficient to generate the real-time dynamic threshold, summarize the ratio of the risk priority number to the real-time dynamic threshold to form ratio data, and perform spatial clustering analysis on the components with ratio data greater than 1 to eliminate isolated abnormal points caused by sensor noise; Based on the component groups formed after clustering, extract the functional relevance parameters of all components within the group, screen the groups where at least two components belong to the same functional chain and the functional relevance parameters exceed the preset threshold, and mark the entire group as a high-risk component set; Perform reverse traceability of the failure mode for each component in the high-risk component set, verify its risk priority number, and calculate the data integrity of the failure propagation link probability weight, impact coefficient, and component criticality factor involved in the process to generate a verified high-risk component list.

[0037] First, a dynamic threshold baseline generation mechanism based on a sliding window is constructed. With a 30-day time window period, the distribution characteristics of the historical risk priority numbers (RPN) of all components are statistically analyzed. The mean μ and standard deviation σ of the RPN within the window are calculated through Gaussian distribution fitting, and the dynamic threshold baseline value is set to μ + 3σ, thus covering 99.7% of the normal fluctuation range and effectively avoiding the interference of occasional outliers. At the same time, an environment condition adaptive model based on the LSTM neural network is built. The input layer receives three parameters: environmental temperature and humidity (normalized to the [0, 1] interval), the current load rate (collected by a current sensor and calculated as the percentage of the rated power), and the continuous operation duration of the device (discretized in hours). The hidden layer adopts a double gated recurrent unit structure, and the output layer generates a dynamic threshold correction coefficient within the range of 0.8 - 1.2 through the Sigmoid function. Through training with historical data, a non-linear mapping relationship between the operating state parameters and the abnormal fluctuations of the RPN is established, enabling the threshold to automatically float by 5% - 8% when the environmental temperature and humidity increase by 10% or the load rate exceeds 85%.

[0038] After calculating the real-time dynamic threshold as the product of the baseline value and the correction coefficient, a spatial clustering analysis is performed on the components with an RPN / threshold ratio > 1. The improved DBSCAN algorithm is used, with the neighborhood radius ε = 0.5 and the minimum number of samples minPts = 3, to perform density clustering in a three-dimensional space coordinate system (the dimensions are the physical location coordinates of the components, the functional hierarchy, and the signal transmission path length). By calculating the silhouette coefficient of each cluster, the isolated points with a coefficient < 0.25 are determined as abnormal data caused by sensor noise and excluded. For the remaining clusters, a functional correlation matrix is constructed, and the functional correlation parameter is defined as the product of the number of overlaps of the signal transmission paths between components and the shared power supply circuit. When the correlation parameter of more than two components within the same cluster in the functional chain > 15, it is determined as a high-risk component set.

[0039] Finally, through the reverse traceability verification mechanism, the integrity of the RPN calculation parameters of each component in the set is verified: check whether the probability weight of the fault propagation link comes from the output of the temporal graph convolutional network, whether the influence coefficient is calculated through double-weight fusion, and whether the component criticality factor is generated based on the functional chain topology analysis. Thus, through the coupled calculation of the three-fold standard deviation baseline value and the LSTM correction coefficient, the threshold can dynamically adapt to equipment aging and environmental mutations. The dual verification mechanisms of spatial clustering and functional correlation can eliminate more than 90% of false alarm cases. At the same time, the reverse traceability verification ensures the integrity of the data traceability chain, improves the recognition accuracy of high-risk components, and reduces the false alarm rate.

[0040] Refer to Figure 7 As shown, the specific update of the fault threshold includes: Obtain the attenuation factor associated with the aging coefficient, take the logarithm after adding 1 to the product of the component criticality factor and the attenuation factor associated with the aging coefficient to obtain an intermediate value, and calculate the product of the intermediate value and the real-time dynamic threshold to obtain the initial threshold adjustment amount; Construct an environment-load correction matrix through the real-time collected environmental temperature and humidity parameters and load volatility, perform a tensor product operation on the initial threshold adjustment amount and the correction matrix to generate a threshold offset amount compensated for environmental conditions. The rows of the environment-load correction matrix are environmental parameters, and the columns are the periods of load volatility; Set the sliding window length to the preset number of days, use the exponentially weighted moving average model to fit the trend of the threshold drift amount within the window, extract the model residuals for normal test. If the residuals follow a normal distribution, construct a confidence interval according to the sliding window distribution. If the residuals are non-normal, directly take the quantiles of the residuals as the upper and lower bounds. When the threshold offset amount exceeds the confidence interval, use the exponential smoothing method to perform dynamic damping processing on the offset amount; Add the processed threshold offset amount to the original fault threshold to generate an updated dynamic fault threshold and synchronize it to the alarm threshold database of the fault self-diagnosis module; Perform cross-device group verification on the updated dynamic fault threshold. Select the threshold median of non-high-risk components in the same model device group as the reference value. When the deviation degree between the updated threshold and the reference value exceeds 20%, trigger threshold feedback and generate a calibration alarm signal.

[0041] Establish a threshold adjustment amount model through the coupled calculation of the component criticality factor and the aging attenuation factor. Define the attenuation factor as the weighted sum of the temperature-humidity-cycle times composite attenuation factor and the friction loss index. Take the natural logarithm after adding 1 to the product of the component criticality factor (such as 0.416) and the attenuation factor (such as 0.85) to obtain an intermediate value (such as ln(0.416×0.85 + 1)=0.28), and multiply it by the real-time dynamic threshold (such as 139) to obtain the initial adjustment amount of 39.0. The environment-load correction matrix is constructed as a 2×3 structure. The row dimension divides the environmental parameter intervals (temperature < 25°C, 25 - 35°C, > 35°C), and the column dimension corresponds to the load volatility intervals (< 5%, 5 - 15%, > 15%). The matrix elements store the empirical coefficients of the historical threshold offset amount. Perform condition compensation through tensor product operation. For example, in the high-temperature and high-load condition, the correction coefficient can reach 1.3, so that the initial adjustment amount 39.0×1.3 = 50.7.

[0042] When using the exponentially weighted moving average model to handle threshold drift, the sliding window is set to 7 days and the forgetting factor λ = 0.2. After trend fitting of the offsets within the window, Anderson-Darling test is used for residual analysis. When the p-value > 0.05, a 95% normal confidence interval is constructed; otherwise, the quantile method is used to determine the upper and lower bounds. When the compensated offset (such as +12.5) exceeds the confidence interval [-5, +8], Holt exponential smoothing method is applied for damping treatment with the smoothing coefficient α = 0.3, so that the offset is adjusted to 12.5 × 0.7 = 8.75. The processed offset is superimposed on the original threshold to generate an updated threshold (such as the original threshold 200 → 208.75).

[0043] In the cross-device verification stage, the median value of the thresholds of non-high-risk components in the same model device group is selected as the benchmark (such as 205). The deviation degree is calculated as |208.75 - 205| / 205 × 100% = 1.83%, which is lower than the 20% threshold, so it is determined to be valid. When the deviation degree exceeds the limit, a calibration signal is triggered and the threshold update is frozen, waiting for manual review. Thus, the accuracy of threshold adjustment can be improved through the dynamic matrix compensation mechanism, the false alarm rate under the condition of sudden change of temperature and humidity can be reduced, and the residual adaptive test mechanism can effectively eliminate abnormal offset interference, significantly improving the reliability of the alarm system.

[0044] Refer to Figure 8 As shown, the establishment of the aging degree quantification model and the calculation of the aging coefficient specifically include: Obtain the cumulative working cycle times of the device, and combine with the temperature and humidity parameters collected by the environmental temperature and humidity sensors to obtain the temperature and humidity - cycle times composite attenuation factor through weighting; Based on the maximum eigenvalue of the bearing temperature gradient matrix, calculate the friction loss index of the rolling elements and the cage, introduce the service time attenuation factor, and construct the exponential attenuation relationship between friction loss and service time; Divide the environmental temperature and humidity parameters into temperature gradient intervals and humidity gradient intervals. In the coordinate system, use the temperature and humidity gradient intervals and the working cycle times as the coordinate axes respectively, calculate the variance value of the bearing temperature gradient within each cubic unit, and generate a risk heat map; Extract the local texture features of the risk heat map through a convolutional neural network, and combine with the working condition correction coefficient to smooth and correct the heat distribution, and output the aging feature vector; Weight the service time attenuation factor, the friction loss index and the aging feature vector to obtain the aging coefficient.

[0045] First, the cumulative working cycle times N are collected through the device controller, and at the same time, the data of the environmental temperature and humidity sensors distributed at the key parts of the device are read. A double-layer weighting strategy is used to generate the temperature and humidity - cycle times composite attenuation factor: in the first layer, Gaussian normalization is performed on the temperature T and humidity H respectively, and their standard scores are calculated (where μ and σ are the standard operating condition parameters of the device); the second layer dynamically couples the normalized temperature and humidity parameters with the number of cycles, and the attenuation factor formula is designed as , where α is the material characteristic coefficient (0.12 for bearing steel and 0.08 for ceramic cage). This formula strengthens the influence of high cycle numbers by introducing a logarithmic function and highlights the accelerated aging effect of high humidity environment by using quadratic terms, which is more in line with the actual aging curve compared with the traditional linear weighting method.

[0046] In the aspect of bearing temperature gradient matrix processing, the principal component analysis method is used to extract the maximum eigenvalue σ of the matrix max , and substitute it into the improved Archard wear formula to calculate the friction loss index: , where K is the material wear coefficient, P is the contact pressure (calculated by Hertz contact theory), H is the material hardness, and β is the service time attenuation factor .

[0047] By introducing an exponential function term, the friction loss calculation considers both the instantaneous temperature gradient characteristics and the cumulative effect of the equipment service time, effectively distinguishing the superposition effect of instantaneous wear caused by sudden overload and long-term fatigue damage.

[0048] In the risk heat map generation link, the environmental temperature and humidity are divided into gradient intervals (the temperature interval is 5°C, and the humidity interval is 10%RH), and the number of working cycles is segmented in units of thousands. After constructing a three-dimensional coordinate system, the joint variance of the bearing axial and radial temperature gradients is calculated in each cubic unit: , and the bicubic interpolation algorithm is used to fill in the missing data to generate a continuously distributed thermal field. Compared with the traditional two-dimensional heat map, the three-dimensional thermal model can simultaneously characterize the interaction between environmental parameters and mechanical loads, and accurately locate the abnormal wear area under high temperature and high humidity conditions.

[0049] The convolutional neural network adopts an improved U-Net architecture, with 4 downsampling layers (each layer contains 3×3 convolution + ReLU + max pooling) in the encoder part, and 4 upsampling layers (bilinear interpolation + skip connection) corresponding in the decoder part. After the output layer, a working condition correction module is connected, and its mathematical expression is: , where W_c and b_c are learnable parameters, and M_env is the environmental mask matrix generated according to the real-time load rate. Thus, the texture feature response of the abnormal area can be enhanced through the frequency domain attention mechanism, and at the same time, the non-related noise can be suppressed by using the working condition parameters, improving the signal-to-noise ratio of the aging feature vector.

[0050] The final aging coefficient calculation adopts a dynamic weight allocation strategy: , the weight coefficients are determined by backpropagation optimization of historical fault data to ensure the balance of the contribution degrees of friction loss, thermal characteristics, and composite attenuation factors. Thus, the prediction error of the remaining bearing life can be reduced, and the prediction stability under high-temperature (>50°C) and high-humidity (>80%RH) working conditions can be improved. By integrating multi-physical field parameters and deep learning features, an accurate quantitative assessment of the equipment aging degree is achieved, providing reliable health status input for subsequent energy consumption optimization.

[0051] Refer to Figure 9 As shown, the real-time monitoring of the operation parameters of the load adjustment device based on the sensor network, combined with the fault threshold and aging coefficient of high-risk components, constructs and develops an energy consumption-reliability joint optimization module, which specifically includes: The motor load rate is monitored in real time through a current sensor, and the Kalman filter is used to predict the load fluctuation to generate a load prediction curve for the next three working cycles; Based on historical task data, load characteristic parameters are extracted, the tasks to be executed by the cleaning device are classified, and the task threshold is input by the user. When the motor power demand and the bearing friction torque volatility are lower than the task threshold, the task is defined as a low-load task, and the remaining tasks are divided into medium and high load levels according to the power range; The sliding time window is used to analyze the time distribution curve of the aging coefficient of high-risk components, and the ratio of the standard deviation to the mean of the aging coefficient within each time window is calculated. When the ratio exceeds the preset threshold, this time period is marked as the peak window of the aging coefficient. During the same time period, low-load tasks preferentially occupy the peak window of the aging coefficient, and medium and high load tasks are not allowed to continuously occupy more than 2 adjacent time windows. The task execution must meet the preset operation cycle deadline; The user inputs the error threshold, and the difference between the real-time motor load monitored by the current sensor and the load prediction curve is used to obtain the real-time load prediction error. When the real-time load prediction error exceeds the error threshold, the sliding window is re-planned, and the task allocation for the next three cycles is recalculated. For the medium and high load tasks scheduled to the peak window, the power output of the peak window is reduced by adjusting the duty cycle of the motor; The optimized task timing queue is output, with low-load tasks concentratedly deployed during the peak period of the aging coefficient, while ensuring that medium and high load tasks are evenly distributed during non-peak periods, finally generating an energy consumption-reliability joint optimization module.

[0052] First, a dynamic prediction framework based on the Kalman filter is built. The motor load rate is used as the state variable (including two dimensions: the current load value and the load change acceleration), and the observation matrix is set as [1 0] to directly obtain the real-time data of the current sensor. The process noise covariance matrix is adaptively calibrated through historical load volatility. Through the time update equation: And the measurement update equation: Iterative calculation is performed to generate the load prediction curves for the next three working cycles. The time step is set to the typical operation cycle of the cleaning device (15 minutes), thereby reducing the prediction error. The task classification module extracts two key parameters by parsing the historical task database: the motor power demand (processed by range normalization) and the bearing friction torque volatility (calculated based on the instantaneous frequency variance using the Hilbert-Huang transform). The user can set task thresholds (default settings are 30% of the rated power and a torque volatility of 0.08). When both parameters are below the thresholds, the task is marked as a low-load task. The remaining tasks are classified according to the power interval division rule: 30% - 60% of the rated power is medium load, and above 60% is high load, forming a three-level task classification system.

[0053] In the aging coefficient time series analysis section, a sliding time window mechanism is constructed (window length 2 hours, step size 15 minutes) to monitor the aging coefficient sequence of high-risk components in real time. Calculate the coefficient of variation of the standard deviation σ and the mean μ of the aging coefficient within each window , when the CV value exceeds the preset threshold of 0.35, this period is marked as the aging coefficient peak window. The task scheduling algorithm adopts a dynamic priority allocation strategy. During the peak window period, low-load tasks are forced to be deployed to this period. Medium and high-load tasks are calculated according to the job cycle deadline constraint through a backtracking algorithm to find the earliest available non-peak period, and the interval between medium and high-load tasks in adjacent time windows is set to be no less than 45 minutes. When the real-time load prediction error (the absolute difference between the measured value and the predicted value) exceeds the user-set error threshold (default 5%), the sliding window replanning mechanism is triggered: re-solve the task queue for the next three cycles. For medium and high-load tasks that have been scheduled to the peak window, the PWM voltage regulation technology is used to dynamically adjust the motor duty cycle, and the power output is limited to the maximum value allowed in the peak window. The calculation formula is At the same time, continuous high-load tasks are decomposed into multiple subtasks and distributed across windows through task splitting technology.

[0054] The finally generated optimized task queue has two characteristics: a low-load task cluster is formed during the aging coefficient peak period (cluster density ≥ 3 tasks / hour), and medium and high-load tasks in non-peak periods are evenly distributed (standard deviation of the interval between adjacent tasks ≤ 8 minutes). This can reduce the peak load exposure time of high-risk components, lower the overall energy consumption, and improve the task deadline achievement rate.

[0055] In summary, the advantages of the present invention are as follows: By integrating multi-source sensor data to construct a dynamic knowledge graph and a time series analysis model, accurate diagnosis of coupled faults and self-adaptive threshold adjustment of cleaning devices are achieved. Combining aging assessment and task scheduling optimization significantly improves energy consumption efficiency, extends the equipment life, and reduces maintenance costs.

[0056] The basic principles, main features and advantages of the present invention have been shown and described above. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, various changes and improvements will occur to the present invention, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection required by the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent management system for energy consumption optimization and fault self-diagnosis of a cleaning device, characterized in that Including: Obtain three groups of core parameters, namely the historical vibration spectrum, motor current harmonic components, and bearing temperature gradient of the cleaning device, construct a dynamic failure mode knowledge graph, and conduct time-series correlation analysis on historical fault data to obtain failure mode analysis data; Obtain the standard data released by the cleaning device manufacturer, establish a multi-dimensional analysis platform. The multi-dimensional analysis platform includes a failure mode analysis module, an aging assessment model, and an energy consumption optimization model. Import the failure mode analysis data into the failure mode analysis module to identify high-risk components and update the fault threshold; Import the historical equipment operation time, number of work cycles, and environmental data into the aging assessment model, introduce a service time decay factor and a working condition correction coefficient to generate a risk heat map. Combine the cumulative number of work cycles of the equipment and the environmental temperature and humidity parameters to establish an aging degree quantification model and calculate the aging coefficient; Based on the real-time monitoring of the load adjustment equipment operation parameters by the sensor network, combine the fault threshold of high-risk components and the aging coefficient to construct and develop an energy consumption-reliability joint optimization module to adjust the equipment operation parameters to reduce energy consumption.

2. The intelligent management system for optimizing energy consumption and self-diagnosing faults of a cleaning device according to claim 1, characterized in that, The specific steps of obtaining the three groups of core parameters, namely the historical vibration spectrum, motor current harmonic components, and bearing temperature gradient of the cleaning device, and constructing a dynamic failure mode knowledge graph are as follows: Obtain the three groups of core parameters, namely the historical vibration spectrum, motor current harmonic components, and bearing temperature gradient of the cleaning device. After aligning the timestamps of the three groups of data, conduct a joint time-frequency domain analysis on the historical vibration spectrum, extract the fundamental frequency, harmonic frequency, and sideband energy distribution feature vectors, calculate the harmonic distortion rate of the motor current harmonic components and quantify the spectral sparsity to generate a current anomaly index; Perform sliding window difference processing on the bearing temperature gradient to construct a temperature gradient tensor with two rows and two columns. The first row is the number of bearing surface temperature sensors, the second row is the number of time segments divided by the sliding window, and the two columns are the axial and radial temperature change rate difference values stored for the corresponding sensors within a specific time window; Input the feature vectors, current anomaly index, and temperature gradient matrix into a graph neural network to establish a dynamic association relationship of the failure mode edges with component status nodes, denoted as the knowledge graph topology structure; Based on the failure mode labels in the historical fault case library, embed the same nodes with the same timestamp into the knowledge graph topology structure to form a dynamic failure mode knowledge graph containing the fault propagation path.

3. An intelligent management system for energy consumption optimization and fault self-diagnosis of a cleaning device according to claim 2, characterized in that, The specific steps of conducting time-series correlation analysis on the historical fault data to obtain failure mode analysis data are as follows: Align the timestamps of the fault event sequences in the historical fault data, extract the fundamental frequency offset of the vibration spectrum, the mutation value of the current harmonic distortion rate, and the extreme value of the bearing temperature gradient corresponding to each fault event, and construct a multi-dimensional time series database; Segment the multi-dimensional time series database using a sliding time window, calculate the variance of the fundamental frequency offset, the change slope of the harmonic distortion rate, and the occurrence frequency of the extreme value of the temperature gradient within the window, summarize to obtain the data of each index. After normalizing the data of each index, obtain the fault correlation intensity index of each time segment through weighted fusion. Use a temporal graph convolutional network to extract spatio-temporal features from a multi-dimensional time series matrix, capture the temporal causal relationships of cross-sensor parameters, and generate the probability weights of the fault propagation link; Input the fault correlation intensity index and the fault propagation link probability weight into a Bayesian network, calculate the conditional probability distribution of different failure modes in the time dimension, and output the failure mode analysis data containing temporal dependence relationships. The failure mode analysis data includes the fault precursor event sequence, the cross-component failure coupling probability, and the dynamic failure threshold drift trend.

4. An intelligent management system for optimizing energy consumption and self-diagnosing faults of a cleaning device according to claim 3, characterized in that, Importing the failure mode analysis data into the failure mode analysis module, identifying high-risk components, and updating the fault threshold specifically includes: Extract the fault propagation link probability weight as the failure mode occurrence probability baseline value, obtain the downtime duration and repair cost caused by the same type of fault in the historical fault case library, and calculate the impact coefficient of each failure mode; Label the role of each component in the function chain, read the number of downstream functions and the total number of functions of each component, divide the number of downstream functions of each component by the total number of functions, and obtain the component criticality factor. Multiply the occurrence probability baseline value, the impact coefficient, and the component criticality factor to get the risk priority number. The number of downstream functions is the number of downstream function modules directly affected when the component fails, and the total number of functions is the total number of all independent function modules in the overall function chain of the cleaning device; Screen the components of the failure mode whose risk priority number exceeds the dynamic threshold, record them as high-risk components, and update the fault threshold.

5. The intelligent management system for optimizing energy consumption and self-diagnosing faults of a cleaning device according to claim 4, characterized in that, The calculation of the impact coefficient of each failure mode specifically includes: Obtain the downtime duration data set and the repair cost data set corresponding to the same type of fault cases in the historical fault case library; Perform logarithmic normalization on the downtime duration data, take the natural logarithm of each downtime duration data, and select the maximum value among them. Calculate the ratio of the natural logarithm of each downtime duration to the maximum value of all natural logarithms, and record it as the normalized downtime impact factor; Perform range normalization on the repair cost data, select the highest and lowest repair costs after normalization, calculate the difference between the highest and lowest repair costs, and record it as the standard range. Divide the difference between each repair cost and the lowest repair cost by the standard range to obtain the normalized cost impact factor; The user inputs the preset downtime duration weight and repair cost weight, perform weighted summation on the normalized downtime impact factor and the normalized cost impact factor to generate the impact factor of each fault case, take the arithmetic mean of the impact factors of the same type of fault cases, and output the impact coefficient of this failure mode.

6. The intelligent management system for optimizing energy consumption and self-diagnosing faults of a cleaning device according to claim 5, characterized in that, The screening of the components of the failure mode whose risk priority number exceeds the dynamic threshold and recording them as high-risk components specifically includes: Obtain the risk priority numbers of all components in the current system, statistically analyze the distribution characteristics of the historical risk priority numbers based on a sliding time window, calculate the mean and standard deviation of the risk priority numbers within the window, and use the sum of the mean of the risk priority numbers and three times the standard deviation as the dynamic threshold baseline value; Real-time collect the operating state parameters of the device, obtain a threshold correction model by training the mapping relationship between the historical operating state and the abnormal fluctuation of the risk priority number through neural network, input the device operating state parameters into the threshold correction model, and output a dynamic threshold correction coefficient. The device operating state parameters include environmental temperature and humidity, current load rate, and device continuous operation duration; Multiply the dynamic threshold baseline value by the dynamic threshold correction coefficient to generate a real-time dynamic threshold. Summarize the ratio of the risk priority number to the real-time dynamic threshold to form ratio data. Conduct spatial clustering analysis on the components with ratio data greater than 1, and eliminate isolated abnormal points caused by sensor noise; Based on the component groups formed after clustering, extract the functional correlation parameters of all components within the group. Screen that there are at least two components within the group belonging to the same functional chain, and the functional correlation parameters exceed the preset threshold, and mark the entire group as a high-risk component set; Conduct reverse traceability of the failure mode for each component in the high-risk component set, verify its risk priority number, and check the data integrity of the fault propagation link probability weight, influence coefficient, and component criticality factor involved in the calculation process to generate a verified high-risk component list.

7. The intelligent management system for optimizing energy consumption and self-diagnosing faults of a cleaning device according to claim 6, characterized in that The specific update of the fault threshold includes: Obtain the attenuation factor associated with the aging coefficient, take the logarithm after adding 1 to the product of the component criticality factor and the attenuation factor associated with the aging coefficient to obtain an intermediate value, and calculate the product of the intermediate value and the real-time dynamic threshold to obtain the initial threshold adjustment amount; Construct an environment-load correction matrix through the real-time collected environmental temperature and humidity parameters and load volatility. Perform a tensor product operation on the initial threshold adjustment amount and the correction matrix to generate a threshold offset after environmental condition compensation. The rows of the environment-load correction matrix are environmental parameters, and the columns are the periods of load volatility; Set the sliding window length to the preset number of days. Use the exponentially weighted moving average model to fit the trend of the threshold drift amount within the window, extract the model residuals for normal test. If the residuals follow a normal distribution, construct a confidence interval according to the sliding window distribution. If the residuals are non-normal, directly take the quantiles of the residuals as the upper and lower bounds. When the threshold offset exceeds the confidence interval, use the exponential smoothing method to perform dynamic damping processing on the offset; Add the processed threshold offset to the original fault threshold to generate an updated dynamic fault threshold, and synchronize it to the alarm threshold database of the fault self-diagnosis module; Conduct cross-device group verification on the updated dynamic fault threshold. Select the threshold median of non-high-risk components in the same model device group as the reference value. When the deviation degree of the updated threshold from the reference value exceeds 20%, trigger threshold feedback and generate a calibration warning signal.

8. The intelligent management system for optimizing energy consumption and self-diagnosing faults of a cleaning device according to claim 7, characterized in that, The establishment of the aging degree quantification model and the calculation of the aging coefficient specifically include: Obtain the cumulative working cycle times of the device, and combine the temperature and humidity parameters collected by the environmental temperature and humidity sensor to obtain a temperature and humidity-cycle times composite attenuation factor through weighting; Based on the maximum eigenvalue of the bearing temperature gradient matrix, calculate the friction loss index of the rolling element and the cage, introduce the service time attenuation factor, and construct an exponential decay relationship between friction loss and service time; Divide the environmental temperature and humidity parameters into temperature gradient intervals and humidity gradient intervals. In the coordinate system, use the temperature and humidity gradient intervals and the number of working cycles as the coordinate axes, calculate the variance value of the bearing temperature gradient in each cubic unit, and generate a risk thermal map; Extract the local texture features of the risk thermal map through a convolutional neural network, combine the working condition correction coefficient to smooth and correct the thermal distribution, and output the aging feature vector; Weight the service time decay factor, friction loss index and aging feature vector to obtain the aging coefficient.

9. The intelligent management system for optimizing energy consumption and self-diagnosing faults of a cleaning device according to claim 8, characterized in that, The real-time monitoring of the operation parameters of the load adjustment device based on the sensor network, combined with the fault threshold and aging coefficient of high-risk components, to construct the development of the energy consumption-reliability joint optimization module specifically includes: Real-time monitor the motor load rate through the current sensor, use the Kalman filter to predict the load fluctuation, and generate the load prediction curve for the next three working cycles; Extract the load characteristic parameters based on the historical task data, classify the tasks to be executed by the cleaning equipment, and input the task threshold by the user. When the motor power demand and the bearing friction torque volatility are lower than the task threshold, this task is defined as a low-load task, and the remaining tasks are divided into medium and high load levels according to the power interval; Adopt a sliding time window to analyze the time distribution curve of the aging coefficient of high-risk components, calculate the ratio of the standard deviation to the mean value of the aging coefficient within each time window. When the ratio exceeds the preset threshold, mark this time period as the peak window of the aging coefficient. Within the same time period, low-load tasks preferentially occupy the peak window of the aging coefficient, and medium and high load tasks are not allowed to continuously occupy more than 2 adjacent time windows. The task execution must meet the preset operation cycle deadline; The user inputs the error threshold, subtract the real-time motor load monitored by the current sensor from the load prediction curve to obtain the real-time load prediction error. When the real-time load prediction error exceeds the error threshold, trigger the sliding window replanning, recalculate the task allocation for the next three cycles, and for the medium and high load tasks scheduled to the peak window, reduce the power output of the peak window by adjusting the duty cycle of the motor; Output the optimized task time sequence queue, concentrate the low-load tasks in the peak period of the aging coefficient, and at the same time ensure that the medium and high load tasks are evenly distributed in the non-peak period, and finally generate the energy consumption-reliability joint optimization module.

Images