Decision-making system for equipment maintenance
By installing sensors on the air compressor and designing maintenance decision-making systems, and establishing a equipment health prediction model, the problem of inappropriate maintenance schedule of the existing technology air compressor is solved, and a more reasonable maintenance strategy and cost reduction are achieved.
Patent Information
- Application Number
- CN202311811644.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-12-11
- Filing Date
- 2023-12-26
- Publication Date
- 2025-06-13
AI Technical Summary
The prior art is difficult to grasp the efficiency of the equipment immediately during maintenance of air compressors, resulting in inappropriate maintenance schedule and increasing unnecessary maintenance costs.
Design a decision-making system for equipment maintenance, detect equipment parameters through multiple sensors, and use the maintenance decision-making department to establish a health prediction model of the equipment, and decide on maintenance schedule and planning based on health scores.
It realizes instant understanding of equipment efficiency and reminds personnel to maintain specific parts, ensures that the maintenance schedule is appropriate and reduces unnecessary maintenance costs.
Smart Images
Figure CN120144629A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a device maintenance technology, and particularly to a decision-making system for device maintenance. Background Art
[0002] With the development of 5G and AIoT technologies, enterprises can obtain a large amount of manufacturing data from various machine sensors. However, if they want to make big data generate practical value, they need to be able to integrate data information and achieve the goals of fault prediction and health management.
[0003] Taking an air compressor as an example, in the semiconductor industry, the maintenance timing and frequency of air compressors mainly refer to the schedule recommended by manufacturers and experts, that is, the so-called preventive maintenance. Although it can effectively prevent machine damage, in order to avoid unexpected failures, a relatively conservative maintenance strategy is usually adopted, resulting in over-maintenance and unnecessary maintenance costs.
[0004] If the efficiency status of the device can be grasped in real time, not only can it remind personnel to perform maintenance on specific parts when the efficiency is too low, but also ensure that the maintenance schedule is planned at an appropriate time to reduce unnecessary maintenance costs. Summary of the Invention
[0005] In view of the above description, the present invention provides a decision-making system for device maintenance, which can establish a health prediction model of the device based on the parameters of the device, and make decisions on the maintenance schedule and plan of the device according to the obtained health score.
[0006] According to an embodiment of the present invention, a decision-making system for device maintenance is provided. The decision-making system for device maintenance includes: a plurality of sensors disposed on the device to be maintained for detecting a plurality of parameters of the device to be maintained; and a maintenance decision-making unit that receives the plurality of parameters and outputs a health score of the device to be maintained. The maintenance decision-making unit further includes: a parameter weight table calculation unit that calculates weights for each of the plurality of parameters to generate weight values for each of the plurality of parameters; a decision model establishment unit that receives the plurality of parameters, generates a health prediction model of the device to be maintained, and outputs a weighted mean square error based on the weight values of each of the plurality of parameters and the health prediction model; and a health score calculation unit that uses a piecewise linear conversion method to convert the weighted mean square error into a health score.
[0007] According to an embodiment of the present invention, in the above-mentioned decision-making system for equipment maintenance, the parameter weight table calculation unit further performs: for each of the multiple parameters, calculate the variance based on the expert-recommended parameter weight, the parameter abnormality probability, and the parameter repair time cost respectively; calculate the parameter index value for each of the multiple parameters based on the maximum variance among the coefficient of variation of the expert-recommended parameter weight, the parameter abnormality probability, and the parameter repair time cost respectively; group the parameter index values of the multiple parameters, and assign corresponding weights to each group.
[0008] According to an embodiment of the present invention, in the above-mentioned decision-making system for equipment maintenance, when the variance is above the first threshold, the parameter weight table calculation unit calculates the parameter index value by the weighted average method for the expert-recommended parameter weight, the parameter abnormality probability, and the parameter repair time cost. When the variance is below the first threshold and above the second threshold, the parameter weight table calculation unit calculates the parameter index value by the multiplication method for the expert-recommended parameter weight, the parameter abnormality probability, and the parameter repair time cost. When the variance is below the second threshold, the parameter weight table calculation unit calculates the parameter index value by the exponential method for the expert-recommended parameter weight, the parameter abnormality probability, and the parameter repair time cost. The first threshold is greater than the second threshold.
[0009] According to an embodiment of the present invention, in the above-mentioned decision-making system for equipment maintenance, the parameter weight table calculation unit groups the parameter index values by a machine learning method. According to an embodiment of the present invention, in the above-mentioned decision-making system for equipment maintenance, the larger the parameter index value, the greater the weight.
[0010] According to an embodiment of the present invention, in the above-mentioned decision-making system for equipment maintenance, the decision model establishment unit uses a machine learning algorithm to establish a decision model for predictive maintenance.
[0011] According to an embodiment of the present invention, in the above-mentioned decision-making system for equipment maintenance, the decision model establishment unit is an unsupervised learning autoencoder. The autoencoder further includes: an encoder, which receives the multiple parameters, compresses the multiple parameters, and presents them in a low-dimensional encoding to extract feature values; and a decoder, which encodes the output of the encoder to perform data reconstruction to establish the health prediction model.
[0012] According to an embodiment of the present invention, in the above-mentioned decision-making system for equipment maintenance, the autoencoder calculates the mean square error based on the weights of the multiple parameters to generate the weighted mean square error.
[0013] According to an embodiment of the present invention, in the decision-making system for equipment maintenance described above, the health score calculation unit further performs: dividing the value range of the weighted mean square error into multiple intervals; in each of the multiple intervals, converting the weighted mean square error into the health score through linear transformation.
[0014] According to an embodiment of the present invention, in the decision-making system for equipment maintenance described above, the multiple intervals are divided by 0, the first quartile, the second quartile, the third quartile, mild outliers, extreme outliers, and the maximum value in the value range of the weighted mean square error.
[0015] According to an embodiment of the present invention, in the decision-making system for equipment maintenance described above, the health score is further compared with a health score threshold, and when the health score is less than the health score threshold, the equipment to be maintained is maintained.
[0016] According to an embodiment of the present invention, in the decision-making system for equipment maintenance described above, the maintenance decision-making unit further includes a false alarm processing unit, and the false alarm processing unit calculates the health score using the moving average method to generate a trend line of the health score.
[0017] According to an embodiment of the present invention, the decision-making system for equipment maintenance described above further includes an output unit, and the output unit is a display for outputting the health score and the trend line of the health score.
[0018] According to an embodiment of the present invention, in the decision-making system for equipment maintenance described above, the display includes: a first area for displaying the health score and the trend line of the health score; and a second area for displaying a control chart of the multiple parameters corresponding to the trend line.
[0019] According to an embodiment of the present invention, in the decision-making system for equipment maintenance described above, the data of the multiple parameters received by the maintenance decision-making unit is deseasonalized, and the deseasonalization process is calculated using the central moving average method.
[0020] Based on the above, using technologies such as data mining, big data analysis, and artificial intelligence, a health prediction system for equipment to be maintained (such as air compressors, etc.) is established, and the analysis results are visually output. In addition to instantaneously transmitting the overall health information of the equipment to be maintained, this system can also report the key parameters affecting the health, enabling plant maintenance personnel to formulate an optimized maintenance strategy and avoid additional costs caused by unexpected failures of the equipment to be maintained. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a schematic diagram of the decision-making architecture for equipment maintenance drawn according to an embodiment of the present invention
[0022] Figure 2 Illustration Figure 1 Schematic diagram of the maintenance decision-making section in
[0023] Figure 3 Schematic diagram of the generation process of the parameter weight table in the maintenance decision-making section illustrated according to an embodiment of the present invention.
[0024] Figure 4 Schematic diagram of the architecture of the model selection section in the maintenance decision-making section illustrated according to an embodiment of the present invention.
[0025] Figure 5 Schematic diagram of the concept of converting weighted mean square error into health score illustrated according to an embodiment of the present invention.
[0026] Figures 6A to 6C Health score trend chart of the false alarm handling method illustrated according to an embodiment of the present invention.
[0027] Figure 7 Schematic diagram of the result presentation of the health score trend chart illustrated according to an embodiment of the present invention.
[0028] Figure 8 Control chart of each parameter affecting the decrease of the health score in the embodiment of the present invention.
[0029] Figure 9 Illustration of the presentation of the health score of the device and the control of each parameter.
[0030]
Symbol Explanation
[0031] 1: Maintenance decision-making system
[0032] 10: Sensor
[0033] 20: Data preprocessing section
[0034] 30: Seasonal change judgment section
[0035] 40: Seasonal de-trending section
[0036] 50: Data standardization section
[0037] 60: Maintenance decision-making section
[0038] 60A: Parameter weight table calculation section
[0039] 60B: Decision model establishment section
[0040] 60B1: Health prediction model establishment section
[0041] 60B2: Weighted mean square error calculation section
[0042] 60B1a: Input data
[0043] 60B1b: Encoder
[0044] 60B1c: Decoder
[0045] 60B1d: Output data
[0046] 60C: Health score calculation unit
[0047] 60D: False alarm handling unit
[0048] 70: Output unit
[0049] 70A: First area
[0050] 70B: Second area Detailed implementation manners
[0051] Figure 1 It is a schematic diagram of the architecture of the decision-making system for equipment maintenance illustrated according to an embodiment of the present invention. Figure 2 Illustration Figure 1 A schematic diagram of the maintenance decision-making unit in the example. In this embodiment, the equipment to be maintained is, for example, factory area equipment that needs regular maintenance or repair, and in the following description, an air compressor will be used as an example of the equipment to be maintained, but it is not intended to limit the implementation object of the present invention.
[0052] As Figure 1 shown, the architecture of this decision-making system 1 can be executed, for example, by a computer having at least a processor and a memory. The decision-making system 1 can include various sensors 10, which are arranged on the air compressor as an application example to collect on-site data. Here, the on-site is, for example, the place where the air compressor is arranged. The sensors 10 can include a frequency sensor, a current sensor, a temperature sensor, an oil pressure sensor, an oil temperature sensor, a pressure sensor, and the like. The frequency sensor can, for example, sense the high and low speed vibration frequencies of the air compressor. The current sensor can sense the operating current. The temperature sensor can sense the inlet temperature, the front and rear bearing temperatures of the motor, and the motor coil temperature. The oil pressure and oil temperature sensors can sense the oil pressure and oil temperature, etc. The pressure sensor can sense the outlet pressure, the final stage exhaust pressure, the system pressure, etc. During implementation, corresponding sensors can be set according to the needs to be monitored of interest to detect various parameter values of the air compressor.
[0053] Next, the decision-making system 1 can include a data preprocessing unit 20. The data preprocessing unit 20 can exclude some abnormal sensed data, such as sensed abnormal data with data values of 0 or null values. These abnormal data come from abnormal sensing of the sensors 10 and will affect the establishment of the health model, so they can be excluded in advance. The data preprocessing unit 20 can screen all the sensed data to exclude abnormal or inapplicable data.
[0054] The decision-making system 1 may further include a seasonal change judgment unit 30. The seasonal change judgment unit 30 determines whether the processed data output by the data preprocessing unit 20 shows periodic changes with seasons. If the data does not show periodic changes with seasons, this data is transmitted to the data standardization unit 50. If the data shows periodic changes with seasons, this data is transmitted to the deseasonalization processing unit 40, where the factors causing seasonal changes are removed. The data after removing the factors causing seasonal changes is then transmitted to the data standardization unit 50.
[0055] The decision-making system 1 may further include a deseasonalization processing unit 40. The deseasonalization processing unit 40 mainly removes the seasonal change factors of the data processed by the data preprocessing unit 20. Here, as an example, the deseasonalization processing unit 40 uses the central point moving average method to perform the operation.
[0056] First, the deseasonalization calculation method of the data (i.e., the data after preprocessing by various sensors 10) is described. The factory process data is mostly affected by climate. In the example of applying the method of the present invention to the predictive maintenance of air compressors, a mathematical method for obtaining the monthly index using the central point moving average method is adopted.
[0057] First, the central point moving average value of the input data is obtained by using CMA(t, n), then the central point moving average is taken once for CMA(t, n) to obtain CMA′(t, n), and finally the temperature X(t) in the t-th month is divided by CMA′(t, n) to obtain the t-th month index, MI(t).
[0058]
[0059]
[0060]
[0061] The decision-making system 1 may further include a data standardization unit 50. The data standardization unit 50 can adopt general statistical calculation methods. For example, the data without seasonal changes at the receiving location can be subtracted by the average value, and then the result is divided by the standard deviation.
[0062] The decision-making system 1 may further include a maintenance decision-making unit 60. The maintenance decision-making unit 60 is the core part of this embodiment. It mainly uses an autoencoder for unsupervised learning of artificial intelligence to form a decision model establishment unit 60B. The decision model establishment unit (autoencoder) 60B will establish a health prediction model for the input parameters (such as the parameter examples shown in Table 1 below). In addition, the maintenance decision-making unit 60 may further include a parameter weight table calculation unit 60A, which can calculate weight values for each input parameter. The decision model establishment unit 60B further calculates the weighted mean square error WMSE based on the parameter weight values generated by the parameter weight table calculation unit 60A and based on the health prediction model generated by the decision model establishment unit 60B. After that, through this weighted mean square error WMSE, the health scores of each parameter can be generated, thereby making the maintenance strategy visual and further making a correct and appropriate maintenance procedure. The maintenance decision-making unit 60 will be further described in detail later.
[0063] The decision-making system 1 may further include an output unit 70. The output unit 70 may be constituted by a display, for example. The display may be any type of display device that can be used to display the above-mentioned decision results (which may include various combinations of graphics, charts, texts, etc.) of existing or future products.
[0064] In addition, the data preprocessing unit 20, the seasonal change judgment unit 30, the deseasonalization processing unit 40, the data standardization unit 50, and the maintenance decision-making unit 60, etc. can all be executed by the above-mentioned processor of the decision-making system 1.
[0065] According to an embodiment of the present invention, the maintenance decision-making unit 60 is an intelligent predictive maintenance decision-making method implemented by a health prediction system. The maintenance decision-making unit 60 includes at least, but is not limited to: a parameter weight table calculation unit 60A, a decision model establishment unit 60B (such as Figure 2 , which may include a health prediction model establishment unit 60B1 and a weighted mean square error calculation unit 60B2), a health score calculation unit 60C, and a false alarm processing unit 60D. In addition, as Figure 2 shown, the decision model establishment unit 60B may further include a health prediction model establishment unit 60B1 and a weighted mean square error calculation unit 60C1, which constitute an autoencoder. The operations and functions of each part will be described in detail below.
[0066] First, the calculation method of the parameter weight table of the parameter weight calculation unit 60A in the maintenance decision-making unit 60 will be described. Generally speaking, based on the operation experience of factory practice, data, and equipment original factory data for reference, the proportion of each parameter type of the air compressor can be defined. Here, the higher the score, the greater the impact on the machine health. In addition, when an abnormality occurs in an item with a higher weight score of the parameter, the deviation amount of the health score calculated by the model is greater. Table 1 lists several parameters of the air compressor and their corresponding weight demonstration examples, that is, the parameter weight table.
[0067] Table 1
[0068]
[0069] Next, the calculation method of the parameter weight table for generating Table 1 will be described. According to the embodiments of the present invention, when calculating the weights of the parameters of the air compressor, at least the following three variables can be considered, namely, the expert recommended parameter weight, the parameter abnormality probability, and the parameter repair time cost, so as to perform the weight calculation.
[0070] First, based on the principle of statistics, the maintenance decision-making unit 60 can calculate the coefficient of variation CV for the above-mentioned expert recommended parameter weight, parameter abnormality probability, and parameter repair time cost respectively, and calculate the parameter indicator value according to different methods based on the calculated coefficient of variation CV. After that, the calculated parameter indicator values are grouped, and corresponding weights are assigned to each group.
[0071] Here, the expert recommended parameter weight refers to the degree of influence of each parameter on the health of the equipment (machine) to be maintained as known through the knowledge of domain experts.
[0072] The parameter abnormality probability means that when collecting the historical data of the parameter, if the parameter distribution conforms to the normal distribution, the upper and lower control limits are set (for example, the mean plus or minus three standard deviations). If the parameter distribution does not conform to the normal distribution, the upper and lower control limits are set to (Q2 plus or minus 3×IQR), and the number of data records exceeding the control limits is calculated as a proportion of the total number of data records to calculate the parameter abnormality probability. Here, Q2 is the second quartile, also known as the median, which is equal to the 50% number after all the values in the sample are arranged from small to large. IQR is the interquartile range, which is a statistical method to determine the difference between the third quartile (Q3, that is, the 75% number after all the values in the sample are arranged from small to large) and the first quartile (Q1, that is, the 25% number after all the values in the sample are arranged from small to large).
[0073] In addition, regarding the parameter repair time cost (or abnormal parameter repair time cost), since the abnormal repair time cost of each parameter is different, the reason for considering the parameter repair time cost is that parameters with longer repair times need to be given higher weights.
[0074] Figure 3It is a schematic diagram of the generation process of the parameter weight table shown according to an embodiment of the present invention. In steps S10 and S12, three variables of the expert-suggested parameter weight, the parameter abnormality probability, and the parameter repair time cost are input, and the coefficient of variation CV of these three variables is calculated. Then, in step S14, based on the calculated coefficient of variation, the maximum value among them is determined.
[0075] In step S16, the maximum coefficient of variation CV determined in step S14 is compared with the first threshold A and the second threshold B respectively, and the parameter index value is calculated based on the comparison results (steps S16A, S16B, S16C). Regarding the first threshold A and the second threshold B, further explanations will be given later.
[0076] When the coefficient of variation CV of the input variable is greater than the first threshold A, step S16A is adopted to calculate the parameter index value corresponding to the input variable in a weighted average manner. In other words, if the variation of the data is too large, in order to avoid the data being amplified too much, the weighted average method is adopted. That is, the following formula can be used for calculation:
[0077] Parameter index value = w1 × parameter abnormality probability + w2 × expert-suggested parameter weight + w3 × parameter repair time cost.
[0078] Where w1, w2, and w3 are the corresponding weighted values corresponding to the parameter abnormality probability, the expert-suggested parameter weight, and the parameter repair time cost respectively.
[0079] When the coefficient of variation CV of the input variable is less than the first threshold A and greater than the second threshold B, step S16B is adopted to calculate the parameter index value corresponding to the input variable in a multiplicative manner. That is, the following formula can be used for calculation:
[0080] Parameter index value = parameter abnormality probability × expert-suggested parameter weight × parameter repair time cost.
[0081] When the coefficient of variation CV of the input variable is less than the second threshold B, step S16C is adopted to calculate the parameter index value corresponding to the input variable in an exponential addition manner. In other words, if the change of the data is small, the gap needs to be amplified, so the exponential method can be used for calculation. That is, the following formula can be used for calculation:
[0082] Parameter index value = exp(parameter abnormality probability × expert-suggested parameter weight × parameter repair time cost). Where exp(x) represents the natural exponential function.
[0083] After that, the parameter index values of each parameter calculated in the above manner generally fall within the range of (0, 100). As an example, such as Figure 3As shown, the parameter index values of the calculated parameters are: 100, 95, 80, 75, 60, 40, 38, 20, 15, 10. The numerical values of the parameter index values cited here are only an exemplary example, and their numerical values will produce different calculated values according to actual applications, and are not used to limit the implementation of the present invention.
[0084] After that, in step S18, further through the method of machine learning, the parameter index values of the above-mentioned parameters are grouped. As Figure 3 shown, for example, the parameter index value of 10 is grouped into 1, the parameter index values of 18 and 20 are grouped into 2, the parameter index values of 38 and 40 are grouped into 3, the parameter index values of 70 and 75 are grouped into 4, and the parameter index values of 100, 65, and 90 are grouped into 5. The grouping here is only an exemplary example, and its numerical values and the number of groups will change according to actual applications, and are not used to limit the implementation of the present invention. Generally speaking, the larger the number, the greater the weight of the parameter. In this way, different weight values can be assigned to various parameters of the air compressor, and the weight value also represents a more critical parameter.
[0085] In addition, according to the embodiment of the present invention, if the variability of the parameter index value is too large, it will lead to too few groups in the grouping. After program simulation, when the variance of the parameter index value is in the range of 0.2 to 0.3 (specified range), it is a reasonable numerical range. Therefore, the embodiment of the present invention sets a group of first threshold A and second threshold B, which can make the variance of the parameter index value fall within the range of 0.2 to 0.3.
[0086] In addition, according to the embodiment of the present invention, as an example, the above-mentioned first threshold A and second threshold B can be determined by using the random search algorithm. First, in step (1), calculate the variance of the abnormal probability, the abnormal parameter repair time cost, and the expert advice weight of each machine tool, and select the largest one, that is, the above-mentioned steps S10, S12, and S14. Then, in step (2), the system randomly generates the first threshold A and the second threshold B.
[0087] After that, in step (3), through the first step to generate the first threshold A and the second threshold B, determine to use one of the above-mentioned weighted average method, multiplication method, and exponential method to calculate the parameter index value, and calculate the parameter index value through the variables owned by each parameter of each device (machine tool) (that is, the above-mentioned parameter abnormal probability, parameter repair time cost, and expert advice parameter weight). That is, corresponding to the above-mentioned step S16.
[0088] After that, in step (4), calculate the variance of the parameter indicators and determine whether all variances fall within the specified range. For example, the range can be set between 0.2 and 0.3. At this time, if one variance is not within this range, it is determined that the set of the first threshold A and the second threshold B is an infeasible solution. When it is determined as an infeasible solution, go back to step (2) again to randomly generate another set of the first threshold A and the second threshold B. At this time, the average value of the variances of the parameter indicator values of each parameter of all machines can be calculated. As a preferred example, 0.2, which is closer to the lower limit of the above-mentioned specified range, is better. In this way, go back to step (2) again and repeatedly perform the above judgment method until the first threshold A and the second threshold B meet the set number of iterations or convergence conditions.
[0089] The following gives an example to illustrate the calculation method of the first threshold A and the second threshold B. In this example, take the low-speed vibration parameter of the air compressor as an example. Suppose the variances of the above three variables for this parameter are as follows respectively, then it can be known that the largest variance is 0.3.
[0090] Variance of the expert-recommended parameter weight = 0.15
[0091] Variance of the parameter anomaly probability = 0.3
[0092] Variance of the parameter repair time cost = 0.2
[0093] In addition, for the low-speed vibration parameter, assume that the expert-recommended parameter weight is 5, the parameter anomaly probability is 0.1, and the parameter repair time cost is 5. These values are only for illustration. In fact, for example, they can be obtained by looking up the lookup table of each parameter of the air compressor.
[0094] Next, the system randomly generates the first threshold A as 0.4 and the second threshold B as 0.2. At this time, it can be judged that the variance 0.3 is between 0.4 and 0.2. Therefore, the multiplication method is used to calculate the parameter indicator value, that is, the parameter indicator value of the low-speed vibration is the product of the expert-recommended parameter weight, the parameter anomaly probability, and the parameter repair time cost. In this example, the parameter indicator value of the low-speed vibration is 5×0.1×5 = 2.5, and the other parameters are calculated in the same way.
[0095] After obtaining the parameter index values of all the parameters of the air compressor, the variance of the parameter index values can be calculated, and it can be determined whether the variances of the parameter index values all fall within the range of 0.2 to 0.3. After that, the above steps are applied to all the machines (because each machine has its own expert advice weight table, probability table of each parameter abnormality, and cost table of abnormality repair time for each parameter). Therefore, if there are 7 machines, 7 variances will be obtained, and the average of these 7 variances is taken. The closer it is to 0.2, the better. If this average variance is the current best solution (the closest to 0.2), then the original best solution is overwritten, which means saving the current set of thresholds (A, B) as the current best solution. After that, a new set of thresholds (A, B) is generated, and the subsequent calculations are continued until the iteration number is reached or convergence occurs.
[0096] Next, as Figure 2 shown, the operation mode of the health prediction model establishment unit 60B1 in the maintenance decision unit 60 is described. According to an embodiment of the present invention, the health prediction model establishment unit 60B1 uses a machine learning algorithm to establish a decision model for predictive maintenance. In the example of predictive maintenance of an air compressor (i.e., the equipment to be maintained), an autoencoder in unsupervised learning is used to identify outliers.
[0097] As Figure 4 shown, this health prediction model establishment unit 60B1 is composed of two neural networks, namely an encoder 60Bb and a decoder 60B1c. The data sensed by the sensor 10 is used as input data [x 1 、x 2 、…、x n 60B1a and input into the encoder 60B1b for compression, and presented in a low-dimensional encoding Z, thereby performing feature extraction. That is, the eigenvalue is extracted from the input data 60B1a.
[0098] After that, the eigenvalue extracted through encoding, that is, the low-dimensional encoding Z, is input into the decoder 60B1c to decode the eigenvalue, thereby performing data reconstruction. After the reconstruction is completed, the decoder 60B1c will output the data [x' 1 、x' 2 、…、x' n 60B1d. When the reconstructed data (output data [x' 1 、x' 2 、…、x' n 60B1d) is close to the original data (input data [x 1 、x 2 、…、x n 60B1a), it indicates that the model has successfully extracted the key features of the original data.
[0099] The system according to the embodiment of the present invention establishes a health prediction model by using this principle. First, it collects equipment health data, and then conducts model training to find the key features of the health data. When the health data enters the model, less error will be obtained. On the contrary, when abnormal data is input, a larger error will be obtained.
[0100] Next, the operation mode of the weighted mean squared error calculation unit 60B2 in the maintenance decision unit 60 will be described. In a typical autoencoder model, the mean squared error (MSE) is mostly used as the loss function to train the model. However, in practice, the influence of each parameter on the health of the equipment is not the same. Therefore, in this embodiment, the original loss function training model is modified, and the weighted mean squared error (WMSE) is used as the loss function to train the model. Thus, the importance of each parameter can be specified, so that the health prediction model can focus on learning the key features of important parameters during training, in order to enhance the identification ability of the health prediction model for the outliers of the parameters.
[0101] Therefore, according to the embodiment of the present invention, since the influence degrees of the parameters on the health of the air compressor are different, a weight w i is introduced, and the calculation method of the mean squared error is corrected, so that the abnormal deviation features can be more easily extracted. The calculation formula of the weighted mean squared error is as follows, and the weight w i is the weight value of each parameter shown in Table 1 above. In addition, in the following formula, i is an integer from 1 to n, and j is an integer from 1 to m, where i represents the number of parameters and j represents the number of data records.
[0102]
[0103] Next, the health score calculation unit 60C in the maintenance decision unit 60 will be described. The health score calculation unit 60C converts the weighted mean squared error (WMSE) output by the above-mentioned weighted mean squared error calculation unit 60B2 into a health score. Figure 5 Combined with the conversion diagram, according to the embodiment of the present invention, as an example, the health score conversion formula is as follows:
[0104] y = (a i ·wnse + b i )·I{r i ≤wmse≤R i}, where I∈{·} = 0 or 1,
[0105] In addition, a i and b irespectively represent the slope and intercept after linear transformation in the i-th interval. As Figure 5 shown, the horizontal axis is the weighted mean square error WMSE, and the vertical axis is the health score. The endpoints of several intervals on the horizontal axis are the first quartile Q1, the second quartile Q2, the third quartile Q3, mild outliers, extreme outliers, and the maximum value (or the fourth quartile Q4). Based on the calculated weighted mean square error, the present invention forms 6 intervals between the above-mentioned interval endpoints, namely interval 1: (0, Q1), interval 2: (Q1, Q2), interval 3: (Q2, Q3), interval 4: (Q3, mild outliers), interval 5: (mild outliers, extreme outliers), and interval 6: (extreme outliers, maximum value). Then, in each of intervals 1 to 6, the weighted mean square error wmse is converted into a health score in a corresponding linear transformation manner.
[0106] As described above, through the health prediction model establishment unit 60B1, the weighted mean square error calculation unit 60B2 can calculate the weighted mean square error (WMSE). Although the similarity between the data and the health data can be understood from the numerical value of the weighted mean square error, the weighted mean square error of these parameters can also be converted into a more intuitive manner that is more in line with the intuitive judgment of the plant personnel. Therefore, the maintenance decision unit 60 of the embodiment of the present invention further uses piece-wise linear transformation to convert the weighted mean square error into a health score. Thereby, on-site personnel can more intuitively identify the health status of the equipment.
[0107] As Figure 5 shown, piece-wise linear transformation is a method of mapping a continuous value range (the value range of wmse) to multiple different intervals 1 to 6, and a linear function is used for transformation within each of intervals 1 to 6. To ensure that this conversion mechanism conforms to practical use, a time period with relatively stable equipment (machine) status must be used as the conversion basis, that is, intervals with long-term abnormalities cannot be included.
[0108] Referring to Figure 5 , first find the quartiles (Q1, Q2, Q3), mild outliers, extreme outliers, and the maximum value of the weighted mean square error, and project each value onto health scores 100, 95, 90, 85, 70, 60, and 0 respectively. Then, establish a relationship between the weighted mean square error and the health score through piece-wise linear transformation.
[0109] Through this method, the original error value can be converted into an easily understandable health score. Additionally, this method can freely define the error range and the corresponding health scores according to requirements. That is to say, after obtaining the health score, the appropriate threshold can be set first through the knowledge of domain experts and fine-tuned according to the actual online operation conditions to meet the practical needs. Here, as the definition of the threshold of the health score, for example, the mild outlier in the health score (health score of 70 points) is used as a warning, the value below the extreme outlier (health score of 60 points) is used as the warning value, and the value below the health score of 65 points is used as the preset maintenance score.
[0110] Therefore, through the above-mentioned health score, the maintenance decision can be designed to be more intuitive and visual. The plant personnel only need to check the health score of the parameter to intuitively know what corresponding actions should be taken for the parameter (such as the above-mentioned warning, alert, maintenance, etc.).
[0111] As Figure 2 shown, the maintenance decision-making department 60 may further include a false alarm processing model 60D. The false alarm processing model 60D presents the trend of the health scores calculated for each of the above parameters. Figures 6A to 6C The horizontal axis of
[0112] represents the data number (index of data). For example, one data is sampled per minute, which means time. The vertical axis represents the corresponding health score. Figure 6A Here,
[0113] As Figure 6B shown, the health scores presented Figure 6A are further calculated by the moving average method to generate the trend line I of the health score. In this example, the moving average value of the health score is calculated with a one-day cycle (n = 1440, n is the number of data points, assuming one data point is extracted per minute) to generate the trend line I. Additionally, Figure 6C the moving average value of the health score is calculated with a one-week cycle (n = 10080, n is the number of data points) to generate the trend line II.
[0114] From Figure 6A the Figure 6C figure, it can be seen that Figure 6B and Figure 6C the trend lines of the health scores presented by both Figure 6A schemes are smoother than the original health score ( Figure 6B and Figure 6CThe two health score trend lines I and II do not have as large a variation in health scores as Figure 6A shown. Therefore, by using the moving average method to calculate the trend line of the health score, the change trend of the health score of the parameter can be more obvious and easier to identify. Moreover, Figure 6B and Figure 6C both schemes can still show the changes in the key trends. This way of presenting the health score not only considers the current health status but also the past health status. Therefore, if you want to focus on the current health status, you can adopt the Figure 6B scheme, and vice versa, adopt the Figure 6C scheme.
[0115] Figure 7 The figure shows a schematic diagram of the result presentation of the health score trend chart drawn according to the embodiments of the present invention. As an example, Figure 7 it presents the health scores of an air compressor in a factory each year and the trend line calculated using the moving average method. In this example, the data in 2021 is used as the training data. It can be seen from the health score trend line that at the time points A, B, and C where the health scores are relatively low, it just corresponds to the annual maintenance time of the air compressor. That is to say, when the health score drops to a relatively low point, the air compressor can be maintained annually. Moreover, it can also be seen from Figure 7 that the health scores were relatively low before the annual maintenance of the air compressor at the time points A, B, and C. Relatively speaking, one to two months after the annual maintenance of the air compressor at the time points A, B, and C, the health score will gradually rise to the healthy score range. Thus, it can be seen that it takes some time for the air compressor to recover to the best state after maintenance.
[0116] In addition, as Figure 7 shown, for example, at the time point C, its health score is around 80, and the state of the air compressor may not be too bad. Therefore, the plant personnel can judge whether to perform maintenance at this time point based on the presented results, and even extend the maintenance period to a time point with a lower health score. In this way, over-maintenance can be carried out when the health state of the air compressor is still acceptable. In addition, if the health score does not rise to the healthy score range as expected after the annual maintenance, it can be initially judged that the maintenance may not be thorough or there are deficiencies, and the plant personnel can then take immediate remedial measures.
[0117] Figure 8 is a control chart showing the parameters that affect the decrease in the health score in the embodiments of the present invention. As Figure 8As shown, nine control charts are presented here, each representing a parameter control chart of an air compressor. Of course, there are multiple parameters for the air compressor, and the user can set the control charts for several parameters that have a greater impact on the health score. Additionally, for each control chart, the center line, upper control line, and lower control line (the three horizontal lines in each control chart in the figure) can be set. Through these three control lines, the plant maintenance personnel can determine the status of each parameter deviating from the control lines. Furthermore, in Figure 8 's example, the left side shows the interval from November 1, 2022 to November 30, 2022, and the right side shows the interval from December 1, 2022 to December 14, 2022. These are only illustrative examples and are not used to limit the implementation of the invention.
[0118] From Figure 8 's left side, it can be known that the main reasons for the sudden drop in the health score are the exhaust pressure at the end section (40%) and high-speed vibration (26%). Similarly, it can be seen from the control lines that the high-speed vibration shows a trend of gradually approaching the control center. Additionally, Figure 8 From the right side of
[0119] Figure 9 it can be seen that the deviation ratio of the high-speed vibration is relatively decreased compared to the left side. Therefore, it can be inferred that in the interval on the right side, the main reason for the increase in the health score is the decrease in high-speed vibration. Figure 9 As shown, the output unit 70 of the decision-making system 1 can be used to present the output result of the maintenance decision-making unit 60 as Figure 9 shown. In the example as
[0120] shown, the first area 70A of the output unit 70 as the display can present the health score chart of the equipment to be maintained, and the second area 70B can present the control charts of the corresponding parameters. Through this side-by-side display method, the plant maintenance personnel can judge the health trend of the equipment to be maintained from the health score, and can further understand the reasons for the decrease or increase in the health score from the control charts of each parameter, and then can adjust the parameter accordingly.
[0120] In summary, according to the embodiments of the present invention, it uses technologies such as data mining, big data analysis, and artificial intelligence to establish a health prediction system for equipment to be maintained (such as an air compressor, etc.), and visually outputs the analysis results. This system can not only immediately transmit the overall health information of the equipment to be maintained, but also report the key parameters affecting the health, enabling the plant maintenance personnel to formulate an optimized maintenance strategy and avoid additional costs caused by unexpected failures of the equipment to be maintained.
Claims
1. A decision-making system for equipment maintenance, comprising: a plurality of sensors disposed on the equipment to be maintained for detecting a plurality of parameters of the equipment to be maintained; and a maintenance decision-making unit that receives the plurality of parameters and outputs a health score of the equipment to be maintained, wherein the maintenance decision-making unit further comprises: a parameter weight table calculation unit that calculates weights for each of the plurality of parameters to generate weight values for each of the plurality of parameters, a decision model establishment unit that receives the plurality of parameters, generates a health prediction model for the equipment to be maintained, and outputs a weighted mean square error based on the weight values of each of the plurality of parameters and the health prediction model, and a health score calculation unit that converts the weighted mean square error into a health score using a piecewise linear conversion method.
2. The decision-making system for equipment maintenance according to claim 1, wherein the parameter weight table calculation unit further performs: for each of the plurality of parameters, calculating variances based on expert-recommended parameter weights, parameter abnormality probabilities, and parameter repair time costs respectively; calculating a parameter index value for each of the plurality of parameters based on the maximum variance among the coefficient of variation of the expert-recommended parameter weights, the parameter abnormality probabilities, and the parameter repair time costs respectively; grouping the parameter index values of the plurality of parameters and assigning corresponding weights to each group.
3. The decision-making system for equipment maintenance according to claim 2, wherein, when the variance is above a first threshold, the parameter weight table calculation unit calculates the parameter index value for the expert-recommended parameter weight, the parameter abnormality probability, and the parameter repair time cost by a weighted average method, when the variance is below the first threshold and above a second threshold, the parameter weight table calculation unit calculates the parameter index value for the expert-recommended parameter weight, the parameter abnormality probability, and the parameter repair time cost by a multiplication method, and when the variance is below the second threshold, the parameter weight table calculation unit calculates the parameter index value for the expert-recommended parameter weight, the parameter abnormality probability, and the parameter repair time cost by an exponential method, wherein the first threshold is greater than the second threshold.
4. The decision-making system for equipment maintenance according to claim 2, wherein the parameter weight table calculation unit groups the parameter index values by a machine learning method.
5. The decision-making system for equipment maintenance according to claim 2, wherein the larger the parameter index value, the greater the weight.
6. The decision-making system for equipment maintenance according to claim 1, wherein the decision model establishment unit uses a machine learning algorithm to establish a decision model for predictive maintenance.
7. The decision-making system for equipment maintenance according to claim 6, wherein the decision model establishment unit is an unsupervised learning autoencoder, and the autoencoder further comprises: an encoder that receives the plurality of parameters, compresses the plurality of parameters, and presents them in a low-dimensional encoding to extract feature values; and a decoder that encodes the output of the encoder to perform data reconstruction to establish the health prediction model.
8. The decision-making system for equipment maintenance as claimed in claim 7, wherein the autoencoder calculates the mean squared error based on the weights of the plurality of parameters to generate the weighted mean squared error.
9. The decision-making system for equipment maintenance as claimed in claim 1, wherein the health score calculation unit further performs: dividing the value range of the weighted mean squared error into a plurality of intervals; in each of the plurality of intervals, converting the weighted mean squared error into the health score by linear transformation.
10. The decision-making system for equipment maintenance as claimed in claim 9, wherein the plurality of intervals are divided by 0, the first quartile, the second quartile, the third quartile, the mild outlier, the extreme outlier, and the maximum value in the value range of the weighted mean squared error.
11. The decision-making system for equipment maintenance as claimed in claim 9, wherein the health score is further compared with a health score threshold, and when the health score is less than the health score threshold, the equipment to be maintained is maintained.
12. The decision-making system for equipment maintenance as claimed in claim 1, wherein the maintenance decision-making unit further includes a false alarm processing unit, and the false alarm processing unit calculates the health score by using a moving average method to generate a trend line of the health score.
13. The decision-making system for equipment maintenance as claimed in claim 11, further comprising an output unit, the output unit being a display for outputting the health score and the trend line of the health score.
14. The decision-making system for equipment maintenance as claimed in claim 13, wherein the display comprises: a first area for displaying the health score and the trend line of the health score; and a second area for displaying a control chart of the plurality of parameters corresponding to the trend line.
15. The decision-making system for equipment maintenance as claimed in claim 1, wherein the data of the plurality of parameters received by the maintenance decision-making unit is deseasonalized, and the deseasonalization is calculated by using a central moving average method.