A knowledge data hybrid driven multi-channel lubricating oil abrasive particle abnormality monitoring method

Abstract

The present application relates to a kind of knowledge data hybrid drive's multi-channel oil abrasive particle abnormal monitoring method, based on expert knowledge to construct abrasive particle confidence rule base;Using operating data is based on triangular membership degree and rule activation formula to activate different rules of abrasive particle confidence rule base, obtain activation weight;Based on activation weight and using the way of evidence reasoning to different rules are fused, to obtain preliminary abrasive particle abnormal monitoring result;The accuracy of preliminary abrasive particle abnormal monitoring result is calculated, and ensure that the accuracy reaches above set value, when monitoring accuracy is less than set value, then input historical data, using sequential least squares method to optimize expert knowledge;Optimized expert knowledge is input into abrasive particle confidence rule base to obtain optimized abrasive particle confidence rule base, to carry out multi-channel oil abrasive particle abnormal monitoring.The present application constructs the hybrid drive model of knowledge drive and data drive, can realize interpretable, high reliability, high performance on-line abrasive particle abnormal monitoring.

Description

Technical Field

[0001] This invention belongs to the field of lubricating oil condition recognition technology, and in particular relates to a multi-channel lubricating oil abrasive particle anomaly monitoring method driven by knowledge data hybridization. Background Technology

[0002] Lubricating oil, as the "blood" of mechanical equipment, contains a wealth of information, among which abrasive information is directly related to the wear condition of the equipment. Real-time monitoring of abrasive information is crucial for understanding the wear status of mechanical equipment. Currently, with the rise of artificial intelligence and the digital transformation of enterprises, there is an urgent need for intelligent automatic monitoring algorithms to replace traditional manual inspection. Fixed thresholds are the simplest monitoring algorithm, but they are only suitable for single-channel monitoring and cannot effectively integrate multi-channel information to reduce information redundancy. Constructing an abrasive confidence rule base can be based on expert knowledge, performing rule reasoning on multi-channel information, and assigning appropriate weights according to the importance of different channels. This can effectively integrate multi-channel information and achieve an overall state judgment output. However, expert knowledge sometimes has limitations and incompleteness, requiring the learning of new knowledge from historical data to compensate for these deficiencies. Therefore, there is an urgent need in this field for a "knowledge- and data-driven dual-drive monitoring model" that can effectively utilize expert knowledge while simultaneously learning new knowledge from data, achieving highly reliable multi-channel abrasive anomaly monitoring of lubricating oil.

[0003] The information disclosed in the background section is only intended to enhance the understanding of the background of the present invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide a knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method, overcoming the shortcomings of over-reliance on expert knowledge or over-reliance on data in industrial scenarios. It utilizes a knowledge-data dual-driven model to achieve effective fusion of multi-channel abrasive information and anomaly monitoring. To achieve the above objective, this invention provides the following technical solution.

[0005] A knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method, characterized in that the method includes the following steps:

[0006] Step S100: Construct a wear particle confidence rule base based on expert knowledge, wherein the wear particle confidence rule base contains different rules;

[0007] Step S200: Load the running data, and use the running data to activate different rules based on triangular membership degree and rule activation formula to obtain activation weights;

[0008] Step S300: Based on the activation weights and using evidence reasoning, different rules are fused to obtain preliminary abrasive anomaly monitoring results;

[0009] Step S400: Calculate the accuracy of the preliminary abrasive anomaly monitoring results and determine whether the accuracy reaches or exceeds the set value; if the accuracy is greater than or equal to the set value, end; if the accuracy is less than the set value, optimize the abrasive confidence rule base to obtain optimization expert knowledge.

[0010] Step S500: Input the optimized expert knowledge back into the abrasive confidence rule base to obtain the optimized abrasive confidence rule base. Based on the optimized abrasive confidence rule base, perform multi-channel lubricating oil abrasive anomaly monitoring to obtain monitoring results.

[0011] Optionally, in step S100, constructing the abrasive grain confidence rule base based on expert experience includes the following steps:

[0012] Step S101: Divide all abrasive channels into three groups: large, medium, and small abrasive channels;

[0013] Step S102: Set the channel weight δ of the i-th channel in the three groups of abrasive channels (large, medium, and small) respectively. i The prerequisite attribute level classification A for the i-th channel i n , n = 1, 2, ..., N, where N represents the number of prerequisite attribute levels, and A represents the number of abrasive grains corresponding to the prerequisite attribute level of the abrasive grain channel; the number of calculation rules L; and the confidence level β of the nth level of the device under the kth rule. n，k Set the rule weight θ of the k-th rule. k .

[0014] Optionally, step S200 specifically includes the following steps:

[0015] Step 201: Load runtime data X;

[0016] Step 202: Calculate the membership degree α between the running data X and the prerequisite attribute level using triangular membership.

[0017] Step 203: The cumulative product α based on the membership degree α k Calculate activation weight w k ;

[0018]

[0019] Among them, w k Let θ represent the activation weight of the k-th rule. k Let α represent the weight of the k-th rule, L represent the total number of rules, and α represent the weight of the k-th rule. k This represents the cumulative product of the membership degree α of the k-th rule.

[0020] Optionally, in step 202, the specific formula for calculating the membership degree α is as follows:

[0021]

[0022] α it =0, t=1, ..., J i , and t≠j,j+1

[0023] Where, α ij A represents the membership degree of the running data of the i-th channel relative to the j-th prerequisite attribute level corresponding to that channel. i(j+1) Indicates the j-th channel of the i-th channel. +1 The corresponding number of abrasive grains for each prerequisite attribute level, γ i A represents a specific data point within the data generated by the i-th channel. ij α represents the number of abrasive grains corresponding to the j-th prerequisite attribute level of the i-th channel; i(j+1) α represents the membership degree of the running data of the i-th channel relative to the (j+1)-th prerequisite attribute level corresponding to that channel; it The degree of membership of the running data of the i-th channel relative to the other prerequisite attribute levels corresponding to that channel is represented by t; the prerequisite attribute level corresponding to that channel is represented by J. i This represents the Jth level of the i-th channel.

[0024] Optionally, in step S300, fusion is performed according to the following formula:

[0025]

[0026] Among them, y j This represents the confidence level of the j-th level of the output, where L is the total number of rules, and w k Let β represent the activation weight of the k-th rule. j，k β represents the confidence level of the j-th conclusion under the k-th rule. i，k Let μ represent the confidence level of the i-th conclusion level under the k-th rule, where μ represents an intermediate quantity with no practical meaning; and N represents the number of conclusion level divisions.

[0027] Optionally, the formula for calculating μ is:

[0028]

[0029] Optionally, in step S400, calculating the accuracy of the abrasive anomaly monitoring results specifically includes:

[0030] The highest confidence level of the abrasive particle anomaly monitoring results is taken as the expected value, and the accuracy is calculated based on the comparison between the expected value and the actual value.

[0031] Optionally, in step S400, if the accuracy is greater than or equal to the set value, the process ends;

[0032] If the accuracy is less than the set value, input training data and apply the sequential least squares method to the abrasive grain confidence rule base parameter θ. k ,β n，k δ i The process of iterative optimization ends when the accuracy reaches a set value or higher, in order to obtain optimization expert knowledge.

[0033] Optionally, the parameters θ of the abrasive grain confidence rule base are determined based on the sequential least squares method. k ,β n，k δ i Iterative optimization, specifically including:

[0034] The model is optimized using a strategy that minimizes the cross-entropy loss. The optimized model is as follows:

[0035] min BCELoss(θ k ,β n，k δ i )

[0036] st.

[0037] 0≤θ k ≤1, k=1, …,L,

[0038] 0≤δ i ≤1, i=1,…,M i ,

[0039] 0≤β n，k ≤1, n=1,…,N, k=1,…,L,

[0040]

[0041] Wherein, BCELoss represents the cross-entropy loss function, with the abrasive confidence rule base parameters θ determined in step S100. k ,β n，k δ i θ serves as the initial input for the sequential least squares method. k δ represents the weight of the k-th rule. i β represents the weight of the i-th channel. n，k This represents the confidence level of the nth conclusion under the kth rule.

[0042] Optionally, after optimization, the accuracy is calculated again using the aforementioned accuracy calculation formula. If the accuracy is greater than or equal to the set value, the task ends; otherwise, training data is added until the accuracy is greater than or equal to the set value, at which point the task ends.

[0043] Compared with existing technologies, this invention has the following beneficial technical effects: The knowledge-driven, data-driven multi-channel lubricating oil abrasive anomaly monitoring method fully utilizes expert knowledge, reduces reliance on data, and has better interpretability; the data-driven approach learns from historical data, compensating for areas overlooked by expert knowledge. This dual-drive approach of knowledge and data allows for complementary advantages between the two methods, achieving more accurate multi-channel abrasive anomaly monitoring. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0045] Figure 1 This is a prediction effect diagram of the expert knowledge rule base of a knowledge data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method in this invention.

[0046] Figure 2 This is a prediction effect diagram of the random rule base of a knowledge data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method in this invention;

[0047] Figure 3 This is a prediction effect diagram of the random rule base optimized for a knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method in this invention.

[0048] Figure 4 This is a flowchart illustrating the overall calculation process of a knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method in this invention.

[0049] Figure 5 This is a comparison diagram of optimized algorithms for a knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method in this invention;

[0050] Figure 6 The abrasive data obtained by the multi-channel lubricating oil abrasive anomaly monitoring method driven by knowledge data hybridization in this invention;

[0051] Figure 7 This is an image showing the anomaly detection results obtained by a knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method according to the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Therefore, the following applies to the appendix Figures 1 to 7 The detailed description of the embodiments of the present invention provided herein is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0054] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0055] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0056] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0057] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0058] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0059] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0060] In one embodiment, the present invention provides a knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method, characterized in that the method includes the following steps:

[0061] Step S100: Construct a wear particle confidence rule base based on expert knowledge, wherein the wear particle confidence rule base contains different rules;

[0062] Step S200: Load the running data, and use the running data to activate different rules based on triangular membership degree and rule activation formula to obtain activation weights;

[0063] Step S300: Based on the activation weights and using evidence reasoning, different rules are fused to obtain preliminary abrasive anomaly monitoring results;

[0064] Step S400: Calculate the accuracy of the preliminary abrasive anomaly monitoring results and determine whether the accuracy reaches or exceeds the set value; if the accuracy is greater than or equal to the set value, end; if the accuracy is less than the set value, optimize the abrasive confidence rule base to obtain optimization expert knowledge.

[0065] Step S500: Input the optimized expert knowledge back into the abrasive confidence rule base to obtain the optimized abrasive confidence rule base. Based on the optimized abrasive confidence rule base, perform multi-channel lubricating oil abrasive anomaly monitoring to obtain monitoring results.

[0066] This embodiment discloses the inventive concept of the present invention, namely, a knowledge- and data-driven multi-channel lubricating oil abrasive anomaly monitoring method. It constructs a hybrid driving model of knowledge-driven and data-driven approaches. The knowledge-driven approach can make full use of expert knowledge, reduce dependence on data, and has good interpretability; the data-driven approach can learn from historical data and make up for the gaps that expert knowledge overlooks. The dual-drive approach of knowledge and data can achieve complementary advantages of the two driving methods, and realize more accurate multi-channel abrasive anomaly monitoring.

[0067] In one embodiment, step S100, constructing the abrasive grain confidence rule base based on expert experience, includes the following steps:

[0068] Step S101: Divide all abrasive channels into three groups: large, medium, and small abrasive channels;

[0069] Step S102: Set the channel weight δ of the i-th channel in the three groups of abrasive channels (large, medium, and small) respectively. i The prerequisite attribute level classification A for the i-th channel i n , n = 1, 2, ..., N, where N represents the number of prerequisite attribute levels, and A represents the number of abrasive grains corresponding to the prerequisite attribute level of the abrasive grain channel; the number of calculation rules L; and the confidence level β of the nth level of the device under the kth rule. n，k Set the rule weight θ of the k-th rule. k .

[0070] In one embodiment, step S200 specifically includes the following steps:

[0071] Step 201: Load runtime data X;

[0072] Step 202: Calculate the membership degree α between the running data X and the prerequisite attribute level using triangular membership.

[0073] Step 203: The cumulative product α based on the membership degree α k Calculate activation weight w k ;

[0074]

[0075] Among them, w k Let θ represent the activation weight of the k-th rule. k Let α represent the weight of the k-th rule, L represent the total number of rules, and α represent the weight of the k-th rule. k This represents the cumulative product of the membership degree α of the k-th rule.

[0076] In one embodiment, the specific formula for calculating the membership degree α in step 202 is as follows:

[0077]

[0078] α it =0, t=1, ..., J i , and t≠j,j+1

[0079] Where, α ij A represents the membership degree of the running data of the i-th channel relative to the j-th prerequisite attribute level corresponding to that channel. i(j+1) Indicates the j-th channel of the i-th channel. +1 The corresponding number of abrasive grains for each prerequisite attribute level, γ i A represents a specific data point within the data generated by the i-th channel. ij α represents the number of abrasive grains corresponding to the j-th prerequisite attribute level of the i-th channel; i(j+1) This indicates that the running data of the i-th channel is relative to the j-th channel corresponding to that channel. +1 Membership degree of each prerequisite attribute level; α it The degree of membership of the running data of the i-th channel relative to the other prerequisite attribute levels corresponding to that channel is represented by t; the prerequisite attribute level corresponding to that channel is represented by J. i This represents the Jth level of the i-th channel.

[0080] In one embodiment, in step S300, fusion is performed according to the following formula:

[0081]

[0082] Among them, y j This represents the confidence level of the j-th level of the output, where L is the total number of rules, and w k Let β represent the activation weight of the k-th rule. j，k β represents the confidence level of the j-th conclusion under the k-th rule. i，k Let μ represent the confidence level of the i-th conclusion level under the k-th rule, where μ represents an intermediate quantity with no practical meaning; and N represents the number of conclusion level divisions.

[0083] In one embodiment, μ is calculated using the following formula:

[0084]

[0085] In one embodiment, step S400, calculating the accuracy of the abrasive particle anomaly monitoring results specifically includes:

[0086] The highest confidence level of the abrasive particle anomaly monitoring results is taken as the expected value, and the accuracy is calculated based on the comparison between the expected value and the actual value.

[0087] In one embodiment, in step S400, if the accuracy is greater than or equal to a set value, the process ends.

[0088] If the accuracy is less than the set value, input training data and apply the sequential least squares method to the abrasive grain confidence rule base parameter θ. k ,β n，k δ i The process of iterative optimization ends when the accuracy reaches a set value or higher, in order to obtain optimization expert knowledge.

[0089] In one embodiment, the parameters θ of the abrasive grain confidence rule base are determined based on the sequential least squares method. k ,β n，k δ i Iterative optimization, specifically including:

[0090] The model is optimized using a strategy that minimizes the cross-entropy loss. The optimized model is as follows:

[0091] min BCELoss(θ k ,β n，k δ i )

[0092] st.

[0093] 0≤θ k ≤1, k=1, …,L,

[0094] 0≤δ i ≤1, i=1,…,M i ,

[0095] 0≤β n，k ≤1, n=1,…,N, k=1,…,L,

[0096]

[0097] Wherein, BCELoss represents the cross-entropy loss function, with the abrasive confidence rule base parameters θ determined in step S100. k ,β n，k δ i θ serves as the initial input for the sequential least squares method. k δ represents the weight of the k-th rule. i β represents the weight of the i-th channel. n，k This represents the confidence level of the nth conclusion under the kth rule.

[0098] In one embodiment, after optimization, the accuracy is calculated again using the accuracy calculation formula. If the accuracy is greater than or equal to the set value, the task ends; otherwise, training data is added until the accuracy is greater than or equal to the set value, at which point the task ends.

[0099] In one embodiment, the present invention provides a knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method, the method comprising the following steps:

[0100] Step 1: Construct a wear particle confidence rule base based on expert experience, specifically including:

[0101] 1. Divide all channels into three abrasive channels: large, medium, and small. If there are 7 channels, divide them into three groups, and then merge each group by adding them together to obtain the three abrasive channels: large, medium, and small.

[0102] For example, one of the three groups contains data from two channels, with the data from the two channels being [1, 2, 3, 4] and [2, 3, 4, 5]. The data from the two channels are added together to form one group, resulting in [3, 5, 7, 9].

[0103] 2. The prerequisite attribute weight δ of the three abrasive channels (large, medium, and small) is set to 1, with two levels: normal and abnormal. The number of abrasive particles corresponding to the normal level of large, medium, and small abrasive particles is zero. The number of abrasive particles corresponding to the abnormal level of large, medium, and small abrasive particles is determined to be 18,000, 60,000, and 150,000 respectively based on expert experience.

[0104] 3. According to the formula The calculated number of rules is 8, where L is the number of rules, n is the number of prerequisite attributes (three types: large, medium, and small abrasive grains), and D is the number of prerequisite attributes. i It is the number of the i-th prerequisite attribute levels, such as large abrasive grains having two levels: normal and abnormal.

[0105] 4. Based on expert experience, anomaly probabilities are assigned to the eight rules. When all three abrasive particles (large, medium, and small) are abnormal, the equipment anomaly probability is 1; when only large and medium abrasive particles are abnormal, the probability is 0.9; when only large and small abrasive particles are abnormal, the probability is 0.8; when only large abrasive particles are abnormal, the probability is 0.7; when only medium and small abrasive particles are abnormal, the probability is 0.6; when only medium abrasive particles are abnormal, the probability is 0.5; when only small abrasive particles are abnormal, the probability is 0.4; and when none of the three abrasive particles are abnormal, the probability is 0. The weight θ of each of the eight rules is set to 1. There are three channels (large, medium, and small), each with two state levels: normal and abnormal. The state levels of the three channels can be combined to obtain eight combinations, which correspond to the eight rules. For example, the actual meaning of a rule can be the equipment anomaly probability when all three channels are abnormal.

[0106] Among them, the abrasive confidence rule base contains valuable expert knowledge and experience, which can be transferred and applied to other multi-channel abrasive data, and has good generalization ability; the sequential least squares method can make full use of historical data to update the posterior probability distribution of model parameters. Compared with mainstream random search algorithms such as particle swarm optimization and genetics, it is a deterministic optimization algorithm with better stability and interpretability, which can meet the high reliability requirements of industrial scenarios.

[0107] Step Two: Input the running data and activate different rules using triangular membership degrees and rule activation formulas, specifically including:

[0108] 1. Load running data X, which comes from the device's running data over a current period of time;

[0109] 2. Calculate the membership degree α between the running data X and the prerequisite attribute level using triangular membership.

[0110]

[0111] α it =0, t=1, ..., J i , and t≠j,j+1

[0112] Where, α ij A represents the membership degree of the running data of the i-th channel relative to the j-th prerequisite attribute level corresponding to that channel. i(j+1) γ represents the number of abrasive grains corresponding to the (j+1)th prerequisite attribute level of the i-th channel. i A represents a specific data point within the data generated by the i-th channel. ij α represents the number of abrasive grains corresponding to the j-th prerequisite attribute level of the i-th channel; i(j+1) α represents the membership degree of the running data of the i-th channel relative to the (j+1)-th prerequisite attribute level corresponding to that channel; it The degree of membership of the running data of the i-th channel relative to the other prerequisite attribute levels corresponding to that channel is represented by t; the prerequisite attribute level corresponding to that channel is represented by J. i This represents the Jth level of the i-th channel;

[0113] 3. Calculate the activation weight w k ;

[0114]

[0115] Among them, w k Let θ represent the activation weight of the k-th rule. k Let α represent the weight of the k-th rule, L represent the total number of rules, and α represent the weight of the k-th rule. k The cumulative product of the membership degree α of the k-th rule is represented by the following formula:

[0116] Where M represents the number of channels. This represents the membership degree of the i-th channel under the k-th rule. This represents the channel weight of the i-th channel after normalization. The normalization formula is: δ i This represents the channel weight of the i-th channel;

[0117] Step 3: Use evidence-based reasoning to fuse different rules and obtain the output result;

[0118] The fusion is performed according to the following formula:

[0119]

[0120] Among them, y j This represents the confidence level of the j-th level of the output, where L is the total number of rules, and w k Let β represent the activation weight of the k-th rule. j，k β represents the confidence level of the j-th level of the output result under the k-th rule. i，k This represents the confidence level of the i-th conclusion under the k-th rule.

[0121] The formula for calculating μ is:

[0122]

[0123] Where N represents the number of grade divisions in the output result;

[0124] Step 4: Calculate the accuracy of the monitoring results;

[0125] The highest confidence level of the monitoring results is taken as the expected value. The accuracy is calculated based on the comparison between the expected value and the operational data labels. The accuracy calculation formula is as follows: Where true represents the number of correct judgments and false represents the number of incorrect judgments.

[0126] The conclusion has two levels: normal and abnormal. The output of the method is the confidence level of the two levels. For example, the confidence level of normal is 0.4 and the confidence level of abnormal is 0.6. Since 0.6 is greater than 0.4, the expected value is abnormal. According to the running data label, i.e. the true value, if the true value is also abnormal, the judgment is correct; otherwise, the judgment is incorrect.

[0127] Step 5: Input training data and iteratively optimize the confidence rule base parameters using the sequential least squares method until the accuracy is greater than or equal to 90%. When the accuracy is greater than or equal to 90%, the optimization is complete and the optimization expert knowledge is obtained. The optimization expert knowledge is then input into the abrasive confidence rule base again to obtain the optimized abrasive confidence rule base. Based on the optimized abrasive confidence rule base, multi-channel lubricating oil abrasive anomaly monitoring is performed to obtain monitoring results.

[0128] Based on experience, an accuracy rate of 90% or higher is generally considered to be a relatively high accuracy rate.

[0129] Input training data, which comes from the historical health data and fault data of the device operation, and calculate the cross-entropy loss between the output result and the label. The formula for calculating the cross-entropy loss is as follows:

[0130]

[0131] Among them, y i This represents the label of sample i, with 1 for positive and 0 for negative. i L represents the probability that sample i is predicted to be of the positive class. i This represents the i-th rule;

[0132] The optimization strategy is determined, and the model optimization is described as follows:

[0133] minBCELoss(θ k ,β n，k δ i )

[0134] st.

[0135] 0≤θ k ≤1, k=1, …,L,

[0136] 0≤δ i ≤1, i=1,…,M i ,

[0137] 0≤β n，k ≤1, n=1,…,N, k=1,…,L,

[0138]

[0139] Where BCELoss represents the cross-entropy loss function, θ k δ represents the weight of the k-th rule. i β represents the weight of the i-th channel. n，k This represents the confidence level of the nth conclusion under the kth rule.

[0140] Using the parameters of the abrasive confidence rule base determined in step one as the initial input for the sequential least squares method, optimization is performed according to the model optimization strategy of minimizing cross-entropy loss. After optimization, the accuracy is calculated again. If the accuracy is greater than or equal to 90%, the task ends; otherwise, training data is added. The training data also comes from the historical health data and fault data of the equipment until the accuracy is greater than or equal to 90%.

[0141] The parameters of the abrasive confidence rule base are iteratively optimized, and the resulting optimized expert knowledge is used to update the existing abrasive confidence rule base, making the expert knowledge more accurate.

[0142] This method effectively integrates abrasive knowledge using evidence-based reasoning, enabling full utilization of expert knowledge. When the expert knowledge is sufficiently accurate, it can reduce data dependence and save data training time. If the expert knowledge is not accurate enough, it can be optimized using a data-driven approach to make the expert knowledge more accurate. This method can achieve interpretable, highly reliable, and high-performance multi-channel abrasive anomaly monitoring.

[0143] To better understand and implement this invention, the following detailed description, in conjunction with the accompanying drawings, outlines a knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method proposed in this invention. The steps are as follows:

[0144] Step 1: Construct a wear particle confidence rule base based on expert experience;

[0145] Step 2: Input the running data, and activate different rules using triangular membership degrees and rule activation formulas;

[0146] Step 3: Use evidence-based reasoning to fuse different rules and obtain the output result;

[0147] Step 4: Calculate the accuracy of the output results. If the accuracy is greater than or equal to 90%, then end; otherwise, continue to step 5.

[0148] Step 5: Input training data and iteratively optimize the confidence rule base parameters using the sequential least squares method until the accuracy is greater than or equal to 90%. When the accuracy is greater than or equal to 90%, the training is complete and the optimized expert knowledge is obtained. The optimized expert knowledge is then input into the abrasive confidence rule base again to obtain the optimized abrasive confidence rule base. Based on the optimized abrasive confidence rule base, multi-channel lubricating oil abrasive anomaly monitoring is performed to obtain monitoring results.

[0149] The abrasive grain confidence rule base constructed based on expert experience in step one is as follows:

[0150] Divide all channels into three abrasive grain channels: large, medium, and small. If there are 7 channels, divide them into three groups, and then merge each group by addition to obtain the three abrasive grain channels: large, medium, and small. Set the channel weight δ for the i-th channel among the three channels (large, medium, and small). i The level classification A of the i-th channel i n n = 1, 2, ..., N, where N represents the number of grade divisions and A represents the number of abrasive grains corresponding to each channel grade. According to the formula... Calculate the number of rules, where L is the number of rules, M is the number of channels, and D... i This refers to the number of level divisions for the i-th channel; it also refers to the confidence level β of the n-th level of the device under the k-th rule. n，k Set the rule weight θ of the k-th rule. k .

[0151] The input running data described in step two, and the activation of different rules using triangular membership degrees and rule activation formulas, are as follows:

[0152] Step 1: Load the running data X.

[0153] Step 2: Calculate the membership degree α between the running data X and the prerequisite attribute level using triangular membership.

[0154]

[0155] α it =0, t=1, ..., J i , and t≠j,j+1

[0156] Where, α ij A represents the membership degree of the running data of the i-th channel relative to the j-th level corresponding to that channel. i(j+1) γ represents the number of abrasive grains corresponding to the (j+1)th level of the i-th channel. i This represents a specific data point in the data generated by the i-th channel.

[0157] Step 3: Calculate the activation weight w

[0158]

[0159] Among them, w k Let θ represent the activation weight of the k-th rule. k Let α represent the weight of the k-th rule, L represent the total number of rules, and α represent the weight of the k-th rule. k The cumulative product of the membership degrees of the k-th rule is represented by the following formula: Where M represents the number of channels. This represents the membership degree of the i-th channel under the k-th rule. This represents the channel weight of the i-th channel after normalization. The normalization formula is:

[0160] The output obtained by fusing different rules using the evidence-based reasoning method described in step three is as follows:

[0161] Fuse according to the formula below.

[0162]

[0163] Among them, y j This represents the confidence level of the j-th level of the output, where L is the total number of rules, and w k Let β represent the activation weight of the k-th rule. j，k β represents the confidence level of the output at the j-th level under the k-th rule. i，k This represents the confidence level of the i-th conclusion under the k-th rule.

[0164] The formula for calculating μ is:

[0165]

[0166] Where N represents the number of grade divisions in the output result.

[0167] The accuracy of the calculation output results described in step four is as follows:

[0168] Take the level with the highest confidence level in the output as the expected value, and calculate the accuracy using the expected value and the true value. The accuracy calculation formula is as follows: Where true represents the number of correct judgments and false represents the number of incorrect judgments.

[0169] The input training data described in step five is used to iteratively optimize the parameters of the confidence rule base using the sequential least squares method until the accuracy is greater than or equal to 90%, as follows:

[0170] Input training data, which comes from the historical health data and fault data of the device operation, and calculate the cross-entropy loss between the output result and the label. The formula for calculating the cross-entropy loss is as follows:

[0171]

[0172] Among them, y i This represents the label of sample i, with 1 for positive and 0 for negative. i This represents the probability that sample i is predicted to be of the positive class.

[0173] The optimization strategy is determined, and the model optimization is described as follows:

[0174] min BCELoss(θk ,β n，k δ i )

[0175] st.

[0176] 0≤θ k ≤1, k=1, …,L,

[0177] 0≤δ i ≤1, i=1,…,M i ,

[0178] 0≤β n，k ≤1, n=1,…,N, k=1,…,L,

[0179]

[0180] Where BCELoss represents the cross-entropy loss function, θ k δ represents the weight of the k-th rule. i β represents the weight of the i-th channel. n，k This represents the confidence level of the nth conclusion under the kth rule.

[0181] The wear grain confidence rule base parameter θ determined in step S100 k ,β n，k δ i The initial values ​​are used as input for sequential least squares, and the model is optimized according to the model optimization strategy of minimizing cross-entropy loss. After optimization, the accuracy is calculated again. If the accuracy is greater than or equal to 90%, the task ends; otherwise, training data is added until the accuracy is greater than or equal to 90%.

[0182] As shown in Table 1, 4μm, 6μm, and 14μm are small, medium, and large abrasive channels, respectively. Each channel has two levels, 0 and 1, where 0 represents normal and 1 represents abnormal. The different levels of the three channels can form eight rules. Each rule has a corresponding conclusion level confidence level. The conclusion level is also divided into normal and abnormal. The sum of the confidence levels of the conclusion level of normal and abnormal under each rule is 1.

[0183] Table 1 Abrasive Confidence Rule Base

[0184]

[0185] like Figure 1 As shown in the figure, this figure shows the monitoring effect based on the abrasive confidence rule base in Table 1. No data-driven approach is needed; knowledge-driven approach is sufficient. It can be seen that a relatively high accuracy rate of 90% or higher can be achieved by relying solely on knowledge-driven approach, indicating that the knowledge is relatively accurate and no data-driven optimization is required.

[0186] like Figure 2 As shown in the image, the parameters of the randomly generated rule base are obtained, meaning that the input knowledge is randomly generated, and the judgment result is obtained based on the randomly generated knowledge. The specific value of the accuracy can be seen in the upper middle of the image, which is approximately 0.002. It can be seen that the accuracy is very low, meaning that most of the predicted values ​​are inconsistent with the actual values.

[0187] like Figure 3 As shown in the figure, the random-generated rule base parameters are used as initial values ​​and optimized by combining training data (historical data) to obtain the monitoring effect of the optimized abrasive confidence rule base. It can be seen that the abrasive knowledge rules optimized by data-driven optimization can achieve higher accuracy, proving the effectiveness of data-driven optimization.

[0188] like Figure 4 As shown in the figure, the overall calculation flowchart is as follows: First, it is determined whether the abrasive rule base composed of expert knowledge can achieve a monitoring accuracy of more than 90%. If it can, no optimization is needed. If not, it is necessary to use a data-driven approach to optimize, make up for the lack of expert knowledge, and achieve a higher monitoring accuracy.

[0189] like Figure 5 As shown in the figure, the accuracy of sequential least squares, particle swarm optimization, and genetic algorithm changes with the number of iterations during optimization. It can be seen that the sequential least squares method has the most stable iteration process, while the particle swarm optimization and genetic optimization algorithms, due to their randomness, cause the accuracy to fluctuate wildly during the iteration process. The change in accuracy of the three methods during the iteration process can reflect the unique advantage of sequential least squares, namely deterministic optimization, which can meet the stable and reliable requirements of industrial scenarios.

[0190] Figure 6 This is abrasive data, from top to bottom representing small abrasive grains, medium abrasive grains, and large abrasive grains. The horizontal axis represents the time point, and the vertical axis represents the number of abrasive grains. You can see that the number of abrasive grains changes over time. Sometimes, an excessively large number of abrasive grains needs to be identified as an anomaly.

[0191] from Figure 7 It can be seen that the method effectively achieves anomaly monitoring, accurately detecting the time point when an anomaly occurs. All anomaly time points are marked with vertical lines, achieving more precise multi-channel abrasive anomaly monitoring.

[0192] The above is only a description of a multi-channel abrasive anomaly monitoring method driven by both knowledge and data in the context of lubricating oil data, but the present invention is not limited to this application scope.

[0193] Finally, it should be noted that the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0194] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A multi-channel lubricating oil abrasive anomaly monitoring method driven by knowledge data hybridization, characterized in that, The method includes the following steps: Step S100: Construct a wear particle confidence rule base based on expert knowledge, wherein the wear particle confidence rule base contains different rules; Step S200: Load the running data, and use the running data to activate different rules based on triangular membership degree and rule activation formula to obtain activation weights; Step S300: Based on the activation weights and using evidence reasoning, different rules are fused to obtain preliminary abrasive anomaly monitoring results; Step S400: Calculate the accuracy of the preliminary abrasive anomaly monitoring results and determine whether the accuracy reaches or exceeds the set value; if the accuracy is greater than or equal to the set value, end; if the accuracy is less than the set value, optimize the abrasive confidence rule base to obtain optimization expert knowledge. Step S500: Input the optimized expert knowledge into the abrasive confidence rule base again to obtain the optimized abrasive confidence rule base. Based on the optimized abrasive confidence rule base, perform multi-channel lubricating oil abrasive anomaly monitoring to obtain monitoring results. In step S100, constructing the abrasive grain confidence rule base based on expert experience includes the following steps: Step S101: Divide all abrasive channels into three groups: large, medium, and small abrasive channels; Step S102: Set the first of the three groups of abrasive channels (large, medium, and small) respectively. Channel weights of each channel , No. Prerequisite attribute level classification for each channel , =1,2,…, ,in This indicates the number of prerequisite attribute level divisions. Indicates the number of abrasive grains corresponding to the abrasive grain channel prerequisite attribute level; the number of calculation rules L; sets the first... Equipment under the rule of Article 1 Confidence levels ; Setting the first Rule weight of a rule ; In step S400, if the accuracy is greater than or equal to the set value, the process ends. If the accuracy is less than the set value, input training data and apply the sequential least squares method to the parameters of the abrasive confidence rule base. The process of iterative optimization ends when the accuracy reaches a set value or higher, in order to obtain optimization expert knowledge.

2. The knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method according to claim 1, characterized in that, Step S200 specifically includes the following steps: Step 201: Load runtime data ; Step 202: Calculate the running data using triangular membership. Membership degree of prerequisite attribute level ; Step 203: Based on membership degree The result of the product Calculate activation weights ; ， in, Indicates the first The activation weight of each rule. Indicates the first The weight of each rule, Indicates the total number of rules. Indicates the first Membership degree of a rule The result of the cumulative product.

3. The knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method according to claim 2, characterized in that, In step 202, the membership degree is calculated. The specific formula is as follows: ， in, Indicates the first The operating data of the first channel is relative to the first channel corresponding to that channel. The membership degree of each prerequisite attribute level, Indicates the first The first channel +1 the number of abrasive grains corresponding to the prerequisite attribute level, Indicates the first A specific data point within the operational data of a single channel; Indicates the first The first channel The number of abrasive grains corresponding to each prerequisite attribute level; Indicates the first The operating data of the first channel is relative to the first channel corresponding to that channel. +1 membership degree of the prerequisite attribute level; Indicates the first The membership degree of the operational data of each channel relative to the other prerequisite attribute levels corresponding to that channel; The prerequisite attribute level for this channel; Indicates the first The first channel Each level.

4. The knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method according to claim 2, characterized in that, In step S300, fusion is performed according to the following formula: ， in, This indicates the first result of the output. Each level of confidence, It is the total number of rules. Indicates the first The activation weight of each rule. Indicates the first Under the first rule, the... The confidence level of each conclusion level, Indicates the first Under the first rule, the... The confidence level of each conclusion level, It represents an intermediate quantity and has no actual meaning. Indicates the number of conclusion levels.

5. The knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method according to claim 4, characterized in that, The calculation formula is: 。 6. A multi-channel lubricating oil abrasive anomaly monitoring method driven by knowledge data hybridization according to any one of claims 1-5, characterized in that, In step S400, calculating the accuracy of the abrasive anomaly monitoring results specifically includes: The highest confidence level of the abrasive particle anomaly monitoring results is taken as the expected value, and the accuracy is calculated based on the comparison between the expected value and the actual value.

7. The knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method according to claim 6, characterized in that, Parameters of the abrasive grain confidence rule base based on sequential least squares method Iterative optimization, specifically including: The model is optimized using a strategy that minimizes the cross-entropy loss. The optimized model is as follows: ， Wherein, BCELoss represents the cross-entropy loss function, using the parameters of the abrasive confidence rule base determined in step S100. , as the initial value input for the sequential least squares method; Indicates the first The weight of each rule, Indicates the first The weight of each channel, Indicates the first Under the rule of Article [number], the first [number] The confidence level of each conclusion.

8. The knowledge-data hybrid-driven multi-channel lubricating oil abrasive anomaly monitoring method according to claim 7, characterized in that, After optimization, the accuracy is calculated again using the aforementioned accuracy calculation formula. If the accuracy is greater than or equal to the set value, the task ends; otherwise, training data is added until the accuracy is greater than or equal to the set value, at which point the task ends.

Images