Equipment abnormity early warning method and system based on machine learning

Abstract

The invention discloses an equipment abnormity early warning method and system based on machine learning, and the method comprises the steps: collecting the working data of equipment, carrying out the preprocessing of the working data, screening the operation data, obtaining abnormal data, importing the abnormal data into an abnormal function, obtaining an abnormal coefficient, and carrying out the early warning of the equipment abnormity. Decomposing the abnormal coefficient and the state data characteristics to obtain a characteristic vector, comprehensively evaluating the state data to obtain a state evaluation value, and calculating the similarity between the abnormal coefficient and the state evaluation value to obtain an abnormal state degree, and according to the feature vector and the abnormal state degree, constructing an equipment abnormity early warning model based on machine learning, inputting to-be-evaluated data into the equipment abnormity early warning model to obtain an equipment abnormity early warning value, and according to the equipment abnormity early warning value, carrying out abnormity early warning. The method not only can improve the precision of the equipment abnormity early warning method, but also has good interpretability, and can be directly applied to an equipment abnormity early warning system.

Description

Technical Field

[0001] The present invention relates to the field of monitoring and early warning, and in particular to a method and system for early warning of equipment abnormality based on machine learning. Background Art

[0002] Today's industrial production environment is highly automated, and the efficient and stable operation of equipment determines production efficiency and product quality. However, the increase in equipment technology complexity and the variability of the operating environment can easily lead to frequent equipment failures, hindering production continuity. Equipment maintenance is required when failures occur frequently. Traditional manual inspections and regular maintenance methods usually affect production efficiency due to untimely warnings, so the method of analyzing equipment operation data for early warning is more efficient. Equipment operation data has the characteristics of complexity, real-time continuity, massiveness, and instability. Equipment abnormality warning technology based on machine learning can better handle such data. Although some equipment abnormality warning technologies based on machine learning have been invented, existing methods still face challenges such as insufficient data preprocessing, reliance on manual abnormal data identification, and the need to improve the accuracy and robustness of warning models. Summary of the invention

[0003] The purpose of the present invention is to provide a device abnormality early warning method and system based on machine learning.

[0004] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0005] The present invention comprises the following steps:

[0006] Collecting working data of the equipment and preprocessing the working data; the working data includes operation data and status data;

[0007] Screening the operating data to obtain abnormal data, adding secondary labels to the abnormal data, and importing the abnormal data into an abnormal function to obtain an abnormal coefficient;

[0008] Decomposing the abnormal coefficient and the state data features to obtain a feature vector, comprehensively evaluating the state data to obtain a state evaluation value, and calculating the similarity between the abnormal coefficient and the state evaluation value to obtain an abnormal state degree;

[0009] A device abnormality warning model based on machine learning is constructed according to the feature vector and the abnormal state degree, the data to be evaluated is input into the device abnormality warning model to obtain a device abnormality warning value, and an abnormality warning is performed according to the device abnormality warning value.

[0010] Furthermore, the method of screening the operating data to obtain abnormal data includes:

[0011] Define the sample to be screened as x, and the training set as U = {y i ∈RW |y1,y2,…,y N}, containing M class labels c1, c2, ..., c M , the corresponding category is c j The training sample set is N, N j Represent the training set U and The number of samples in, W represents the data dimension, k represents the number of neighbors, and the number of samples to be screened x to The Euclidean distance of:

[0012]

[0013] Sort categories c in ascending order of Euclidean distance j , take the first k nearest neighbors The corresponding category c of the sample x to be screened j The first i local mean vectors of For x in c j The local mean vector set of From x to c j Euclidean distance to local means;

[0014] Assign weights to local mean vectors Calculate various types of c j Pseudo-neighbors of

[0015]

[0016] The class labels of the samples to be screened are:

[0017]

[0018] By obtaining the class label c of the data to be filtered and the corresponding category data proportion w c Identify abnormal data:

[0019]

[0020] Where A represents the judgment result of abnormal data, ρ represents the feature vector of the corresponding category, and θ represents the change parameter of the corresponding category data. represents the time required to identify abnormal data, c x Indicates the class label of the sample to be screened, w c represents the proportion of the corresponding category, and t represents the weight assignment of the corresponding category;

[0021] Label different categories of abnormal data according to the label class number.

[0022] Furthermore, the method of importing the abnormal data into the abnormal function to obtain the abnormal coefficient includes:

[0023] Construct an abnormal function, import the abnormal data x corresponding to the abnormal time t into the abnormal function to calculate the abnormal coefficient Y(t):

[0024]

[0025] Where Y(t) is the abnormal coefficient of the abnormal data x corresponding to the abnormal time t, is the normal distribution expression of the state density function of the scale standard ξ, N is the number of abnormal data, l = N + 1 is the distribution position of the abnormal data weight, l i ∈[1,N] is the distribution location of abnormal data observations, x SD is the standard deviation of abnormal data, MAXx SD is the maximum standard deviation in the abnormal data set, x SK is the skewness of the abnormal data.

[0026] Furthermore, the method of decomposing the abnormal coefficient and the state data feature to obtain a feature vector includes:

[0027] Sort the anomaly numbers according to the time series, and match the status data and anomaly data according to the time points to form a vector set;

[0028] Define the four-dimensional vector space with three imaginary parts and one real part as the four-dimensional number field One of the vectors in the four-dimensional number field is q=q0+iq1+jq2+kq3, where i, j, and k are the coefficients of the three imaginary parts, and the conjugate calculation and involution calculation of the vector are defined as follows:

[0029]

[0030] q β =-βqβ,β=i,j,k

[0031] in and q β They are the conjugate calculation and involution calculation of the vector respectively. Combining the conjugate calculation and involution calculation of the vector, we can get the α-type Hermitian operation of the vector.

[0032] Generate sampling matrix Where Y = B -1 AΩ,A is a Hermitian matrix, is a Gaussian matrix, n is the dimension of the quaternion vector, k is the expected value, and p is the oversampling parameter; Y is decomposed based on the inner product of B to obtain Y = QR, where is an n-dimensional unitary matrix, Is an upper triangular matrix, the identity matrix I = Q H BQ;

[0033] Generate the projection matrix and perform eigendecomposition:

[0034]

[0035] Where Λ is the eigenvalue, S is the corresponding eigenvector, and the first k largest eigenvalues ​​and corresponding eigenvectors are retained as

[0036] Furthermore, the method for comprehensively evaluating the state data to obtain a state evaluation value includes:

[0037] The state data is divided into input indicators, expected output indicators and unexpected output indicators. The input indicator matrix is ​​defined as X, x ik is the value of the kth decision unit of the ith input indicator, and the expected output indicator matrix is ​​defined as Y d , is the value of the kth decision unit of the rth expected output indicator, and the matrix of undesired output indicators is defined as Y u , is the value of the kth decision unit of the rth non-expected output indicator;

[0038] Construct the objective function to minimize the inefficiency value of the decision-making unit and obtain the state evaluation value ρ k :

[0039]

[0040] where ρ k is the state evaluation value of the kth group of data, m1, m2 and m3 are the number of input indicators, expected output indicators and unexpected output indicators, respectively. and They are the slack variables of input index, expected output index and unexpected output index respectively;

[0041] The corresponding constraints are:

[0042]

[0043] where λ j (j≠k) is the weight vector,

[0044] Furthermore, the method of calculating the similarity between the abnormal coefficient and the state evaluation value to obtain the abnormal state degree includes:

[0045] Define the abnormal coefficient set A = {a1, a2, ..., a m}, m is the number of abnormal coefficients, the state evaluation value is matched with the abnormal coefficient, and m state evaluation values ​​are obtained. The state evaluation value set B = {b1, b2, …, b m};

[0046] The abnormal coefficient and status evaluation value are the scores of different dimensions of equipment working conditions. The score weights are defined as:

[0047]

[0048] Where w1(i) is the score weight of the i-th group of data, i∈(1,m), a i and b i are the abnormal coefficient and status evaluation value of the i-th group respectively;

[0049] Convert the abnormal coefficient set A and the state evaluation value set B into vector form to construct the similarity matrix R m×m , define the similarity weight between two vectors as:

[0050]

[0051] Where w2(i) is the similarity weight of the i-th group of data, and are the average values ​​of abnormal coefficients and the average values ​​of state evaluation values ​​in abnormal coefficient set A and state evaluation value set B respectively;

[0052] Get the comprehensive weight:

[0053]

[0054] Calculate the data similarity between the abnormal coefficient set A and the state evaluation value set B to obtain the abnormal state degree:

[0055]

[0056] Where sim(i) is the abnormal state degree of the i-th group of data, and It is the weighted average value of the abnormal coefficient and the status evaluation value.

[0057] Furthermore, a method for constructing a device abnormality warning model based on machine learning according to the feature vector and the abnormal state degree includes:

[0058] The abnormal state degree and the feature vector are matched and combined according to the time points to obtain a comprehensive feature vector, and the comprehensive feature vector is divided into a training set and a test set;

[0059] Construct the BP neural network structure, add the genetic algorithm GA and the ant colony algorithm ACO into the neural network, and input the training set into the GAACO-BP neural network for training;

[0060] The genetic algorithm digitally encodes the potential solution of the problem, establishes the mapping relationship between phenotype and genotype, randomly generates the initial population, distinguishes the quality of individuals through the fitness function, generates a new population through selection, crossover, and mutation operations, and determines the initial weight and bias parameters;

[0061] The ant colony algorithm uses pheromones to mark the path, guides ants to choose the path through pheromone concentration and heuristic information, and constructs an abnormal warning function by adjusting the initial weight and bias parameters through the movement of ants in the solution space;

[0062] The cross-validation method is used to evaluate the test set data, adjust the abnormal warning function, and build the equipment abnormality warning model.

[0063] Furthermore, the method of triggering different levels of warnings according to the abnormal warning of the equipment includes:

[0064] Input the data to be evaluated into the device abnormality warning model to obtain the device abnormality warning value and abnormality identification label, and compare the device abnormality warning value with the warning threshold to issue a device abnormality warning:

[0065]

[0066] Where y is the abnormal warning value of the equipment, θ min is the initial warning threshold, θ low is the low warning threshold, θ high High warning threshold.

[0067] In the second aspect, a device abnormality early warning system based on machine learning includes:

[0068] Data acquisition module: used to collect the working data of the equipment, pre-process the working data, and transmit the working data to the data processing module; the working data includes operation data and status data;

[0069] Data processing module: used for processing the working data to obtain feature vectors and abnormal state degrees;

[0070] Machine learning module: used for inputting the feature vector and abnormal state degree into the GAACO-BP neural network to obtain an equipment abnormality warning model;

[0071] Abnormal warning module: used to manage the working data and the data to be evaluated, and to issue abnormal warnings to users according to the abnormal warning values ​​of the equipment.

[0072] The beneficial effects of the present invention are:

[0073] The present invention is a method and system for early warning of equipment abnormality based on machine learning. Compared with the prior art, the present invention has the following technical effects:

[0074] The present invention can improve the accuracy of equipment abnormality warning by screening abnormal data, obtaining abnormal coefficients, feature decomposition, obtaining state evaluation values, obtaining abnormal state degrees, and constructing model steps, thereby improving the accuracy of equipment abnormality warning, optimizing equipment abnormality warning, greatly saving resources, and improving work efficiency. It can realize equipment abnormality warning, real-time data monitoring and abnormality warning of the equipment working status, which is of great significance for abnormal warning of the equipment working status, can adapt to different equipment abnormality warning systems, and equipment abnormality warning needs of different users, and has a certain universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 This is a flow chart of the steps of a device abnormality warning method based on machine learning in the present invention. DETAILED DESCRIPTION

[0076] The present invention is further described below by means of specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0077] The present invention provides a device abnormality early warning method and system based on machine learning, comprising the following steps:

[0078] like Figure 1 As shown, in this embodiment, the following steps are included:

[0079] Collecting working data of the equipment and preprocessing the working data; the working data includes operation data and status data;

[0080] Screening the operating data to obtain abnormal data, adding secondary labels to the abnormal data, and importing the abnormal data into an abnormal function to obtain an abnormal coefficient;

[0081] Decomposing the abnormal coefficient and the state data features to obtain a feature vector, comprehensively evaluating the state data to obtain a state evaluation value, and calculating the similarity between the abnormal coefficient and the state evaluation value to obtain an abnormal state degree;

[0082] A device abnormality warning model based on machine learning is constructed according to the feature vector and the abnormal state degree, the data to be evaluated is input into the device abnormality warning model to obtain a device abnormality warning value, and an abnormality warning is performed according to the device abnormality warning value.

[0083] In this embodiment, the method of screening the operating data to obtain abnormal data includes:

[0084] Define the sample to be screened as x, and the training set as U = {y i ∈R W |y1,y2,…,y N}, containing M class labels c1, c2, ..., c M , the corresponding category is c j The training sample set is N, N j Represent the training set U and The number of samples in, W represents the data dimension, k represents the number of neighbors, and the number of samples to be screened x to The Euclidean distance of:

[0085]

[0086] Sort categories c in ascending order of Euclidean distance j , take the first k nearest neighbors The corresponding category c of the sample x to be screened j The first i local mean vectors of For x in c j The local mean vector set of From x to c j Euclidean distance to local means;

[0087] Assign weights to local mean vectors Calculate various types of c j Pseudo-neighbors of

[0088]

[0089] The class labels of the samples to be screened are:

[0090]

[0091] By obtaining the class label c of the data to be filtered and the corresponding category data proportion w c Identify abnormal data:

[0092]

[0093] Where A represents the judgment result of abnormal data, ρ represents the feature vector of the corresponding category, and θ represents the change parameter of the corresponding category data. represents the time required to identify abnormal data, c x Indicates the class label of the sample to be screened, w c represents the proportion of the corresponding category, and t represents the weight assignment of the corresponding category;

[0094] Label different types of abnormal data according to label class numbers;

[0095] In the actual evaluation, XX company's baler was used as the research object to give 5 sets of working data. The operating data (packaging frequency, conveyor belt speed) are: 60 times / min, 2m / s; 70 times / min, 2.5m / s; 50 times / min, 1.8m / s; 80 times / min, 3m / s; 65 times / min, 2.2m / s;

[0096] Status data (preset packing quantity, equipment current, equipment temperature, working noise, vibration signal, actual packing quantity, loss, equipment wear value) are: 500 pcs / hour, 15A, 35℃, 70dB, 3G, 480 pcs / h, 10 pcs / h, 0.5%; 600 pcs / hour, 18A, 40℃, 75dB, 4G, 590 pcs / h, 5 / h, 0.3%; 400 pcs / hour, 12A, 30℃, 65dB, 2G, 410 pcs / h, 3 / h, 0.4%; 700 pcs / hour, 20A, 45℃, 80dB, 5G, 680 pcs / h, 10 / h, 0.6%; 550 pcs / hour, 17A, 38℃, 72dB, 3.5G, 560 pcs / h, 2 / h, 0.2%;

[0097] According to the above method, the fourth and fifth groups of operating data are screened out as abnormal data, and the corresponding abnormal situation labels are packaging component abnormality and conveyor belt abnormality, respectively.

[0098] In this embodiment, the method of importing the abnormal data into the abnormal function to obtain the abnormal coefficient includes:

[0099] Construct an abnormal function, import the abnormal data x corresponding to the abnormal time t into the abnormal function to calculate the abnormal coefficient Y(t):

[0100]

[0101] Where Y(t) is the abnormal coefficient of the abnormal data x corresponding to the abnormal time t, is the normal distribution expression of the state density function of the scale standard ξ, N is the number of abnormal data, l = N + 1 is the distribution position of the abnormal data weight, li ∈[1,N] is the distribution location of abnormal data observations, x SD is the standard deviation of abnormal data, MAXx SD is the maximum standard deviation in the abnormal data set, x SK is the skewness of abnormal data;

[0102] In the actual evaluation, the two sets of screened abnormal data were imported into the abnormal function and the abnormal coefficients were calculated to be approximately 11.78 and 8.95 respectively.

[0103] In this embodiment, the method of decomposing the abnormal coefficient and the state data feature to obtain a feature vector includes:

[0104] Sort the anomaly numbers according to the time series, and match the status data and anomaly data according to the time points to form a vector set;

[0105] Define the four-dimensional vector space with three imaginary parts and one real part as the four-dimensional number field One of the vectors in the four-dimensional number field is q=q0+iq1+jq2+kq3, where i, j, and k are the coefficients of the three imaginary parts, and the conjugate calculation and involution calculation of the vector are defined as follows:

[0106]

[0107] q α =-αqα,α=i,j,k

[0108] in and q α They are the conjugate calculation and involution calculation of the vector respectively. Combining the conjugate calculation and involution calculation of the vector, we can get the α-type Hermitian operation of the vector.

[0109] Generate sampling matrix Where Y = B -1 AΩ,A is a Hermitian matrix, is a Gaussian matrix, n is the dimension of the quaternion vector, k is the expected value, and p is the oversampling parameter;

[0110] Perform QR decomposition on Y based on the inner product of B to obtain Y = QR, where is an n-dimensional unitary matrix, Is an upper triangular matrix, the identity matrix I = Q H BQ;

[0111] Generate the projection matrix and perform eigendecomposition:

[0112]

[0113] Where Λ is the eigenvalue, S is the corresponding eigenvector, and the first k largest eigenvalues ​​and corresponding eigenvectors are retained as

[0114] In the actual evaluation, the abnormal coefficient is combined with the state data to form a new vector q = (abnormal coefficient, preset packing amount, equipment current, equipment temperature, working noise, vibration signal, actual packing amount, loss amount, equipment wear value), and the new vectors q4 = [11.78, 700, 20, 45, 80, 5, 680, 10, 0.6] and q5 = [8.95, 500, 17, 38, 72, 3.5, 560, 2, 0.2] are obtained. Since the feature decomposition process requires the construction of a vector matrix to complete, multiple sets of working data are randomly given again to repeat the above steps to obtain 7 new vectors: [8.68, 650, 17, 38, 65, 3.5, 640, 12, 0.2], [8.59, 580, 18, 42, 70, ,4,560,7,0.3], [8.32,520,15,30,68,2.8,508,15,0.5], [9.26,680,19,44,75,4.2,660,6,0.1], [9.38,700,20,45,72,5,685,10,0.4], [8.11,540,16,33,62,3,530,9,0.6], [8.67,560,17.5,39,69,3.2,550,8,0.25], and the corresponding abnormal labels are [conveyor belt abnormality, conveyor belt abnormality, packaging component abnormality, packaging component abnormality, packaging component abnormality, packaging component abnormality, conveyor belt abnormality]. The new square matrix is ​​formed and then feature decomposition is performed.

[0115] In this embodiment, the method of comprehensively evaluating the state data to obtain a state evaluation value includes:

[0116] The state data is divided into input indicators, expected output indicators and unexpected output indicators. The input indicator matrix is ​​defined as X, x ik is the value of the kth decision unit of the ith input indicator, and the expected output indicator matrix is ​​defined as Y d , is the value of the kth decision unit of the rth expected output indicator, and the matrix of undesired output indicators is defined as Y u , is the value of the kth decision unit of the rth non-expected output indicator;

[0117] Construct the objective function to minimize the inefficiency value of the decision-making unit and obtain the state evaluation value ρ k :

[0118]

[0119] where ρ k is the state evaluation value of the kth group of data, m1, m2 and m3 are the number of input indicators, expected output indicators and unexpected output indicators, respectively. and They are the slack variables of input index, expected output index and unexpected output index respectively;

[0120] The corresponding constraints are:

[0121]

[0122] where λ j (j≠k) is the weight vector,

[0123] In the actual evaluation, based on the 9 sets of data obtained, the corresponding input indicators, expected output indicators and unexpected output indicators were input, and m1=5, m2=1 and m3=2 were set, and 9 corresponding state evaluation values ​​{-0.89, -0.92, -0.91, -0.93, -0.94, -0.90, -0.87, -0.92, -0.91} ​​were calculated.

[0124] In this embodiment, the method of calculating the similarity between the abnormal coefficient and the state evaluation value to obtain the abnormal state degree includes:

[0125] Define the abnormal coefficient set A = {a1, a2, ..., a m}, m is the number of abnormal coefficients, the state evaluation value is matched with the abnormal coefficient, and m state evaluation values ​​are obtained. The state evaluation value set B = {b1, b2, …, b m};

[0126] The abnormal coefficient and status evaluation value are the scores of different dimensions of equipment working conditions. The score weights are defined as:

[0127]

[0128] Where w1(i) is the score weight of the i-th group of data, i∈(1,m), a i and b i are the abnormal coefficient and status evaluation value of the i-th group respectively;

[0129] Convert the abnormal coefficient set A and the state evaluation value set B into vector form to construct the similarity matrix R m×m , define the similarity weight between two vectors as:

[0130]

[0131] Where w2(i) is the similarity weight of the i-th group of data, and are the average values ​​of abnormal coefficients and the average values ​​of state evaluation values ​​in abnormal coefficient set A and state evaluation value set B respectively;

[0132] Get the comprehensive weight:

[0133]

[0134] Calculate the data similarity between the abnormal coefficient set A and the state evaluation value set B to obtain the abnormal state degree:

[0135]

[0136] Where sim(i) is the abnormal state degree of the i-th group of data, and It is the weighted average value of the abnormal coefficient and the status evaluation value;

[0137] In the actual evaluation, the calculation is small. By normalizing the data and calculating the z-score of each data, the abnormal state correction value sim = {0.58, 5.07, -0.69, 1.02, 2.87, -0.74, 0.67, 0.82, 0} is obtained.

[0138] In this embodiment, a method for constructing a device abnormality warning model based on machine learning according to the feature vector and the abnormal state degree includes:

[0139] The abnormal state degree and the feature vector are matched and combined according to the time points to obtain a comprehensive feature vector, and the comprehensive feature vector is divided into a training set and a test set;

[0140] Construct the BP neural network structure, add the genetic algorithm GA and the ant colony algorithm ACO into the neural network, and input the training set into the GAACO-BP neural network for training;

[0141] The genetic algorithm digitally encodes the potential solution of the problem, establishes the mapping relationship between phenotype and genotype, randomly generates the initial population, distinguishes the quality of individuals through the fitness function, generates a new population through selection, crossover, and mutation operations, and determines the initial weight and bias parameters;

[0142] The ant colony algorithm uses pheromones to mark the path, guides ants to choose the path through pheromone concentration and heuristic information, and constructs an abnormal warning function by adjusting the initial weight and bias parameters through the movement of ants in the solution space;

[0143] The cross-validation method is used to evaluate the test set data, adjust the abnormal warning function, and build the equipment abnormal warning model;

[0144] In the actual evaluation, three sets of data to be evaluated are given (packaging frequency, conveyor belt speed, preset packaging amount, equipment current, equipment temperature, working noise, vibration signal, actual packaging amount, loss amount, equipment wear value): (55, 2.2, 450, 16, 40, 72, 3.5, 440, 8, 0.4), (65, 2.7, 550, 17, 42, 78, 4.2, 530, 15, 0.5), (75, 3.1, 650, 19, 47, 82, 4.8, 630, 15, 0.7). The data to be evaluated are input into the equipment abnormality warning model to output the equipment abnormality warning value y = {-0.05, 3.3, 6.66}, and the corresponding abnormal identification situation label is [conveyor belt abnormality, conveyor belt abnormality, packaging component abnormality].

[0145] In this embodiment, the method of triggering different levels of warnings according to the device abnormality warning includes:

[0146] Input the data to be evaluated into the device abnormality warning model to obtain the device abnormality warning value and abnormality identification label, and compare the device abnormality warning value with the warning threshold to issue a device abnormality warning:

[0147]

[0148] Where y is the abnormal warning value of the equipment, θ min is the initial warning threshold, θ low is the low warning threshold, θ high High warning threshold;

[0149] In the actual evaluation, the warning threshold θ is set according to the actual working conditions of the baler. min =0,θ low =5 and θ high =8, the triggered warning conditions are:

[0150]

[0151] The equipment abnormality warning value y={-0.05,3.3,6.66}, and the triggered warning conditions are: no warning, conveyor belt level 1 warning and packaging component level 2 warning.

[0152] In the second aspect, a device abnormality early warning system based on machine learning includes:

[0153] Data acquisition module: used to collect the working data of the equipment, pre-process the working data, and transmit the working data to the data processing module; the working data includes operation data and status data;

[0154] Data processing module: used for processing the working data to obtain feature vectors and abnormal state degrees;

[0155] Machine learning module: used for inputting the feature vector and abnormal state degree into the GAACO-BP neural network to obtain an equipment abnormality warning model;

[0156] Abnormal warning module: used to manage the working data and the data to be evaluated, and to issue abnormal warnings to users according to the abnormal warning values ​​of the equipment.

[0157] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A device abnormality early warning method based on machine learning, characterized in that: The following steps are involved: S1. Collecting working data of the equipment and preprocessing the working data; the working data includes operation data and status data; S2, screening the operating data to obtain abnormal data, adding secondary labels to the abnormal data, and importing the abnormal data into an abnormal function to obtain an abnormal coefficient; S3, decomposing the abnormal coefficient and the state data features to obtain a feature vector, comprehensively evaluating the state data to obtain a state evaluation value, and calculating the similarity between the abnormal coefficient and the state evaluation value to obtain an abnormal state degree; S4. Construct a device abnormality warning model based on machine learning according to the feature vector and the abnormal state degree, input the data to be evaluated into the device abnormality warning model to obtain a device abnormality warning value, and perform an abnormality warning according to the device abnormality warning value.

2. According to the device abnormality early warning method based on machine learning in claim 1, it is characterized in that: The method of screening the operating data to obtain abnormal data comprises: Define the sample to be screened as x, and the training set as U = {y i ∈R W |y1,y2,…,y N }, containing M class labels c1, c2, ..., c M , the corresponding category is c j The training sample set is N, N j Represent the training set U and The number of samples in, W represents the data dimension, k represents the number of neighbors, and the number of samples to be screened x to The Euclidean distance of: Sort categories c in ascending order of Euclidean distance j , take the first k nearest neighbors The corresponding category c of the sample x to be screened j The first i local mean vectors of For x in c j The local mean vector set of From x to c j Euclidean distance to local means; Assign weights to local mean vectors Calculate various types of c j Pseudo-neighbors of The class labels of the samples to be screened are: By obtaining the class label c of the data to be filtered and the corresponding category data proportion w c Identify abnormal data: Where A represents the judgment result of abnormal data, ρ represents the feature vector of the corresponding category, and θ represents the change parameter of the corresponding category data. represents the time required to identify abnormal data, c x Indicates the class label of the sample to be screened, w c represents the proportion of the corresponding category, and t represents the weight assignment of the corresponding category; Label different categories of abnormal data according to the label class number.

3. According to the device abnormality early warning method based on machine learning in claim 1, it is characterized in that: The method of importing the abnormal data into the abnormal function to obtain the abnormal coefficient includes: Construct an abnormal function, import the abnormal data x corresponding to the abnormal time t into the abnormal function to calculate the abnormal coefficient Y(t): Where Y(t) is the abnormal coefficient of the abnormal data x corresponding to the abnormal time t, is the normal distribution expression of the state density function of the scale standard ξ, N is the number of abnormal data, l = N + 1 is the distribution position of the abnormal data weight, l i ∈[1,N] is the distribution location of abnormal data observations, x SD is the standard deviation of abnormal data, MAXx SD is the maximum standard deviation in the abnormal data set, x SK is the skewness of the abnormal data.

4. According to the device abnormality early warning method based on machine learning in claim 1, it is characterized in that: The method of decomposing the abnormal coefficient and the state data feature to obtain a feature vector comprises: Sort the anomaly numbers according to the time series, and match the status data and anomaly data according to the time points to form a vector set; Define the four-dimensional vector space with three imaginary parts and one real part as the four-dimensional number field One of the vectors in the four-dimensional number field is q=q0+iq1+jq2+kq3, where i, j, and k are the coefficients of the three imaginary parts, and the conjugate calculation and involution calculation of the vector are defined as follows: q α =-αqα,α=i,j,k in and q α They are the conjugate calculation and involution calculation of the vector respectively. Combining the conjugate calculation and involution calculation of the vector, we can get the α-type Hermitian operation of the vector. Generate sampling matrix Where Y = B -1 AΩ,A is a Hermitian matrix, is a Gaussian matrix, n is the dimension of the quaternion vector, k is the expected value, and p is the oversampling parameter; Perform QR decomposition on Y based on the inner product of B to obtain Y = QR, where is an n-dimensional unitary matrix, Is an upper triangular matrix, the identity matrix I = Q H BQ; Generate the projection matrix and perform eigendecomposition: Where Λ is the eigenvalue, S is the corresponding eigenvector, and the first k largest eigenvalues ​​and corresponding eigenvectors are retained as 5. According to the device abnormality early warning method based on machine learning in claim 1, it is characterized in that: The method for comprehensively evaluating the state data to obtain a state evaluation value comprises: The state data is divided into input indicators, expected output indicators and unexpected output indicators. The input indicator matrix is ​​defined as X, x ik is the value of the kth decision unit of the ith input indicator, and the expected output indicator matrix is ​​defined as Y d , is the value of the kth decision unit of the rth expected output indicator, and the matrix of undesired output indicators is defined as Y u , is the value of the kth decision unit of the rth non-expected output indicator; Construct the objective function to minimize the inefficiency value of the decision-making unit and obtain the state evaluation value ρ k : where ρ k is the state evaluation value of the kth group of data, m1, m2 and m3 are the number of input indicators, expected output indicators and unexpected output indicators, respectively. and They are the slack variables of input index, expected output index and unexpected output index respectively; The corresponding constraints are: where λ j (j≠k) is the weight vector, 6. According to the device abnormality early warning method based on machine learning in claim 1, it is characterized in that: The method for calculating the similarity between the abnormal coefficient and the state evaluation value to obtain the abnormal state degree includes: Define the abnormal coefficient set A = {a1, a2, ..., a m }, m is the number of abnormal coefficients, the state evaluation value is matched with the abnormal coefficient, and m state evaluation values ​​are obtained. The state evaluation value set B = {b1, b2, ..., b m }; The abnormal coefficient and status evaluation value are the scores of different dimensions of the equipment working condition. The score weights are defined as: Where w1(i) is the score weight of the i-th group of data, i∈(1,m), a i and b i are the abnormal coefficient and status evaluation value of the i-th group respectively; Convert the abnormal coefficient set A and the state evaluation value set B into vector form to construct the similarity matrix R m×m , define the similarity weight between two vectors as: Where w2(i) is the similarity weight of the i-th group of data, and are the average values ​​of abnormal coefficients and the average values ​​of state evaluation values ​​in abnormal coefficient set A and state evaluation value set B respectively; Get the comprehensive weight: Calculate the data similarity between the abnormal coefficient set A and the state evaluation value set B to obtain the abnormal state degree: Where sim(i) is the abnormal state degree of the i-th group of data, and It is the weighted average value of the abnormal coefficient and the status evaluation value.

7. According to the machine learning-based equipment abnormality early warning method of claim 1, it is characterized in that: The method for constructing a device abnormality warning model based on machine learning according to the feature vector and the abnormal state degree includes: The abnormal state degree and the feature vector are matched and combined according to the time points to obtain a comprehensive feature vector, and the comprehensive feature vector is divided into a training set and a test set; Construct the BP neural network structure, add the genetic algorithm GA and the ant colony algorithm ACO into the neural network, and input the training set into the GAACO-BP neural network for training; The genetic algorithm digitally encodes the potential solution of the problem, establishes the mapping relationship between phenotype and genotype, randomly generates the initial population, distinguishes the quality of individuals through the fitness function, generates a new population through selection, crossover, and mutation operations, and determines the initial weight and bias parameters; The ant colony algorithm uses pheromones to mark the path, guides ants to choose the path through pheromone concentration and heuristic information, and constructs an abnormal warning function by adjusting the initial weight and bias parameters through the movement of ants in the solution space; The cross-validation method is used to evaluate the test set data, adjust the abnormal warning function, and build the equipment abnormality warning model.

8. According to the device abnormality early warning method based on machine learning in claim 1, it is characterized in that: The method of triggering different levels of warnings according to the abnormal warning of the equipment includes: Input the data to be evaluated into the device abnormality warning model to obtain the device abnormality warning value and abnormality identification label, and compare the device abnormality warning value with the warning threshold to issue a device abnormality warning: Where y is the abnormal warning value of the equipment, θ min is the initial warning threshold, θ low is the low warning threshold, θ high High warning threshold.

9. A device abnormality warning system based on machine learning, characterized in that: include: Data acquisition module: used to collect working data of the equipment, pre-process the working data, and transmit the working data to the data processing module; The working data includes operation data and status data; Data processing module: used for processing the working data to obtain feature vectors and abnormal state degrees; Machine learning module: used for inputting the feature vector and abnormal state degree into the GAACO-BP neural network to obtain an equipment abnormality warning model; Abnormal warning module: used to manage the working data and the data to be evaluated, and to issue abnormal warnings to users according to the abnormal warning values ​​of the equipment.

Images