Vehicle-mounted fan state evaluation method based on multi-information fusion

Abstract

The invention relates to the technical field of fan quality evaluation, in particular to a vehicle-mounted fan state evaluation method based on multi-information fusion, and the evaluation method comprises a quality detection step which comprises the following steps: collecting thermotechnical data, image information, three-phase current, vibration speed, sound and vibration acceleration of a vehicle-mounted fan; calculating the fan power according to the thermotechnical data; extracting blade turning information according to the image information; current unbalance is extracted according to the three-phase current; calculating an RMS value according to the vibration speed; calculating sensitive characteristics according to the sound and vibration acceleration; and judging whether the vehicle-mounted fan is qualified or not based on the fan power, the blade steering information, the current unbalance amount, the RMS value and the sensitive characteristics, comprehensively realizing the quality evaluation of the factory product by combining the industry standard and the machine learning method aiming at the obtained multiple physical quantity signals of the factory vehicle-mounted fan, and further performing grade evaluation. And finally, fault identification is realized based on an evaluation result, and the accuracy and the practicability are high.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine quality assessment, and in particular to a vehicle-mounted wind turbine state assessment method based on multi-information fusion. Background Art

[0002] The onboard fan is a key device for forced ventilation and cooling of the power system of the carriage, and plays a vital role in ensuring the safe and stable operation of the locomotive. Due to factors such as assembly and manufacturing, the quality of the onboard fan products is very likely to be poor, resulting in a short service time after installation. Therefore, it is necessary to study the quality assessment method of onboard fans and their key components.

[0003] The existing quality assessment methods for vehicle-mounted fans and their key components mainly rely on industry standards and are implemented through single physical quantity spot inspection. The current methods have the defects of large manual operation influence and single physical quantity dimension, and whether there is abnormal information in the product cannot be fully assessed.

[0004] Although many studies have conducted detailed analysis and identification of abnormal signal characteristics of different physical quantities, there is no wind turbine quality assessment method that can connect the complete process of abnormal identification-qualified level assessment-fault location and comprehensively consider the multi-information fusion of existing industry standards. At the same time, few existing methods consider the scarcity of prior sample size in actual sites, making it difficult to apply various algorithms in practice.

[0005] Therefore, it is urgent to develop a practical wind turbine condition assessment method based on existing industry standards and methods. Summary of the invention

[0006] The purpose of the present invention is to provide a vehicle-mounted fan status assessment method based on multi-information fusion. The fan quality assessment method is based on the multi-physical quantity signals of the vehicle-mounted fan obtained from the factory, and combines industry standards and machine learning methods to comprehensively determine whether the factory product quality is qualified, and then further evaluates the grade of qualified products, and finally realizes fault identification based on the evaluation results. The specific technical solution is as follows:

[0007] A vehicle-mounted fan state assessment method based on multi-information fusion includes a quality detection step, which includes:

[0008] Step 1.1, collect thermal data, image information, three-phase current, vibration speed, sound and vibration acceleration of the vehicle-mounted fan;

[0009] Step 1.2: Process the thermal data, image information, three-phase current, vibration velocity, sound and vibration acceleration obtained in step 1.1 as follows:

[0010] Calculate fan power based on thermal data; extract blade steering information based on image information; extract current imbalance based on three-phase current; calculate RMS value based on vibration velocity; calculate sensitive features based on sound and vibration acceleration;

[0011] Step 1.3: Make a judgment. The judgment criteria are as follows:

[0012] The fan power comparison is as follows: the calculated fan power is compared with the rated power. If the power error is within the error range, it is qualified; otherwise, it is unqualified.

[0013] Comparison of blade steering information: specifically, comparing the extracted blade steering information with production requirements. If the steering is consistent, it is qualified; otherwise, it is unqualified.

[0014] Comparison of current unbalance, specifically: comparing the extracted current unbalance with the error range required by the industry, if it is within the range, it is qualified, otherwise it is unqualified;

[0015] RMS value comparison, specifically: compare the calculated RMS value with the industry standard, if it is below the limit value, it is qualified, otherwise it is unqualified;

[0016] A multi-sensitive feature set comparison is jointly constructed, specifically: the sound and vibration acceleration are jointly used based on a single-class support vector machine algorithm to identify whether the quality is qualified. If no abnormality is detected, it is qualified, otherwise it is unqualified.

[0017] Preferably, the fan power P is calculated using the following formula based on thermal data:

[0018]

[0019] Where: U is voltage, I is current, cosφ is power factor;

[0020] The current unbalance is extracted according to the three-phase current using the following formula:

[0021]

[0022] Where: D amp is the current amplitude imbalance; D phs is the current phase angle imbalance; amp A is the stator current amplitude corresponding to the current transformer A in the three-phase motor; amp B is the stator current amplitude corresponding to the current transformer B in the three-phase motor; amp C is the stator current amplitude corresponding to the current transformer C in the three-phase motor; phs A is the phase angle corresponding to the current transformer A in the three-phase motor; phs B is the phase angle corresponding to the current transformer B in the three-phase motor; phsC is the phase angle corresponding to the current transformer C in the three-phase motor;

[0023] The RMS value is calculated based on the vibration velocity using the following formula:

[0024]

[0025] Where: x i Represents the amplitude of the vibration velocity; i represents the i-th vibration velocity value in the signal, i=1,2,...,N.

[0026] Preferably, in step 1.3:

[0027] After the wind turbine power comparison is qualified, the blade steering information comparison is carried out; after the blade steering information comparison is qualified, the current imbalance comparison is carried out; after the current imbalance comparison is qualified, the RMS value comparison is carried out; after the RMS value comparison is qualified, the joint construction of multiple sensitive feature set comparison is carried out.

[0028] Preferably, the quality inspection step further includes:

[0029] Step 1.4: The vehicle-mounted fan that has been qualified in step 1.3 is evaluated for its qualification level through the extreme learning machine algorithm.

[0030] Preferably, performing qualification level assessment by extreme learning machine algorithm comprises the following steps:

[0031] Get a single-class support vector machine model;

[0032] The sound and vibration acceleration of the vehicle-mounted fan collected in step 1.1 are input into the single-class support vector machine model, and the qualified level is output.

[0033] Preferably, the process of acquiring the single-class support vector machine model includes:

[0034] The sensitive eigenvalues ​​are calculated using the existing sound and vibration acceleration physical quantities. A training set containing multi-dimensional features of multiple physical quantities is established based on the extended feature merging method.

[0035] Build the initial SVM model;

[0036] The training set is sent to the initial SVM model for training to obtain a single-class support vector machine model.

[0037] Preferably, the single-class support vector machine model does not need to classify the output of the training set, and when all non-abnormal samples are positive categories, a hypersphere is used for division, as follows:

[0038] Assume that the parameters of the generated hypersphere are the center o and the corresponding hypersphere radius r>0, and the volume of the hypersphere V(r) is minimized, requiring that the distance from all training data points z to the center is strictly less than r; at the same time, construct a slack variable ξ with a penalty coefficient C y , the optimization problem is as follows:

[0039]

[0040] ||z y -o||2≤r+ξ y , y=1,2,3,...,m;

[0041] ξ y ≥0;

[0042] After solving the Lagrangian duality, determine whether the new data point z is inside the hypersphere. If the distance from z to the center is less than or equal to the radius r, it is not an outlier. If it is outside the hypersphere, it is an outlier.

[0043] Preferably, the method further includes a fault location step;

[0044] The fault location steps include:

[0045] According to the unqualified result of the fan power comparison, the output motor is abnormal;

[0046] According to the unqualified result of the blade steering information comparison, the motor assembly error is output;

[0047] According to the unqualified result of the current unbalance comparison, the output fan current is unbalanced;

[0048] The vibration velocity, sound and vibration acceleration of the vehicle-mounted fan collected in step 1.1 are input into the deep branch convolutional network fan fault diagnosis model, and the fault location result is output; the fault location result includes dynamic balance abnormality and / or bearing fault.

[0049] Preferably, the process of acquiring the fan fault diagnosis model includes:

[0050] The three physical quantity data of vibration velocity, sound and vibration acceleration that have been collected are converted into RGB color images, where each color represents a physical quantity, thereby constructing a training set consisting of RGB color images;

[0051] Get the initial DBCN model;

[0052] The training set is sent to the initial DBCN model for three-channel convolution feature extraction; the initial three-channel feature mapping value is fully connected to complete the feature fusion process;

[0053] After multiple iterations, a deep branched convolutional network fan fault diagnosis model was obtained.

[0054] Preferably, the fan fault diagnosis model is an extreme learning machine model, which includes an input layer, a hidden layer and an output layer, each layer is composed of neurons, and the neurons in the hidden layer have a threshold value b n , layers are connected by neurons, and neurons in different layers are connected by weight matrices ω and β;

[0055] Assume that given P different data samples (s a , t a ), where S a =[s a1 ,s a2 ,s a3 , ..., s aP ], t a =[t a1 , t a2 , t a3 ,…,t aP ],s a ,t a ∈RP, assuming that the activation function is g(s) and the number of hidden layer nodes is n, the output of its single hidden layer feedforward neural network is as follows:

[0056]

[0057] Among them: C refers to the nodes of the hidden layer; b a Refers to the hidden layer threshold randomly generated in ELM; ω a refers to the weight matrix connecting the input layer and the hidden layer, β a refers to the weight matrix connecting the hidden layer and the output layer, ω a With β a The definition is as follows:

[0058]

[0059] If the activation function is infinitely differentiable, the true output value without error is obtained:

[0060]

[0061] Simplify the above formula to:

[0062] Hβ=T;

[0063] Where: T refers to the true value of the output; H refers to the hidden layer output matrix, and H is defined as follows:

[0064]

[0065] Then we have:

[0066]

[0067] Where: c refers to the regularization coefficient.

[0068] HH T Replaced by the kernel function matrix Ω ELM , then:

[0069]

[0070] The application of the technical solution of the present invention has the following beneficial effects:

[0071] (1) The vehicle-mounted fan status assessment method based on multi-information fusion of the present invention includes a quality detection step, which specifically includes: collecting thermal data, image information, three-phase current, vibration speed, sound and vibration acceleration of the vehicle-mounted fan; calculating the fan power according to the thermal data; extracting blade steering information according to the image information; extracting current unbalance according to the three-phase current; calculating the RMS value according to the vibration speed; calculating sensitive features according to the sound and vibration acceleration; judging whether the vehicle-mounted fan is qualified based on the fan power, blade steering information, current unbalance, RMS value and sensitive features. The vehicle-mounted fan status assessment method of the present invention integrates multiple physical information to comprehensively evaluate the quality of the vehicle-mounted fan, with high accuracy and strong practicality.

[0072] (2) The judgment rule in the present invention is: after the wind turbine power comparison is qualified, the blade steering information comparison is performed; after the blade steering information comparison is qualified, the current imbalance comparison is performed; after the current imbalance comparison is qualified, the RMS value comparison is performed; after the RMS value comparison is qualified, the joint construction of the multi-sensitive feature set comparison is performed. The judgment is performed in sequence in the form of current, vibration velocity, vibration acceleration, and thermal data. The judgment process fully considers the existing national standards and meets the actual needs.

[0073] (3) The quality inspection step in the present invention also includes evaluating the qualified vehicle-mounted fan based on the sound and vibration acceleration through the extreme learning machine algorithm to obtain the precise qualified grade of the vehicle-mounted fan, and the qualified grades include excellent qualified, good qualified and medium qualified.

[0074] (4) The vehicle-mounted fan status assessment method of the present invention also includes a fault location step, which accurately locates the fault of the vehicle-mounted fan by combining the simple physical quantity of the fault mapping and the complex physical quantity of the fault mapping.

[0075] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0077] Figure 1 It is a flow chart of a vehicle-mounted fan state assessment method based on multi-information fusion in a preferred embodiment of the present invention;

[0078] Figure 2 is a schematic diagram of the structure of an extreme learning machine in a preferred embodiment of the present invention;

[0079] Figure 3 It is a schematic diagram of the feature extraction process of the deep network model in the preferred embodiment of the present invention;

[0080] Figure 4 is a schematic diagram of the structure of the deep branch in the preferred embodiment of the present invention;

[0081] Figure 5 It is a structural diagram of a DBCN model in a preferred embodiment of the present invention;

[0082] Figure 6 1 is a schematic diagram of the fault characteristic analysis of the bearing of a THTF4.5 fan with a rotation speed of 2400 r / min in an embodiment of the present invention, wherein: (a) is the fault characteristic analysis of the inner ring of the bearing of fan 1; (b) is the fault characteristic analysis of the outer ring of the bearing of fan 1; (c) is the fault characteristic analysis of the rolling element of fan 1; (d) is the fault characteristic analysis of the inner ring of the bearing of fan 2; (e) is the fault characteristic analysis of the outer ring of the bearing of fan 2; (f) is the fault characteristic analysis of the rolling element of fan 2;

[0083] Figure 7 1 and 1 are fault identification results of two models in an embodiment of the present invention, where (a) is an ELM model and (b) is a Muti-class SVM model. DETAILED DESCRIPTION

[0084] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0085] Example:

[0086] A vehicle-mounted wind turbine condition assessment method based on multi-information fusion, see Figure 1 , including quality inspection steps and fault location steps. Before quality inspection, it is necessary to collect parameters of the vehicle-mounted fan as a whole. The details are as follows:

[0087] The quality inspection steps include:

[0088] Step 1.1: Collect thermal data, image information, three-phase current, vibration velocity, sound, and vibration acceleration of the vehicle-mounted fan.

[0089] Step 1.2: Process the thermal data, image information, three-phase current, vibration velocity, sound and vibration acceleration obtained in step 1.1 as follows:

[0090] Calculate fan power based on thermal data; extract blade steering information based on image information; extract current imbalance based on three-phase current; calculate RMS value based on vibration velocity; calculate sensitive features based on sound and vibration acceleration;

[0091] Specifically:

[0092] ①、Calculation of fan power P:

[0093] According to the thermal data, the fan power P is calculated using the following formula:

[0094]

[0095] Where: U is voltage, I is current, cosφ is power factor;

[0096] ②. Current unbalance:

[0097] Current detection is the most commonly used motor diagnostic technology. The motor used to test the vehicle-mounted fan is a three-phase motor, and the stator current of each phase can be analyzed for abnormalities by calculating the changes in amplitude and phase. Assuming that the stator current of each phase is obtained through current transformers A, B, and C, the detailed analysis process is as follows:

[0098] Perform FFT transformation on the collected stator current to obtain the stator current amplitude amp and phase angle phs of this phase:

[0099]

[0100] Where: Q represents the amount of stator current; J represents the imaginary unit; i[q] represents the qth value in the stator current; imag() represents taking the imaginary part; real() represents taking the real part; e represents the base of the natural logarithm; π represents pi; k' represents the k'th frequency component;

[0101] The stator current quantity Q is determined by the following formula:

[0102]

[0103] Where: f s represents the preset sampling frequency; f min Indicates the minimum frequency value of the external interference signal collected synchronously with the stator current; ppm indicates the preset diagnostic accuracy; maxrepresents the maximum available hardware resource of the motor controller; O(1) represents the hardware resource of the motor controller occupied when the preset number is 1;

[0104] Calculate the stator current amplitude difference of each phase as the current amplitude imbalance D amp , calculate the phase angle difference of each phase stator current as the current phase angle imbalance D phs ,as follows:

[0105]

[0106] Where: amp A is the stator current amplitude corresponding to the current transformer A in the three-phase motor; amp B is the stator current amplitude corresponding to the current transformer B in the three-phase motor; amp C is the stator current amplitude corresponding to the current transformer C in the three-phase motor; phs A is the phase angle corresponding to the current transformer A in the three-phase motor; phs B is the phase angle corresponding to the current transformer B in the three-phase motor; phs C It is the phase angle corresponding to the current transformer C in the three-phase motor.

[0107] In this embodiment, the reference value of the current amplitude imbalance is generally the average of the three phases, and the reference value of the current phase angle imbalance is 120°. By comparing the amplitude and phase angle difference between the stator currents of each phase with the reference values, the abnormality of the motor can be directly reflected. At the same time, by comparing the imbalance of the normal motor and the abnormal motor, the current difference when the motor is abnormal can be further reflected.

[0108] ③. Calculate the RMS value based on the vibration velocity:

[0109] The vibration velocity mainly reflects the vibration magnitude of the equipment. According to national and industry testing standards, the RMS value of the vibration velocity is usually calculated for analysis, specifically:

[0110]

[0111] Where: x i Represents the amplitude of the vibration velocity; i represents the i-th vibration velocity value in the signal, i=1,2,...,N.

[0112] ④. Sensitive features corresponding to sound and vibration acceleration calculations:

[0113] Sound and vibration acceleration are among the most popular physical quantities for abnormal judgment and diagnosis. They have many sensitive features that can be used as evaluation criteria, and the optimal sensitive features need to be selected in combination with actual fan products. It is recommended to use psychoacoustic indicators for sound physical quantities: loudness, fluctuation, roughness, sharpness, and annoyance, and to use impact characteristics, energy characteristics, and waveform characteristics for vibration acceleration physical quantities: mean, peak-to-peak value, absolute minimum value, margin factor, and other characteristics.

[0114] Step 1.3, make a judgment. Evaluate the corresponding physical quantities based on the calculated characteristics, consider the complexity of different physical quantities for judging the quality of the fan and whether there are relevant industry standards, and evaluate them in the order of simple-complex, with standards-without standards: thermal data, image information, three-phase current, vibration speed, sound and vibration acceleration, that is: see Figure 1 After the fan power comparison is qualified, the blade steering information comparison is performed; after the blade steering information comparison is qualified, the current imbalance comparison is performed; after the current imbalance comparison is qualified, the RMS value comparison is performed; after the RMS value comparison is qualified, the joint construction of multi-sensitive feature set comparison is performed. If one of the physical quantity judgment results is unqualified, the fan quality is judged to be unqualified and no further judgment operation is performed. The judgment criteria are as follows:

[0115] The fan power comparison is specifically: compare the calculated fan power with the rated power. If the power error is within the error range, it is qualified, otherwise it is unqualified. Here, if the power deviates from 13kw, it is considered unqualified.

[0116] The blade steering information is compared, specifically: the extracted blade steering information is compared with the production requirements, if the steering is consistent, it is qualified, otherwise it is unqualified.

[0117] The current unbalance comparison is specifically: comparing the extracted current unbalance with the error range required by the industry. If it is within the range, it is qualified, otherwise it is unqualified; that is, if the balance amount is greater than the set threshold, it is judged as unqualified.

[0118] RMS value comparison, specifically: compare the calculated RMS value with the industry standard, if it is below the limit value, it is qualified, otherwise it is unqualified.

[0119] A multi-sensitive feature set comparison is jointly constructed, specifically: the sound and vibration acceleration are jointly used based on the one-class support vector machine algorithm (i.e., One-Class Support Vector Machine, One-Class SVM) to identify whether the quality is qualified. If no abnormality is detected, it is qualified, otherwise it is unqualified.

[0120] In this embodiment: the SVM One-Class algorithm is specifically:

[0121] SVM One-Class is different from the traditional classification and regression SVM based on supervised learning. It is an unsupervised learning method, which does not require the output of the training set to be labeled. When all samples that are not abnormal are positive categories, a hypersphere is used instead of a hyperplane for division. The algorithm obtains the spherical boundary around the data in the feature space, hoping to minimize the volume of this hypersphere, thereby minimizing the impact of abnormal point data. The acquisition process of the one-class support vector machine model described in this embodiment includes: calculating sensitive eigenvalues ​​using the existing collected sound and vibration acceleration physical quantities; forming a training set containing multi-dimensional features of multiple physical quantities based on the extended feature merging method; constructing an initial SVM model; sending the training set to the initial SVM model for training to obtain a single-class support vector machine model. Further preferred: the one-class support vector machine model does not require the output of the training set to be labeled, and when all samples that are not abnormal are positive categories, a hypersphere is used for division, as follows:

[0122] Assume that the parameters of the generated hypersphere are the center o and the corresponding hypersphere radius r>0, and the volume of the hypersphere V(r) is minimized, requiring that the distance from all training data points z to the center is strictly less than r; at the same time, construct a slack variable ξ with a penalty coefficient C y , the optimization problem is as follows:

[0123]

[0124] ||z y -o||2≤r+ξ y ,y=1,2,3,...,m;

[0125] ξ y ≥0, y=1,2,...,m;

[0126] After solving the Lagrangian duality, determine whether the new data point z is inside the hypersphere. If the distance from z to the center is less than or equal to the radius r, it is not an outlier. If it is outside the hypersphere, it is an outlier.

[0127] Step 1.4: Evaluate the qualified grade of the fans with qualified quality. Specifically, the qualified car fans determined in step 1.3 are evaluated by the extreme learning machine algorithm (i.e., ELM). The extreme learning machine (ELM) is a type of machine learning system or method based on the feedforward neural network (FNN), which is suitable for supervised learning and unsupervised learning problems. It has the advantages of fast operation speed and providing global optimal solutions. Compared with the single layer perceptron and the support vector machine (SVM), the extreme learning machine has more advantages in learning rate and generalization ability. When facing the early small sample car fan motor data, the model has the advantages of high efficiency and automation compared with deep learning or manual judgment.

[0128] This embodiment preferably includes the following steps:

[0129] Get a single-class support vector machine model;

[0130] The sound and vibration acceleration of the vehicle-mounted fan collected in step 1.1 are input into the single-class support vector machine model, and the qualified level is output.

[0131] The wind turbine fault diagnosis model is an extreme learning machine model, which includes an input layer, a hidden layer and an output layer. Figure 2 , each layer is composed of neurons, and the hidden layer neurons have a threshold b n , layers are connected by neurons, and neurons in different layers are connected by weight matrices ω and β;

[0132] Assume that given P different data samples (S a , t a ), where sa = [s a1 ,s a2 ,s a3 , ..., s aP ], t a =[t a1 , t a2 , t a3 ,…,t aP ],S a ,t a ∈RP, assuming that the activation function is g(s) and the number of hidden layer nodes is n, the output of its single hidden layer feedforward neural network is as follows:

[0133]

[0134] Among them: C refers to the nodes of the hidden layer; b a Refers to the hidden layer threshold randomly generated in ELM; ω a refers to the weight matrix connecting the input layer and the hidden layer, β a refers to the weight matrix connecting the hidden layer and the output layer, ω a With β a The definition is as follows:

[0135]

[0136] If the activation function is infinitely differentiable, the true output value without error is obtained:

[0137]

[0138] Simplify the above formula to:

[0139] Hβ=T;

[0140] Where: T refers to the true value of the output; H refers to the hidden layer output matrix, and H is defined as follows:

[0141]

[0142] Then we have:

[0143]

[0144] Where: c refers to the regularization coefficient.

[0145] In order to further improve the generalization ability and stability of the ELM model, HH T Replaced by the kernel function matrix Ω ELM , then:

[0146]

[0147] In this embodiment, considering the complexity of the mapping relationship between different physical quantities and faults, fault location is performed on physical quantities with simple and complex fault mappings respectively. Therefore, the fault location steps in this embodiment include:

[0148] (1) Fault mapping simple physical quantities, including:

[0149] According to the unqualified result of the fan power comparison, the output motor is abnormal;

[0150] According to the unqualified result of the blade steering information comparison, the motor assembly error is output;

[0151] According to the unqualified result of the current unbalance comparison, the output fan current is unbalanced;

[0152] (2) Fault mapping of complex physical quantities, including:

[0153] The vibration velocity, sound and vibration acceleration of the vehicle-mounted fan collected in step 1.1 are input into the deep branch convolutional network fan fault diagnosis model, and the fault location result is output. That is, for the physical quantities of vibration velocity, sound and vibration acceleration, a deep branch convolutional network model (Deep Branch Convolutional Networks, DBCN) is proposed for more accurate fault identification. This model can diagnose abnormal dynamic balance and bearing faults of the fan. Considering the possible unknown fault forms, samples that do not belong to these two categories are diagnosed as other fault types. The specific principle and calculation process of the DBCN model are as follows:

[0154] Deep network models represented by CNN (Convolutional Neural Networks, CNN) have been widely studied because of their powerful adaptive feature extraction capabilities compared to shallow machine learning models, especially in the field of fault diagnosis. Most of the latest research results are related to this type of method. The feature extraction method of deep network models is layer-by-layer progressive. The feature extraction results of the current layer are based on the previous layer. As the number of network layers increases, the extracted features gradually change from "generalization" to "specialization", and are increasingly conducive to the classification of data samples of different categories. This process is like Figure 3 shown.

[0155] Considering the small number of samples of measured data in the target domain during actual on-site fault location, the traditional network model feature extraction structure is single. When faced with small sample target domain data tasks, it is very easy to produce poor accuracy and low effectiveness of extracted features, resulting in poor diagnostic and recognition effects when applied. In this regard, based on the principle of deep network model feature extraction, a deep branch network structure is proposed to enhance the deep feature extraction capability of the target domain measured data to improve the diagnostic and recognition effect. The specific model structure is as follows Figure 4 This structure only appears in the deeper layers of the network model. It extracts useful information at multiple scales by broadening the width of the key feature extraction layer, ensuring that the model can efficiently improve its feature extraction capability when facing small sample data.

[0156] The DBCN training process is roughly divided into forward propagation and backward propagation. Forward propagation mainly performs multi-scale feature extraction and mapping, and backward propagation is to build a loss function based on the training output results and update the network parameters layer by layer through reverse propagation. The DBCN model structure is as follows Figure 5As shown in the figure, the feature extraction layer of the model consists of three parts. The first two parts are shallow / middle-level feature extraction modules composed of convolution and pooling processes, and the third part is a deep-level feature extraction module composed of multiple convolution and pooling branches of different scales. This module uses convolution kernels of different scales to finely extract deep features of the data, further finely explore the characteristics of small sample data in the target domain, and improve the effect of migration diagnosis. In the specific implementation of the model, considering the problem of feature splicing after multiple convolutions, an additional DepthConcatenation layer suitable for deep feature splicing is added after the deep feature extraction module. Finally, there is the fully connected layer, which maps and represents the extracted deep features, and implements sample classification through the Softmax function.

[0157] The training process of the DBCN model is divided into forward propagation and back propagation. In the forward propagation process, it is assumed that the original data is D = [d1, d2, ..., d u ], the deep feature extraction module can be expressed as:

[0158] F deep =[e1,e2,...,e v ];

[0159] The feature extraction process is a convolution process, and the specific calculation is:

[0160]

[0161] Among them: F w is the wth feature obtained by convolution; F l is the lth convolved local area in the original image; v is the vth input image; K is the convolution kernel; b w is the bias of the wth feature; f is the activation function.

[0162] The pooling layer is mainly used to reduce the size of the feature map obtained by convolution. By reducing the features, the number of fully connected layer parameters is reduced to prevent model overfitting. Commonly used pooling forms include: maximum pooling, mean pooling, random pooling, etc. The specific calculation of the pooling process is:

[0163]

[0164] Where: ρ is the pooling function, for example: maximum pooling is to take the maximum value of the pooling area, and mean pooling is to calculate the mean of the pooling area.

[0165] The fully connected layer is consistent with the conventional network structure, in which each neuron between layers is connected. In the fully connected layer, the image features obtained by the convolution and pooling process will be integrated, and different activation functions will be used to further achieve operational goals such as image classification.

[0166] The commonly used activation function of convolutional networks is Relu. Furthermore, in order to enhance the learning ability and generalization of the model, the activation function is improved to Leaky Relu. The activation function is shown in the following formula, which can reduce the neuron necrosis phenomenon that may occur in the Relu function due to unreasonable parameter initialization.

[0167] After the fully connected layer, the output data is converted into a probability in the range of [0,1] through the Softmax function, so as to achieve the classification of different samples. The Softmax function is defined as:

[0168]

[0169] In the formula, γ p is the output value of the pth node; γ u is the output value of the u-th node; U is the number of output nodes.

[0170] During the back-propagation process, the loss function uses the cross entropy loss function shown below:

[0171]

[0172] In the formula, Y and are the real fault label and the predicted label respectively; k is the sample category, k = 1, 2, ..., K; y k is the probability that the sample belongs to the kth class. The updated network parameter ψ can be obtained through training:

[0173]

[0174] Then the iterative update process of the network parameters by gradient descent method is:

[0175]

[0176] In the formula, λ is the learning rate of the network model.

[0177] Assume that the number of deep branches is G, and the corresponding network parameters are ψ d1 , d2 , ..., ψ dG . Then the network parameters of each branch are synchronously updated as follows:

[0178]

[0179] When the reverse update is performed to the middle-level feature extraction module, the process can be described as a many-to-one process, so the gradient calculation process can be expressed as:

[0180]

[0181] Where:m are the parameters of the middle-level feature extraction module.

[0182] The fault location result includes dynamic balance abnormality and / or bearing fault.

[0183] Preferably, the acquisition process of the fan fault diagnosis model includes: converting the three physical quantity data of vibration velocity, sound and vibration acceleration that have been collected into RGB color images, wherein: each color represents a physical quantity, thereby constructing a training set consisting of RGB color images; obtaining an initial DBCN model; sending the training set into the initial DBCN model for three-channel convolution feature extraction; the initial three-channel feature mapping value fully connected layer to complete the feature fusion process; and obtaining a deep branched convolutional network fan fault diagnosis model after multiple iterations.

[0184] The technical solution of this embodiment is specifically:

[0185] 1. Test content: Conduct 50Hz working condition multi-physical quantity and multi-fault type prototype test on the traction fan, model: THTF5.5, test 2 units (i.e. fan 1 and fan 2).

[0186] 2. Test process: This test uses vibration acceleration sensors, sound sensors, and vibration velocity sensors to collect vibration and sound data of the traction fan and its motor fault prototype.

[0187] The multi-physical quantity information is monitored and recorded accordingly through the DEWETRON multi-channel data acquisition device and the supporting host computer software, see Table 1 for details.

[0188] Online real-time statistical analysis of time domain waveforms to analyze and identify typical fault types of traction fans and their motors.

[0189] Table 1 Statistics of state type and working conditions, test physical quantity, test times, acquisition time, and sampling frequency

[0190]

[0191] 3. Data Analysis:

[0192] The bearing model in this test is SKF 6310. The fault characteristic frequency at a speed of 3000r / min is calculated based on the corresponding parameters as shown in Table 2:

[0193] Table 2 Characteristic frequency of rolling bearing (speed 3000r / min)

[0194] Bearing components Bearing inner ring Bearing outer ring Rolling element Characteristic frequency (Hz) 247.6 152.4 198.1

[0195] When the fan speed is 3000r / min, the vibration acceleration of the measured data of different bearing components failure is demodulated to obtain the envelope spectrum as shown below: Figure 6 As shown, where:

[0196] (a) is the fault characteristic analysis of the inner ring of the bearing of fan 1;

[0197] (b) is the fault characteristic analysis of the outer ring of the bearing of fan 1;

[0198] (c) is the fault characteristic analysis of the rolling element of fan 1;

[0199] (d) is the fault characteristic analysis of the inner ring of the bearing of fan 2;

[0200] (e) is the fault characteristic analysis of the outer ring of the bearing of fan 2;

[0201] (f) is the failure characteristic analysis of the rolling element of fan 2.

[0202] Combination Figure 6 Observations:

[0203] 1) The peak spectrum of 244Hz can be seen on the envelope amplitude diagram of the bearing inner ring fault, which is very close to the theoretical value of 247.6Hz;

[0204] 2) The peak spectrum of 151Hz can be seen on the bearing outer ring fault envelope amplitude diagram, which is very close to the theoretical value of 152.4Hz;

[0205] 3) In the bearing rolling element fault envelope amplitude diagram, the peak spectrum of 195Hz and 200Hz can be seen, which is very close to the theoretical value of 198.1Hz.

[0206] 4. Fan diagnostic test:

[0207] Here, the training samples with the constructed dimension of [180,8] are input into ELM and Multi-class SVM for training. The fault recognition results of the two models are shown in Figure 7 As shown in the figure, (a) is the ELM model and (b) is the Muti-class SVM model. Figure 7 It can be seen that the total recognition accuracy of the ELM model is 81.11%, and the total recognition accuracy of the SVM model is 82.22%. When there is only one fan sample, the SVM model is better. The ELM model has a better recognition effect on abnormal dynamic balance and bearing failure, and the SVM model is the same, but it has a better recognition effect on normal samples.

[0208] Furthermore, the proposed method was tested, and the test result statistics are shown in Table 3. According to Table 3, the total recognition accuracy of the DBCN model (i.e., the scheme of this embodiment) is 100%, and this is true for both the fan and the motor. The recognition accuracy of other methods has decreased due to the strong noise contained in the fan test data. This result shows the superiority of the proposed DBCN model for fan fault location.

[0209] Table 3 Comparison of recognition results of various schemes

[0210]

[0211] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A vehicle-mounted fan status assessment method based on multi-information fusion, characterized in that: The method comprises a quality inspection step, wherein the quality inspection step comprises: Step 1.1, collect thermal data, image information, three-phase current, vibration speed, sound and vibration acceleration of the vehicle-mounted fan; Step 1.2: Process the thermal data, image information, three-phase current, vibration velocity, sound and vibration acceleration obtained in step 1.1 as follows: Calculate fan power based on thermal data; extract blade steering information based on image information; extract current imbalance based on three-phase current; calculate RMS value based on vibration velocity; calculate sensitive features based on sound and vibration acceleration; Step 1.3: Make a judgment. The judgment criteria are as follows: The fan power comparison is as follows: the calculated fan power is compared with the rated power. If the power error is within the error range, it is qualified; otherwise, it is unqualified. Comparison of blade steering information: specifically, comparing the extracted blade steering information with production requirements. If the steering is consistent, it is qualified; otherwise, it is unqualified. Comparison of current unbalance, specifically: comparing the extracted current unbalance with the error range required by the industry, if it is within the range, it is qualified, otherwise it is unqualified; RMS value comparison, specifically: compare the calculated RMS value with the industry standard, if it is below the limit value, it is qualified, otherwise it is unqualified; A multi-sensitive feature set comparison is jointly constructed, specifically: the sound and vibration acceleration are jointly used based on a single-class support vector machine algorithm to identify whether the quality is qualified. If no abnormality is detected, it is qualified, otherwise it is unqualified.

2. The vehicle-mounted fan status assessment method according to claim 1, characterized in that: According to the thermal data, the fan power P is calculated using the following formula: Where: U is voltage, I is current, cosφ is power factor; The current unbalance is extracted according to the three-phase current using the following formula: Where: D amp is the current amplitude imbalance; D phs is the current phase angle imbalance; amp A is the stator current amplitude corresponding to the current transformer A in the three-phase motor; amp B is the stator current amplitude corresponding to the current transformer B in the three-phase motor; amp C is the stator current amplitude corresponding to the current transformer C in the three-phase motor; phs A is the phase angle corresponding to the current transformer A in the three-phase motor; phs B is the phase angle corresponding to the current transformer B in the three-phase motor; phs C is the phase angle corresponding to the current transformer C in the three-phase motor; The RMS value is calculated based on the vibration velocity using the following formula: Where: x i Represents the amplitude of the vibration velocity; i represents the i-th vibration velocity value in the signal, i=1,2,...,N.

3. The vehicle-mounted fan status assessment method according to claim 1, characterized in that: In step 1.3: After the wind turbine power comparison is qualified, the blade steering information comparison is carried out; after the blade steering information comparison is qualified, the current imbalance comparison is carried out; after the current imbalance comparison is qualified, the RMS value comparison is carried out; after the RMS value comparison is qualified, the joint construction of multiple sensitive feature set comparison is carried out.

4. The vehicle-mounted fan status assessment method according to claim 3, characterized in that: The quality inspection step also includes: Step 1.4: The vehicle-mounted fan that has been qualified in step 1.3 is evaluated for its qualification level through the extreme learning machine algorithm.

5. The vehicle-mounted fan status assessment method according to claim 4, characterized in that: The qualification level assessment by extreme learning machine algorithm includes the following steps: Get a single-class support vector machine model; The sound and vibration acceleration of the vehicle-mounted fan collected in step 1.1 are input into the single-class support vector machine model, and the qualified level is output.

6. The vehicle-mounted fan status assessment method according to claim 5, characterized in that: The acquisition process of the single-class support vector machine model includes: The sensitive eigenvalues ​​are calculated using the existing sound and vibration acceleration physical quantities. A training set containing multi-dimensional features of multiple physical quantities is established based on the extended feature merging method. Build the initial SVM model; The training set is sent to the initial SVM model for training to obtain a single-class support vector machine model.

7. The vehicle-mounted fan status assessment method according to claim 6, characterized in that: In the single-class support vector machine model, there is no need to label the output of the training set, and when all non-abnormal samples are positive categories, a hypersphere is used for division, as follows: Assume that the parameters of the generated hypersphere are the center o and the corresponding hypersphere radius r>0, and the volume of the hypersphere V(r) is minimized, requiring that the distance from all training data points z to the center is strictly less than r; at the same time, construct a slack variable ξ with a penalty coefficient C y , the optimization problem is as follows: ||z y -o||2≤r+ξ y ,y=1,2,3,...,m; x y ≥0,; After solving the Lagrangian duality, determine whether the new data point z is inside the hypersphere. If the distance from z to the center is less than or equal to the radius r, it is not an outlier. If it is outside the hypersphere, it is an outlier.

8. The vehicle-mounted fan status assessment method according to any one of claims 1 to 7, characterized in that: It also includes a fault location step; The fault location steps include: According to the unqualified result of the fan power comparison, the output motor is abnormal; According to the unqualified result of the blade steering information comparison, the motor assembly error is output; According to the unqualified result of the current unbalance comparison, the output fan current is unbalanced; The vibration velocity, sound and vibration acceleration of the vehicle-mounted fan collected in step 1.1 are input into the deep branch convolutional network fan fault diagnosis model, and the fault location result is output; the fault location result includes dynamic balance abnormality and / or bearing fault.

9. The vehicle-mounted fan status assessment method according to claim 8, characterized in that: The process of acquiring the fan fault diagnosis model includes: The three physical quantity data of vibration velocity, sound and vibration acceleration that have been collected are converted into RGB color images, where each color represents a physical quantity, thereby constructing a training set consisting of RGB color images; Get the initial DBCN model; The training set is sent to the initial DBCN model for three-channel convolution feature extraction; the initial three-channel feature mapping value is fully connected to complete the feature fusion process; After multiple iterations, a deep branched convolutional network fan fault diagnosis model was obtained.

10. The vehicle-mounted fan status assessment method according to claim 9, characterized in that: The fan fault diagnosis model is an extreme learning machine model, which includes an input layer, a hidden layer and an output layer. Each layer is composed of neurons. The neurons in the hidden layer have a threshold value b. n , layers are connected by neurons, and neurons in different layers are connected by weight matrices ω and β; Assume that given P different data samples (s a , t a ), where s a =[s a1 ,s a2 , S a3 , ..., S aP ], t a =[t a1 , t a2 , t a3 ,…,t aP ],s a ,t a ∈RP, assuming that the activation function is g(s) and the number of hidden layer nodes is n, the output of its single hidden layer feedforward neural network is as follows: Among them: C refers to the nodes of the hidden layer; b a Refers to the hidden layer threshold randomly generated in ELM; ω a refers to the weight matrix connecting the input layer and the hidden layer, β a refers to the weight matrix connecting the hidden layer and the output layer, ω a With β a The definition is as follows: If the activation function is infinitely differentiable, the true output value without error is obtained: Simplify the above formula to: Hβ=T. Where: T refers to the true value of the output; H refers to the hidden layer output matrix, and H is defined as follows: Then we have: Where: c refers to the regularization coefficient. HH T Replaced by the kernel function matrix Ω ELM , then:

Images