Fan drive chain man-machine cooperation health assessment method

Abstract

The present application belongs to the technical field of wind power equipment intelligent operation and maintenance, and discloses a wind turbine transmission chain man-machine collaborative health evaluation method, which comprises the following steps: simulating the fault mode of the wind turbine transmission chain, obtaining the vibration signal under different fault modes, and training the constructed neural network model; obtaining the evaluation set of BtO and OtW about the risk degree of each fault mode; constructing a Bayesian hierarchical model, converting the risk weight of the fault mode and the evaluation set into a probability distribution form, determining the aggregated risk weight of each fault mode in the evaluation set based on the Bayesian hierarchical model; using the trained neural network model to evaluate the fault state membership degree of the wind turbine transmission chain, fusing the fault state membership degree and the aggregated risk weight to calculate the health degree index, and evaluating the health state of the wind turbine transmission chain. The present application combines monitoring big data and operation and maintenance experience knowledge, and can realize accurate quantitative evaluation of the health state of the wind turbine transmission chain.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent operation and maintenance of wind power equipment, and more specifically, relates to a human-machine collaborative health assessment method for wind turbine transmission chains. Background Technology

[0002] Wind power equipment operates in complex and variable environments, constantly exposed to extreme conditions such as hurricanes, thunderstorms, and high humidity. This makes its drivetrain components susceptible to varying degrees and forms of fatigue, wear, fracture, and corrosion, gradually reducing their functional characteristics and operating efficiency, severely impacting their long-term stable and reliable operation. Furthermore, due to limitations in the accessibility of wind power equipment maintenance, wind turbine maintenance cannot be carried out as regularly as with conventional equipment. Limited accessibility, harsh working environments, and finite maintenance resources make wind turbine drivetrain components typical examples of multi-constraint, limited-access repair devices. With the deepening of Industry 4.0 and the continuous breakthroughs in emerging technologies such as multimodal sensing, industrial big data, and artificial intelligence, intelligent operation and maintenance technology is considered a key entry point and breakthrough for improving the reliability, maintainability, and safety of wind turbine drivetrains.

[0003] Research revealed that most current wind turbine drivetrain maintenance technologies are based on monitoring data and model-driven approaches under known ideal conditions. However, due to the extremely harsh working environment, complex and variable operating conditions, uncertain damage and failure mechanisms, and a scarcity of fault samples in wind turbine drivetrains, existing technologies struggle to meet the accuracy requirements of health assessments. In recent years, the rapid development of smart IoT, new sensing and communication technologies, and their application in wind power equipment have enabled the timely collection and aggregation of operational monitoring data, working environment and health status monitoring data for wind turbine drivetrains. The data transmission frequency is on the order of seconds, with approximately 1.5TB of new data added daily.

[0004] Existing methods utilize big data from wind turbine drivetrain monitoring to effectively diagnose drivetrain failure modes and conduct drivetrain health assessments. However, current assessment methods are typically qualitative and cannot achieve precise quantitative assessments of drivetrain health status. Summary of the Invention

[0005] To address the above-mentioned deficiencies or improvement needs of existing technologies, this invention provides a human-machine collaborative health assessment method for wind turbine drive chains, which solves the problem that existing wind turbine drive chain health assessment methods are usually qualitative and cannot achieve accurate quantitative assessment of the health status of the drive chain.

[0006] To achieve the above objectives, according to the present invention, a human-machine collaborative health assessment method for wind turbine drivetrain is provided, comprising:

[0007] Offline training and data processing phase:

[0008] The fault modes of the wind turbine drive chain are identified and simulated. Vibration signals of the wind turbine drive chain during operation under different fault modes are obtained and used to train the constructed neural network model. The neural network model is then used to evaluate the membership degree of the fault state of the wind turbine drive chain.

[0009] Obtain the evaluation sets of BtO and OtW for the risk level of each failure mode; construct a Bayesian hierarchical model to transform the risk weights of failure modes and the evaluation sets into a probability distribution form; and determine the aggregated risk weight of each failure mode that integrates all risk opinions in the evaluation set based on the Bayesian hierarchical model.

[0010] Online assessment phase:

[0011] The vibration signal during the operation of the wind turbine drive chain is input into the above-mentioned offline trained neural network model to obtain the fault state membership degree of the wind turbine drive chain; the health index is calculated by fusing the fault state membership degree and the aggregated risk weight based on the ratio system method, and the health status of the wind turbine drive chain is evaluated based on the health index.

[0012] According to the wind turbine drivetrain human-machine collaborative health assessment method provided by the present invention, the neural network model is a multifunctional parallel network, which includes a counting self-attention module and a localization self-attention module, wherein:

[0013] The counting self-attention module is used to calculate the number of faults, and the positioning self-attention module is used to calculate the fault probability distribution.

[0014] The multifunctional parallel network is used to combine the number of faults calculated by the counting self-attention module with the fault probability distribution result calculated by the positioning self-attention module to obtain the fault state membership degree.

[0015] According to the wind turbine drivetrain human-machine collaborative health assessment method provided by the present invention, the membership degree of the fault state is calculated according to the following formula:

[0016]

[0017] Where p is the sample number, the sample is a vibration signal, p = 1, 2, ..., P, and P is the number of samples; f ( p ) The time and frequency domain characteristics of the vibration signal; FSM(f ( p ) ) represents the membership degree of the fault state; For the predicted number of faults; FPDj(f ( p )) represents the failure probability distribution; θj represents the model parameters for each failure mode category; j is the failure mode number; N is the maximum number among the failure mode numbers; L (p) Number and label the fault modes; The calculation results of the localization self-attention module in the layer before the output layer are applicable to fault mode j.

[0018] According to the wind turbine drivetrain human-machine collaborative health assessment method provided by the present invention, the evaluation set is obtained through expert evaluation by an expert panel, and obtaining the evaluation set specifically includes:

[0019] Obtain evaluation information from any expert in the expert panel regarding the most severe and least severe failure modes;

[0020] Based on the evaluation information of any expert, obtain the BtO vector and OtW vector of the risk level of each failure mode corresponding to the evaluation information of any expert, and summarize the BtO vector and OtW vector of all experts in the expert group to form the evaluation set.

[0021] According to the wind turbine drivetrain human-machine collaborative health assessment method provided by the present invention, the risk weights of the failure modes and the evaluation set in the Bayesian hierarchical model are specifically transformed into a probability distribution form as follows:

[0022] The failure mode risk weights and the evaluation set are transformed into a joint probability distribution using the Bayesian hierarchical model. The failure mode risk weights include individual risk weights for each failure mode determined based on the evaluation information of a single expert, and aggregated risk weights for each failure mode determined by integrating the opinions of all experts in the expert group.

[0023] According to the wind turbine drivetrain human-machine collaborative health assessment method provided by the present invention, the aggregated risk weight for each failure mode, which incorporates the opinions of all experts in the expert panel, specifically includes the following in S2:

[0024] Based on the evaluation information of individual experts in the expert group, obtain the single risk weight of each failure mode determined based on the evaluation information of individual experts.

[0025] The vectors in the evaluation set and the individual risk weights of each failure mode are transformed into multinomial distributions, and the individual risk weights and aggregate risk weights of each failure mode are transformed into Dirichlet distributions. Combined with the Bayesian hierarchical model, the aggregate risk weights are determined.

[0026] According to the wind turbine drivetrain human-machine collaborative health assessment method provided by the present invention, the Bayesian hierarchical model is specifically as follows:

[0027]

[0028] Where P(·) represents the form of the probability distribution; and These are the BtO and OtW vectors in the fault risk assessment set provided by K experts. It is a single fault risk weight vector derived from the evaluation information of K experts. The single risk weight of failure mode j is derived based on the evaluation information of the kth expert. It is the aggregate risk weight of failure mode j; and The failure risk vectors BtO and OtW are provided for the k-th expert; K is the total number of experts.

[0029] According to the wind turbine drivetrain human-machine collaborative health assessment method provided by the present invention, the single risk weight of failure mode j is determined based on the evaluation information of the k-th expert. Specifically, it is calculated using the following formula:

[0030]

[0031] in, and The single risk weights for the most severe and least severe failure modes, determined based on the opinion of the kth expert; N is the maximum number among the failure mode numbers. The severity of the most severe failure mode provided by the k-th expert relative to the j-th failure mode. The severity of the j-th failure mode provided by the k-th expert relative to the least severe failure mode.

[0032] According to the wind turbine drivetrain human-machine collaborative health assessment method provided by the present invention, the polynomial distribution is performed according to the following relationship:

[0033]

[0034] in, and The fault risk vectors BtO and OtW are provided for the k-th expert; It is the single risk weight of failure mode j derived from the evaluation information of the kth expert; K is the total number of experts;

[0035] The Dirichlet distribution is determined according to the following relationship:

[0036]

[0037] in, is the aggregate risk weight of failure mode j; Dir(·) is the Dirichlet distribution; γ is a non-negative parameter with a GAMMA distribution.

[0038] According to the wind turbine drivetrain human-machine collaborative health assessment method provided by the present invention, the health index is determined according to the following formula:

[0039]

[0040] Among them, HI is the health index; DI is the damage index; AC is the amplification factor; and Acc is the diagnostic accuracy of the neural network model. is the aggregate risk weight of fault mode j; FSMj is the j-th element in the fault state membership; z is the benefit criterion number; It is the number of faults predicted by the counting self-attention module.

[0041] Overall, compared with the prior art, the human-machine collaborative health assessment method for wind turbine drivetrain provided by this invention offers the following advantages:

[0042] 1. A neural network model is constructed to calculate the membership degree of equipment failure status. Then, based on the evaluation set obtained by the BWM evaluation method and the Bayesian hierarchical model, the aggregated risk weight of the failure mode is calculated and determined. Subsequently, the failure status membership degree and aggregated risk weight are integrated based on the ratio system method to calculate the health index and conduct a health assessment of the wind turbine drive chain. This method combines monitoring big data and operation and maintenance experience knowledge to achieve accurate quantitative assessment of the health status of the wind turbine drive chain.

[0043] 2. By constructing a multifunctional parallel network, the number of faults in the wind turbine drive train and the probability of each fault occurring can be determined, and this can be combined with the membership degree of the fault status of the output equipment to accurately quantify the evolution stage of the detected faults.

[0044] 3. Constructing the GCD method: First, based on the BWM method, K experts evaluate the risk level of each failure mode to obtain an evaluation set. Then, a Bayesian hierarchical model is constructed. By combining the multinomial distribution function and the Dirichlet distribution function, the aggregated failure risk weights are finally determined. The failure risk preferences of all experts can be comprehensively considered to give the optimal aggregated failure risk weight evaluation result.

[0045] 4. An extended Ratio system method is proposed to consider the impact of fault diagnosis model accuracy on result confidence. Based on this, the membership degree of wind turbine drivetrain fault state and fault aggregation risk weight are combined to give the final wind turbine drivetrain health assessment result. Attached Figure Description

[0046] Figure 1 This is a flowchart of the human-machine collaborative health assessment method for wind turbine transmission chains provided by the present invention;

[0047] Figure 2 This is a schematic diagram of a human-machine collaborative health assessment method for a wind turbine drivetrain constructed according to a specific embodiment of the present invention;

[0048] Figure 3 yes Figure 2 A schematic diagram of the human-machine collaborative health assessment framework for the wind turbine drive chain;

[0049] Figure 4 yes Figure 2 A schematic diagram of the Versatile parallel network in the human-machine collaborative health assessment method for wind turbine drivetrain;

[0050] Figure 5 These are the equipment health assessment results under various fault conditions constructed according to specific embodiments of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0052] Please see Figure 1 This embodiment provides a human-machine collaborative health assessment method for wind turbine drivetrain, the method including:

[0053] Offline training and data processing phase:

[0054] The fault modes of the wind turbine drive chain are identified and simulated. Vibration signals of the wind turbine drive chain during operation under different fault modes are obtained and used to train the constructed neural network model. The neural network model is then used to evaluate the membership degree of the fault state of the wind turbine drive chain.

[0055] Obtain evaluation sets of BtO (Best-to-other) and OtW (Other-to-worst) for the risk level of each failure mode; construct a Bayesian hierarchical model to transform the risk weights of failure modes and the evaluation sets into a probability distribution form; based on the Bayesian hierarchical model, determine the aggregated risk weight of each failure mode that integrates all risk opinions in the evaluation set; the evaluation set can be obtained by evaluating the risk level of each failure mode based on the BWM (Best to Worst) method to form an evaluation set including BtO and OtW evaluation information; the evaluation set gathers relevant experience and knowledge of wind turbine operation and maintenance, and the aggregated risk weights obtained based on the evaluation set are more objective and reliable;

[0056] Online assessment phase:

[0057] The vibration signal during the operation of the wind turbine drive chain is input into the above-mentioned offline trained neural network model to obtain the fault state membership degree of the wind turbine drive chain; the health index is calculated by fusing the fault state membership degree and the aggregated risk weight based on the ratio system method, and the health status of the wind turbine drive chain is evaluated based on the health index.

[0058] In some specific embodiments, the neural network model is a multifunctional parallel network, which includes a counting self-attention module and a localization self-attention module, wherein:

[0059] The counting self-attention module is used to calculate the number of faults, and the positioning self-attention module is used to calculate the fault probability distribution.

[0060] The multifunctional parallel network is used to combine the number of faults calculated by the counting self-attention module with the fault probability distribution result calculated by the positioning self-attention module to obtain the fault state membership degree.

[0061] refer to Figure 2 The wind turbine drivetrain human-machine collaborative health assessment method provided in this embodiment mainly includes the following steps:

[0062] Step 1: Determine the typical fault modes of the wind turbine drive chain to be simulated, analyze the vibration response characteristics of the drive chain under different fault modes, and then deploy vibration sensors to collect vibration signals during the operation of the drive chain under different fault models.

[0063] Specifically, the vibration signal of the wind turbine drivetrain was acquired by an NI-cDAQ-9174 / 9234 vibration sensor with a sampling frequency of 10240Hz. The sampling duration was set to 100s, and 1,024,000 data points were obtained for each fault mode. 510 samples from each fault mode were evenly selected for training the neural network model. The wind turbine drivetrain is typically a gearbox, which mainly contains gears, bearings, and couplings. The vibration sensor can be placed in the middle of the base of the wind turbine gearbox to effectively reflect its vibration.

[0064] Step 2: The vibration signals collected under each fault mode are used as samples to train the constructed multifunctional parallel network. The trained Versatile parallel network is then used to evaluate the fault state membership of the wind turbine drive train.

[0065] Specifically, refer to Figure 3 and Figure 4The constructed multifunctional parallel network employs a self-attention neural network. The Versatileparallel network includes a counter self-attention module and a localizer self-attention module.

[0066] The calculation of the membership degree of the fault state includes the following steps:

[0067] The vibration signal is labeled with the number of faults, and its time domain and frequency domain features are extracted. The number of faults is used as the module output, and the extracted features are used as the input to train the Counter self-attention module. The number of faults is determined through this Locator self-attention module.

[0068] The vibration signal is labeled with a fault mode tag, and its time domain and frequency domain features are extracted. The fault mode is used as the module output, and the extracted features are used as the input to train the Locator self-attention module. The SoftMax activation function in the Locator self-attention module is modified to the SoftProb function so that its output is the fault probability distribution.

[0069] The number of faults calculated by the Counter self-attention module is combined with the fault probability distribution calculated by the Locator self-attention module to obtain the fault state membership degree.

[0070] Specifically, the Counter self-attention module is used to obtain the number of equipment faults. First, time-domain and frequency-domain features are extracted from the vibration signal. Then, a label u with the corresponding number of faults is generated. (p) Time-domain and frequency-domain features f of ∈(1,...,U) (p) (p = 1, ..., P) This is input into the Counter self-attention module for model training, where U is the number of faults in the fault mode with the most faults. The fault number prediction result is:

[0071]

[0072] in, This is the result of the fault quantity prediction, W1 C and W2 C These are the weights of the Counter self-attention module, b1 C and It is the bias of the Counter self-attention module.

[0073] The Locator self-attention module is used to obtain the fault probability distribution. This involves assigning corresponding fault category labels L... p Time-domain and frequency-domain features f of ∈(1,..NN,) (p) (p = 1, ..., P) Train the Locator self-attention module, where N is the maximum fault mode number, and the fault category label can be the fault mode number. The fault category classification result is:

[0074]

[0075] in, This is the failure mode classification result, W1 L and W2 L These are the weights of the Locator self-attention module, b1 L and It is the bias of the Locator self-attention module.

[0076] To enable the model to recognize multiple fault modes simultaneously, the objective function of the Locator self-attention module is modified:

[0077]

[0078] Where P is the total number of samples, and if sample p belongs to fault mode j, E ij The value is 1 if it is 1, otherwise it is 0.

[0079] Subsequently, the failure probability distribution FPD(f) of each sample was obtained through the SoftProb module. ( p ) )as follows:

[0080]

[0081] Among them, FPDj(f ( p ) ) represents the failure probability distribution; j is the failure mode number; L ( p ) Number and label the failure modes; θ j These are the model parameters for the j-th fault category; The calculation result of the localization self-attention module applicable to fault mode j in the layer before the output layer; N is the maximum number in the fault mode number; The score for the j-th type of fault mode is the result after normalization by the Softprob module.

[0082] The results from the fault counting module and the fault probability distribution calculation module are combined to finally obtain the fault state membership degree. The fault state membership degree is calculated according to the following formula:

[0083]

[0084] Where p is the sample number, the sample is a vibration signal, p = 1, 2, ..., P, and P is the number of samples; f ( p ) The time and frequency domain characteristics of the vibration signal; FSM(f ( p ) ) represents the membership degree of the fault state; The predicted number of failures.

[0085] Step 3, refer to Figure 3 In some specific embodiments, the evaluation set is obtained through expert evaluation. Specifically, a failure mode risk assessment expert group can be constructed, and the evaluation information of the risk level of each failure mode by the experts using the BWM method (Best to Worst method) can be obtained, forming the failure mode risk Best-to-other and Other-to-worst evaluation sets. K experts can be invited to form a wind turbine drive chain failure mode risk assessment expert group. The experts in the expert group use the BWM method to evaluate the risk level of each failure mode to the equipment based on their own knowledge and experience using natural language processing, and then convert it into a failure risk BtO / OtW evaluation set.

[0086] Specifically, obtaining the evaluation set includes:

[0087] Obtain the evaluation information of any expert in the expert panel on the most severe failure mode and the least severe failure mode; that is, each of the K invited experts determines the most severe failure and the least severe failure.

[0088] Based on the evaluation information of any expert, obtain the Best-to-Others (BtO) and Others-to-Worst (OtW) risk vectors for each failure mode corresponding to the evaluation information of any expert. Then, summarize the BtO and OtW vectors of all experts in the expert group to form the evaluation set. That is, use the Best to worst method to give the Best-to-Others (BtO) and Others-to-Worst (OtW) failure risk vectors, and summarize the two types of vectors from K experts to form the evaluation set.

[0089] In other embodiments, the evaluation set can also be obtained in other ways, such as directly selecting relevant evaluation sets from existing studies, training an evaluation model based on existing data, and using the evaluation model to evaluate and obtain the evaluation set, or other artificial intelligence methods. The specific method of obtaining the evaluation set is not limited.

[0090] Step four: Based on the evaluation set provided by the expert panel, a Bayesian hierarchical model is constructed to transform the evaluation opinions of each expert and the risk weights of each failure mode into probability distributions. By comprehensively applying the multinomial distribution function and the Dirichlet distribution function, the aggregated failure risk weights are finally determined.

[0091] Specifically, in the Bayesian hierarchical model, the risk weights of the failure modes and the evaluation set are transformed into a probability distribution form as follows:

[0092] The Bayesian hierarchical model transforms the failure mode risk weights and the evaluation set into a joint probability distribution; that is, the Bayesian hierarchical model establishes a functional relationship between the failure mode risk weights and the vectors in the evaluation set in the form of a probability distribution. The failure mode risk weights include individual risk weights determined based on the evaluation information of a single expert, and aggregated risk weights determined by integrating the opinions of all experts in the expert panel. In the joint probability distribution, the failure mode risk weights include both individual and aggregated risk weights for each failure mode; the evaluation set includes the BtO vector and OtW vector of K experts.

[0093] Furthermore, the aggregated risk weights for each failure mode, which incorporate the opinions of all experts in the expert panel, are specifically determined in S2 as follows:

[0094] Based on the evaluation information of individual experts in the expert group, the single risk weight of each failure mode determined by the evaluation information of individual experts can be obtained. Specifically, based on the evaluation information of individual experts, namely the BtO vector and OtW vector corresponding to the evaluation information of individual experts, the single risk weight of each failure mode corresponding to the evaluation information of individual experts can be calculated.

[0095] Then, the vectors in the evaluation set and the individual risk weights of each failure mode are transformed into multinomial distributions. The individual risk weights and aggregate risk weights of each failure mode are then transformed into Dirichlet distributions. Combined with the Bayesian hierarchical model, the aggregate risk weight is determined. After determining the individual risk weights of each failure mode corresponding to the evaluation information of each expert, the only unknown variable in the joint probability distribution determined by the Bayesian hierarchical model is the aggregate risk weight. Therefore, by combining the multinomial and Dirichlet distributions, the aggregate risk weight can be obtained.

[0096] Specifically, a Bayesian hierarchical model is constructed to convert the fault risk BtO / OtW evaluation set provided by K experts into a probability distribution form; the Bayesian hierarchical model is as follows:

[0097]

[0098] Where P(·) represents the form of the probability distribution; and These are the BtO and OtW vectors in the fault risk assessment set provided by K experts. It is a single fault risk weight vector derived from the evaluation information of K experts. The single risk weight of failure mode j is derived based on the evaluation information of the kth expert. It is the aggregate risk weight of failure mode j; and The failure risk vectors BtO and OtW are provided for the k-th expert; K is the total number of experts.

[0099] The single risk weight of failure mode j determined based on the evaluation information of the k-th expert. Specifically, it is calculated using the following formula:

[0100]

[0101] in, and The single risk weights for the most severe and least severe failure modes, determined based on the opinion of the kth expert; N is the maximum number among the failure mode numbers. The severity of the most severe failure mode provided by the k-th expert relative to the j-th failure mode. The severity of the j-th failure mode provided by the k-th expert relative to the least severe failure mode.

[0102] The multinomial distribution is determined according to the following relationship:

[0103]

[0104] in, and The fault risk vectors BtO and OtW are provided for the k-th expert; It is the single risk weight of failure mode j derived from the evaluation information of the kth expert; K is the total number of experts;

[0105] The Dirichlet distribution is determined according to the following relationship:

[0106]

[0107] in, is the aggregate risk weight of failure mode j; Dir(·) is the Dirichlet distribution; γ is a non-negative parameter with a GAMMA distribution.

[0108] This embodiment proposes a Group Confidence Decision (GCD) method. This GCD method first obtains an evaluation set by having K experts assess the risk level of each failure mode based on the Bayesian Multivariate Model (BWM). Then, it constructs a Bayesian hierarchical model, combining multinomial and Dirichlet distribution functions to finally determine the aggregated failure risk weights. The proposed GCD method can synthesize the opinions of K experts, providing a comprehensive and reliable failure risk weight assessment result.

[0109] Step 5: By fusing fault state membership and aggregated risk weights using the extended Ratio system method, the equipment health index is calculated, thereby providing a comprehensive and quantitative assessment of the wind turbine drivetrain health. After obtaining the state membership and aggregated risk weights, the equipment health index is obtained through analysis using the extended Ratio system method. The health index is calculated according to the following formula:

[0110]

[0111] Wherein, HI is the health index; DI is the damage index; AC is the amplification factor; and Acc is the diagnostic accuracy of the neural network model, which is the ratio of the number of correctly diagnosed samples to the total number of diagnosed samples. is the aggregate risk weight of fault mode j; FSMj is the j-th element in the fault state membership; z is the benefit criterion number; This is the number of faults predicted by the self-attention counting module. When the network diagnostic accuracy is low, an amplification factor is introduced to increase the damage index, reflecting the impact of the uncertainty of fault occurrence risk on the health status of the wind turbine.

[0112] To further illustrate the embodiments of the present invention in detail, in some specific embodiments, wind turbine gearbox fault simulation experimental data are used to verify the method. The wind turbine gearbox fault simulation experimental platform consists of two ABB MV1008-225 motors (1.2kW), a gearbox, a flywheel, a data acquisition system, and a computer. One motor acts as the prime mover driving the multi-stage gearbox, while the other acts as an asynchronous generator simulating various resistance torques. In the experiment, the input speed of the device is set to 1400 rpm, and the speeds of the two meshing gear sets in the gearbox are 1184 rpm and 840 rpm, respectively.

[0113] The health of the wind turbine gearbox needs to be evaluated under fifteen fault modes, namely one health state (H1), six single fault conditions (IFS1~IFS6), and eight compound fault conditions (CFS1~CFS8). Specific fault types and corresponding labels are shown in Table 1. In this specific embodiment, the six single fault conditions are numbered, i.e., N=6. The compound fault conditions can be a combination of the single fault condition numbers. The six common single fault conditions involved in this experiment include tooth breakage, tooth plate detachment, tooth plate cracking, coupling loosening, bearing rolling element wear, and bearing outer ring wear.

[0114] Table 1

[0115]

[0116]

[0117] The gearbox vibration signal was acquired by an NI-cDAQ-9174 / 9234 vibration sensor with a sampling frequency of 10240Hz. The sampling duration was set to 100s, which means that 1,024,000 data points were obtained for each fault state, i.e., fault mode (i.e., one sensor × 100s × 10240Hz). 510 samples were selected from the data points under each fault state, for a total of 7650 samples as the dataset. After uniform mixing, the samples were divided into a 60% training set, a 10% validation set, and a 30% test set.

[0118] To verify the effectiveness of the Versatile parallel fault state membership evaluation network, this paper uses eight methods, i.e., eight different networks, for comparative analysis based on the same dataset mentioned above, which are recorded as Method-1 to Method-8, as follows.

[0119] Method-1: pairwise probabilistic multi-label classification (PPMLC)

[0120] Method-2: multilabel decision tree (MDT)

[0121] Method-3: gradient boosting decision tree (GBDT)

[0122] Method 4: random k labelsets (RAKEL)

[0123] Method-5: dual-extreme learning machine (Dual-ELM)

[0124] Method-6: multi-label radial basis function (ML-RBF)

[0125] Method-7: multi-label convolutional neural network with wavelettransform (WT-MLCNN)

[0126] Method-8: The multifunctional parallel network proposed in this invention

[0127] To eliminate the randomness of the comparison results, thirty repeated experiments were performed. The training and testing diagnostic accuracy of the eight methods are recorded in Table 2. The table shows that the diagnostic accuracy of the network of this invention is significantly higher than that of the other seven methods. This further demonstrates the strong inventiveness and applicability of this invention, making it suitable for practical industrial applications.

[0128] Table 2

[0129]

[0130]

[0131] Table 3 shows the number of faults and the fault probability distribution obtained by inputting the test set into the Versatile parallel network. Table 3 shows that most samples corresponding to the 15 fault conditions can be correctly identified. Combining the number of faults predicted by the Counterself-attention module and the fault probability distribution calculated by the Locator self-attention module, the equipment health index can be accurately assessed.

[0132] Table 3

[0133]

[0134]

[0135] The risk weights for each individual failure case were determined using GCD. CL was the most severe failure with a risk weight as high as 0.3138, followed by LT (0.3109), BT (0.1447), CT (0.1332), and BRW (0.0974).

[0136] By extending the Ratio system method to combine fault state membership with fault risk weights, the overall equipment health under different fault conditions is obtained, such as... Figure 5 As shown.

[0137] This embodiment collects vibration signals generated by the wind turbine drivetrain, constructs a Versatile parallel network, and determines the membership degree of equipment fault states. It invites experts in intelligent operation and maintenance to form a wind turbine drivetrain fault risk assessment expert group. Experts in this group use the Group Confidence Decision (GCD) method based on the Best to Worst (BWM) method to assess fault risks, constructing a Bayesian hierarchical model to transform the experts' assessment opinions into probability distributions. Combining multinomial and Dirichlet distribution functions, the aggregated fault risk weights are determined. An extended ratio system is used to define the equipment health index formula, calculating the equipment health index to achieve wind turbine drivetrain health assessment. This embodiment solves the problem that existing wind turbine health assessment methods cannot comprehensively consider the opinions of all experts in the expert group, achieving the goal of comprehensive health assessment.

[0138] The wind turbine drivetrain human-machine collaborative health assessment method provided by this invention uses a versatile parallel network to calculate the membership degree of equipment fault states. Then, a GCD model is used to determine the fault risk weights. Finally, based on an extended ratio system method, the fault state membership degree and fault risk weights are fused to calculate a health index, thereby assessing the health of the wind turbine drivetrain. This method measures the impact of different fault conditions on equipment behavior with finer granularity, thus accurately reflecting the current health status of the equipment. It provides guidance for subsequent maintenance and post-event monitoring strategies, improves maintenance efficiency, and offers good flexibility.

[0139] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for human-machine collaborative health assessment of a wind turbine drivetrain, characterized in that, include: Offline training and data processing phase: The failure modes of the wind turbine drive train are identified and simulated. Vibration signals during the operation of the wind turbine drive train under different failure modes are obtained. These vibration signals are then used to train the constructed neural network model. The neural network model is used to evaluate the membership degree of the failure state of the wind turbine drive train. The neural network model is a multi-functional parallel network. Obtain the evaluation sets of BtO and OtW for the risk level of each failure mode; construct a Bayesian hierarchical model to transform the risk weights of failure modes and the evaluation sets into a probability distribution form; and determine the aggregated risk weight of each failure mode that integrates all risk opinions in the evaluation set based on the Bayesian hierarchical model. Online assessment phase: The vibration signal during the operation of the wind turbine drive chain is input into the above-mentioned offline trained neural network model to obtain the fault state membership degree of the wind turbine drive chain. The health index is calculated by fusing the fault state membership degree and the aggregated risk weight based on the ratio system method, and the health status of the wind turbine drive chain is evaluated based on the health index. The evaluation set is obtained through expert evaluation by a panel of experts. The specific steps involved in obtaining the evaluation set are as follows: Obtain evaluation information from any expert in the expert panel regarding the most severe and least severe failure modes; Based on the evaluation information of any expert, obtain the BtO vector and OtW vector of each failure mode risk level corresponding to the evaluation information of any expert, and summarize the BtO vector and OtW vector of all experts in the expert group to form the evaluation set. The Bayesian hierarchical model transforms the risk weights of failure modes and the evaluation set into a probability distribution form as follows: The failure mode risk weights and the evaluation set are transformed into a joint probability distribution using the Bayesian hierarchical model. The failure mode risk weights include individual risk weights for each failure mode determined based on the evaluation information of a single expert, and aggregated risk weights for each failure mode determined by integrating the opinions of all experts in the expert group. Based on the evaluation information of individual experts in the expert group, obtain the single risk weight of each failure mode determined based on the evaluation information of individual experts. The vectors in the evaluation set and the individual risk weights of each failure mode are transformed into multinomial distributions, and the individual risk weights and aggregate risk weights of each failure mode are transformed into Dirichlet distributions. Combined with the Bayesian hierarchical model, the aggregate risk weights are determined.

2. The wind turbine drivetrain human-machine collaborative health assessment method as described in claim 1, characterized in that, The multifunctional parallel network includes a counting self-attention module and a localization self-attention module, wherein: The counting self-attention module is used to calculate the number of faults, and the positioning self-attention module is used to calculate the fault probability distribution. The multifunctional parallel network is used to combine the number of faults calculated by the counting self-attention module with the fault probability distribution result calculated by the positioning self-attention module to obtain the fault state membership degree.

3. The wind turbine drivetrain human-machine collaborative health assessment method as described in claim 2, characterized in that, The membership degree of the fault state is calculated according to the following formula: ； ； in, p The sample number indicates that the sample represents a vibration signal. p =1,2,…, P , P The number of samples; f （p） The time and frequency domain characteristics of the vibration signal; The membership degree of the fault state; The predicted number of failures; This represents the failure probability distribution. For each fault mode category, the model parameters are: j It is the fault mode number; N The highest number among the fault mode numbers; L （p） Number and label the fault modes; To be applicable to failure modes j The localization self-attention module calculates the results of the layer before the output layer.

4. The wind turbine drivetrain human-machine collaborative health assessment method as described in claim 1, characterized in that, The Bayesian hierarchical model is specifically as follows: ； in, P (·) represents the form of the probability distribution; and yes K The BtO vector and OtW vector in the fault risk assessment set provided by the experts. It is based on K The fault single risk weight vector derived from expert evaluation information. It is based on the first k The failure modes derived from the evaluation information by the experts j Single risk weight, It is a failure mode j Aggregate risk weights; and For the first k The fault risk BtO vector and OtW vector provided by experts; K It is the total number of experts.

5. The wind turbine drivetrain human-machine collaborative health assessment method as described in claim 1, characterized in that, Based on the k Failure mode determined by the evaluation information of the experts j Single risk weight Specifically, it is calculated using the following formula: ； in, and According to the first k Single risk weights for the most severe and least severe failure modes as determined by expert opinions; N The highest number among the fault mode numbers; For the first k The most severe failure mode provided by the experts is relative to the [number]th [number]. j The severity of each failure mode For the first k The expert provided the first j The severity of each failure mode relative to the least severe failure mode.

6. The wind turbine drivetrain human-machine collaborative health assessment method as described in claim 1, characterized in that, The multinomial distribution is determined according to the following relationship: ； in, and For the first k The fault risk BtO vector and OtW vector provided by experts; It is based on the first k The failure modes derived from the evaluation information by the experts j Single risk weight; K It is the total number of experts; The Dirichlet distribution is determined according to the following relationship: ； in, It is a failure mode j Aggregate risk weights; It is a Dirichlet distribution; It is a non-negative parameter with a GAMMA distribution.

7. The wind turbine drivetrain human-machine collaborative health assessment method as described in claim 2, characterized in that, The health indicators are determined according to the following formula: ； Among them, HI is the health index; DI is the damage index; and AC is the amplification factor. Acc It refers to the diagnostic accuracy of the neural network model; It is a failure mode j Aggregate risk weights; It is the first in the membership degree of the fault state. j One element; z It is the Benefit criterion number; It is the number of faults predicted by the counting self-attention module.

Images