Reference conversion neural network-based rotating machine health index construction method under time-varying rotating speed condition

Abstract

A method for constructing a health index of a rotating machine based on a reference conversion neural network under a time-varying rotation speed condition, which belongs to the technical field of evaluating the health state of a rotating machine.SOLUTION: This includes steps S1 to S8. S1: collecting data; S2: pre-processing the collected vibration signal of the entire life cycle; S3: Determine a baseline rotation speed band; S4: Establish a performance degradation model in the baseline rotation speed band; S5: After obtaining the degradation model parameters in the baseline rotation speed band, the non-reference observation values are interpolated according to the established double exponential degradation model to obtain a reference health index, which is a reference state observation sequence that approximates the entire life cycle. S6-S7: constructing and training and optimizing the reference conversion neural network; S8: Health Indicator Construction: Input the test data to the running machinery and obtain the health indicators of the rotating machinery using the optimized reference transformation neural network.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention belongs to the technical field of health state assessment of rotating machinery, and relates to a method for constructing a health index of rotating machinery based on a norm transformation neural network under time-varying rotation speed conditions. [Background technology]

[0002] Rotating machinery is widely used in production and daily life, and gears, bearings, and other components are important components of its transmission chain. Key components, such as gears and bearings, are prone to defects such as pitting and breakage due to harsh operating environments, including varying rotational speeds, high loads, high temperatures, and corrosion. The expansion of these defects can disrupt smooth operation of equipment, increase noise and vibration amplitude, reduce equipment accuracy and reliability, and ultimately lead to equipment failure and even casualties. Therefore, quantitative assessment of the health of rotating machinery during operation is crucial. During normal operation, machinery should not be shut down frequently to monitor its health and detect the extent of deterioration. As an important bridge between machinery health monitoring and remaining life prediction, health indicators not only quantitatively reveal the operating health of equipment, but also enable specific prediction methods to estimate its remaining useful life.

[0003] However, existing health index construction methods are typically only applicable to machinery operating at a constant rotational speed. In practical engineering, production requirements often necessitate changing the operating conditions of rotating machinery, such as changing the gears in an automobile, changing the joint motion of an industrial robot, or changing the pitch of a wind turbine. As a result, the collected condition monitoring signal is disturbed by amplitude modulation caused by changes in rotational speed. This prevents existing health index construction methods from clearly and accurately revealing the equipment's health degradation process, affecting the remaining lifespan. If all condition observations could be unified to a set baseline rotational speed, interference from rotational speed changes could be eliminated. Therefore, it is important to research methods for constructing health indexes for machinery under time-varying rotational speed conditions. Summary of the Invention [Problem to be solved by the invention]

[0004] Therefore, the present invention aims to provide a method for constructing a health index of a rotating machine based on a reference transformation neural network under time-varying rotational speed conditions, and to improve the prediction accuracy of the remaining useful life of the rotating machine. [Means for solving the problem]

[0005] To achieve the above objectives, the present invention provides the following technical solutions:

[0006] A method for constructing a health index for rotating machinery based on a reference transformation neural network under time-varying rotational speed conditions, specifically The condition monitoring system monitors vibration signals X=(x1,x2,…,x N ) T where N is the number of samples over the life cycle,

number

number

[0007] Furthermore, in step S2, the data preprocessing specifically involves filtering the collected whole life cycle vibration signals to reduce noise, remove outliers, and match the rotation speed. In order to reduce the data dimension, the RMS value or other trend statistical features of the vibration signals are applied to the state observation variables F=(f1, f2, ..., f N ) T and extracting the result as

[0008] Furthermore, in step S3, the baseline rotation speed band is determined by specifically defining s∈[(1−δ)×s B ,(1+δ)×s B ], where s B represents the rated rotation speed of the equipment, and δ is the coefficient of variation. The coefficient of variation δ can be set based on engineering experience or data analysis results.

[0009] Furthermore, step S4 specifically includes selecting a double exponential model as a degradation fitting model of the state value in the baseline rotational speed band, and the formula representing the double exponential model is Y(t)=α1exp(β1t)+α2exp(β2t)+ε, where Y(t) is an approximation of the reference state observation value at time t, α1, α2, β1, and β2 are state parameters;

number

[0010] Furthermore, in step S5, after obtaining sufficient state observation variables, the reference observation sequence {f *} and estimate the parameters of the degradation model based on

number

number

[0011] Furthermore, in step S6, the constructed baseline transformation neural network includes an input layer, an output layer, and N hidden layers, where the input layer has three neurons, and the three parameters that need to be input during training and testing are the original state observation variable f, the baseline rotation speed ratio R, s , and the deterioration time ratio R t The predicted values ​​in the output layer are virtual baseline observations converted into the baseline rotation speed band, Baseline rotational speed ratio R s is the difference between the equipment's current rotation speed s(t) and the baseline rotation speed s B The deterioration time ratio R tis the current operating time of the equipment t k and its life cycle T, which includes the operating time information of the equipment, during training and testing, R t is calculated differently, namely:

number

[0012] Further, in step S7, training the optimized reference transformation neural network includes: BTNN Assuming that represents the weight and bias parameters of the reference transformation neural network, during the network training process, the network's predicted value

number

number

number

number

number

number

[0013] Furthermore, in step S8, the health index is constructed by applying test data {f}, {R s}, {R t} to θ BTNN and predict the health index values ​​converted to baseline rotation speeds in real time, and generate the whole life cycle health index vector H=[h1,h2,…,h N ] T This will reveal the performance degradation process and lay the foundation for subsequent fault warning and remaining life prediction. [Effects of the Invention]

[0014] The beneficial effects of the present invention are as follows: The present invention can evaluate the health state of a rotating machine under time-varying rotational speed conditions, improve the accuracy of predicting the remaining useful life of the rotating machine, and reduce the maintenance costs of mechanical equipment.

[0015] Other advantages, objects, and features of the present invention will be set forth in part in the specification which follows, and in part will become apparent to those skilled in the art upon examination of the following, or may be learned from the practice of the invention. The objectives and other advantages of the present invention will be realized and obtained by the following specification. [Brief explanation of the drawings]

[0016] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention is preferably described in detail below with reference to the accompanying drawings.

[0017] [Figure 1]1 is a flowchart of a method for constructing a rotating machinery health index based on a reference transformation neural network according to the present invention. [Figure 2] FIG. 1 is a schematic diagram of a reference transformation neural network structure. [Figure 3] FIG. 1 is a simplified diagram of the structure of a turbofan engine. [Figure 4] Test engine health indicator. [Figure 5] 1 is a schematic diagram of a gearbox structure of a wind turbine. [Figure 6] Comparison of different health index construction methods for two test bearings. DETAILED DESCRIPTION OF THE INVENTION

[0018] Hereinafter, embodiments of the present invention will be described through specific examples. Those skilled in the art will easily understand other advantages and effects of the present invention from the contents disclosed herein. The present invention may be implemented or applied in other different specific embodiments, and the details of the present specification may be modified or changed in various ways based on different perspectives and applications without departing from the spirit of the present invention. It should be noted that the figures provided in the following examples are only used to roughly explain the basic concept of the present invention, and the following examples and features of the examples may be combined with each other without contradiction.

[0019] The drawings are for illustrative purposes only, and are not actual drawings but are only schematic views, and should not be understood as limiting the present invention. In order to better explain the embodiments of the present invention, some parts of the drawings are omitted, enlarged, or reduced in size, and do not represent the actual size of the product. Those skilled in the art will understand that some well-known structures and their descriptions may be omitted in the drawings.

[0020] In the drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. When terms such as "upper," "lower," "left," "right," "front," and "rear" are used to indicate directions or positional relationships in the description of the present invention, they are based on the directions or positional relationships shown in the drawings. It should be understood that this is for the convenience of explaining and simplifying the description of the present invention, and does not suggest or imply that devices or elements must have a specific orientation or be constructed or operate in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are used only for illustrative purposes and should not be understood as limiting the present invention. Those skilled in the art can understand the specific meaning of the above terms based on the specific situation.

[0021] 1 to 6, the present invention provides a method for constructing a health index of a rotating machine based on a reference transformation neural network in a time-varying rotation speed. As shown in Fig. 1, the specific process is as follows: (1) Data collection: The condition monitoring system collects vibration signals X = (x1, x2, ..., x) throughout the life cycle of the machinery. N ) T where N is the number of samples over the life cycle,

number

number

number

number

[0022] Based on the above considerations, and taking advantage of the strong nonlinear approximation ability of deep neural networks, a reference transformation neural network is constructed to implicitly model the complex nonlinear mapping relationship between the original state variables and the reference observation values. As shown in Figure 2, the network structure of the reference transformation neural network is mainly composed of an input layer, an output layer, and N hidden layers. The input layer of the reference transformation neural network has three neurons, and the three parameters that need to be input during training and testing are the original state observation variable f, the baseline rotation speed ratio R, s , and the deterioration time ratio R t The predicted values ​​of the network output layer are the virtual baseline observations transformed into the baseline rotation speed band.

[0023] Baseline rotational speed ratio R sis the difference between the equipment's current rotation speed s(t) and the baseline rotation speed s B It contains information about the actual rotation speed of the equipment and the predetermined baseline rotation speed. t is the current operating time of the equipment t k and its entire life cycle T, and includes the operating time information of the equipment. During training and testing, R t is calculated differently, namely:

number

number

number

number

[0024] The above is the method proposed by the present invention for constructing a healthy finger of a rotating machine based on a reference transformation neural network under time-varying rotation speed conditions, and the effectiveness of this method will be demonstrated by the following specific experiments.

[0025] Example 1: As shown in Figure 3, NASA Ames Research Center used the C-MAPSS simulation model to simulate damage propagation paths within an aircraft turbofan engine and generate 21 sensor variables reflecting engine degradation. The C-MAPSS dataset consisted of four subsets, FD001-FD004, generated under various combinations of six operating conditions and two failure modes. Subsets FD001 and FD003 were for a single operating condition, while subset FD004 was affected by the operating rotational speed and had two failure modes, although the failure modes were outside the scope of the study. Therefore, subset FD002, which is affected by the time-varying rotational speed, was selected for validation. To improve computational efficiency, five engine unit datasets were randomly selected to form the dataset, three of which were used for training and two for testing.

[0026] The final turbofan engine health index is shown in Figure 4. Compared with the original state variables, the health index constructed based on the proposed method significantly reduced irregular fluctuations caused by changes in engine speed and exhibited a monotonically increasing trend, effectively reflecting the deterioration of engine health over time. In the early stages, the turbofan engine had a low number of operating cycles and was in good health. As the operating time increased, the health index value gradually increased, indicating that the turbofan engine's health gradually deteriorated with increasing operating time. Finally, the health index value reached its maximum value, indicating that the turbofan engine was not operating normally due to some serious fault.

[0027] Comparative Test: To fully demonstrate the superiority of this method, the health index constructed by the proposed method (BTNN-HI) was compared with the health index based on the original observed variables and the reference transformation function (BTF-HI), respectively.

[0028] In the present invention, the scale conversion coefficient and the position conversion coefficient of the reference conversion function are constructed by selecting an exponential expression and a linear expression, respectively, for constructing the reference conversion function. y(t)=γexp(s(t)-s B )y * (t)+ρ(s(t)-s B ) where γ and ρ were parameters that needed to be fitted based on known data. Conversely, after determining the specific form of the reference transformation function, the state observations at any rotation speed can be mapped to the reference state values ​​through the following inverse function:

number

number

[0029] Note that the reference observations were only observable at the baseline rotation speed, and the reference observations obtained at other rotation speeds were virtual values. For the virtual observations that could not be obtained directly, the reference observations were approximated by interpolation by establishing a degradation model at the baseline rotation speed. Therefore, the parameters γ and ρ can be estimated by minimizing the sum of squared errors between the interpolated approximate reference observations and the transformed virtual reference values.

number

[0030] Furthermore, the Levenberg-Marquardt algorithm or the trust-region reflection algorithm could be used to solve the optimized estimated parameters to obtain the desired reference transformation function.

[0031] To comprehensively evaluate the performance of health indicators, the above three health indicators were compared in terms of monotonicity (Mon), trend (Tre), robustness (Rob), and comprehensive evaluation index (CI).

[0032] To calculate these evaluation indices, the health index needs to be decomposed into a trend term and a random term. H(t k )=H T (t k )+H R (t k ) Here, H(t k ) is time t k represents the health index value at T (t k ) is its average tendency, H R (t k ) represented random fluctuations. Based on the obtained trend term and random term, the calculation formulas for monotonicity, trend, robustness, and comprehensive indexes are as follows: (1) Monotonicity:

number

number

number

number

[0033] Table 1. Performance comparison results of various health indicators between test engines E4 and E5 [Table 1]

[0034] The performance evaluation results of various health indicators for two test turbofan engines are shown in Table 1. Table 1 shows that the proposed method for constructing a machine health indicator based on a baseline transformation neural network significantly improves overall performance and is superior to the method for constructing a health indicator based on original observations and baseline transformation functions. However, the performance of BTF-HI varies depending on the test engine and is not necessarily superior to the original observation variables. Therefore, the proposed method is more suitable for constructing a health indicator for rotating machinery with time-varying rotational speeds.

[0035] Example 2: The gearbox structure of a 2MW wind turbine at a domestic wind farm is shown in Figure 5. Its transmission system primarily consists of one planetary stage and two-stage parallel gear transmission. To monitor the vibration of the low-speed shaft bearing during fan operation, an accelerometer was attached to the bearing seat housing near the low-speed shaft bearing. The accelerometer sampling frequency was 25.6 kHz, with each sampling time being 1.28 seconds. A rotational speed sensor was also attached to the main shaft to obtain input rotational speed information. Because the wind turbine gearbox bearing (WTGB) data set was obtained from a real engineering environment, the original vibration signals contained a lot of measurement noise and interference from other components. Therefore, filtering and noise reduction were performed. Preprocessing, such as removing abnormal data and matching rotational speed data, was also performed. Finally, five sets of life-cycle data sets were obtained, as shown in Table 2.

[0036] Table 2 Details of the WTGB dataset [Table 2]

[0037] Since the data in this study were actually collected vibration signals, in addition to the original observed RMS values ​​and BTF-HI, dimensionless health indicators such as skewness, kurtosis, and shape factor were also extracted as a comparison method. Figure 6 shows the health indicators of test bearings B4 and B5 constructed using different methods. Figure 6 shows that BTNN-HI, RMS, and BTF-HI can better reveal the deterioration relationship of bearing health over time compared with indicators such as skewness and kurtosis waveform coefficient. While BTF-HI can reduce the interference of time-varying rotational speed to a certain extent, it may also introduce new jumps. BTNN-HI not only largely eliminates the effects of rotational speed changes, but also provides clearer degradation paths and more consistent end-of-life failure thresholds, making it particularly important for improving the accuracy of machine health assessment and RUL prediction.

[0038] Table 3. Performance comparison results of various health indicators for bearing B4 [Table 3]

[0039] Table 4. Performance comparison results of various health indicators for bearing B5 [Table 4]

[0040] The quantitative evaluation results of the specific performance of various health indicators for the two test wind turbine bearings are listed in Tables 3 and 4, respectively. The tables show that the health indicator constructed using the proposed method exhibited the greatest improvement in monotonicity compared to the other five indicators, which is consistent with intuitive observations of the health indicator curves. Furthermore, the trend and robustness were also significantly improved, especially for bearing B4. Although skewness, kurtosis, and shape coefficient are dimensionless indicators, their performance evaluation results were relatively poor and they were unable to clearly characterize the deterioration state of wind turbine bearings. In summary, BTNN-HI exhibited better overall performance than other health indicators and was more useful for monitoring the health status of machinery and predicting remaining service life at time-varying rotational speeds.

[0041] From the above two applications, it can be seen that compared with health indices such as RMS, BTF-HI, skewness, kurtosis, given RMS, BTF-HI, bias, and shape factor, BTNN-HI can more clearly reveal the deterioration trend of rotating machinery under time-varying rotational speed, making it easier to determine a unified fault threshold and reducing experimental costs. Therefore, compared with health index construction methods based on baseline transformation functions and dimensionlessness, the baseline transformation neural network is more suitable for constructing health indices for rotating machinery under time-varying rotational speed conditions.

[0042] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the present invention, and they should be included in the claims of the present invention.

Claims

1. 1. A method for constructing a health index of a rotating machine based on a reference transformation neural network under time-varying rotational speed conditions, the method comprising: The condition monitoring system measures the vibration signal X = (x １ , x ２ , ..., x Ｎ ) Ｔ where N is the number of samples over the life cycle, [Equation 1] , M is the number of sampling points of a single sample, and at the same time, the real-time rotation speed of the equipment S = (s １ , s ２ , ..., s Ｎ ) Ｔ a data collection step S1 of collecting a data pre-processing step S2 of pre-processing the collected vibration signals throughout the life cycle; a step S3 of determining a baseline rotational speed band; a step S4 of establishing a performance degradation model in the baseline rotation speed band, in which after determining the baseline rotation speed band, an observation value is defined, and it is determined whether the observation value is a reference observation value according to the rotation speed information, and for observation values ​​outside the baseline rotation speed band, a performance degradation model is established that interpolates to approximate the reference observation; After obtaining the degradation model parameters at the baseline rotational speed band, the non-reference observations are interpolated based on the established double exponential degradation model to approximate the reference condition observation sequence throughout the life cycle. [Equation 2] a non-reference observation interpolation step S5 to obtain a reference health index, Step S6 of constructing a reference transformation neural network that transforms the original state observation variables into virtual reference observations at the baseline rotational speed; a step S7 of training and optimizing the reference transformation neural network; a health index construction step S8 for inputting test data for a machine in operation and obtaining a health index for the rotating machine using an optimized reference transformation neural network; A method for constructing a health index for a rotating machine, comprising:

2. In step S2, the data preprocessing specifically includes filtering the collected whole life cycle vibration signals to reduce noise, removing outliers, and matching the rotation speed. In order to reduce the data dimension, the RMS value or other trend statistical features of the vibration signals are used as the state observation variable F=(f １ , f ２ , ..., f Ｎ ) Ｔ and extracting the health index of a rotating machine as the health index of the rotating machine.

3. In step S3, the baseline rotation speed band is determined by specifically setting s∈[(1−δ)×s Ｂ , (1 + δ) × s Ｂ ], where s Ｂ 2. The method for constructing a health index for a rotating machine according to claim 1, wherein δ represents the rated rotational speed of the equipment, and δ is a coefficient of variation.

4. Specifically, step S4 includes selecting a double exponential model as a deterioration model of the state value in the baseline rotational speed band; The formula for the double exponential model is Y(t) = α １ exp(β １ t) + α ２ exp(β ２ t) + ε where Y(t) is the approximation of the reference state observation at time t, and α １ , α ２ , β １ , and β ２ is a state parameter, [Equation 3] 3. The method for constructing a health index for a rotating machine according to claim 2, wherein: is a Gaussian noise term.

5. In step S5, after acquiring sufficient state observation variables, the reference observation sequence {f ＊ }, and estimating parameters of the degradation model based on [Equation 4] where f ｉ ＊ represents the state observation obtained in the baseline rotational speed band at time ti, [Equation 5] 5. The method of claim 4, wherein: ##EQU1## represents an estimate of a state parameter.

6. In step S6, the constructed reference transformation neural network includes an input layer, an output layer, and N hidden layers, where the input layer has three neurons, and the three parameters that need to be input during training and testing are the original state observation variable f, the baseline rotation speed ratio R ｓ , and the deterioration time ratio R ｔ The predicted values ​​in the output layer are virtual baseline observations converted into the baseline rotation speed band, Baseline rotational speed ratio R ｓ is the difference between the equipment's current rotation speed s(t) and the baseline rotation speed s Ｂ and includes information about the actual rotation speed of the equipment and the predetermined baseline rotation speed, and the deterioration time ratio R ｔ is the current operating time of the equipment t ｋ and its life cycle T, which includes the operation time information of the equipment, during training and testing, R ｔ is calculated differently as follows: [Equation 6] where n is the number of samples of the same type of equipment whose entire life cycle is known, and T ｊ The method for constructing a health index for a rotating machine according to claim 1, characterized in that: j is the entire life cycle of the j-th sample of the same type of equipment.

7. In step S7, training the optimized reference transformation neural network includes: θ ＢＴＮＮ Assuming that represents the weight and bias parameters of the reference transformation neural network, during the network training process, the network's predicted values [Equation 7] and the reference state observation value f approximated over the entire life cycle Ｂ ＊ Use the mean squared error between as the optimization target, [Equation 8] Here, n ｂａｔｃｈ is the batch size of the input data during training, [Equation 9] is the objective function, and the training strategy uses a mini-batch stochastic gradient descent algorithm to optimize the network parameters, and a backpropagation algorithm to calculate the optimized parameters by minimizing the objective function. [Equation 10] where lr is the network learning rate and θ ＢＴＮＮ are the network parameters, [0011] are the optimized network parameters, The method for constructing a health index for a rotating machine according to claim 6, characterized in that the training process is repeated to obtain an optimized reference transformation neural network.

8. In step S8, the health index is constructed by applying test data {f}, {R ｓ }, {R ｔ } to θ ＢＴＮＮ and predict the health index value converted to the baseline rotation speed in real time, and generate the whole life cycle health index vector H = [h １ , h ２ , ..., h Ｎ ] Ｔ The method for constructing a health index for a rotating machine according to claim 7, further comprising obtaining:

Images