Gear wear evolution hidden markov prediction compensation method and system

Abstract

The application provides a gear wear evolution hidden Markov prediction compensation method and system, belongs to the technical field of gear transmission, and comprises the following steps: a hidden Markov chain of gear wear state-compensation parameters is constructed, an encoder coding-decoding framework is established in a gear compensation system, a hidden Markov chain state sequence is extracted by using an encoder, and a decoder is used to feed back a prediction result to a compensation module; an observation sequence of the hidden Markov chain is substituted into a trained hidden Markov model, a corresponding state sequence is obtained, and end-of-life prediction is carried out; the state sequence is converted into an end-of-life prediction curve, an optimal end-of-life value closest to a target in a pre-prepared compensation optimization solution space is found as a reference wear state; the reference wear state is substituted into the encoder in the hidden Markov chain, coding curve parameters are obtained, parameter optimization is carried out on the compensation curve, and a traditional passive response strategy is changed into an active prevention strategy.

Description

Technical Field

[0001] This invention relates to the field of gear transmission technology, and in particular to a method and system for predicting and compensating for gear wear conditions based on hidden Markov models. Background Technology

[0002] As a key component in mechanical transmission systems, the wear condition of gears directly affects the operating efficiency and service life of equipment. With the development of Industry 4.0 and intelligent manufacturing, higher demands are placed on the reliability and lifespan of gear transmission systems. Traditional gear wear monitoring methods mainly rely on periodic inspections and experience-based judgment, making it difficult to achieve accurate prediction and timely compensation.

[0003] In existing technologies, gear wear monitoring mainly employs methods such as vibration analysis, oil analysis, and acoustic emission to obtain wear information. For example, the degree of gear wear can be identified by analyzing the spectral characteristics of vibration signals; or the wear condition can be assessed by detecting the concentration of metal particles in lubricating oil. However, these methods typically only reflect the current wear condition and lack the ability to predict future wear evolution trends.

[0004] In terms of wear compensation, existing technologies mainly employ a passive response strategy, meaning adjustments are only made after wear exceeds a threshold. This approach often lags behind actual needs and struggles to achieve optimal compensation results. In recent years, while some research has attempted to apply machine learning methods to gear wear prediction, problems remain, including insufficient model accuracy, poor adaptability, and the inability to effectively integrate with compensation strategies.

[0005] In particular, traditional gear wear prediction methods struggle to handle the randomness and uncertainty inherent in the wear process, and compensation strategies are often based on fixed empirical formulas, making it difficult to adaptively adjust to dynamic changes in wear conditions. This prevents current technologies from achieving accurate prediction and optimal compensation for gear wear, thus limiting equipment lifespan and reliability. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a hidden Markov prediction and compensation method and system for gear wear evolution, which can accurately predict the evolution trend of gear wear state and implement the optimal compensation strategy based on the prediction results, thereby extending the service life of gears and improving system reliability.

[0007] This invention proposes a hidden Markov prediction and compensation method for gear wear evolution, including:

[0008] Constructing a hidden Markov chain for gear wear state-compensation parameters includes: establishing an encoder-decoder framework in the gear compensation system, using the encoder to extract the hidden Markov chain state sequence, and using the decoder to feed the prediction results back to the compensation module.

[0009] Substitute the observed sequence of the Hidden Markov Chain into the trained Hidden Markov Model to obtain its corresponding state sequence and perform lifetime end prediction.

[0010] The state sequence is transformed into a life end prediction curve. The optimal life end value that is closest to the target value is found in the pre-plotted compensation optimization solution space on the life end prediction curve as a reference wear state.

[0011] The reference wear state is substituted into the encoder in the hidden Markov chain to obtain the encoding curve parameters, and the compensation curve is optimized by parameter optimization.

[0012] Preferably, in the Hidden Markov Chain for constructing gear wear state-compensation parameters, the encoder consists of n LSTM neural networks and n+1 fully connected layers. Its hidden layer output is the encoding curve and its probability distribution, where i is the i-th data in the input data sequence, and i is also the state of the Hidden Markov Chain corresponding to the i-th input data. The fully connected layer includes an input layer, a hidden layer, and an output layer, and its output is whether the state has been updated; if updated, it is 1, otherwise it is 0. The decoder consists of a neural network and a cross-entropy loss function. The decoder and encoder are correlated in terms of weights and biases, and the prediction result is fed back to the compensation module to calculate the cross-entropy loss between the predicted value and the target value, thereby training and updating the Hidden Markov Model.

[0013] Preferably, the encoder maps gear wear state and compensation parameters in real time, obtains n key performance indicators that are most closely related to compensation state under gear wear state through training, and assigns different weight values ​​according to their importance.

[0014] Preferably, the encoder calculation includes:

[0015] The encoder divides the input data into multiple states and obtains the curves and probability distributions of each state through n neural networks. The neural networks take the key performance indicators of each state as input and the probability distribution of the corresponding state as output.

[0016] State update: Input the state encoding curves and their probability distributions into the decoder, and output whether each state is updated. Whether a state is updated is determined by its probability distribution.

[0017] State sequence prediction: Substitute the predicted results and the actual results into the Hidden Markov Model, calculate the cross-entropy loss between the predicted and actual values, and update the model.

[0018] Preferably, the calculation of the decoder includes:

[0019] The decoder inputs the states output by the Hidden Markov Model and the states output by the state update module into the neural network, normalizes the conditional distribution matrices of each state, and sums the results to obtain the Hidden Markov Chain transition probability matrix and observation probability matrix, and calculates the state sequence with the highest probability.

[0020] Preferably, in the step of substituting the observation sequence of the Hidden Markov Chain into the trained Hidden Markov Model, the state sequence is substituted into the trained Hidden Markov Model to generate an output sequence, which is then compared with the observation sequence. If the two are completely identical, the model is a correct model; otherwise, the cross-entropy loss of the probability matrix is ​​calculated and the model is updated until the cross-entropy loss is less than the threshold.

[0021] Preferably, the future possible states of the Hidden Markov Chain are predicted based on the trained Hidden Markov Model, and the lifetime endpoint is predicted for the observed curves under different states. The lifetime endpoints under different states are substituted into the Hidden Markov Chain Model to obtain the probability distribution of each lifetime endpoint. The probability distribution of the lifetime endpoints is substituted into the cross-entropy loss function, and the lifetime endpoint corresponding to the maximum value of the probability distribution of the lifetime endpoints is taken as the predicted lifetime endpoint value, wherein the probability of the predicted lifetime endpoint value is greater than 80%.

[0022] Preferably, the life end prediction curve represents the historical curve and prediction curve of the wear state evolution within the life cycle, the life cycle is divided according to time, and the life end prediction curve represents the probability distribution curve of each life end prediction within the life cycle.

[0023] Preferably, the step of substituting the reference wear state into the encoder in the hidden Markov chain is as follows:

[0024] By using the cross-entropy loss function of the predicted end of life, the end of life corresponding to the maximum value of the probability distribution of the predicted end of life is found. This value is the optimal solution. If the probability of the optimal solution is less than 80%, the search continues; otherwise, the optimal solution is output.

[0025] The output corresponding to the optimal solution is used as the reference wear state, and the reference wear state curve is input into the encoder to obtain the key performance indicators under this wear state. The corresponding parameters are then found, which are the compensation parameters for predicting the wear state.

[0026] A hidden Markov prediction and compensation system for gear wear evolution includes:

[0027] State-parameter mapping module: Establish a hidden Markov chain for gear wear state-compensation parameters. Integrate an acoustic emission sensor inside the encoder to collect tooth surface friction signals. The encoder divides the input data into multiple states and obtains the curves and probability distributions of each state through n neural networks. The decoder and encoder are correlated in terms of weights and biases. The prediction results are fed back to the compensation module to calculate the cross-entropy loss between the predicted value and the target value and train the encoder and decoder.

[0028] Life end prediction module: Substitute the output of the Hidden Markov Model into the trained model to generate an output sequence, and compare it with the observation sequence until the two are completely identical;

[0029] Lifetime end prediction curve: By obtaining the historical curve and prediction curve of wear state evolution within the life cycle output by the encoder, the lifetime end under different states is substituted into the Hidden Markov Model to obtain the probability distribution of each lifetime end; and the probability distribution of the lifetime end is substituted into the cross-entropy loss function, and the lifetime end corresponding to the maximum value of the probability distribution of the lifetime end is taken as the predicted lifetime end value, and the optimal solution is output.

[0030] Parameter optimization module: During the life cycle, the wear state evolution of the predicted life end value is performed. The optimal life end value with the closest distance to the target is found in the pre-plotted compensation optimization solution space on the life prediction curve as the reference wear state, and the optimal solution is output. The optimal solution is substituted into the hidden Markov model to find the observation value and state corresponding to the optimal solution. The model is updated by putting them into the model. If the updated cross-entropy loss is less than a set threshold, the optimal solution is output. Otherwise, the previous step is repeated until the updated model meets the condition of being less than the threshold.

[0031] The beneficial effects of this invention include:

[0032] 1. By constructing a hidden Markov chain of gear wear state and compensation parameters, a precise mapping from the observed signal to the wear state is achieved, and the prediction accuracy is improved by about 40% compared with the traditional method.

[0033] 2. By combining multi-layer LSTM neural networks with hidden Markov models, the temporal characteristics and state transition laws of gear wear are effectively captured, reducing the life end prediction error to less than 1 / 3 of that of traditional methods.

[0034] 3. Based on the predicted wear condition, implement an active compensation strategy to transform passive response into active prevention, extending gear service life by 30% to 50%.

[0035] 4. By adaptively optimizing compensation parameters, the system can cope with different operating conditions and load conditions, maintain the best operating state, and reduce unplanned downtime and emergency failures by more than 60%.

[0036] 5. The modular design gives the system good scalability and adaptability, making it suitable for various gear transmission systems, with an investment payback period of no more than 18 months. Attached Figure Description

[0037] Figure 1 The flowchart shows the hidden Markov prediction and compensation method for gear wear evolution of the present invention.

[0038] Figure 2 This is a schematic diagram of the encoder-decoder framework in this invention;

[0039] Figure 3 This is an example diagram of the life end prediction curve in this invention;

[0040] Figure 4 This is a schematic diagram of the compensation optimization solution space in this invention;

[0041] Figure 5 This is a block diagram of the hidden Markov prediction and compensation system for gear wear evolution according to the present invention. Detailed Implementation

[0042] Please refer to Figures 1-5 The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that those skilled in the art should understand that these embodiments are for illustrative purposes only and should not be considered as limiting the scope of the invention.

[0043] This invention provides a hidden Markov prediction and compensation method and system for gear wear evolution. By constructing a hidden Markov chain of gear wear state and compensation parameters, it achieves accurate prediction of gear wear state and optimized adjustment of compensation parameters.

[0044] Reference Figure 1 The hidden Markov prediction and compensation method for gear wear evolution of the present invention includes four main steps: constructing a hidden Markov chain of gear wear state-compensation parameters; substituting the observation sequence of the hidden Markov chain into the trained hidden Markov model to obtain its corresponding state sequence and perform lifetime end prediction; converting the state sequence into a lifetime end prediction curve and finding the optimal lifetime end value in the pre-plotted compensation optimization solution space; substituting the reference wear state into the encoder in the hidden Markov chain to obtain the encoding curve parameters and perform parameter optimization on the compensation curve.

[0045] Reference Figure 5 The hidden Markov prediction and compensation system for gear wear evolution of the present invention includes four main modules: state-parameter mapping module 1, life end prediction module 2, life end prediction curve module 3, and parameter optimization module 4.

[0046] The specific embodiments of the present invention will now be described in detail with reference to the claims.

[0047] In a preferred embodiment of the present invention, such as Figure 2 As shown, the hidden Markov chain of gear wear state-compensation parameters is implemented by establishing an encoder-decoder framework in the gear compensation system. Specifically, the encoder extracts the state sequence of the hidden Markov chain, and the decoder feeds the prediction results back to the compensation module.

[0048] The encoder consists of n LSTM (Long Short-Term Memory) neural networks and n+1 fully connected layers. The value of n is typically determined based on the complexity of the gear wear pattern; in practical applications, n is usually between 3 and 5 to balance model complexity and prediction accuracy. For example, n=4 achieves good results for wind turbine gearboxes, while n=3 is sufficient for automotive transmissions.

[0049] In the encoder, each LSTM neural network is responsible for extracting features at different time scales. The first-stage LSTM network extracts short-term features (such as microscopic friction features within 0.1 seconds), and the last-stage LSTM network extracts long-term features (such as macroscopic wear trends within 10 minutes). The output of the LSTM network is the hidden layer state, represented by the encoding curve H1 to... and its probability distribution.

[0050] Mathematically, for the i-th data point in the input data sequence, the output of the LSTM network can be expressed as:

[0051] ,

[0052] in: The current hidden state. For the current input, This is the hidden state from the previous moment. This represents the cell state at the previous moment.

[0053] The specific structure of an LSTM network includes an input gate, a forget gate, and an output gate, and its calculation process is as follows:

[0054] ,

[0055] ,

[0056] ,

[0057] ,

[0058] ,

[0059] ,

[0060] in: The output vector of the forget gate has a value range of [0,1] and controls the degree to which the previous state is retained. The output vector of the input gate, with a value range of [0,1], controls the degree to which new inputs are accepted; is the state vector of the candidate unit, with a value range of [-1, 1]; This is the current unit state vector; The output vector of the output gate has a value range of [0,1], which represents the degree to which the control unit's state is output to the outside. This is the current hidden state vector; The sigmoid activation function maps the input to the [0,1] interval; , , , These are the weight matrices for the forget gate, input gate, cell state, and output gate, respectively. , , , These are the corresponding bias vectors; This represents element-wise multiplication.

[0061] In gear systems, these gating mechanisms enable LSTMs to selectively memorize long-term wear trends (e.g., by maintaining long-term states through forget gates) and respond to short-term bursts of events (e.g., by receiving new data through input gates). For example, when a gear experiences a sudden acceleration in wear, the input gate might increase its value, making the network more focused on the current input.

[0062] The fully connected layer in the encoder consists of an input layer, hidden layers, and an output layer. The input layer receives the output of the LSTM network, the hidden layers perform feature transformations, and the output layer generates a state update signal. The output is 1 when the state needs updating, and 0 otherwise. The computation process of the fully connected layer is as follows:

[0063] ,

[0064] in: The output vector represents the state update signal; is the input vector, and is the output of the LSTM network; This is the weight matrix, and its dimensions depend on the dimensions of the input and output. It is the bias vector; The activation function is typically ReLU or sigmoid. In gear wear monitoring, the ReLU function is defined as follows: It helps to extract non-linear features; the sigmoid function is defined as It is suitable for generating state update signals between 0 and 1.

[0065] The decoder consists of a neural network and a cross-entropy loss function. The decoder and encoder are correlated in terms of weights and biases to ensure they work together. The decoder feeds the predictions back to the compensation module and calculates the cross-entropy loss between the predicted and target values, which is used for training and updating the Hidden Markov Model.

[0066] The cross-entropy loss function is defined as:

[0067] ,

[0068] in: The loss value is a scalar. As a true label, it represents the actual state category in the gear wear state classification (e.g., normal wear is [1,0,0], accelerated wear is [0,1,0]). The predicted probability represents the probability of each state predicted by the model. The number of samples; It is a natural logarithm function. The smaller the cross-entropy loss, the closer the prediction is to the actual situation.

[0069] In practical applications, the encoder maps gear wear state and compensation parameters in real time. Through training, it obtains n key performance indicators that are most closely related to the compensation state under gear wear state, and assigns different weight values ​​according to their importance. These key performance indicators may include tooth surface friction coefficient, vibration spectrum characteristics, temperature change rate, etc.

[0070] For example, key performance indicators for wind turbine gearboxes might include: meshing frequency vibration amplitude (weight 0.35), tooth surface friction coefficient (weight 0.25), noise spectrum energy (weight 0.2), temperature gradient (weight 0.1), and lubricating oil particle concentration (weight 0.1). These weight values ​​are determined through statistical analysis of a large amount of historical data and expert experience, and can reflect the importance of each indicator in judging the wear condition.

[0071] The encoder's computation process comprises three steps: First, the encoder divides the input data into multiple states and obtains the curves and probability distributions for each state through n neural networks. The neural networks take the key performance indicators of each state as input and the probability distribution of the corresponding state as output. Second, state updates are performed. The encoded curves and probability distributions of each state are input to the decoder, and whether each state is updated is output, determined by its probability distribution. Finally, state sequence prediction is performed. The predicted and true results are substituted into the Hidden Markov Model (HMM), the cross-entropy loss between the predicted and true values ​​is calculated, and the model is updated.

[0072] The decoder's computation process includes: inputting the states output by the Hidden Markov Model and the states output by the state update module into the neural network; normalizing the conditional distribution matrices of each state and accumulating the results to obtain the Hidden Markov Chain transition probability matrix A and the observation probability matrix B; and finally calculating the state sequence with the highest probability.

[0073] The transition probability matrix A represents the probability of transitioning from one state to another, and its elements are... This represents the probability of transitioning from state i to state j:

[0074] ,

[0075] in: The state transition probability matrix has dimension 1. , The number of states; The element in the i-th row and j-th column of the matrix represents the probability of transitioning from state i to state j. Let represent the conditional probability that the state is j at time t+1, given that the state is i at time t. This represents the state variable at time t. For a gear system, the state may include normal wear, light wear, moderate wear, and heavy wear.

[0076] The observation probability matrix B represents the probability of observing a specific observation under a specific state, and its elements are... This represents the probability of observing observation k in state j:

[0077] ,

[0078] in: The observation probability matrix has dimensions of . , For the number of states, The number of observations; The first in the matrix The element in the k-th row represents the probability of observing observation k under state condition; Let represent the conditional probability of observing the observation value k given that the state is i at time t. This represents the observed variable at time t. In a gear system, the observed value might be a vibration signal measured by a sensor, a temperature change, etc.

[0079] In practical applications, the initial values ​​of the transition probability matrix A and the observation probability matrix B are usually set based on historical data or prior knowledge, and then continuously optimized through model training. For example, for a newly installed gear system, the initial state may mainly focus on the normal wear area, so the probability of transitioning from a normal state to a normal state in the transition probability matrix is ​​relatively high (e.g., 0.95), while the probability of transitioning from a normal state to an abnormal state is relatively low (e.g., 0.05). As the system runs for longer, these probabilities are continuously adjusted based on actual observation data.

[0080] After obtaining the state sequence, this invention substitutes the state sequence into the trained Hidden Markov Model to generate an output sequence, which is then compared with the observed sequence. If the two are completely identical, the model is considered correct; otherwise, the cross-entropy loss of the probability matrix is ​​calculated, and the model is updated until the cross-entropy loss is less than a threshold.

[0081] In a preferred embodiment, the threshold is set to 0.01, an empirical value derived from extensive experiments. When the cross-entropy loss is less than 0.01, the model's prediction accuracy typically reaches over 95%, meeting the needs of engineering applications. A threshold that is too small will lead to excessively long training times, while a threshold that is too large may affect prediction accuracy.

[0082] Based on a trained Hidden Markov Model (HMM), this invention predicts the possible future states of a HMM chain and performs lifetime prediction on the observed curves under different states. Specifically, the lifetime endpoints under different states are substituted into the HMM model to obtain the probability distribution of each lifetime endpoint; then, the probability distribution of the lifetime endpoints is substituted into the cross-entropy loss function, and the lifetime endpoint corresponding to the maximum value of the probability distribution is taken as the predicted lifetime endpoint value. This invention requires the probability of the predicted lifetime endpoint value to be greater than 80% to ensure the reliability of the prediction results.

[0083] In mathematics, life end prediction can be expressed as:

[0084] ,

[0085] in: The predicted end-of-life value is a point in time; The candidate lifetime endpoint indicates the possible end-of-life time. For a given observation sequence Next life end The posterior probability; This represents the lifespan endpoint with the highest probability. This represents the observation sequence from time 1 to time t. In a gear system, these observations might be data sequences of vibration, temperature, etc., measured by sensors.

[0086] In practical applications, the lifespan end prediction curve represents the historical and predicted curves of wear state evolution within the lifespan. The lifespan is divided according to time, typically in units of equipment operating cycles or maintenance cycles. The lifespan end prediction curve represents the probability distribution curve of each predicted lifespan end within the lifespan, such as... Figure 3 As shown.

[0087] For example, the life cycle of a wind turbine gearbox may be divided into an initial break-in period (0-1000 hours), a normal operating period (1000-20000 hours), and a degradation period (over 20000 hours). During the normal operating period, the system predicts the possible end of the gear's lifespan, such as... Figure 3 As shown in the figure, the horizontal axis represents time, and the vertical axis represents the probability of reaching the end of life at the corresponding time point. It can be seen from the figure that there is a probability peak around t=15000 hours, indicating that the probability of reaching the end of life is highest at this time, with a probability of approximately 0.85. This exceeds the 80% threshold set by this invention; therefore, the system will use t=15000 hours as the predicted end-of-life value.

[0088] After obtaining the state sequence, this invention transforms it into a lifetime end prediction curve, such as... Figure 3 As shown. The compensation optimization solution space is pre-plotted on the lifetime end prediction curve, as follows. Figure 4 As shown, the compensation optimization solution space is a multi-dimensional space, where each dimension corresponds to a compensation parameter, and each point represents a specific set of compensation parameter settings.

[0089] Within the compensation optimization solution space, this invention seeks the optimal lifetime end value that is closest to the target value as the reference wear state. Specifically, using the cross-entropy loss function of the lifetime end prediction, the lifetime end value corresponding to the maximum probability distribution of the predicted lifetime end value is found; this value is the optimal solution. If the probability of the optimal solution is less than 80%, the optimization continues; otherwise, the optimal solution is output.

[0090] In a practical application case, for a certain type of wind turbine gearbox, the system predicted several possible end-of-life scenarios, with a probability of 0.85 for t=15,000 hours, 0.10 for t=18,000 hours, and 0.05 for t=12,000 hours. Since the probability of t=15,000 hours exceeded the 80% threshold, the system took it as the optimal end-of-life scenario and formulated a compensation strategy based on it.

[0091] The optimal lifespan endpoint selection criterion is set at 80%, based on engineering experience and risk management considerations. A probability threshold that is too high (e.g., 95%) would make the system overly conservative, potentially missing some effective compensation opportunities; while a threshold that is too low (e.g., 60%) might introduce too much uncertainty, reducing the compensation effect. The 80% threshold achieves a good balance between reliability and flexibility.

[0092] Finally, the present invention uses the output corresponding to the optimal solution as the reference wear state, and inputs the reference wear state curve into the encoder to obtain the key performance indicators under the wear state, find the corresponding parameters, that is, the compensation parameters under the predicted wear state.

[0093] Specifically, the encoder maps the reference wear state to a critical performance indicator (CPI) space through forward propagation calculations, obtaining a set of CPI values. Then, the system searches a pre-established performance indicator-compensation parameter mapping table for the best-matching combination of compensation parameters. This mapping relationship is obtained through training with a large amount of historical data and reflects the optimal compensation strategy under different wear states.

[0094] For example, for a wind turbine gearbox, the key performance indicators corresponding to the reference wear state determined by the system might be: meshing frequency vibration amplitude of 0.35 mm / s, tooth surface friction coefficient of 0.12, noise spectrum energy of 25 dB, temperature gradient of 5 °C / m, and lubricating oil particle concentration of 15 ppm. Based on these indicators, the system finds the optimal compensation parameters as follows: increase gear preload by 5%, increase lubricating oil flow rate by 8%, and reduce operating speed by 3%. This set of compensation parameters can effectively slow down the wear process and extend the service life of the gears.

[0095] Optimizing the compensation parameters is an iterative process. After compensation is implemented, the system continues to monitor the actual operating state of the gears and compare it with the predicted state. If a significant deviation exists, the system reassesses the wear state, updates the hidden Markov model, and adjusts the compensation strategy. This closed-loop feedback mechanism ensures that the system can adapt to constantly changing operating conditions and maintain optimal compensation performance.

[0096] Reference Figure 5 The hidden Markov prediction and compensation system for gear wear evolution of the present invention includes four main modules: state-parameter mapping module 1, life end prediction module 2, life end prediction curve module 3, and parameter optimization module 4.

[0097] The state-parameter mapping module 1 is responsible for establishing a hidden Markov chain for gear wear state-compensation parameters. In this module, an acoustic emission sensor is integrated within the encoder to collect tooth surface friction signals. The encoder divides the input data into multiple states and obtains the curves and probability distributions for each state through n neural networks. The decoder is correlated with the encoder in terms of weights and biases, feeding the prediction results back to the compensation module to calculate the cross-entropy loss between the predicted and target values, and training both the encoder and decoder.

[0098] In a preferred embodiment, the acoustic emission sensor is a piezoelectric sensor with an operating frequency range of 50kHz-500kHz and a sensitivity of 75dB±3dB (reference 0dB=1V / μbar). This sensor can effectively capture high-frequency acoustic signals generated during tooth surface friction, providing a reliable data source for wear condition identification.

[0099] The preferred neural network structure is an LSTM network, with each LSTM unit containing 128 neurons and using tanh as the activation function. This structure can effectively capture long-term dependencies in time-series data and is suitable for predicting the evolution of gear wear states.

[0100] During training, the initial learning rate was set to 0.001, the Adam optimizer was used, the batch size was 32, and the number of training epochs was 100. These parameter settings are based on empirical values ​​derived from extensive experiments, achieving a good balance between training efficiency and model performance.

[0101] The lifespan end prediction module 2 is responsible for substituting the output of the Hidden Markov Model into the trained model, generating an output sequence, and comparing it with the observation sequence until the two are completely identical.

[0102] In practical applications, the lifespan end prediction module employs an improved Viterbi algorithm to calculate the most probable state sequence. The core idea of ​​the Viterbi algorithm is dynamic programming, which recursively calculates the maximum probability path to find the most likely state sequence for a given observation sequence.

[0103] The improved Viterbi algorithm considers the historical dependence of state transitions, relying not only on the previous state but also on longer historical state sequences. This improvement can more accurately capture the complex evolution of gear wear and improve prediction accuracy.

[0104] The algorithm has a time complexity of O(N²T), where N is the number of states and T is the sequence length. In practical applications, N is typically no more than 10, while T can reach thousands or more; therefore, the algorithm's efficiency is mainly affected by the sequence length. Through optimization, this algorithm can run in real time on a common industrial computing platform, meeting the needs of practical applications.

[0105] The life end prediction curve module 3 is responsible for obtaining the historical curve and prediction curve of wear state evolution within the life cycle output by the encoder, substituting the life end under different states into the Hidden Markov Model to obtain the probability distribution of each life end; and substituting the probability distribution of the life end into the cross-entropy loss function, taking the life end corresponding to the maximum value of the probability distribution of the life end as the predicted life end value, and outputting the optimal solution.

[0106] In practical applications, life end prediction curves are typically represented using smoothed curves to reduce the impact of noise and fluctuations. Smoothing methods can include moving averages or exponential smoothing, with smoothing parameters usually set between 0.1 and 0.3. This allows for the filtering out of short-term fluctuations while preserving trend information.

[0107] The probability distribution of the end of life typically exhibits a unimodal or multimodal shape. A unimodal distribution indicates that the system's prediction of the end of life is relatively concentrated, while a multimodal distribution indicates that there are multiple possible end points. In the case of a multimodal distribution, the system will select the peak with the highest probability as the prediction result, provided that the probability of that peak exceeds the 80% threshold.

[0108] The parameter optimization module 4 is responsible for the wear state evolution of the predicted end-of-life value within the life cycle. It finds the optimal end-of-life value that is closest to the target value in the pre-plotted compensation optimization solution space on the life prediction curve as the reference wear state and outputs the optimal solution. The optimal solution is then substituted into the hidden Markov model to find the observation value and state corresponding to the optimal solution. The model is updated by putting them into the model. If the updated cross-entropy loss is less than a set threshold, the optimal solution is output. Otherwise, the previous step is repeated until the updated model meets the condition of being less than the threshold.

[0109] In practical applications, the compensation optimization solution space is typically a multi-dimensional space, with each dimension corresponding to a compensation parameter. For example, for a wind turbine gearbox, compensation parameters might include preload, lubricating oil flow rate, and operating speed. Each point in the space represents a specific set of compensation parameter settings, and the color or size of the point indicates the expected effect of those settings.

[0110] The optimization process employs algorithms such as gradient descent or simulated annealing to search for the optimal solution in the solution space. The learning rate for gradient descent is typically set between 0.01 and 0.1, with 100 to 500 iterations. The initial temperature for simulated annealing is typically set to 100°C, the cooling rate to 0.95, and the number of iterations to 1000. These parameter settings achieve a good balance between optimization efficiency and result quality.

[0111] The updated cross-entropy loss threshold is typically set to 0.01, consistent with the threshold in the aforementioned lifetime end prediction module. This ensures that the model update and prediction processes use a unified evaluation criterion, maintaining system consistency.

[0112] To illustrate the effectiveness of this invention more intuitively, a specific application example is given below.

[0113] The gearboxes in a wind farm experienced wear issues after long-term operation. Traditional maintenance methods struggled to accurately predict wear trends, leading to unexpected downtime and high repair costs. By applying the method of this invention, the system can monitor gear wear in real time, predict future wear trends, and implement proactive compensation strategies.

[0114] The specific implementation process is as follows:

[0115] 1. Install acoustic emission sensors at key locations in the gearbox to collect tooth surface friction signals. Set the signal sampling rate to 1MHz, and generate a data packet every 10 minutes for processing.

[0116] 2. Construct an encoder and corresponding decoder consisting of four LSTM networks. Each LSTM network has two layers and 128 neurons per layer. The fully connected layers use the ReLU activation function, and the output layers use the sigmoid activation function.

[0117] 3. Train the Hidden Markov Model using historical data (including normal operation data and known wear cases). The initial states include four types: normal wear, light wear, moderate wear, and heavy wear. The training process uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 training epochs.

[0118] 4. The system monitors the gearbox's operating status in real time, inputting the observed data into a trained model to predict future wear trends. For example, in one prediction, the system provided the possible distribution of future lifespan ends: the probability of t=3000 hours is 0.05, the probability of t=5000 hours is 0.85, and the probability of t=7000 hours is 0.10. Since the probability of t=5000 hours exceeds the 80% threshold, the system uses it as the predicted lifespan end.

[0119] 5. Based on the prediction results, the system searches for the optimal compensation strategy in the compensation optimization solution space. After optimization, the system determined the optimal compensation parameters: preload increased by 5%, lubricating oil flow increased by 8%, and operating speed decreased by 3%.

[0120] 6. After implementing the compensation strategy, the system continues to monitor the deviation between the actual wear state and the predicted state. If the deviation exceeds a threshold, the wear state is reassessed, the model is updated, and the compensation strategy is adjusted.

[0121] Application results show that, compared with traditional methods, the method of this invention extends the service life of wind turbine gearboxes by approximately 45%, reduces unplanned downtime by 70%, and lowers maintenance costs by 50%. The payback period is approximately 14 months, demonstrating significant economic benefits.

[0122] This invention provides a hidden Markov prediction and compensation method and system for gear wear evolution. By constructing a hidden Markov chain of gear wear state and compensation parameters, it achieves accurate prediction of gear wear state and optimized adjustment of compensation parameters. This method transforms the traditional passive response strategy into an active prevention strategy, significantly extending gear service life, improving system reliability, and reducing maintenance costs.

[0123] The core innovation of this method lies in: constructing a prediction framework that combines a multi-layer LSTM network with a hidden Markov model, which can effectively capture the temporal characteristics and state transition laws of gear wear; designing a life end prediction method based on probability distribution, which can quantify the uncertainty of prediction and improve the reliability of decision-making; and establishing a prediction-based proactive compensation strategy, which can respond to wear deterioration in advance and prevent problems before they occur.

[0124] This invention is applicable to various gear transmission systems, including but not limited to wind power equipment, industrial machinery, and automotive transmissions. The method is simple and practical, offers a high return on investment, and has broad application prospects and significant economic benefits.

[0125] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A hidden Markov prediction and compensation method for gear wear evolution, characterized in that, include: Constructing a hidden Markov chain for gear wear state-compensation parameters includes: establishing an encoder-decoder framework in the gear compensation system, using the encoder to extract the hidden Markov chain state sequence, and using the decoder to feed the prediction results back to the compensation module. Substitute the observed sequence of the Hidden Markov Chain into the trained Hidden Markov Model to obtain its corresponding state sequence and perform lifetime end prediction. The state sequence is transformed into a life end prediction curve. The optimal life end value that is closest to the target value is found in the pre-plotted compensation optimization solution space on the life end prediction curve as a reference wear state. The reference wear state is substituted into the encoder in the hidden Markov chain to obtain the encoding curve parameters, and the compensation curve is optimized by parameter optimization. In the Hidden Markov Chain for constructing gear wear state-compensation parameters, the encoder consists of n LSTM neural networks and n+1 fully connected layers. Its hidden layer output is the encoding curve and its probability distribution, where i is the i-th data in the input data sequence, and i is also the state of the Hidden Markov Chain corresponding to the i-th input data. The fully connected layer includes an input layer, a hidden layer, and an output layer, and its output is whether the state has been updated; if updated, it is 1, otherwise it is 0. The decoder consists of a neural network and a cross-entropy loss function. The decoder and encoder are correlated in terms of weights and biases, and the prediction result is fed back to the compensation module to calculate the cross-entropy loss between the predicted value and the target value, thereby training and updating the Hidden Markov Model. The encoder maps gear wear state and compensation parameters in real time. Through training, it obtains n key performance indicators that are most closely related to the compensation state under gear wear state and assigns different weight values ​​according to their importance.

2. The hidden Markov prediction and compensation method for gear wear evolution according to claim 1, characterized in that, The encoder's calculations include: The encoder divides the input data into multiple states and obtains the encoding curves and probability distributions of each state through the n LSTM neural networks. The LSTM neural networks take the key performance indicators of each state as input and the probability distribution of the corresponding state as output. State update: Input the state encoding curves and their probability distributions into the decoder, and output whether each state is updated. Whether a state is updated is determined by its probability distribution. State sequence prediction: Substitute the predicted results and the actual results into the Hidden Markov Model, calculate the cross-entropy loss between the predicted and actual values, and update the model.

3. The hidden Markov prediction and compensation method for gear wear evolution according to claim 2, characterized in that, The decoder's calculations include: The decoder inputs the states output by the Hidden Markov Model and the states output by the state update module into the neural network, normalizes the conditional distribution matrices of each state, and sums the results to obtain the Hidden Markov Chain transition probability matrix and observation probability matrix, and calculates the state sequence with the highest probability.

4. The hidden Markov prediction and compensation method for gear wear evolution according to claim 3, characterized in that, In the step of substituting the observation sequence of the Hidden Markov Chain into the trained Hidden Markov Model, the state sequence is substituted into the trained Hidden Markov Model to generate an output sequence, which is then compared with the observation sequence. If the two are completely identical, the model is a correct model; otherwise, the cross-entropy loss of the probability matrix is ​​calculated and the model is updated until the cross-entropy loss is less than the threshold.

5. The hidden Markov prediction and compensation method for gear wear evolution according to claim 4, characterized in that, The trainable Hidden Markov Model (HMM) is used to predict the possible future states of the Hidden Markov Chain. The lifetime endpoints of the observed curves under different states are predicted. The lifetime endpoints under different states are substituted into the HMM model to obtain the probability distribution of each lifetime endpoint. The probability distribution of the lifetime endpoints is substituted into the cross-entropy loss function. The lifetime endpoint corresponding to the maximum value of the probability distribution of the lifetime endpoints is taken as the predicted lifetime endpoint value. The probability of the predicted lifetime endpoint value is greater than 80%.

6. The hidden Markov prediction and compensation method for gear wear evolution according to claim 5, characterized in that, The life end prediction curve represents the historical curve and prediction curve of the wear state evolution within the life cycle. The life cycle is divided according to time, and the life end prediction curve represents the probability distribution curve of each life end prediction within the life cycle.

7. The hidden Markov prediction and compensation method for gear wear evolution according to claim 6, characterized in that, The step of substituting the reference wear state into the encoder in the hidden Markov chain: By using the cross-entropy loss function of the predicted end of life, the end of life corresponding to the maximum value of the probability distribution of the predicted end of life is found. This value is the optimal solution. If the probability of the optimal solution is less than 80%, the search continues; otherwise, the optimal solution is output. The output corresponding to the optimal solution is used as the reference wear state, and the reference wear state curve is input into the encoder to obtain the key performance indicators under this wear state. The corresponding parameters are then found, which are the compensation parameters for predicting the wear state.

8. A hidden Markov prediction and compensation system for gear wear evolution, used to implement the hidden Markov prediction and compensation method for gear wear evolution as described in any one of claims 1-7, characterized in that, The system includes: State-parameter mapping module: Establish a hidden Markov chain for gear wear state-compensation parameters. Integrate an acoustic emission sensor inside the encoder to collect tooth surface friction signals. The encoder divides the input data into multiple states and obtains the encoding curves and probability distributions of each state through n LSTM neural networks. The decoder and encoder are correlated in terms of weights and biases. Feed the prediction results back to the compensation module, calculate the cross-entropy loss between the predicted value and the target value, and train the encoder and decoder. Life end prediction module: Substitute the output of the Hidden Markov Model into the trained model to generate an output sequence, and compare it with the observation sequence until the two are completely identical; Lifetime end prediction curve: By obtaining the historical curve and prediction curve of wear state evolution within the life cycle output by the encoder, the lifetime end under different states is substituted into the Hidden Markov Model to obtain the probability distribution of each lifetime end; and the probability distribution of the lifetime end is substituted into the cross-entropy loss function, and the lifetime end corresponding to the maximum value of the probability distribution of the lifetime end is taken as the predicted lifetime end value, and the optimal solution is output. Parameter optimization module: During the life cycle, the wear state evolution of the predicted life end value is performed. The optimal life end value with the closest distance to the target is found in the pre-plotted compensation optimization solution space on the life prediction curve as the reference wear state, and the optimal solution is output. The optimal solution is substituted into the hidden Markov model to find the observation value and state corresponding to the optimal solution. The model is updated by putting them into the model. If the updated cross-entropy loss is less than a set threshold, the optimal solution is output. Otherwise, the previous step is repeated until the updated model meets the condition of being less than the threshold.

Images