A machine tool feed shaft precision degradation predictive maintenance method based on continuous optimization of mamba network

Abstract

The present application belongs to the technical field of industrial equipment health monitoring and predictive maintenance, and discloses a machine tool feed shaft precision degradation predictive maintenance method based on continuous optimization Mamba network. The machine tool feed shaft vibration signal is collected by a three-way vibration sensor, and after baseline correction, filtering and noise reduction, and self-encoder feature extraction, a low-dimensional feature vector is constructed. The Mamba network model is pre-trained using the bearing full life cycle data of the test platform, the state evolution law is captured by combining the dynamic parameter generation mechanism, and the state space equation is discretized by the zero-order holding rule to realize efficient time series modeling. In actual monitoring, the model predicts the future feature sequence based on part of the input data, calculates the health index and compares it with the multi-level threshold, and realizes the four-stage state discrimination of health, sub-health, early degradation and serious degradation. Further, a continuous learning mechanism is introduced, and the model parameters are dynamically updated through experience replay and small learning rate fine-tuning to adapt to working condition changes and new fault modes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial equipment health monitoring and predictive maintenance, and relates to a machine tool feed shaft precision degradation predictive maintenance method based on a continuously optimized Mamba network. BACKGROUND

[0002] In modern manufacturing systems, machine tools as core processing equipment, their stable operation plays an important role in ensuring product quality and improving production efficiency. Among them, the feed shaft is a key component for machine tools to achieve high-precision machining, and its performance changes will directly affect the machining precision.

[0003] Traditional maintenance methods generally use post-maintenance and preventive maintenance methods, but both methods have obvious limitations: preventive maintenance may cause over-maintenance or under-maintenance, resulting in serious waste of resources; post-maintenance is often accompanied by unplanned downtime, high maintenance costs, and low equipment utilization. And traditional predictive maintenance methods focus more on health state identification, making it difficult to effectively depict the precision degradation process, especially in the early stages of precision decline when there are no obvious signs of failure. Therefore, it is urgent to build an intelligent predictive maintenance method for key component state to realize dynamic identification of feed shaft system operating state and forward-looking judgment of fault trend, and improve the self-maintenance ability and operation stability of the equipment.

[0004] In recent years, the research on machine tool state monitoring and predictive maintenance based on deep learning has been continuously deepened, especially the technology of fusing multi-source signals and time series modeling has made certain progress in the industrial field. In the patent "Machine tool vibration state monitoring system and fault prediction method thereof" (CN118386028A), the use of vibration historical data and workpiece quality evaluation for state discrimination is emphasized, and a vibration-process data mapping mechanism is established to improve the correlation of the prediction dimension. In the patent "Edge-cloud collaborative numerical control machine tool predictive maintenance system and method" (CN118011951A), an edge-cloud platform collaborative structure is constructed, and through the modeling of multiple sensor signals such as vibration, temperature, and current, distributed deployment of state monitoring and model training is realized. In the patent "Machine tool feed shaft health monitoring method based on optimized condition autoencoder" (CN119691690A), a wavelet packet decomposition and condition variational autoencoder structure are combined to realize the state identification of the feed shaft based only on health data, alleviating the problem of data scarcity.

[0005] The above research demonstrates that predictive maintenance methods have achieved significant success in the field of machine tool condition monitoring. However, in practical applications, traditional maintenance strategies still suffer from issues such as untimely maintenance and unexpected downtime, which waste maintenance resources. To address this issue, the present invention, by introducing the Mamba network architecture, can achieve highly accurate classification of the current equipment status (healthy / subhealthy / initial degradation / severe degradation) of machine tool feed axes, predict future degradation trends, and output the degradation stage of the current state, even while acquiring only a partial vibration signal. Furthermore, by incorporating a continuous learning mechanism, the model dynamically adapts to evolving equipment states, further enhancing prediction accuracy and robustness. This invention aims to optimize traditional maintenance strategies, shifting from passive response and fixed-cycle maintenance to proactive prediction and efficient resource allocation, effectively reducing the risk of unexpected downtime and improving equipment availability and production efficiency. Summary of the Invention

[0006] The purpose of this invention is to provide a predictive maintenance method for machine tool feed axis accuracy degradation based on continuously optimized Mamba networks. Without relying on a complete time series, it can realize the identification and trend prediction of the current operating status, especially the accurate identification of the "transition state" in the early stage of degradation, and provide a more intelligent maintenance solution for industrial equipment.

[0007] The technical solution of the present invention:

[0008] A predictive maintenance method for machine tool feed axis accuracy degradation based on continuous optimization of Mamba network is proposed. The steps are as follows:

[0009] 1) First, the vibration signal during the processing is collected through a three-axis vibration sensor. 2) Then, the collected original vibration signal is preprocessed, including baseline correction, DC offset removal, filtering and noise reduction operations; then, a sliding window and autoencoder are used to extract typical characteristic parameters representing the signal energy, impact and spectral structure from the preprocessed signal to construct a feature vector. 3) Subsequently, the Mamba network model is trained using the feature vectors containing the degradation process of the bearing throughout its life cycle measured on the test bench, so that it can learn and capture the entire evolution law of the bearing from health to wear and damage. Based on the monitoring data of the feed axis of a specific machine tool, the training model is fine-tuned to adapt to its specific working conditions and bearing characteristics. In the actual monitoring process, when only partial monitoring data is available, the trained Mamba network model is used to predict the subsequent feed axis feature vectors. 4) The feature vector or prediction result output by the model representing the current operating status of the feed axis is output in the form of a health index HI. This index can be mapped to a trajectory deviation index to judge the accuracy degradation trend. According to the HI value and the preset threshold , , , determines the current health stage of the feed axis (healthy, subhealthy, initial degradation, severe degradation). Predictive results drive early maintenance decisions, enabling the foreseeing and intervening in potential degradation or failures, thereby implementing predictive maintenance based on state evolution trends. 5) Based on the Mamba model's predictions of bearing status and newly generated vibration signal samples during actual machine tool operation, a continuous learning approach is employed. This not only focuses on the accuracy of state category recognition, but also dynamically evaluates the model's prediction stability for the feed axis's accuracy trajectory. If the accumulated prediction deviation exceeds the set range, model retraining will be triggered to ensure high sensitivity to accuracy trends after long-term operation.

[0010] The specific steps are as follows:

[0011] The first step is vibration signal collection;

[0012] Deploy a vibration sensor on the feed axis of the machine tool to collect the vibration signal of the machine tool, and convert the mechanical vibration into an analog electrical signal through the vibration sensor, which is then digitized synchronously by a multi-channel data acquisition card; wherein, the sampling frequency is set , each channel at time The vibration signal sampled internally is expressed as , forming the original vibration signal through continuous collection;

[0013] The second step is the original vibration signal preprocessing and feature extraction;

[0014] Perform baseline correction and DC offset removal on the original vibration signal of each channel:

[0015]

[0016] in, is the sampling time, is the vibration signal after removing DC, is the average value of the currently sampled vibration signal;

[0017] The vibration signal after DC bias is removed Perform bandpass filtering to remove irrelevant interference from environmental noise and obtain the filtered vibration signal :

[0018]

[0019] in, is the unit pulse sampling response of the bandpass filter; is the filtered vibration signal; N is the filter length, and the order is N-1; k is the kth sampling point of the impulse response of the bandpass filter; is the sampling period;

[0020] From the filtered vibration signal Constructing time series data samples through sliding windows :

[0021]

[0022] in, represents the w-th sliding window sample, L represents the length of the sliding window, and w represents the time step index of the window;

[0023] Then use the autoencoder to learn and extract time series data samples A series of time domain eigenvalues ​​and frequency domain eigenvalues ​​in the autoencoder are used to generate a low-dimensional feature vector Z; the autoencoder consists of an encoder and a decoder. The encoder maps the input time series data samples to a low-dimensional space and outputs a low-dimensional feature vector Z; the decoder accepts the low-dimensional feature vector Z and the speed N of the machine tool feed axis to reconstruct the original input ;

[0024]

[0025]

[0026]

[0027] in, , is the encoder weight, is the encoder bias, is the decoder weight, is the decoder bias;

[0028] The third step is Mamba model training and full life cycle simulation;

[0029] The training strategy of the Mamba model is given: the vibration signals collected over the entire life cycle of the bearing are processed through a sliding window and the encoder part of the autoencoder to obtain a standardized low-dimensional feature vector sequence: , as the training data set of the Mamba model, T is the total length of the normalized low-dimensional feature vector sequence; the training data set is divided into 80% for training and 20% for testing;

[0030] The core of the Mamba model is the Selective State Space Model (SSM), which is used to dynamically model the input normalized low-dimensional feature vector sequence in the time dimension. Its continuous time form is:

[0031]

[0032] in, represents the state vector of the Mamba model in continuous time; represents the time derivative of the state vector, represents the state transition matrix, represents the input matrix, represents the output matrix, is the normalized low-dimensional feature vector, Represents the output signal of the system;

[0033] In order to improve computational efficiency, the continuous-time system is discretized using the zero-order hold rule, and we get:

[0034]

[0035] in, represents the state vector of the Mamba model in discrete time, 、 Because A, B, The derived discretization parameters, is the input feature vector of the discrete time step;

[0036] In the framework of the Mamba model, 、 , C are generated by the dynamic parameter network according to the current input standardized low-dimensional feature vector Real-time generation enables the state update process to adapt to changes in the input feature vector.

[0037]

[0038] in is a learnable parameter of MLP. This design enables the model to adapt to the input feature structure under different working conditions and enhance the modeling ability of non-stationary time series.

[0039] In order to further mine the standardized low-dimensional feature vector of the current input The dependency relationship between dimensions is modeled dynamically in the time dimension. The Mamba state space modeling process based on the variable dimension is introduced. Specifically, the low-dimensional feature vector extracted by the autoencoder is converted into According to the importance of features, they are divided into key variables and auxiliary variables, namely:

[0040]

[0041] Among them, t is the index of the current time step, f is the dimension of the feature vector, is the i-th characteristic component at time t, ;

[0042] Key variables refer to the characteristic components that are most sensitive to the health status of the machine tool feed axis and directly reflect the degradation evolution trend; auxiliary variables refer to secondary characteristic components that indirectly affect the health status or couple with noise; in the variable-dimensional Mamba state space modeling, the input feature vector updates the hidden state in the order from key variables to auxiliary variables, thereby capturing the dependencies between different characteristic components.

[0043] The state update process along the variable dimension is described by the following formula:

[0044]

[0045] in, represents the hidden state of the f-th input feature vector component of the variable dimension; Indicates the The hidden state of the input feature vector components; Represents the value of the fth eigenvector component at time step t; F represents the total number of feature dimensions; 、 are the state transition matrix and input matrix of the variable dimension Mamba model respectively;

[0046] During the training phase, the standardized low-dimensional feature vector output by the autoencoder As a supervisory signal, a loss function based on mean square error is constructed in the following form:

[0047]

[0048] in, is the complete feature vector of the next moment predicted by the Mamba model, is the true target feature, Measures the Mamba model's ability to fit the time evolution trend of low-dimensional feature vectors;

[0049] The gradient of the network parameters is calculated by the error back propagation algorithm:

[0050]

[0051] in, represents the set of parameters in the network, represents the weight matrix, b represents the bias vector;

[0052] Update the parameters using the Adam optimizer:

[0053]

[0054] Among them, the learning rate [1 ~1 ] range, dynamically adjusted according to the accuracy of the validation set; The loss function parameter gradient;

[0055] After training, the Mamba model is deployed in the actual feed axis condition monitoring task. During actual online monitoring, when the Mamba model obtains the standardized feature sequence for the current period:

[0056]

[0057] in, represents the standardized feature sequence at the current moment, Represents the normalized feature vector at time t, where t is the current time index;

[0058] The Mamba model uses the overall change trends and sequence patterns learned from the bearing life cycle data to derive characteristic sequences for a long period of time in the future:

[0059]

[0060] in, is the future feature vector sequence predicted by the Mamba model based on the current input feature vector, The first The feature vector of time, K represents the predicted duration;

[0061] This process is equivalent to simulating the evolution path of the vibration signal characteristics of the machine tool feed shaft bearing from its current state to the end of its life cycle, thus making up for the lack of complete life cycle data in actual data. Based on these derived future feature vectors, a comprehensive health index (HI) is calculated.

[0062]

[0063]

[0064] Among them, L is the prediction weight, e is the prediction bias, is the activation function;

[0065] Step 4: Status prediction and early warning;

[0066] Thresholds are set based on historical sample statistical analysis and expert experience, where is the health threshold, above which the person is considered to be fully healthy; is the warning threshold, between and In between is sub-health; is the degradation threshold, between and Between is the initial degradation, below The value is severely degraded;

[0067] According to the health index HI and set threshold , , The relationship between the feed axis and the feed axis is used to determine the current health stage (completely healthy, sub-healthy, initial degradation, severe degradation). Specific judgment results are generated and maintenance suggestions are generated through the interface.

[0068] Step 5: Continuous learning and dynamic model optimization;

[0069] First, the system continuously monitors the predictive and discriminative criteria of the Mamba model on the latest standardized feature sequence. When the prediction error or health discrimination accuracy deteriorates for a period of time and exceeds the preset health threshold, or when a certain number of new data samples that have not been trained are accumulated, the Mamba model is automatically updated. An incremental learning strategy combining experience replay with small learning rate fine-tuning is used to avoid the "catastrophic forgetting" problem. Experience replay pool: maintains a fixed-size buffer R to store a small number of representative standardized feature sequence samples from different historical health stages; mixed training batch: in each incremental training, a batch of samples (denoted as ), and extract a batch of old samples from the buffer R (denoted as ), the two are mixed to form a new training batch . Further, use Fine-tune the Mamba model to ensure that the model effectively consolidates old knowledge while learning new knowledge.

[0070] Beneficial effects of the present invention:

[0071] (1) The Mamba model in this paper is pre-trained using data from the entire life cycle of general bearings and then fine-tuned for specific machine tool feed shaft bearing data. This strategy effectively compensates for the scarcity of specific machine tool life data and significantly enhances the model's generalization ability under different operating conditions, achieving efficient data utilization.

[0072] (2) This invention introduces a continuous learning mechanism that enables the Mamba model to dynamically adapt to new operating conditions and potential new failure modes online. This ensures that the model can maintain high accuracy over the long term in complex and changing industrial environments, avoiding performance degradation caused by environmental changes.

[0073] (3) The present invention provides maintenance personnel with sufficient warning time through precise status identification and a multi-level early warning mechanism. This enables machine tool maintenance to shift from passive response to active prevention, effectively avoiding unplanned downtime, significantly reducing maintenance costs, and improving machine tool equipment utilization and production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 Flowchart of the predictive maintenance method for machine tool feed axis accuracy degradation based on continuous optimization of Mamba network.

[0075] Figure 2 This is a schematic diagram of the Mamba network model. DETAILED DESCRIPTION

[0076] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0077] The predictive maintenance method for machine tool feed axis accuracy degradation based on continuous optimization of the Mamba network has the following steps:

[0078] Taking the X-axis of a certain model of three-axis vertical machining center as an example, the embodiment of the present invention is described in detail. The travel range of the three-axis vertical machining center is 0 ~ 850mm, the maximum feed speed is 20000mm / min, and the process is as follows Figure 1 shown.

[0079] First, the vibration sensor is used to collect the vibration signal of the bearing throughout its life cycle on the test bench. At the same time, vibration sensors are arranged at the proximal bearing and distal bearing of the X-axis motor of the machine tool to actually collect the vibration signal of the X-axis of the vertical machining center when it is running in a healthy state. , the sampling frequency is 50000Hz, the machine tool movement range is 0 ~850mm, the feed speed is 16000mm / min, and the acquisition time is 3min.

[0080] Then the original signal preprocessing stage is carried out, and the above-mentioned collected signals are and Perform baseline correction and remove DC bias respectively to obtain the vibration signal after preliminary processing and ; Then, the low-frequency mechanical vibration and high-frequency background noise that are irrelevant to the accuracy change are filtered out through a band-pass filter to obtain the purified vibration signal and , the formula is as follows:

[0081]

[0082] in, is the sampling time, is the vibration signal after removing DC component, is the average value of the current sampling signal.

[0083]

[0084] wherein, is the unit impulse response of the band-pass filter, is the filtered vibration signal, N is the filter length (number of coefficients), the order is N-1, k is the kth sampling point of the band-pass filter impulse response, is the sampling period;

[0085] from the filtered vibration signal a time series data sample is constructed by a sliding window , the formula is as follows:

[0086]

[0087] wherein, each is a 64-dimensional standardized low-dimensional feature vector, represents the wth sliding window sample, and the sliding window length L=100.

[0088]

[0089]

[0090]

[0091] wherein, is the encoder weight, is the encoder bias, is the decoder weight, is the decoder bias.

[0092] Subsequently, the bearing full-life vibration data collected by the test platform is processed by a sliding window and a feature encoder to construct a standardized low-dimensional feature vector sequence as the training data set of the Mamba model. In order to improve the generalization ability and training stability of the model, the data set is divided before modeling, and 80% and 20% of the data set are respectively used for training and testing. The model is described by the following state transition and output equations:

[0093]

[0094]

[0095] wherein, , respectively represent the discretized time series Mamba parameters, , 、 、C , dynamically generated matrix, constructed by multi-layer perceptron (MLP):

[0096]

[0097] The structure is as follows: input layer dimension: 64; hidden layer dimension: 256; output dimension: 128*128=16384.

[0098] In order to further explore the input feature vector After the dynamic modeling of the time dimension is completed, the dependency relationship between the dimensions is introduced into the Mamba state space modeling process based on the variable dimension; the low-dimensional feature vector extracted by the autoencoder is converted into According to the importance of features, they are divided into key variables and auxiliary variables, namely: The input feature vector is used to update the hidden state in order from key variables to auxiliary variables, thereby capturing the dependencies between different feature components.

[0099] The state update process of the Mamba model based on variable dimensions is described by the following formula:

[0100]

[0101] in, represents the hidden state of the f-th input feature vector component of the variable dimension; Indicates the The hidden state of the input feature vector components; Represents the value of the fth eigenvector component at time step t; F represents the total number of feature dimensions; 、 are the state transition matrix and input matrix of the variable dimension Mamba model respectively;

[0102] During the training phase, the model is iteratively optimized using the mean square error as the loss function:

[0103]

[0104] Use error back propagation to find the derivative: ;

[0105] The optimizer uses Adam: ;

[0106] in, represents the set of parameters in the network, represents the weight matrix, b represents the bias vector;

[0107] After the model is trained, in actual application, the standardized feature sequence obtained after pre-processing is input:

[0108]

[0109] The future feature trend output by the model is as follows:

[0110]

[0111] wherein, is the feature sequence predicted by the model for the next K=20 steps based on the current input feature vector, .

[0112] Based on the future feature sequence derived by the model, the comprehensive health index HI is calculated, which is between 0 and 1:

[0113]

[0114] wherein, is the prediction weight, and e is the prediction bias.

[0115] According to expert experience, set the threshold value , , , =0.8 is the health threshold, and a value higher than this is considered completely healthy; =0.65 is the early warning threshold, and a value between and is sub-healthy; =0.45 is the degradation threshold, and a value between and is early degradation, and a value lower than is severe degradation. The health index HI obtained by collecting signals under the current health status of the machine tool is significantly higher than 0.8, indicating that the machine tool is currently in a healthy stage.

[0116] In addition, the experience replay and fine-tuning mechanism is introduced to realize continuous learning: when the error increases or a new working condition appears, representative samples are automatically selected from the historical sample pool to construct an incremental training set with the current data, and the Mamba model is fine-tuned online to ensure its long-term stability and wide adaptability.

Claims

1. A predictive maintenance method for machine tool feed axis accuracy degradation based on continuous optimization of Mamba network, characterized in that: Here are the steps: The first step is vibration signal collection; Deploy a vibration sensor on the feed axis of the machine tool to collect the vibration signal of the machine tool, and convert the mechanical vibration into an analog electrical signal through the vibration sensor, which is then digitized synchronously by a multi-channel data acquisition card; wherein, the sampling frequency is set , each channel at time The vibration signal sampled internally is expressed as , forming the original vibration signal through continuous collection; The second step is the original vibration signal preprocessing and feature extraction; Perform baseline correction and DC offset removal on the original vibration signal of each channel: ； in, is the sampling time, is the vibration signal after removing DC, is the average value of the currently sampled vibration signal; The vibration signal after DC bias is removed Perform bandpass filtering to remove irrelevant interference from environmental noise and obtain the filtered vibration signal : ； in, is the unit sample response of the bandpass filter; is the filtered vibration signal; N is the filter length, and the order is N-1; k is the kth sampling point of the impulse response of the bandpass filter; is the sampling period; From the filtered vibration signal Constructing time series data samples through sliding windows : ； in, represents the w-th sliding window sample, L represents the length of the sliding window, and w represents the time step index of the window; Then use the autoencoder to learn and extract time series data samples A series of time domain eigenvalues ​​and frequency domain eigenvalues ​​in the , generate a low-dimensional feature vector Z; the autoencoder consists of an encoder and a decoder, the encoder takes the input time series data sample Mapped into a low-dimensional space, the low-dimensional feature vector Z is output; the decoder accepts the low-dimensional feature vector Z and the speed N of the machine tool feed axis to reconstruct the original input ; ； ； ； in, , is the encoder weight, is the encoder bias, is the decoder weight, is the decoder bias; The third step is Mamba model training and bearing life cycle simulation; The training strategy of the Mamba model is to process the vibration signals collected over the entire life cycle of the bearing through a sliding window and an autoencoder to obtain a standardized low-dimensional feature vector sequence: , as the training data set of the Mamba model, T is the total length of the normalized low-dimensional feature vector sequence; the training data set is divided into 80% for training and 20% for testing; The Mamba model is a selective state space model used to dynamically model the input normalized low-dimensional feature vector sequence in the time dimension. Its continuous time form is: ； in, represents the state vector of the Mamba model in continuous time; represents the time derivative of the state vector, represents the state transition matrix, represents the input matrix, represents the output matrix, is the normalized low-dimensional feature vector, Represents the output signal of the system; In order to improve computational efficiency, the continuous-time system is discretized using the zero-order hold rule, and we get: ； in, represents the state vector of the Mamba model in discrete time, 、 Because A, B, The derived discretization parameters, is the input feature vector of the discrete time step; In the framework of the Mamba model, 、 , C are generated by the dynamic parameter network according to the current input standardized low-dimensional feature vector Real-time generation enables the state update process to adapt to changes in the input feature vector; ； in, is the learnable parameter of MLP; In order to further mine the standardized low-dimensional feature vector of the current input After the dependencies between dimensions are dynamically modeled in the time dimension, the Mamba state space modeling process based on the variable dimension is introduced: the low-dimensional feature vector extracted by the encoder is converted into According to the importance of features, they are divided into key variables and auxiliary variables, namely: ； Among them, t is the index of the current time step, f is the dimension of the normalized low-dimensional feature vector of the current input, is the i-th characteristic component at time t, ; Key variables are the characteristic components that are most sensitive to the health status of the machine tool feed axis and directly reflect the degradation evolution trend; auxiliary variables are secondary characteristic components that indirectly affect the health status or couple with noise. In the variable-dimensional Mamba state space modeling, the input feature vectors update the hidden state in the order from key variables to auxiliary variables, thereby capturing the dependencies between different input feature components. The state update process along the variable dimension is described by the following formula: ； in, represents the hidden state of the f-th input feature vector component of the variable dimension; Indicates the The hidden state of the input feature vector components; Represents the value of the fth eigenvector component at time step t; F represents the total number of feature dimensions; 、 are the state transition matrix and input matrix of the variable dimension Mamba model respectively; During the training phase, the standardized low-dimensional feature vector output by the autoencoder As a supervisory signal, a loss function based on mean square error is constructed in the following form: ； in, is the complete feature vector of the next moment predicted by the Mamba model, is the true target feature, Measures the Mamba model's ability to fit the time evolution trend of low-dimensional feature vectors; The gradient of the network parameters is calculated by the error back propagation algorithm: ； in, represents the set of parameters in the network, represents the weight matrix, b represents the bias vector; Update the parameters using the Adam optimizer: ； Among them, the learning rate Set to Range 1 ~1 Range, dynamically adjusted according to the accuracy of the validation set; The loss function parameter gradient; After training, the Mamba model is deployed in the actual machine tool feed axis condition monitoring task. During actual online monitoring, when the Mamba model obtains the standardized low-dimensional feature vector sequence for the current period: ； in, Represents the normalized low-dimensional feature vector sequence at the current moment, Represents the normalized low-dimensional feature vector at time t; The Mamba model uses the overall change trends and sequence patterns learned from the bearing life cycle data to derive characteristic sequences for a long period of time in the future: ； in, is the future feature vector sequence predicted by the Mamba model based on the current input feature vector, The first The feature vector of time, K represents the predicted duration; Based on the derived future eigenvectors, a comprehensive health index HI is calculated; ； ； Among them, L is the prediction weight, e is the prediction bias, is the activation function; Step 4: Status prediction and early warning; Based on historical sample statistical analysis and combined with expert experience is the health threshold, above which the person is considered to be fully healthy; is the warning threshold, between and In between is sub-health; is the degradation threshold, between and Between is the initial degradation, below The value is severely degraded; According to the health index HI and set health threshold , , The relationship between the feed axis and the machine tool is used to determine the health stage of the machine tool feed axis, generate specific judgment results and generate maintenance suggestions through the interface; Step 5: Continuous learning and dynamic model optimization; First, the predictive and discriminative criteria of the Mamba model on the latest standardized low-dimensional feature vector sequence are continuously monitored. When the prediction error or health discrimination accuracy deteriorates for a period of time and exceeds the preset health threshold, or a certain number of new data samples that have not been trained are accumulated, the Mamba model is automatically updated. An incremental learning strategy combining experience replay with small learning rate fine-tuning is adopted. The experience replay pool maintains a fixed-size buffer R to store a small number of representative standardized low-dimensional feature vector sequence samples from different historical health stages. Mixed training batches are extracted from the newly collected and labeled feature data in each incremental training, which are recorded as , and extract a batch of old samples from the buffer R, recorded as , the two are mixed to form a new training batch ;use Fine-tune the Mamba model.

Images