Industrial mother machine electric spindle service life prediction model method based on improved iTransform

Abstract

The invention discloses an industrial mother machine electric spindle service life prediction method based on an improved iTransform model, and relates to the field of predictive maintenance, and the method comprises the steps: carrying out the preprocessing of operation state original data and multi-dimensional influence factor data, complementing the missing data, and obtaining model data; performing feature extraction and analysis on the model data to obtain time sequence features; performing hierarchical information processing on the time sequence features, and outputting fused features; compiling and predicting the fused features to obtain a motorized spindle life prediction result; a scene generator is embedded in the iTransform, the scene generator is optimized, and the life variation amplitude under each scene is generated; the sensitive degree of each variable to the service life is obtained through decomposition processing of the service life of the industrial mother machine electric spindle. The advantages of a physical model, a statistical model and a machine learning model are combined, the prediction precision and robustness are improved, and more reliable support is provided for intelligent operation and maintenance of the motorized spindle.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of predictive maintenance, in particular to an improved iTransformer-based industrial machine electric spindle life prediction model method. BACKGROUND

[0002] With the acceleration of the intelligent transformation of the manufacturing industry, the performance and reliability requirements of industrial mother machines, as the core equipment of modern manufacturing, are increasingly improved. As a key component of industrial mother machines, electric spindles directly determine the machining precision and efficiency, and their running state has an important influence on production quality and cost. However, electric spindles are prone to performance degradation or even sudden failure due to factors such as wear and fatigue when running at high speed and high load for a long time, which may cause equipment downtime and production interruption. Traditional post-maintenance and regular maintenance modes cannot meet the needs of modern manufacturing for efficient and accurate operation and maintenance. Therefore, how to monitor the running state of electric spindles in real time, predict their remaining life by combining historical data and intelligent algorithms, and develop maintenance plans in advance to avoid unplanned downtime and reduce operation and maintenance costs has become a key task in the intelligent operation and maintenance and energy saving and efficiency improvement of the current industrial mother machine field.

[0003] Currently, common electric spindle life prediction methods include physical modeling methods, statistical model methods, and machine learning methods. Among them, machine learning methods have been widely used in life prediction due to their strong non-linear fitting ability and adaptability to complex data. For example, traditional machine learning algorithms such as support vector machines and random forests can effectively process electric spindle operation data, but they have certain limitations when processing high-dimensional time series data. In recent years, deep learning models such as long short-term memory networks and convolutional neural networks have shown significant advantages in electric spindle life prediction due to their ability to capture long-term dependencies and spatial features of time series data. However, electric spindle life is influenced by multiple factors such as load, speed, temperature, and vibration, and a single model often cannot fully reflect its degradation pattern.

[0004] The prior art has not yet proposed an effective solution to the problems in the related art. SUMMARY

[0005] To solve the problems in the related art, the present application proposes an improved iTransformer-based industrial machine electric spindle life prediction model method to overcome the above technical problems existing in the prior art.

[0006] To this end, the specific technical solutions adopted by the present application are as follows:

[0007] An improved iTransformer-based industrial machine electric spindle life prediction model method, comprising:

[0008] The running state original data and multi-dimensional influence factor data of the industrial mother machine electric spindle are preprocessed, and the missing data is supplemented by using gray dynamic correlation analysis to obtain model data;

[0009] The model data is extracted and analyzed by using the parallel processing of the extended long short-term memory network, and the output features are fused to obtain the time sequence features; the hierarchical information processing of the time sequence features is performed through the attention mechanism to realize the dual capture of local details and global trend, and the fused features are output;

[0010] The fused features are compiled and predicted by iTransformer to obtain the life prediction result of the industrial mother machine electric spindle;

[0011] When the error of the prediction result exceeds the prediction error setting value, the model parameters of iTransformer are fine-tuned to output the optimal model, and the life of the industrial mother machine electric spindle is predicted based on the optimal model;

[0012] The scenario generator is embedded in iTransformer, and the scenario generator is optimized by total deviation and diversity loss function to cover multi-scenario disturbance data of different working conditions and environmental factors, and the life change amplitude under each scenario is generated; the sensitivity of each variable to the life is obtained by decomposing the life of the industrial mother machine electric spindle.

[0013] Further, the running state original data and multi-dimensional influence factor data of the industrial mother machine electric spindle are preprocessed, and the missing data is supplemented by using gray dynamic correlation analysis to obtain model data, including:

[0014] The running state original data and multi-dimensional influence factor data of the industrial mother machine electric spindle are obtained and preprocessed;

[0015] The gray dynamic correlation degree between the candidate variables and the missing variables in the running state original data and the multi-dimensional influence factor data is calculated, and the variables with a gray dynamic correlation degree greater than a preset value are selected to form a complementary variable cluster;

[0016] According to the gray dynamic correlation degree of the complementary variables in the complementary variable cluster, the dynamic weight of the complementary variables is calculated, and the filling value of the missing variables is calculated in combination with the complementary variables;

[0017] The running state original data and multi-dimensional influence factor data are supplemented by using the filling value to obtain the model data.

[0018] Further, the model data is extracted and analyzed by using the parallel processing of the extended long short-term memory network, and the output features are fused to obtain the time sequence features, including:

[0019] The parallel processing structure of the extended long short-term memory network including a single-gate long short-term memory network and a matrix long short-term memory network;

[0020] The long-term dependency and nonlinear dynamic change in the degradation process of the motorized spindle are captured from the model data by the parallel processing structure of the extended long short-term memory network, and the output time sequence features are fused.

[0021] Further, the long-term dependency and nonlinear dynamic change in the degradation process of the motorized spindle are captured from the model data by the parallel processing structure of the extended long short-term memory network, and the output time sequence features are fused, including:

[0022] The input information is dynamically selected and weighted by the multi-head attention mechanism, and the corresponding attention weight is obtained; the output of each head in the multi-head attention mechanism is obtained by weighted summation, and the outputs of all heads are connected to obtain the gating information;

[0023] The model data is processed by combining the gating information and the parallel processing structure of the single-gate long short-term memory network and the matrix long short-term memory network to obtain the feature representation corresponding to the single-gate long short-term memory network and the feature representation corresponding to the matrix long short-term memory network;

[0024] The feature representation corresponding to the single-gate long short-term memory network and the feature representation corresponding to the matrix long short-term memory network are fused by weighted summation to obtain the time sequence feature.

[0025] Further, the time sequence feature is processed by the attention mechanism for hierarchical information processing to realize dual capture of local details and global trends, and the fused feature is output, including:

[0026] The attention mechanism containing local details and global trends is constructed, and the local feature and the global trend feature are captured based on the high-frequency branch and the low-frequency branch of the attention mechanism, respectively;

[0027] The local feature and the global trend feature are mapped to the encoder by the embedding layer, and the dimension is adjusted by the attention layer to output the fused feature.

[0028] Further, the fused feature is compiled and predicted by the iTransformer to obtain the industrial mother machine motorized spindle life prediction result, including:

[0029] The encoder output is obtained by compiling the fused feature by the iTransformer.

[0030] The encoder output is predicted by using the full connection layer and the regression layer to obtain the industrial mother machine motorized spindle life prediction result.

[0031] Furthermore, a scenario generator is embedded in iTransformer and optimized using total deviation and diversified loss functions to cover multi-scenario perturbation data under different operating conditions and environmental factors, generating lifetime variation amplitudes under each scenario, including:

[0032] Embed the scenario generator into the iTransformer model; obtain the working conditions and thresholds at all times to calculate the total deviation; calculate the diversification loss based on the probability of different working conditions occurring; optimize the scenario generator using the diversification loss and the total deviation.

[0033] After generating multiple scenario environments through the scenario generator, the basic operating conditions and environmental factors are combined to generate the perturbed operating conditions; and the operating condition environment combination is generated based on the perturbed operating conditions.

[0034] Based on the logarithmic mean of the life under various operating conditions, the standard deviation of the natural logarithm of the electric spindle life, and the life of the electric spindle under the corresponding scenario, the predicted life of the electric spindle under any scenario is calculated.

[0035] The magnitude of life variation is calculated based on the predicted life of the electric spindle under arbitrary scenarios and the predicted life of the electric spindle under standard operating conditions.

[0036] Furthermore, by decomposing the lifespan of the industrial machine tool's electric spindle, the sensitivity of each variable to lifespan was obtained, including:

[0037] The lifespan of the industrial machine tool's main spindle is decomposed, and the expected lifespan is obtained.

[0038] Based on expected lifespan, spindle lifespan under various operating conditions, the variance of spindle lifespan, and the variance of each variable, the sensitivity of each variable to lifespan is calculated.

[0039] Furthermore, the formulas for calculating the sensitivity of each variable to lifespan are as follows:

[0040] ;

[0041] ;

[0042] In the formula, The variance of the spindle life. For the first Lifespan under various operating conditions and environmental combinations For life expectancy, For a single input The difference between the other party The degree of sensitivity, In order to be in The variance under the given conditions is N.

[0043] Further, the improved iTransformer-based industrial mother machine electric spindle life prediction model method further comprises: calculating the variance elasticity of the predicted life of each group of scenarios according to the variance of the predicted life of the industrial mother machine electric spindle and the corresponding different disturbances.

[0044] The beneficial effects of the present application are:

[0045] (1) Accurate life prediction: through the extended memory mechanism of xLSTM and the hierarchical feature extraction of HiLo attention mechanism, the residual useful life (RUL) of the electric spindle can be accurately quantified, and the shortcomings of traditional methods in life prediction accuracy and long-term dependence modeling are solved.

[0046] (2) Efficient feature extraction: the HiLo attention mechanism divides the electric spindle operation data into high-frequency and low-frequency parts, respectively capturing detailed information and global trends, significantly improving the efficiency of feature extraction and the attention of the model to key information.

[0047] (3) Model covers multiple scenarios: by embedding a multi-scenario generator in iTransformer, it automatically covers multi-scenario disturbance data of different working conditions and environmental factors, generates life change amplitude under each scenario, quantifies the sensitivity of key variable changes to life length, solves the problem of data scarcity and imbalance of electric spindles, and enhances the robustness of the model under uncertain and variable working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0049] Figure 1 is a flowchart according to an embodiment of the present application;

[0050] Figure 2 is an xLSTM exponential gating mechanism diagram according to an embodiment of the present application;

[0051] Figure 3 is a Hi-Lo attention mechanism diagram according to an embodiment of the present application;

[0052] Figure 4 is a flowchart of an improved iTransformer-based industrial mother machine electric spindle life prediction model method according to an embodiment of the present application;

[0053] Figure 5It is a flow chart for generating life change amplitude under each scenario in an improved iTransformer-based industrial host machine electric spindle life prediction model method according to an embodiment of the present application. DETAILED DESCRIPTION

[0054] To further illustrate the embodiments, the present application provides drawings which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can be used to explain the operating principle of the embodiments in conjunction with the related description of the specification. Those skilled in the art should understand other possible implementations and advantages of the present application by referring to these contents. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0055] According to an embodiment of the present application, an improved iTransformer-based industrial host machine electric spindle life prediction model method is provided.

[0056] The present application will be further described in conjunction with the drawings and specific embodiments, as shown in Figures 1-2 The improved iTransformer-based industrial host machine electric spindle life prediction model method according to an embodiment of the present application includes:

[0057] S1, pre-process the running state original data and multi-dimensional influence factor data of the industrial host machine electric spindle, and use gray dynamic correlation analysis to complete the missing data to obtain model data.

[0058] S2, use the parallel processing of the extended long short-term memory network to extract and analyze the features of the model data, and fuse the output features to obtain the time sequence features; through the attention mechanism, the time sequence features are processed in a hierarchical manner to realize the dual capture of local details and global trends, and the fused features are output.

[0059] S3, compile and predict the fused features through iTransformer to obtain the industrial host machine electric spindle life prediction result.

[0060] S4, when the error of the prediction result exceeds the prediction error setting value, the model parameters of iTransformer are fine-tuned to output the optimal model, and the life of the industrial host machine electric spindle is predicted based on the optimal model.

[0061] S5, embed a scenario generator in iTransformer, and optimize the scenario generator through total deviation and diversification loss function to cover multi-scenario disturbance data of different working conditions and environmental factors, generate life change amplitude under each scenario, and use decomposition processing of the industrial host machine electric spindle life to obtain the sensitivity of each variable to the life.

[0062] In one embodiment, the running state raw data and multi-dimensional influence factor data of the industrial mother machine electric spindle are preprocessed, and the missing data is complemented by using gray dynamic correlation analysis to obtain model data including:

[0063] The running state raw data and multi-dimensional influence factor data of the industrial mother machine electric spindle are obtained and preprocessed; the gray dynamic correlation degree between the candidate variables and the missing variables in the running state raw data and multi-dimensional influence factor data is calculated, and the variables with a gray dynamic correlation degree greater than a preset value are screened out to form a complementary variable cluster; the dynamic weight of the complementary variables in the complementary variable cluster is calculated according to the gray dynamic correlation degree of the complementary variables, and the filling value of the missing variables is calculated in combination with the complementary variables; the running state raw data and multi-dimensional influence factor data are complemented by using the filling value to obtain model data.

[0064] In one embodiment, the model data is feature extracted and analyzed by using parallel processing of an extended long short-term memory network, and the output features are fused to obtain time sequence features including:

[0065] The parallel processing structure of the extended long short-term memory network including a single-gate long short-term memory network and a matrix long short-term memory network is configured; the long-term dependence and nonlinear dynamic change in the degradation process of the electric spindle are captured from the model data by the parallel processing structure of the extended long short-term memory network, and the output time sequence features are fused.

[0066] In one embodiment, the long-term dependence and nonlinear dynamic change in the degradation process of the electric spindle are captured from the model data by the parallel processing structure of the extended long short-term memory network, and the output time sequence features are fused including:

[0067] The input information is dynamically selected and weighted by a multi-head attention mechanism, and the corresponding attention weight; the output of each head in the multi-head attention mechanism is obtained by weighted summation, and the outputs of all heads are connected to obtain gating information; the model data is processed by combining the gating information and the parallel processing structure of the single-gate long short-term memory network and the matrix long short-term memory network to obtain the feature representation corresponding to the single-gate long short-term memory network and the feature representation corresponding to the matrix long short-term memory network; the feature representation corresponding to the single-gate long short-term memory network and the feature representation corresponding to the matrix long short-term memory network are fused by weighted summation to obtain time sequence features.

[0068] In one embodiment, the time sequence features are processed by the attention mechanism in a hierarchical manner to realize dual capture of local details and global trends, and the fused features are output including:

[0069] The attention mechanism containing local details and global trends is constructed, and based on the high-frequency branch and the low-frequency branch of the attention mechanism, local features and global trend features are captured respectively; the local features and the global trend features are mapped to the encoder through an embedding layer, and the dimensions are adjusted through an attention layer to output fused features.

[0070] In one embodiment, the fused features are compiled and predicted by the iTransformer to obtain the industrial mother machine electric spindle life prediction result, including:

[0071] The fused features are compiled by the iTransformer to obtain the encoder output; the encoder output is predicted by using a fully connected layer and a regression layer to obtain the industrial mother machine electric spindle life prediction result.

[0072] In one embodiment, a scenario generator is embedded in the iTransformer, and the scenario generator is optimized by a total deviation degree and a diversification loss function to cover multi-scenario disturbance data of different working conditions and environmental factors to generate the life change amplitude in each scenario, including:

[0073] The scenario generator is embedded in the iTransformer model; the working conditions and the threshold values at all times are obtained to calculate the total deviation degree; the diversification loss is calculated based on the probability of occurrence of different working conditions; the scenario generator is optimized by the diversification loss and the total deviation degree; after the multi-scenario environment is generated by the scenario generator, the disturbed working conditions are generated by combining the basic working conditions and the environmental factors; the working condition environment combination is generated according to the disturbed working conditions; the predicted life of the electric spindle in any scenario is calculated based on the logarithmic life mean value, the standard deviation of the natural logarithm of the electric spindle life and the life of the electric spindle in the corresponding scenario; the life change amplitude is calculated based on the predicted life of the electric spindle in any scenario and the predicted life of the electric spindle in the standard working condition.

[0074] In one embodiment, the sensitivity of each variable to the life is obtained by decomposing the industrial mother machine electric spindle life, including:

[0075] The industrial mother machine electric spindle life is decomposed, and the life expectation is obtained; the sensitivity of each variable to the life is calculated based on the life expectation, the electric spindle life under the working condition environment combination, the variance of the spindle life and the variance under each variable condition.

[0076] In one embodiment, the calculation formula of the sensitivity of each variable to the life is:

[0077] ;

[0078] ;

[0079] In the formula, variance of the main shaft life, life under the combination of the working conditions and the environment, life expectancy, sensitivity of the variance, sensitivity of the variance, sensitivity of the variance, sensitivity of the variance, variance under the condition, variance under the condition,

[0080] In one embodiment, the improved iTransformer-based industrial parent machine electric main shaft life prediction model method further comprises: calculating the variance elasticity of the life prediction of each group of scenarios according to the variance of the life prediction of the industrial parent machine electric main shaft and the corresponding different disturbances.

[0081] In order to facilitate the understanding of the above technical solutions of the present application, the working principle of the present application in the actual process will be described in detail below.

[0082] The present application fuses time series data with multi-dimensional influencing factors such as operating conditions, load characteristics and environmental parameters, overcomes the limitations of traditional models in modeling complex time-dependent relationships. The xLSTM network is used to extract time series features, the Hi-Lo attention mechanism is used to realize the dual capture of local details and global trends, an end-to-end life prediction framework is constructed, the life of the electric main shaft is predicted by the improved iTransformer, and the amplitude of life change is evaluated. The present application accurately models and predicts the running state of the electric main shaft, so that the equipment maintenance personnel can identify potential failure risks in advance, take preventive maintenance measures at the early stage of performance degradation, reasonably plan the spare parts replacement cycle, and realize efficient management of the whole life cycle of the equipment.

[0083] The present application aims at the problem of insufficient modeling of complex time-dependent relationships in traditional models for industrial mother machine electric spindle life prediction, and proposes a life prediction model based on hierarchical time series feature extraction and multi-scale attention mechanism. The method fuses time series data with multi-dimensional influencing factors such as operating conditions, load characteristics and environmental parameters, overcoming the limitations of traditional models in modeling complex time series relationships. The xLSTM network is used to extract time series features, and the Hi-Lo attention mechanism is used to capture local details and global trends simultaneously, building an end-to-end life prediction framework. The present application includes: 1) In view of the problem that traditional models are difficult to fully reflect the degradation law of electric spindles, xLSTM is introduced to realize multi-dimensional influence factor fusion, improving the adaptability of the model to complex conditions; 2) Through the Hi-Lo attention mechanism, local details and global trends are captured, solving the problem of important information loss in long time series data. The present application can accurately predict the life of the electric spindle, help equipment maintenance personnel identify fault risks in advance, optimize maintenance strategies, and realize efficient management of the whole life cycle of the equipment. 3) By improving iTransformer, it can not only output the life of the electric spindle, but also evaluate the change range of the life, and quantify the sensitivity of variables to the length of life.

[0084] The steps include: 1: obtaining electric spindle running state original data, including the vibration signal, temperature data and current data of the electric spindle to be predicted within a period of time; in addition to the above data, multi-dimensional influence factor data such as device load condition, environmental temperature and running time are also obtained for joint analysis. 2: Data preprocessing and feature analysis are performed on the collected data, dynamic feature correlation analysis method is used to analyze the correlation of multiple variables, by comparing the feature correlation degree at different time points, the change trend of the correlation degree between variables with time is understood, and the influence factors with greater correlation with the life of the electric spindle are selected as variables for subsequent life prediction. 3: The xLSTM-HiLo method is used to establish the multi-dimensional joint learning electric spindle life prediction model. 4: The training data is input into the improved iTransformer model for training and real-time life prediction. 5: Calculate the prediction error, if the prediction error exceeds the set prediction error value, improve the life prediction model and repeat step 4. 6: According to the life prediction result of the model, the maintenance period of the electric spindle is reasonably arranged. As shown in the figure, it specifically includes: Figures 1-3

[0085] I. Obtain electric spindle running state original data, including the vibration signal, temperature data and current data of the electric spindle to be predicted within a period of time; in addition to the above data, multi-dimensional influence factor data such as device load condition, environmental temperature and running time are also obtained for joint analysis.

[0086] ​II. Data preprocessing and feature analysis of collected data. Due to the long service life of the motorized spindle, it is difficult to obtain its full life cycle data. There are problems such as serious data missing, difficulty in ensuring the consistency of filled data and original data, unreasonable variable weight distribution, etc. By constructing a complementary variable cluster and calculating the gray dynamic correlation degree, the missing data is screened and filled, the data integrity is improved, and the quality of the filled data under the condition of small sample is ensured. The correlation analysis of multiple variables is carried out by using the complementary-gray dynamic feature correlation analysis method. By constructing a complementary variable cluster, the trend of the correlation degree between missing variables with time is understood. Since there are missing variables in the collected data , such as vibration data, there are other variable sets , such as temperature, load, current data, etc. Calculate the gray dynamic correlation degree of the alternative variables and the missing variables , the formula is:

[0087] ;

[0088] Among them, , is the resolution coefficient, is the comparison sequence number, is the time sequence. The selected variable constitutes a complementary variable cluster , where , the dynamic weight formula of the complementary variable is:

[0089] ;

[0090] Among them, is the dynamic weight of the complementary variable , is the gray dynamic correlation degree of the complementary variable . Then the filling value of the missing variable at time is . In order to verify whether the filled data meets the correlation consistency, the deviation must meet the formula:

[0091] ;

[0092] The filled data is used as the input of the life prediction model of the motorized spindle of the numerical control machine tool.

[0093] Through the above preprocessing steps, the quality of the data can be effectively improved, the noise interference can be reduced, and the expression ability of the features can be enhanced, thereby providing high-quality input data for subsequent model training and prediction. The preprocessed data will be input into the xLSTM module for time series feature extraction, and the Hi-Lo attention mechanism will be further combined to capture multi-scale features and improve the accuracy of prediction.

[0094] III. Adopting the xLSTM-HiLo method, a multi-dimensional joint learning electric spindle life prediction model is established. The extended memory mechanism of xLSTM is used to extract and analyze the running data of the electric spindle, thereby establishing a health state prediction model of the electric spindle. The running data of the electric spindle is divided into multi-dimensional features such as vibration signal, temperature, speed and load, and the vibration spectrum feature, time domain statistical feature and temperature change trend are selected as key analysis factors. Through the memory enhancement ability of xLSTM, the model can effectively capture the long-term dependence and nonlinear dynamic change in the degradation process of the electric spindle, thereby realizing accurate prediction of the remaining useful life of the electric spindle. The input gate is introduced into the multi-head attention mechanism. When the input gate is processed through the multi-head attention mechanism, the calculation of the input gate is no longer dependent on simple linear transformation and sigmoid function, but through the multi-head attention mechanism to dynamically select and weight the input information. Specifically, in the gating process, the input gate is processed through the multi-head attention mechanism, and its formula is:

[0095] ;

[0096] In the formula: 、 、 query, key and value, is the input data, 、 、 is the weight matrix that maps the input sequence to query, key and value, and the attention weight is calculated by the formula:

[0097] ;

[0098] In the formula: is the attention score matrix, is the probability distribution function, is the scaling factor. The output of each head is obtained by weighted summation, and the outputs of all heads are connected to obtain the final output of the input gate as shown in the formula:

[0099] ;

[0100] In the formula: is the activation function, is the input matrix of the input gate, is the output projection matrix of the input gate. The forget gate is calculated as:

[0101] ;

[0102] where, is the input matrix of the forget gate, is the output projection matrix of the forget gate. This is the calculation process of the two gates (input gate and forget gate) under the assignment. Another path directly enters the state gate and output gate, and the state gate, forget gate are as shown in the formula:

[0103] ;

[0104] ;

[0105] where, is the normalization processing of , and are the weight matrices of the state gate and the output gate respectively, and are the bias vectors of the state gate and the output gate respectively.

[0106] After calculating the gating information, the sLSTM (single gate long short-term memory network) integrates the time sequence information through the enhanced state update mechanism. The state includes the update of the cell state and the hidden state, and considers the spatial information, and the formula is:

[0107] ;

[0108] ;

[0109] where, is the state of the current time step, , , are the outputs of the forget gate, the input gate, and the output gate respectively, is the new information of the current time, is the hidden state of the current time, is the normalized feature, is the scaled feature, is the activated input feature, is the normalization function, is the projection matrix, is the activation function, and the output feature of this part is obtained by linear transformation .

[0110] In mLSTM (Matrix Long Short-Term Memory Network), layer normalization is first applied to eliminate differences in feature scales, and feature amplitudes are adjusted using projection factors. Branches are then processed in parallel. In this case, the formulas for calculating the query, key, and value are:

[0111] ;

[0112] In the formula: , , They are respectively in Queries, keys, and values. The feature vector at the current position. , , These are the weight matrices for the query, key, and value, respectively. Size of the hidden layer , , These are the bias terms for the query, key, and value, respectively.

[0113] The formulas for calculating the input gate and forget gate through attention projection are as follows:

[0114] ;

[0115] In the formula: , , These are query, key, and value, respectively. , , They are respectively in Queries, keys, and values. , , The input sequence is respectively Weight matrices mapped to queries, keys, and values. The attention score matrix, Let be the probability distribution function. It is a scaling factor. For the first The output matrix of each attention head, This is for concatenating functions. For the final multi-head attention output matrix, For the final output, This is the final output projection weight matrix.

[0116] The stable state is updated through gating, and the covariance matrix and normalized state are also updated. Finally, the updated result is output. For the query input section of all headers:

[0117] ;

[0118] ;

[0119] wherein: is the state variable at the current time step , is the forget gate state at time step , is the input gate state at time step , is a hyper-parameter, is the input gate state at time step , and are the main memory state at the current time and the main memory state at the previous time, respectively, and are the normalized state at the current time and the normalized state at the previous time, respectively, and are the value and the interaction information of the bond, respectively, is the hidden state at time step is the weight matrix of the output gate, is the bias vector of the output gate, is the query vector, denotes diagonal matrix.

[0120] The application group normalization is applied, and the standard feature fusion is performed with the branch. The generated feature map is multiplied point by point with the Swish branch, wherein is the number of skip channels, and Swish is an activation function, and the output is obtained by dimension reduction projection.

[0121] After parallel processing of the sLSTM and the mLSTM, the weights of the two output features are fused, the advantages of the two kinds of LSTM are combined, and the processing capability for the time sequence features is enhanced, as shown in the formula:

[0122] ;

[0123] wherein: is the fused feature representation. , are the weights corresponding to the sLSTM and the mLSTM, respectively. , are the feature representations corresponding to the sLSTM and the mLSTM, respectively. is a layer normalization function.

[0124] Four, build a local detail and global trend HiLo attention mechanism. HiLo attention mechanism by dividing the input features into high and low frequency two parts, respectively processing different levels of information, so as to optimize the model's attention to key features in the case of limited computing resources. High frequency part captures detailed information, while low frequency part focuses on global trends, through this hierarchical processing, the model can more effectively allocate computing resources, avoid task overload problem. HiLo attention mechanism draws on the hierarchical processing mode of human visual system, can automatically filter out important information and ignore secondary information in complex data environment, so as to improve the sensitivity and prediction accuracy of the model to key features. In Hi-Lo attention mechanism, the total number of heads is allocated in proportion The high frequency branch captures local features through window attention. Given the input , where is the relative position bias, B represents the batch size, T represents the length of the input sequence, and D represents the feature dimension. Reshape it to , assuming , convert one-dimensional time series to two-dimensional format, so that the module can capture the spatio-temporal features of the data. The number of windows in the height and width directions is as follows:

[0125] ;

[0126] In the formula: , are the number of windows in height and width, , are the height and width of the input data, is the size of the window.

[0127] Reshape the input to adapt to the window size. Then, obtain QKV through the high frequency projection matrix of the head, as shown in the formula:

[0128] ;

[0129] In the formula: is the concatenated query key value matrix, is the high frequency projection matrix, is the windowed input feature.

[0130] At the end of the high frequency part, calculate the score of the attention mechanism, as follows:

[0131] ;

[0132] In the low frequency part, first perform query projection:

[0133] ;

[0134] where: is the high-pass filtered feature of the layer feature, is the high-pass filtered feature of the attention head, , , are the query matrix, key matrix, value matrix of the th head respectively, is the transpose matrix of , is the key vector dimension of the th head, is the projection of the query matrix, is the query weight matrix of the th head of the low frequency branch. The shape of the query is reshaped to fit the size of the low frequency part, and the joint projection of the key and value is performed as shown in the formula:

[0135] ;

[0136] where: and are the concatenated key and value matrix respectively, is the KV weight matrix of the th head of the low frequency branch.

[0137] The shape of the KV is reshaped, a linear layer is used to calculate K and V simultaneously, and slicing is performed on the third dimension. At the end of the low frequency part, the attention score is calculated, and the formula is as follows:

[0138] ;

[0139] where: is the low-pass filtered feature of the layer feature, is the low-pass filtered feature of the attention head, is the key vector dimension of the th head.

[0140] After explaining the part of the Hi-Lo attention mechanism, the input of the feature fusion is mapped to the encoder through the embedding layer. After passing through the attention layer and adjusting the dimension:

[0141] ;

[0142] where: is the final feature after fusion, , are the processing results of the high frequency branch and the low frequency branch on respectively. The output of is output through the feedforward neural network and the linear layer. The decoder is roughly the same as the encoder, and the final output is , where This is the final output of the feedforward neural network and the linear layer. This is the final output of the decoder.

[0143] V. Scenario Generator Embedded with iTransformer Model. Input basic operating conditions are obtained through data integration. Environmental factors The formula is: , .

[0144] in, For the coefficients of the parameter matrix, for The fused feature vector at time step For bias terms, The surface temperature of the equipment. The thermal resistivity of the equipment. This refers to the device's power consumption.

[0145] Basic working conditions Environmental factors and features extracted by xLSTM Basic working conditions Environmental factors Encoding as latent variables The formula is:

[0146] , ;

[0147] in, The scenario is generated through diffusion, using a random factor. Simultaneously, physical constraints are strengthened to limit the range of operating conditions; the formula is as follows:

[0148]

[0149] in, Total deviation for Operating conditions at any given time , These are the upper and lower thresholds for the optimal predicted lifetime of the electric spindle, respectively, and N is the number of electric spindle samples.

[0150] Maximum entropy regularization encourages the generation of different scenarios, as shown in the formula:

[0151] ;

[0152] in, For diversified loss functions, This represents the probability of different operating conditions occurring. (Through...) + Optimization generator, make sure to generate a reasonable variety of scenarios, after generating multiple scenarios, the basic working condition and environmental factors Combined with the working condition after generating disturbance, the formula is:

[0153] ;

[0154] Where, is the working condition after disturbance at time t, is the reference working condition at time t, is the environmental factor at time t, is the element-wise multiplication. Generate group working condition-environment combination Spindle life obeys a parametric distribution, and its conditional probability density function formula is: ;

[0155] ;

[0156] For the first scenario, the expected formula of predicting life is:

[0157] ;

[0158] Where, is the life of the spindle, is the life of the spindle in the first scenario, is the mean of the logarithmic life under the working condition-scenario , is the standard deviation of the natural logarithm of life. Then the amplitude formula of life change is:

[0159] , where is the life under the standard working condition. Spindle life

[0160] is a function of the basic working condition and environmental factors , that is . Let , that is, life can be decomposed as:

[0161] ;

[0162] Where, is the expected life, is the output change part contributed by the basic working condition, ​​​The output change part contributed by environmental factors, The output part contributed by the interaction of base working conditions and environmental factors, d is the number of dimensions of the feature space. For each variable The formula of its sensitivity to life:

[0163] ;

[0164] ;

[0165] Wherein, The variance of the main shaft life, The life under the i-th working condition-environment combination, The life expectancy, The single input The sensitivity to the variance The variance under the i-th condition, The number of motor shaft samples. Six, input the training data into the iTransformer model for training and real-time life prediction. The final output is obtained by the forward neural network and linear layer output from the four , through multi-layer iTransformer compilation, the encoder output is obtained, the formula is:

[0166] , and the obtained output is input into the full connection layer, the formula is: , wherein, and are the weights and biases of the full connection layer respectively. The result is input into the regression layer, and the life prediction result is output, the formula is: , wherein and are the weights and biases of the regression layer respectively. The variance elasticity of the predicted life of each group of scenarios is calculated , wherein, The variance of the predicted life, The corresponding different perturbations. In addition, the results of parameter setting and ablation experiment are shown in Table 1 and Table 2. Wherein α is the learning rate, ɛ is the constant for numerical stability, AdamW is the optimizer for parameter tuning, and the loss function

[0167] Table 1 Parameter setting

[0168] Table 2 Ablation experiment

[0169]

[0170] ​​In summary, the xLSTM-HiLo attention mechanism and the improved iTransformer are combined to realize the accurate prediction of the life of the electric spindle. Through the hierarchical feature extraction of the HiLo attention mechanism and the extended memory capacity of the xLSTM, the accuracy and robustness of the life prediction are significantly improved. Through the improved iTransformer for life prediction, the remaining useful life of the electric spindle is effectively quantified, and the variance elasticity is obtained. A reasonable maintenance strategy is proposed to improve the efficiency and economy of the equipment operation, and to provide a scientific basis for the intelligent operation and maintenance of industrial equipment.

[0171] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An improved iTransformer-based industrial mother machine electric spindle life prediction model method, characterized in that, The application relates to a method for predicting the service life of an industrial mother machine electric spindle. The method comprises the following steps: Preprocessing the running state original data and multi-dimensional influence factor data of the industrial mother machine electric spindle, and using grey dynamic correlation analysis to complete the missing data to obtain model data; Using parallel processing of an extended long short-term memory network to extract and analyze the features of the model data, and fusing the output features to obtain time sequence features; Processing the time sequence features through an attention mechanism to realize dual capture of local details and global trends, and outputting the fused features; Using iTransformer to compile and predict the fused features to obtain the service life prediction result of the industrial mother machine electric spindle; When the prediction error exceeds the set prediction error value, the model parameters of the iTransformer are fine-tuned to output an optimal model, and the service life of the industrial mother machine electric spindle is predicted based on the optimal model; 2. The improved iTransformer-based industrial mother machine electric spindle life prediction model method according to claim 1, characterized in that, Embedding a scenario generator in the iTransformer, and optimizing the scenario generator through total deviation and diversity loss functions to cover multi-scenario disturbance data of different working conditions and environmental factors, and generating the service life change amplitude under each scenario; and using decomposition processing of the service life of the industrial mother machine electric spindle to obtain the sensitivity of each variable to the service life. The preprocessing of the running state original data and multi-dimensional influence factor data of the industrial mother machine electric spindle, and the use of grey dynamic correlation analysis to complete the missing data to obtain model data comprises the following steps: Acquiring and preprocessing the running state original data and multi-dimensional influence factor data of the industrial mother machine electric spindle; Calculating the grey dynamic correlation degree between the candidate variables and the missing variables in the running state original data and the multi-dimensional influence factor data, and screening out variables with a grey dynamic correlation degree greater than a preset value to form a complementary variable cluster; According to the grey dynamic correlation degree of the complementary variables in the complementary variable cluster, calculating the dynamic weight of the complementary variables, and combining the complementary variables to calculate the filling value of the missing variables; 3. The improved iTransformer-based industrial mother machine electric spindle life prediction model method according to claim 1, wherein, Using the filling value to complete the running state original data and the multi-dimensional influence factor data to obtain model data. The use of parallel processing of an extended long short-term memory network to extract and analyze the features of the model data, and the fusion of the output features to obtain time sequence features comprises the following steps: Configuring a parallel processing structure of an extended long short-term memory network comprising a single-gate long short-term memory network and a matrix long short-term memory network; 4. The improved iTransformer-based industrial mother machine electric spindle life prediction model method according to claim 3, characterized in that, Using the parallel processing structure of the extended long short-term memory network to capture the long-term dependence relationship and nonlinear dynamic change in the degradation process of the electric spindle from the model data, and fusing and outputting the time sequence features. The use of the parallel processing structure of the extended long short-term memory network to capture the long-term dependence relationship and nonlinear dynamic change in the degradation process of the electric spindle from the model data, and the fusion and output of the time sequence features comprises the following steps: Using a multi-head attention mechanism to dynamically select and weight the input information and the corresponding attention weight; using weighted summation to obtain the output of each head in the multi-head attention mechanism, and connecting the outputs of all heads to obtain gating information; The model data is processed through the combination of the gating information and the parallel processing structure of the single-gate long short-term memory network and the matrix long short-term memory network to obtain a feature representation corresponding to the single-gate long short-term memory network and a feature representation corresponding to the matrix long short-term memory network; The feature representation corresponding to the single-gate long short-term memory network and the feature representation corresponding to the matrix long short-term memory network are fused through a weighted summation manner to obtain the time sequence feature.

5. The improved iTransformer-based industrial mother machine electric spindle life prediction model method according to claim 1, wherein, The hierarchical information processing of the time sequence feature through the attention mechanism realizes the dual capture of local details and global trends, and outputs the fused feature, which includes: The attention mechanism containing the local details and the global trends is constructed, and the local feature and the global trend feature are captured based on the high-frequency branch and the low-frequency branch of the attention mechanism; The local feature and the global trend feature are mapped into the encoder through an embedding layer, and the dimension is adjusted through an attention layer to output the fused feature.

6. The improved iTransformer-based industrial mother machine electric spindle life prediction model method according to claim 1, wherein, The fused feature is compiled and predicted through the iTransformer to obtain the industrial mother machine electric spindle life prediction result, which includes: The fused feature is compiled through the iTransformer to obtain the encoder output; The encoder output is predicted through a fully connected layer and a regression layer to obtain the industrial mother machine electric spindle life prediction result.

7. The improved iTransformer-based industrial mother machine electric spindle life prediction model method according to claim 1, characterized in that, The scenario generator is embedded in the iTransformer, and the scenario generator is optimized through the total deviation degree and the diversification loss function to cover multi-scenario disturbance data of different working conditions and environmental factors to generate the life change amplitude in each scenario, which includes: The scenario generator is embedded in the iTransformer model; the working conditions and the threshold values at all times are obtained to calculate the total deviation degree; the diversification loss is calculated based on the probability of occurrence of different working conditions; the scenario generator is optimized through the diversification loss and the total deviation degree; After generating the multi-scenario environment through the scenario generator, the disturbed working conditions are generated by combining the basic working conditions and the environmental factors; the working condition environment combination is generated according to the disturbed working conditions; Based on the log life mean, the standard deviation of the natural logarithm of the electric spindle life, and the life of the electric spindle in the corresponding scenario under the working condition environment combination, the prediction life of the electric spindle in any scenario is calculated; Based on the prediction life of the electric spindle in any scenario and the prediction life of the electric spindle under the standard working condition, the amplitude of the life change is calculated.

8. The improved iTransformer-based industrial mother machine electric spindle life prediction model method according to claim 1, wherein, The decomposition processing of the industrial mother machine electric spindle life is used to obtain the sensitivity of each variable to the life, which includes: The industrial mother machine electric spindle life is decomposed to obtain the life expectation; Based on the life expectation, the electric spindle life under the working condition environment combination, the variance of the spindle life, and the variance of each variable, the sensitivity of each variable to the life is calculated.

9. The improved iTransformer-based industrial mother machine electric spindle life prediction model method according to claim 8, characterized in that, The calculation formula of the sensitivity of each variable to the life is: ； ； In the formula, The variance of the spindle life. For the first Lifespan under various operating conditions and environmental combinations For life expectancy, For a single input The difference between the other party The degree of sensitivity, In order to be in The variance under the given conditions, where N is the number of electric spindle samples.

10. The improved iTransformer-based industrial mother machine electric spindle life prediction model method according to claim 1, characterized in that, It also includes: According to the variance of the industrial mother machine electric spindle prediction life and the corresponding different disturbances, the variance elasticity of the prediction life of each group of scenarios is calculated.

Images