CN-121256736-B - Fault prediction method and system for correcting deviation of laser die-cutting and winding integrated machine

CN121256736BCN 121256736 BCN121256736 BCN 121256736BCN-121256736-B

Abstract

The invention discloses a fault prediction method and a system for correcting a laser die-cutting winding integrated machine, which belong to the technical field of intelligent manufacturing and industrial equipment fault prediction and comprise the steps of collecting and preprocessing multidimensional sensor data of a correcting mechanism; the method and the system for predicting the faults of the laser die-cutting winding integrated machine, disclosed by the invention, can reduce prediction errors, improve early warning advance and accuracy, reduce redundant calculation, improve model efficiency to adapt to industrial deployment, realize potential fault identification early warning, reduce equipment fault rate and shutdown risk and have remarkable economic and social benefits.

Inventors

PENG ZHANGLIN
Zhao Shantao
ZHANG QIANG
YANG SHANLIN
LU XIAONONG
XU JIAWEN
ZHU KEYU

Assignees

合肥工业大学

Dates

Publication Date: 20260508
Application Date: 20250903

Claims (7)

1. The fault prediction method for correcting the laser die-cutting and winding integrated machine is characterized by comprising the following steps of: s1, receiving original offset time series data acquired by a multidimensional sensor of a deviation correcting mechanism and preprocessing the original offset time series data; s2, evaluating equipment stability, constructing sliding window samples, calculating variability coefficients of sequences in each sliding window to quantify local fluctuation level, setting a dynamic threshold, marking the sliding window samples as a high-risk running state when the variability coefficient values exceed the dynamic threshold, and screening abnormal time periods; s3, constructing an improved transducer-based prediction model, wherein the model comprises an input embedded layer, a dynamic expansion attention mechanism, an encoder and a decoder, and predicting the preprocessed data input model corresponding to the abnormal time period obtained by S2 screening; S31, constructing an input embedding layer of feature codes, linearly connecting and mapping 5-dimensional original features of a sensor into 512-dimensional embedding vectors, generating position codes by calculating sine and cosine functions of different frequencies, embedding the 512-dimensional embedding vectors, extracting and embedding time features, and mapping the time features into the 512-dimensional embedding vectors to finish construction of the input embedding layer; s32, constructing a dynamic expansion attention mechanism, dividing an input sequence into a plurality of non-overlapping processing sections through a fixed sliding window, endowing each attention head with different expansion rates, extracting local and global model features by the different attention heads, and summarizing and splicing the local and global model features into a multi-head dynamic attention mechanism; The deduction process for constructing the dynamic expansion attention mechanism is as follows: a. Given a set of time sequences Wherein Is the length of the sequence and, Is the dimension of the model, and the fixed sliding window is of the size of And is divided into Non-overlapping processing segments; (13) (14) b. Projecting the time series to have Dimension of In the attention header, the formula is as follows: (15) Wherein, the In order to query the matrix, In the form of a matrix of keys, In the form of a matrix of values, For the corresponding attention head(s), Feature mapping for a particular head to be learnable; c. for each sliding time period And each head Dispensing an expansion rate The calculation amount is reduced by skipping the sampling time index, and the formula is as follows: (16) Wherein, the Is the first Time period of (a) The set of valid time intervals after the individual head skip samples, In order to process the time frame of the segment, To the expansion rate The generated equidistant time steps; d. time series segment The calculated attention of the user is collected and spliced with all non-overlapped time periods and attention heads to form a multi-head dynamic attention mechanism; (17) (18) Wherein, the 、 And Respectively represent for time series fragments Index set selected in (a) Is represented by a query, a key, and a value, For the output of the multi-headed dynamic attention mechanism, Outputting a projection matrix; s33, constructing an encoder, wherein the encoder comprises an embedded layer, a dynamic expansion attention layer, an encoder layer and a normalization layer; S34, constructing a decoder, wherein the decoder is formed by stacking a plurality of decoder layers, and each layer comprises a mask type MDDA module, a cross attention module and a normalization and output mapping layer; S4, if the future deviation track trend predicted by the model approaches or exceeds a set deviation correction threshold, generating an early warning signal, and marking the early warning signal as a potential fault state; s5, adopting a dynamic expansion attention mechanism optimization and segmentation reasoning mechanism, and deploying the model on an edge computing node or an industrial server for prediction.
2. The fault prediction method for correcting a laser die-cutting and winding integrated machine according to claim 1, wherein the specific steps of S1 are as follows: s11, data cleaning is carried out, outlier data is removed by adopting a 6 sigma normal distribution assumed outlier detection method, and the specific process is as follows: a. Setting the original time series data as The mean and standard deviation of the population are calculated as follows: (1) (2) Wherein, the For the total length of the time series, Is the mean value of the two values, Is the standard deviation; b. constructing an abnormal threshold interval, and setting a reserved interval as Wherein Is a constant; c. Judging and processing the abnormal data, and removing the abnormal data from the time sequence or performing interpolation restoration by using front-back average; s12, performing signal smoothing processing, adopting a Kalman filter to inhibit high-frequency noise and keep signal trend, wherein the specific process is as follows: a. initializing state estimation values Matrix of the coordination equation Let the initial value be , B. The recursion calculation of the smooth estimated value is divided into two stages of prediction and updating; The calculation formula of the state prediction stage is as follows: (3) (4) Wherein, the To at the same time Time of day based on All information at the moment and before, pair A predicted estimate of the time of day state, Is that The state estimate after the time of day has been updated, Is that Covariance matrix corresponding to the time prediction estimation value, Is that State estimation value after time update Is used for the co-variance matrix of (a), Covariance matrix of process noise; The calculation formula of the state updating stage is as follows: (5) (6) (7) Wherein, the Is that The kalman gain at the moment in time, In order to observe the covariance matrix of the noise, Is that The actual observations of the time of day sensor, Is that The state estimation value updated by Kalman filtering at the moment, Is that State estimation value after time update Covariance matrix of (2) C. after Kalman filtering smoothing, the output smoothed time sequence is 。
3. The fault prediction method for correcting a laser die-cutting and winding integrated machine according to claim 1, wherein the specific steps of S2 are as follows: S21, setting the length of the sliding window as Dividing the original time sequence into a plurality of continuous time periods, and calculating a mean value and a standard deviation for each time period, wherein the calculation formula is as follows: (8) (9) S22, calculating the variability coefficient in each time period window The calculation formula is as follows: (10) S23, judging the state of the equipment by comparing the distribution change of the variability coefficient values of all time periods, setting a dynamic threshold value by combining the working condition map and the operation history, marking the variability coefficient values as high-risk operation states when the variability coefficient values exceed the dynamic threshold value, and screening out abnormal time periods.
4. The fault prediction method for correcting a laser die-cutting and winding integrated machine according to claim 1, wherein a calculation formula for generating a position code is as follows: (11) Wherein, the To at the same time Time dimension Is used for the position-coding of the values of (c), To be about A function of the time of day, Is of the same dimension as The frequency parameter of the interest is set to be the same, For the dimension of the position-coding vector, Is an index variable; the calculation formula of the mapping time feature is as follows: (12) Wherein, the In order to obtain the time characteristic value after the coding, As the data of the original time unit, The maximum possible value of the current time unit is taken.
5. The fault prediction method for correcting the laser die-cutting winding all-in-one machine according to claim 1 is characterized in that the specific structure of the encoder in S33 is that an embedding layer maps original time sequence data and timestamp characteristics thereof to a high-dimensional representation space, the embedding operation comprises scalar embedding, time embedding and position embedding, a dynamic expansion attention layer introduces a dynamic expansion attention mechanism of S32, different expansion rates are distributed for different attention heads, attention scores are normalized by a Softmax function to generate attention weights, the encoder layer comprises an attention structure based on skip sampling of the different expansion rates and a feedforward network based on a one-dimensional convolution structure, the feedforward network comprises two 1-dimensional convolution layers, the middle is connected by an activation function, an output result is processed through residual connection and normalization operation, the encoder layer is repeatedly stacked twice, and the normalization layer is arranged at the tail end of an encoder module for normalization.
6. The fault prediction method for correcting the laser die-cutting winding integrated machine according to claim 1 is characterized in that the specific structure of the decoder in S34 is that a mask MDDA module introduces a causal shielding mechanism in attention calculation to realize autoregressive prediction modeling, a cross attention module fuses encoder output and decoder input based on a standard full-connection attention mechanism to establish a corresponding relation between a current predicted value and a historical representation, and a normalization and output mapping layer is used for stabilizing network output, wherein the normalization module is used for mapping the decoder output to a prediction target dimension space.
7. A fault prediction system for correcting a laser die-cutting and winding integrated machine, which is applied to a fault prediction method for correcting a laser die-cutting and winding integrated machine as claimed in any one of claims 1 to 6, and is characterized by comprising: The data preprocessing module is used for receiving and preprocessing original offset data from a sensor of the deviation correcting mechanism, constructing a sliding window sample based on a time sequence, eliminating high-frequency fluctuation and noise by adopting Kalman filtering, and introducing a missing value interpolation mechanism to ensure data continuity; The equipment stability evaluation module is used for calculating continuous variation coefficients of corresponding sequences in each sliding window, setting a dynamic threshold value, and marking a high-risk running state when the continuous variation coefficient values exceed the threshold value, so as to preliminarily screen abnormal time periods and reduce redundant calculation; The deviation rectifying track prediction module is constructed based on an improved transducer model and comprises an input embedded layer, a dynamic expansion attention mechanism, a Encoder-Decoder framework and a prediction objective function, wherein the input embedded layer fuses time indexes, sensor position codes and data values to form input features, the dynamic expansion attention mechanism distributes different expansion rates for each attention head, the model carries out long-range modeling on a receptive field with different scales on a time sequence, the Encoder framework takes historical deviation data as coding input, the deviation track of a plurality of time steps in the future is predicted by taking the past time sequence and the output of Encoder as the input of a Decoder, and the prediction objective function adopts weighted mean square error as an optimization target; the fault judging module judges faults according to the future deviation track predicted by the deviation correcting track predicting module, generates an early warning signal mark as a potential fault state if the predicted track trend approaches or exceeds a set deviation correcting threshold value, and feeds back the early warning signal mark to the upper control system in real time.

Description

Fault prediction method and system for correcting deviation of laser die-cutting and winding integrated machine Technical Field The invention relates to the technical field of intelligent manufacturing and industrial equipment fault prediction, in particular to a fault prediction method for correcting a laser die-cutting and winding integrated machine. Background Along with continuous optimization of the lithium battery manufacturing process, the laser die cutting and winding integrated machine becomes key equipment in battery cell production, and the running stability of the integrated machine directly influences the consistency and the yield of the battery cells. The deviation correcting mechanism in the device is used for adjusting the centering precision of the pole piece or the diaphragm material in the transmission process in real time so as to ensure the winding quality. However, under long-time high-speed operation, faults such as sensor drift, execution clamping stagnation, mechanism imbalance and the like easily occur in the deviation correcting mechanism, and once the deviation exceeds a threshold value, material deviation, rolling deviation or equipment shutdown are easily caused, so that production interruption and material waste are caused. At present, an industrial field often adopts a mode of triggering an alarm based on manual inspection or a fixed threshold to monitor the running state of equipment. The method depends on manual experience or static rules, can not effectively identify tiny abnormal movement of equipment before failure occurs, causes failure early warning hysteresis, and is difficult to meet the requirement of intelligent manufacturing on predictive maintenance of the equipment. In recent years, some researches try to introduce traditional machine learning algorithms (such as SVM and random forest) to perform modeling analysis on data of the deviation correcting sensor. However, these methods generally rely on a fixed feature extraction process, which is difficult to accommodate for the complex and varying industrial conditions and long-range dependence characteristics in the time series data. In addition, the traditional model has obvious bottlenecks in modeling capability, training efficiency and real-time performance when facing large-scale and high-dimensional industrial sensor data. In view of the above problems, studies have been started focusing on a time-series modeling method based on deep learning. For example, some documents attempt to predict the state of equipment by using an LSTM model, but because of its serial computing structure, training takes a long time and is difficult to process the efficient modeling requirement of millions of industrial data, while standard transformers have strong global modeling capability, but have O (n 2) as the computing complexity, and are easy to cause excessive resource consumption when processing long-time sequences. Therefore, how to combine the dynamic expansion sampling modeling concept, and to improve the model reasoning efficiency on the premise of ensuring the prediction accuracy, to construct a set of efficient, real-time and self-adaptive fault prediction method suitable for the deviation correcting mechanism of the laser die-cutting and winding integrated machine becomes a key problem to be solved urgently. Disclosure of Invention The invention aims to provide a fault prediction method and a system for correcting a laser die-cutting winding integrated machine, which are based on a high-efficiency transducer structure and can be used for accurately modeling and predicting risks on offset time series data acquired by a correction sensor. In order to achieve the above purpose, the invention provides a fault prediction method for correcting a laser die-cutting and winding integrated machine, which comprises the following steps: s1, receiving original offset time series data acquired by a multidimensional sensor of a deviation correcting mechanism and preprocessing the original offset time series data; S2, evaluating equipment stability, constructing sliding window samples, calculating variability coefficients (CoefficientofVariation, CV) of sequences in each sliding window to quantify local fluctuation level, setting a dynamic threshold, marking as a high-risk running state when a CV value exceeds the dynamic threshold, and screening abnormal time periods; S3, constructing an improved transducer-based prediction model, wherein the model comprises an input embedded layer, a dynamic expansion attention mechanism (Multi-ScaleDynamicDilatedAttention), an encoder and a decoder, and the preprocessed data is input into the model for prediction; S4, if the future deviation track trend predicted by the model approaches or exceeds a set deviation correction threshold, generating an early warning signal, and marking the early warning signal as a potential fault state; s5, adopting a dynamic expansion attention mechanism optimizatio