A fault prediction method and system for deviation correction of a laser die-cutting and winding all-in-one machine

Abstract

The application discloses a kind of for laser die cutting winding integrated machine deviation correction fault prediction method and system, belong to intelligent manufacturing and industrial equipment fault prediction technical field, including the multi-dimensional sensor data of deviation correction mechanism and pre-processing;Through sliding window calculation sequence variability coefficient, combine dynamic threshold to filter high-risk abnormal time period;Offset trajectory prediction is carried out based on improved Transform prediction model is constructed;Fault judgment is carried out, and model is deployed in edge computing device or server to realize efficient prediction;The application adopts the above-mentioned one for laser die cutting winding integrated machine deviation correction fault prediction method and system, can reduce prediction error, 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 failure rate and downtime risk, with significant economic and social benefits.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial equipment fault prediction technology, and in particular to a fault prediction method for correction of laser die-cutting and winding integrated machines. Background Technology

[0002] With the continuous optimization of lithium battery manufacturing processes, laser die-cutting and winding integrated machines have become key equipment in cell production, and their operational stability directly affects the consistency and yield of cells. The alignment mechanism in this type of equipment is used to adjust the alignment accuracy of electrode or separator materials during transmission in real time to ensure winding quality. However, under prolonged high-speed operation, the alignment mechanism is prone to malfunctions such as sensor drift, actuator jamming, and mechanism misalignment. Once the deviation exceeds the threshold, it can easily lead to material misalignment, winding deviation, or equipment shutdown, causing production interruptions and material waste.

[0003] Currently, industrial sites often monitor equipment operating status using methods based on manual inspections or fixed threshold-triggered alarms. These methods rely on human experience or static rules, and cannot effectively identify subtle anomalies in equipment before a failure occurs, resulting in delayed fault warnings and failing to meet the predictive maintenance requirements of smart manufacturing.

[0004] In recent years, some studies have attempted to introduce traditional machine learning algorithms (such as SVM and random forest) to model and analyze data from bias correction sensors. However, these methods typically rely on fixed feature extraction processes, making them difficult to adapt to complex and variable industrial conditions and the long-range dependencies in time-series data. Furthermore, when faced with large-scale, high-dimensional industrial sensor data, traditional models exhibit significant bottlenecks in modeling capabilities, training efficiency, and real-time performance.

[0005] To address the aforementioned issues, existing research has begun to focus on time series modeling methods based on deep learning. For example, some literature has attempted to use LSTM models to predict equipment status, but due to their serial computation structure, training is time-consuming and difficult to handle the efficient modeling requirements of millions of industrial data points; while the standard Transformer, although possessing powerful global modeling capabilities, has a computational complexity of O(n^2). 2 This can easily lead to excessive resource consumption when processing long-term sequences.

[0006] Therefore, how to combine the idea of ​​dynamic expansion sampling modeling to improve the model inference efficiency while ensuring prediction accuracy, and to build a set of efficient, real-time, and adaptive fault prediction methods suitable for the correction mechanism of laser die-cutting and winding integrated machines, has become a key problem that urgently needs to be solved. Summary of the Invention

[0007] The purpose of this invention is to provide a fault prediction method and system for correction of laser die-cutting and winding integrated machines. Based on a high-efficiency Transformer structure, it can accurately model and predict risks from the offset time series data collected by the correction sensor.

[0008] To achieve the above objectives, the present invention provides a fault prediction method for correction in a laser die-cutting and winding integrated machine, comprising the following steps:

[0009] S1. Receive and preprocess the raw offset time series data collected by the multi-dimensional sensor of the correction mechanism;

[0010] S2. Conduct equipment stability assessment, construct sliding window samples, calculate the coefficient of variation (CV) of the sequence within each sliding window to quantify the local fluctuation level, set a dynamic threshold, and mark the operation as high-risk when the CV value exceeds the dynamic threshold, and screen abnormal time periods.

[0011] S3. Construct an improved Transformer prediction model, which includes an input embedding layer, a multi-scale dynamic dilated attention mechanism, an encoder, and a decoder, and input the preprocessed data into the model for prediction.

[0012] S4. If the trend of the future deviation trajectory predicted by the model approaches or exceeds the set correction threshold, an early warning signal is generated and marked as a potential fault state.

[0013] S5. Employs a dynamic expansion attention mechanism and a segmented inference mechanism to deploy the model on edge computing nodes or industrial servers for prediction.

[0014] Preferably, the specific steps of S1 are as follows:

[0015] S11. Perform data cleaning, using outlier detection methods based on the 6σ normal distribution assumption to remove outlier data. The specific process is as follows:

[0016] a. Set the original time series data as X i =x 1 ,x 2 ,...,x t ,...,x T The population mean and standard deviation are calculated using the following formulas:

[0017]

[0018] Where T is the total length of the time series, u is the mean, and σ is the standard deviation;

[0019] b. Construct an anomaly threshold interval, and set the retention interval as [μ-kσ,μ+kσ], where k is a constant;

[0020] c. Identify and process outlier data, removing outlier data from the time series or using interpolation to repair it using the average of the previous and subsequent data.

[0021] S12. Perform signal smoothing processing, using a Kalman filter to suppress high-frequency noise and preserve signal trend. The specific process is as follows:

[0022] a. Initialize state estimates With the coequation matrix P, let the initial value be... P0 = 1;

[0023] b. The smoothed estimate is calculated recursively, and is divided into two stages: prediction and update.

[0024] The calculation formula for the state prediction stage is:

[0025]

[0026] P t|T-1 =P t-1 +Q (4)

[0027] in, Given the information from time t-1 and earlier, this is a predicted estimate of the state at time t. Let P be the updated state estimate at time t-1. t|t-1 Let P be the covariance matrix corresponding to the predicted estimate at time t. t-1 The updated state estimate at time t-1 The covariance matrix of , where Q is the covariance matrix of the process noise;

[0028] The formula for calculating the state update phase is:

[0029]

[0030] P t = (1-K) t )·P t|t-1 (7)

[0031] Among them, K t Let x be the Kalman gain at time t, R be the covariance matrix of the observation noise, and x be the Kalman gain at time t. t The actual observation value of the sensor at time t. Let P be the state estimate after Kalman filtering at time t, and P be the updated state estimate at time t. covariance matrix

[0032] c. After Kalman filtering and smoothing, the output smoothed time series is:

[0033] Preferably, the specific steps of S2 are as follows:

[0034] S21. Set the sliding window length to W, divide the original time series into multiple consecutive time periods, and calculate the mean and standard deviation for each time period. The calculation formula is as follows:

[0035]

[0036] S22. Calculate the coefficient of variation (cv) within each time window. The calculation formula is as follows:

[0037]

[0038] S23. By comparing the distribution changes of the coefficient of variation values ​​in different time periods, the status of the equipment is determined. A dynamic threshold is set by combining the operating condition map and the operating history. When the coefficient of variation value exceeds the dynamic threshold, it is marked as a high-risk operating state, and abnormal time periods are screened out.

[0039] Preferably, the specific steps of S3 are as follows:

[0040] S31. Construct the input embedding layer for feature encoding. Linearly connect and map the 5-dimensional original features of the sensor into a 512-dimensional embedding vector. Generate position codes by calculating sine and cosine functions of different frequencies and embed them into the 512-dimensional embedding vector. Extract and embed time features. Map the time features to the 512-dimensional embedding vector to complete the construction of the input embedding layer.

[0041] S32. Construct a dynamic expansion attention mechanism, which divides the input sequence into multiple non-overlapping processing segments through a fixed sliding window, assigns different expansion rates to each attention head, extracts local and global model features from different attention heads, and summarizes and splices them into a multi-head dynamic attention mechanism.

[0042] S33. Construct the encoder, including the embedding layer, the dynamically expanded attention layer, the encoder layer, and the normalization layer;

[0043] S34. Construct a decoder, which consists of several stacked decoder layers. Each layer includes a masked MDDA module, a cross-attention module, and a normalization and output mapping layer.

[0044] Preferably, the formula for generating the location code is as follows:

[0045]

[0046] in, Let f(t) be the position value encoded in dimension i at time t. (i) Let be a function of time t. Here, d is the frequency parameter associated with dimension i, d is the dimension of the position encoding vector, and k is the index variable;

[0047] The formula for calculating the mapping time feature is:

[0048]

[0049] Among them, time feature The encoded time feature value is represented by Time Unit Value, which is the original time unit data, and Max Unit Value, which is the maximum possible value for the current time unit.

[0050] Preferably, the derivation process of S32 constructing the dynamic expansion attention mechanism is as follows:

[0051] a. Given a set of time series Where L is the sequence length, d is the model dimension, the sliding window size is fixed at S, and it is divided into N non-overlapping processing segments;

[0052]

[0053] b. Project the time series onto a plane with d h =For the H attention heads in the d / H dimension, the formula is as follows:

[0054]

[0055] Among them, Q h For querying the matrix, K h V is the key matrix. h Let h be the value matrix, and h be the corresponding attention head. To be learnable, feature mappings are applied to a specific head.

[0056] c. For each sliding time interval X (i) and each head h i Assign an expansion rate r h The computational cost is reduced by skipping the sampling time index, as shown in the following formula:

[0057]

[0058] in, Let [0, S) be the set of valid time intervals for skip sampling of the h-th head in the i-th time interval, and let [0, S) be the time range of the processing segment, {0, r h ,2r h ...} represents the expansion rate r hGenerate equally spaced time steps;

[0059] d. Time series segment X (i) The computational attention mechanism aggregates and splices together all non-overlapping time periods and attention heads to form a multi-head dynamic attention mechanism.

[0060]

[0061] in, and These represent time series segments X. (i) The selected set of indices The query, key, and value representations, MDDA(X) is the output of the multi-head dynamic attention mechanism. This is for outputting the projection matrix.

[0062] Preferably, the encoder structure in S33 is as follows: the embedding layer maps the original time series data and its timestamp features to a high-dimensional representation space, and the embedding operation includes a combination of scalar embedding, temporal embedding and positional embedding; the dynamically expanded attention layer introduces the dynamic expanded attention mechanism of S32, assigns different expansion rates to different attention heads, and uses the Softmax function to normalize the attention scores to generate attention weights; the encoder layer includes an attention structure based on skip sampling with different expansion rates and a feedforward network based on a one-dimensional convolutional structure. The feedforward network contains two one-dimensional convolutional layers connected by an activation function in the middle. The output results are processed by residual connection and normalization operation. The encoder layer is stacked twice; the normalization layer is set at the end of the encoder module to perform normalization processing.

[0063] Preferably, the specific structure of the decoder in S34 is as follows: the masked MDDA module introduces a causal masking mechanism in the attention calculation to realize autoregressive prediction modeling; the cross-attention module is based on the standard fully connected attention mechanism, which fuses the encoder output and the decoder input to establish the correspondence between the current prediction value and the historical representation; in the normalization and output mapping layer, the normalization module is used to stabilize the network output, and the linear mapping layer maps the decoder output to the prediction target dimension space.

[0064] A fault prediction system for correction in a laser die-cutting and winding integrated machine includes:

[0065] The data preprocessing module receives raw offset data from the sensors of the correction mechanism and preprocesses it, constructs a time-series-based sliding window sample, uses Kalman filtering to eliminate high-frequency fluctuations and noise, and introduces a missing value interpolation mechanism to ensure data continuity.

[0066] The equipment stability assessment module calculates the continuous coefficient of variation of the corresponding sequence within each sliding window, sets a dynamic threshold, and marks a high-risk operating state when the continuous coefficient of variation value exceeds the threshold. This is used to initially screen abnormal time periods and reduce redundant calculations.

[0067] The trajectory prediction module is built on an improved Transformer model, including an input embedding layer, a dynamic expansion attention mechanism, an Encoder-Decoder architecture, and a prediction objective function. The input embedding layer integrates time indices, sensor location codes, and data values ​​to form input features. The dynamic expansion attention mechanism assigns different expansion rates to each attention head, allowing the model to perform long-range modeling of receptive fields with different scales for time series. The Encoder architecture uses historical offset data as encoded input and predicts the offset trajectory for several future time steps by using past time series and the output of the Encoder as input to the Decoder. The prediction objective function uses weighted mean square error as the optimization objective.

[0068] The fault judgment module determines the fault based on the future offset trajectory predicted by the correction trajectory prediction module. If the predicted trajectory trend approaches or exceeds the set correction threshold, an early warning signal is generated to mark it as a potential fault state, and feedback is sent to the upper control system in real time.

[0069] Therefore, the fault prediction method and system for laser die-cutting and winding integrated machines described above have the following beneficial effects:

[0070] (1) By introducing a dynamic expansion attention mechanism, the model can effectively capture the dependencies at different time scales. Compared with traditional LSTM and GRU models, the average prediction error on multi-condition datasets is reduced by 12% to 18%. Experiments show that the multi-head dynamic attention mechanism can predict the trajectory approaching the correction limit 1 to 5 minutes before the equipment deviates abnormally, significantly improving the advance warning and enhancing the accuracy and foresight of fault prediction.

[0071] (2) The dynamic expansion attention mechanism significantly reduces redundant attention computation, reducing the overall computational complexity of the model from O(n^2) to O(n^2). 2 The inference speed is reduced to O(n), which is about 3 to 10 times faster than the multi-head dynamic attention mechanism under millions of time series data. It is significantly better than the standard multi-head dynamic attention mechanism Transformer. It can also run in real time on edge computing devices, meet the deployment requirements of industrial scenarios, and significantly improve model processing efficiency and scalability.

[0072] (3) By accurately modeling and dynamically evaluating the offset trajectory of the correction sensor, this invention can achieve early identification of potential faults and early warning of abnormal states, effectively reducing the equipment emergency shutdown rate caused by winding misalignment and material misalignment. In actual production line testing, the equipment failure rate was reduced by approximately 26.7% by the multi-head dynamic attention mechanism, indirectly increasing the cell yield by approximately 3% to 5% by the multi-head dynamic attention mechanism, and reducing the risk of unplanned equipment downtime.

[0073] (4) The prediction framework is applicable to various types of winding / correction industrial equipment. The model structure supports flexible adaptation to different equipment and sensor configurations, which is convenient for promotion and application in intelligent manufacturing scenarios such as lithium batteries, packaging, and printing. It has good versatility and engineering application value.

[0074] (5) By reducing equipment downtime and manual inspection costs, this invention helps improve production cycle time and ensure product consistency. It also assists enterprises in building intelligent operation and maintenance systems, promoting the intelligent upgrading of the manufacturing industry. If deployed on a winding production line producing 1 million battery cells annually, it is estimated that annual savings in operation and maintenance and loss costs will exceed RMB 300,000, demonstrating significant economic and social benefits.

[0075] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0076] Figure 1 This is a flowchart of a fault prediction method for laser die-cutting and winding integrated machine according to the present invention.

[0077] Figure 2 This is a schematic diagram of a fault prediction method for correction of a laser die-cutting and winding integrated machine, based on the present invention. The left side is a cross-sectional view, and the right side is a top view.

[0078] Figure 3 This invention provides a fault prediction method for correction in a laser die-cutting and winding integrated machine, using a thermal image of the positive electrode sensor of the die-cutting and winding machine.

[0079] Figure 4 The present invention provides a fault prediction method for correction of a laser die-cutting and winding integrated machine, using a thermal diagram of the negative electrode sensor of the cutting and winding machine;

[0080] Figure 5 This is a schematic diagram of a fault prediction method for laser die-cutting and winding integrated machines based on a multi-scale dynamic expansion attention efficient Transformer architecture. Detailed Implementation

[0081] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0082] Example

[0083] like Figure 1 As shown, this invention provides a fault prediction method for correction in a laser die-cutting and winding integrated machine, comprising the following steps:

[0084] S1. Receive and preprocess the raw offset time series data collected by the multi-dimensional sensors of the correction mechanism to ensure the accuracy and stability of subsequent modeling. The sensor positions are as follows: Figure 2 As shown.

[0085] S11. Perform data cleaning, using outlier detection methods based on the 6σ normal distribution assumption to remove outlier data. The specific process is as follows:

[0086] a. Set the original time series data as X i =x 1 ,x 2 ,...,x t ,...,x T Where i = 5, representing 5 sensors (negative or positive poles), the population mean and standard deviation are calculated using the following formula:

[0087]

[0088] Where T is the total length of the time series, μ is the mean, and σ is the standard deviation;

[0089] b. Construct an anomaly threshold interval, and set the retention interval as [μ-kσ,μ+kσ], where k is a constant;

[0090] c. Identify and remove or repair abnormal data. The judgment rule is: if |x t -μ|>kσ, then x t If an outlier is identified, it should be removed from the time series or repaired by interpolation using the average of the preceding and following points.

[0091] S12. Perform signal smoothing processing, using a Kalman filter to suppress high-frequency noise and preserve signal trend. The specific process is as follows:

[0092] a. Initialize state estimates With the coequation matrix P, let the initial value be... P0 = 1;

[0093] b. The smoothed estimate is calculated recursively, and is divided into two stages: prediction and update.

[0094] The calculation formula for the state prediction stage is:

[0095]

[0096] P t|t-1 =P t-1 +Q (4)

[0097] in, Given the information from time t-1 and earlier, this is a predicted estimate of the state at time t. Let P be the updated state estimate at time t-1. t|t-1 Let P be the covariance matrix corresponding to the predicted estimate at time t. t-1 The updated state estimate at time t-1 The covariance matrix of , where Q is the covariance matrix of the process noise;

[0098] The formula for calculating the state update phase is:

[0099]

[0100] P t = (1-K) t )·P t|t-1 (7)

[0101] Among them, K t Let x be the Kalman gain at time t, R be the covariance matrix of the observation noise, and x be the Kalman gain at time t. t The actual observation value of the sensor at time t. Let P be the state estimate after Kalman filtering at time t, and P be the updated state estimate at time t. covariance matrix

[0102] c. After Kalman filtering and smoothing, the output smoothed time series is: This smoothing sequence effectively suppresses high-frequency disturbances and more accurately reflects the trend changes in sensor measurements.

[0103] S2. Conduct equipment stability assessment by analyzing the data from each correction sensor throughout the day. Using a 10-minute sliding window, calculate the coefficient of variation (CV) for that time period to quantify the level of local fluctuation. The CV value reflects the degree of data fluctuation within that window: the higher the CV value, the greater the system volatility, the lower the stability, and the higher the probability of failure; conversely, the lower the CV value, the more stable the system operation. Set a dynamic threshold, and when the CV value exceeds the dynamic threshold, mark it as a high-risk operating state to filter abnormal time periods.

[0104] S21. Set the sliding window length to W (e.g., 10 minutes), divide the original time series into multiple consecutive time periods, and calculate the mean and standard deviation for each time period. The calculation formula is as follows:

[0105]

[0106] S22. Calculate the coefficient of variation (cv) within each time window. The calculation formula is as follows:

[0107]

[0108] S23. By comparing the distribution changes of CV values ​​across different time periods, it is possible to determine in which time periods the equipment experiences significant fluctuations. Combining the operating condition graph and operational history to set dynamic thresholds helps identify fault risk areas near the winding station, aiding in early warning and scheduling. When the coefficient of variability exceeds the dynamic threshold, it is marked as a high-risk operating state, thus filtering out abnormal time periods. Figure 3 and Figure 4 It can be concluded that the sensors qj0475 and qj0482, located near the winding station, are prone to malfunctions that could cause equipment downtime. Figure 3 and Figure 4 The X-axis represents a 10-minute time window, the Y-axis represents the sensor number, and the rightmost color band represents the corresponding coefficient of variability. The higher the system instability, the higher the coefficient of variability, and the brighter the color.

[0109] S3. Construct an improved Transformer prediction model, which includes an input embedding layer, a dynamically expanded attention mechanism, an encoder for extracting long-range dependencies in time series, and a decoder with a masked multi-head attention mechanism. The preprocessed data is input into the model for prediction. The efficient Transformer architecture based on multi-scale dynamically expanded attention constructed in this implementation is as follows: Figure 5 As shown.

[0110] S31. Construct the input embedding layer for feature encoding. Extract and embed the features of the sensor process data. Taking the negative electrode as an example, this embodiment selects five sensors, qj0471-qj0475, to predict the trajectory of sensor qj0475. Linearly connect the 5-dimensional original features of the sensor and map them into a 512-dimensional embedding vector.

[0111] Encoding and embedding include:

[0112] Positional encoding and embedding are necessary because Transformers lack an inherent mechanism for handling sequence order. Positional encoding is introduced to preserve temporal order information in time series data. Positional_Embedding allows the model to infer positional information from these periodic representations by calculating sine and cosine functions at different frequencies, as shown in the formula below.

[0113] The formula for generating the position code is:

[0114]

[0115] in, Let f(t) be the position value encoded in dimension i at time t. (i) Let be a function of time t. Here, d is the frequency parameter associated with dimension i, d is the dimension of the position encoding vector, and k is the index variable;

[0116] Temporal Feature Extraction and Embedding: In time series prediction, temporal features often contain important information, such as seasonality and periodicity, reflecting the regularity and repetition of time series at different time scales. Extracting these features and incorporating them into the model helps the model better understand the dynamic characteristics of the time series. For the current task, the device sampling frequency is on the order of seconds; therefore, this embodiment uses four dimensions—day, hour, minute, and second—to encode the time series information. The specific encoding rule is: Second of minute, minute of hour, hour of day, day of week, as shown in the formula below, encoding the time unit into values ​​in the range [-0.5, 0.5]; and mapping the day, hour, minute, and second time features to a 512-dimensional high-dimensional space.

[0117] The formula for calculating the mapping time feature is:

[0118]

[0119] Among them, time feature The encoded time feature value is represented by Time Unit Value, which is the original time unit data, and Max Unit Value, which is the maximum possible value for the current time unit.

[0120] Taking the 20th second relative to the minute as an example, 20s = 20 / 59 - 0.5 = -0.161.

[0121] S32. Construct a dynamic expansion attention mechanism, which divides the input sequence into multiple non-overlapping processing segments through a fixed sliding window, assigns different expansion rates to each attention head, and extracts local and global model features from different attention heads, which are then aggregated and spliced ​​into a multi-head dynamic attention mechanism.

[0122] In traditional self-attention mechanisms, each query vector is scaled with every key-value pair in the sequence to generate attention weights. This results in a computational complexity of O(L2) quadratic when Transformer processes long sequences. Below is the standard Self-Attention expression (kernel smoothing form).

[0123]

[0124] The kernel function is:

[0125]

[0126] Therefore, this embodiment constructs a dynamic expanded attention mechanism specifically for large-scale industrial time series data. In this mechanism, each attention head is assigned a different expansion rate. By dividing the processing window, different attention heads can extract local and global pattern features. This design enables the aggregation of features from multiple receiving regions, dynamically controlling the balance between modeling capability and computational efficiency. Through expanded skip computation, it effectively reduces the number of keys and values ​​involved in each calculation, thereby improving efficiency and reducing complexity. The derivation process for constructing the dynamic expanded attention mechanism is as follows:

[0127] a. Given a set of time series Where L is the sequence length, d is the model dimension, the sliding window size is fixed at S, and it is divided into N non-overlapping processing segments;

[0128]

[0129] b. Project the time series onto a plane with d h =For the H attention heads in the d / H dimension, the formula is as follows:

[0130]

[0131] Among them, Q h For querying the matrix, K h V is the key matrix. h Let h be the value matrix, and h be the corresponding attention head. To be learnable, feature mappings are applied to a specific head.

[0132] c. For each sliding time interval X (i) and each head h i Assign an expansion rate r h In this way, each attention head only focuses on a specific time interval r. h The tokens within the tokens reduce computation by skipping sampling time indices, as shown in the following formula:

[0133]

[0134] in, Let [0, S) be the set of valid time intervals for skip sampling of the h-th head in the i-th time interval, and let [0, S) be the time range of the processing segment, {0, r h ,2r h ...} represents the expansion rate r h Generate equally spaced time steps;

[0135] d. Time series segment X (i) The computational attention mechanism aggregates and splices together all non-overlapping time periods and attention heads to form a multi-head dynamic attention mechanism.

[0136]

[0137] in, and These represent time series segments X. (i) The selected set of indices The query, key, and value representations, MDDA(X) is the output of the multi-head dynamic attention mechanism. This is for outputting the projection matrix.

[0138] This mechanism, when the dataset is large enough and an incremental growth rate is given, such as [2,2,4,4,8,8,16,16], can reduce the complexity from O(n^2) to O(n^2). 2 The value decreases to O(n).

[0139] S33. Construct the encoder. The encoder structure in this embodiment is used to efficiently model long-distance dependencies in multivariate time series while maintaining low computational complexity. The encoder includes an embedding layer, a dynamically expanded attention layer, an encoder layer, and a normalization layer.

[0140] The embedding layer maps the raw time-series data and its timestamp features to a high-dimensional representation space. Embedding operations include combinations of scalar embedding, temporal embedding, and positional embedding. The output of this layer provides the input basis for subsequent encoder layers.

[0141] The dynamically expanded attention layer introduces a dynamic attention expansion mechanism. This mechanism assigns different expansion rates to different attention heads, enabling the model to simultaneously capture local details and global dependencies for long-range modeling. Subsequently, the attention scores are normalized using the Softmax function to generate attention weights. While effectively reducing the computational cost of model training, this mechanism still maintains good modeling capabilities and prediction accuracy.

[0142] The encoder layer comprises two main modules: an attention structure based on dynamic skip sampling and a feedforward network based on a one-dimensional convolutional structure. The feedforward network consists of two one-dimensional convolutional layers connected by an activation function, and the output is processed through residual connections and normalization. This encoder layer structure is stacked twice to enhance modeling capabilities.

[0143] A normalization layer (LayerNorm) is set at the end of the entire encoder module to stabilize the training process and improve the quality of downstream feature representation.

[0144] S34. Constructing the decoder: The decoder structure in this embodiment is used to generate predicted values ​​for the target time series based on encoded historical information and previous prediction results. The decoder consists of several stacked decoder layers, each layer including a masked MDDA module, a cross-attention module, and a normalization and output mapping layer.

[0145] The masked MDDA module introduces a causal masking mechanism in attention computation to ensure that each position in the sequence only focuses on its historical position, thereby avoiding access to future information and achieving autoregressive prediction modeling.

[0146] The Cross-Attention module is based on the standard fully connected attention mechanism. It fuses the encoder output and the decoder input to establish a correspondence between the current prediction and the historical representation, so as to achieve long-distance dependency modeling across sequences.

[0147] The decoder structure is equipped with a normalization module and a linear mapping layer at the end. The normalization module is used to stabilize the network output, and the linear mapping layer maps the decoder output to the prediction target dimension space.

[0148] S4. If the trend of the future deviation trajectory predicted by the model approaches or exceeds the set correction threshold, an early warning signal is generated and marked as a potential fault state.

[0149] S5 employs Dynamic Expanded Attention (MDDA) optimization and segmented inference mechanisms to deploy the model on edge computing nodes or industrial servers for prediction, effectively adapting to the processing needs of millions of sensor data and ensuring real-time performance and computational resource efficiency.

[0150] A fault prediction system for correction in a laser die-cutting and winding integrated machine includes:

[0151] The data preprocessing module receives raw offset data from the sensors of the correction mechanism and preprocesses it, constructs a time-series-based sliding window sample, uses Kalman filtering to eliminate high-frequency fluctuations and noise, and introduces a missing value interpolation mechanism to ensure data continuity.

[0152] The equipment stability assessment module calculates the continuous coefficient of variation of the corresponding sequence within each sliding window, sets a dynamic threshold, and marks a high-risk operating state when the continuous coefficient of variation value exceeds the threshold. This is used to initially screen abnormal time periods and reduce redundant calculations.

[0153] The trajectory prediction module is built on an improved Transformer model and includes an input embedding layer, a dynamic expansion attention mechanism, an Encoder-Decoder architecture, and a prediction objective function. The input embedding layer integrates time index (Time2Vec encoding), sensor position encoding, and data values ​​to form input features. The dynamic expansion attention mechanism assigns different expansion rates and receptive fields to each attention head for long-range dependency modeling. The Encoder architecture uses historical offset data as encoded input and the decoder predicts the offset trajectory for several future time steps. The prediction objective function uses weighted mean square error as the optimization objective, taking into account the relative distance between the predicted value and the sensor's physical limit (upper limit of the correction).

[0154] The fault judgment module determines the fault based on the future offset trajectory predicted by the correction trajectory prediction module. If the predicted trajectory trend approaches or exceeds the set correction threshold, an early warning signal is generated to mark it as a potential fault state, and feedback is sent to the upper control system in real time to realize pre-fault intervention.

[0155] Therefore, the present invention adopts the above-mentioned fault prediction method and system for correction of laser die-cutting and winding integrated machine. By introducing a dynamic expansion attention mechanism, it effectively extracts multi-scale time-dependent features, and significantly improves the training and inference efficiency of the model under large-scale industrial data while ensuring prediction accuracy. The method can realize early identification and real-time warning of potential equipment faults, reduce the risk of unplanned downtime, and improve the operational stability and product yield in the battery cell manufacturing process.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fault prediction method for correction in a laser die-cutting and winding integrated machine, characterized in that, Includes the following steps: S1. Receive and preprocess the raw offset time series data collected by the multi-dimensional sensor of the correction mechanism; S2. Conduct equipment stability assessment, construct sliding window samples, calculate the coefficient of variation of the sequence within each sliding window to quantify the level of local fluctuations, set a dynamic threshold, and mark the operation as high-risk when the coefficient of variation exceeds the dynamic threshold, and screen abnormal time periods. S3. Construct an improved Transformer prediction model, which includes an input embedding layer, a dynamic expansion attention mechanism, an encoder, and a decoder. Input the preprocessed data corresponding to the abnormal time periods obtained by S2 into the model for prediction. S31. Construct the input embedding layer for feature encoding. Linearly connect and map the 5-dimensional original features of the sensor into a 512-dimensional embedding vector. Generate position codes by calculating sine and cosine functions of different frequencies and embed them into the 512-dimensional embedding vector. Extract and embed time features. Map the time features to the 512-dimensional embedding vector to complete the construction of the input embedding layer. S32. Construct a dynamic expansion attention mechanism, which divides the input sequence into multiple non-overlapping processing segments through a fixed sliding window, assigns different expansion rates to each attention head, extracts local and global model features from different attention heads, and summarizes and splices them into a multi-head dynamic attention mechanism. The derivation process for constructing the dynamic expansion attention mechanism is as follows: a. Given a set of time series ,in It is the sequence length. It is the model dimension, with a fixed sliding window size. and divided into Non-overlapping processing segments; （13） （14） b. Project the time series onto a graph with... Dimensional In each attention focus, the formula is as follows: （15） in, For querying the matrix, The key matrix, For value matrices, For the corresponding attention head, To be learnable, feature mappings are applied to a specific head. c. For each sliding time interval and each head Assign an expansion rate The computational cost is reduced by skipping the sampling time index, as shown in the following formula: （16） in, For the first The time period, the first The set of effective time intervals after head skipping sampling The time range for processing segments. According to the expansion rate Generate equally spaced time steps; d. Time series segments The computational attention mechanism aggregates and splices together all non-overlapping time periods and attention heads to form a multi-head dynamic attention mechanism. （17） （18） in, , and These represent time series segments. The selected set of indices Queries, key and value representations, For the output of the multi-head dynamic attention mechanism, To output the projection matrix; S33. Construct the encoder, including the embedding layer, the dynamically expanded attention layer, the encoder layer, and the normalization layer; S34. Construct a decoder, which consists of several stacked decoder layers. Each layer includes a masked MDDA module, a cross-attention module, and a normalization and output mapping layer. S4. If the trend of the future deviation trajectory predicted by the model approaches or exceeds the set correction threshold, an early warning signal is generated and marked as a potential fault state. S5. Employs a dynamic expansion attention mechanism and a segmented inference mechanism to deploy the model on edge computing nodes or industrial servers for prediction.

2. The fault prediction method for laser die-cutting and winding integrated machine according to claim 1, characterized in that, The specific steps of S1 are as follows: S11. Perform data cleaning, using outlier detection methods based on the 6σ normal distribution assumption to remove outlier data. The specific process is as follows: a. Set the original time series data as The population mean and standard deviation are calculated using the following formulas: （1） （2） in, The total length of the time series. The mean, Standard deviation; b. Construct an anomaly threshold range and set the retention range as follows. ,in It is a constant; c. Identify and process outlier data, removing outlier data from the time series or using interpolation to repair it using the average of the previous and subsequent data. S12. Perform signal smoothing processing, using a Kalman filter to suppress high-frequency noise and preserve signal trend. The specific process is as follows: a. Initialize state estimates With co-equation matrix Let the initial value be , b. The smoothed estimate is calculated recursively, and is divided into two stages: prediction and update. The calculation formula for the state prediction stage is: （3） （4） in, In order to be in Time, based on All information up to and including the time, for The predicted estimate of the state at time step. for The updated state estimate at time step. for The covariance matrix corresponding to the predicted value at time step. for State estimate after time-update The covariance matrix, Let be the covariance matrix of the process noise; The formula for calculating the state update phase is: （5） （6） （7） in, for Kalman gain at time step The covariance matrix of the observed noise, for The actual observations of the time sensor for The state estimate updated by Kalman filtering at time step [time]. for State estimate after time-update covariance matrix c. After Kalman filtering and smoothing, the output smoothed time series is: .

3. The fault prediction method for laser die-cutting and winding integrated machine according to claim 1, characterized in that, The specific steps of S2 are as follows: S21. Set the length of the sliding window to... The original time series is divided into multiple consecutive time periods, and the mean and standard deviation are calculated for each time period. The calculation formulas are as follows: （8） （9） S22. Calculate the coefficient of variation within each time window. The calculation formula is as follows: （10） S23. By comparing the distribution changes of the coefficient of variation values ​​in different time periods, the status of the equipment is determined. A dynamic threshold is set by combining the operating condition map and the operating history. When the coefficient of variation value exceeds the dynamic threshold, it is marked as a high-risk operating state, and abnormal time periods are screened out.

4. The fault prediction method for laser die-cutting and winding integrated machine according to claim 1, characterized in that, The formula for generating the position code is: （11） in, In order to be in Time dimension Location code value, For about A function of time, For dimension Relevant frequency parameters, The dimension of the position encoding vector. For index variables; The formula for calculating the mapping time feature is: （12） in, These are the encoded time feature values. Data in raw time units, The maximum possible value for the current time unit.

5. The fault prediction method for laser die-cutting and winding integrated machine according to claim 1, characterized in that, The specific structure of the encoder in S33 is as follows: The embedding layer maps the original time series data and its timestamp features to a high-dimensional representation space. The embedding operation includes a combination of scalar embedding, temporal embedding, and positional embedding. The dynamically dilated attention layer introduces the dynamic dilated attention mechanism of S32, assigns different dilation rates to different attention heads, and uses the Softmax function to normalize the attention scores to generate attention weights. The encoder layer includes an attention structure based on skip sampling with different dilation rates and a feedforward network based on a one-dimensional convolutional structure. The feedforward network contains two one-dimensional convolutional layers connected by an activation function. The output is processed by residual connection and normalization operation. The encoder layer is stacked twice. The normalization layer is set at the end of the encoder module to perform normalization processing.

6. The fault prediction method for correction of a laser die-cutting and winding integrated machine according to claim 1, characterized in that, The specific structure of the decoder in S34 is as follows: the masked MDDA module introduces a causal masking mechanism in the attention calculation to realize autoregressive prediction modeling; the cross-attention module is based on the standard fully connected attention mechanism, which fuses the encoder output and the decoder input to establish the correspondence between the current prediction value and the historical representation. In the normalization and output mapping layer, the normalization module is used to stabilize the network output, and the linear mapping layer maps the decoder output to the prediction target dimension space.

7. A fault prediction system for correction of a laser die-cutting and winding integrated machine, applied to the fault prediction method for correction of a laser die-cutting and winding integrated machine as described in any one of claims 1-6, characterized in that, include: The data preprocessing module receives raw offset data from the sensors of the correction mechanism and preprocesses it, constructs a time-series-based sliding window sample, uses Kalman filtering to eliminate high-frequency fluctuations and noise, and introduces a missing value interpolation mechanism to ensure data continuity. The equipment stability assessment module calculates the continuous coefficient of variation of the corresponding sequence within each sliding window, sets a dynamic threshold, and marks a high-risk operating state when the continuous coefficient of variation value exceeds the threshold. This is used to initially screen abnormal time periods and reduce redundant calculations. The trajectory prediction module is built on an improved Transformer model, including an input embedding layer, a dynamic expansion attention mechanism, an Encoder-Decoder architecture, and a prediction objective function. The input embedding layer integrates time indices, sensor location codes, and data values ​​to form input features. The dynamic expansion attention mechanism assigns different expansion rates to each attention head, allowing the model to perform long-range modeling of receptive fields with different scales for time series. The Encoder architecture uses historical offset data as encoded input and predicts the offset trajectory for several future time steps by using past time series and the output of the Encoder as input to the Decoder. The prediction objective function uses weighted mean square error as the optimization objective. The fault judgment module determines the fault based on the future offset trajectory predicted by the correction trajectory prediction module. If the predicted trajectory trend approaches or exceeds the set correction threshold, an early warning signal is generated to mark it as a potential fault state, and feedback is sent to the upper control system in real time.

Images