Data enhancement method for industrial field slitting knife wear

Abstract

The present invention discloses a method for enhancing data on the wear of industrial field slitting knives, comprising collecting periodic data on the wear of slitting knives from a plurality of industrial fields, aggregating the periodic data on the wear of slitting knives from each industrial field in chronological order, and obtaining wear time series data corresponding to each industrial field; grouping the wear time series data corresponding to each industrial field according to grouping variables, and obtaining multiple groups of grouped time series data; selecting full sampling data from each group of grouped time series data, and replacing part of the sampling data in the full sampling data with sampling data from other grouped time series data, and obtaining multiple groups of composite time series data as training data; performing mask value replacement on data features of each single dimension in the training data, and obtaining masked data; and performing regression training on a language model using the masked data, and generating a label data set for the wear of slitting knives. The present invention can enhance the data on the wear of slitting knives in industrial fields.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery manufacturing, and in particular to a method for enhancing wear data of industrial on-site slitting knives. Background Art

[0002] The monitoring of slitting knife wear is limited by on-site production requirements. It is difficult to collect tool wear at a regular and high frequency, and it is impossible to establish a correlation between wear and other characteristics (meters) such as runout amplitude. Therefore, monitoring slitting knife wear is a major problem in the lithium battery manufacturing industry.

[0003] The traditional method is to remove the slitting knife after using it for a period of time and test the wear under a microscope, such as Figure 1 As shown in 2, in the actual production environment of a factory, the cutter is generally replaced every 400,000 meters. The test requires removing the tool holder, which takes a long time (1 hour) and seriously affects the production progress. Moreover, the wear amount can only be collected when the tool is replaced. The wear amount at other times is unknown, so the collected wear amount data is missing. Therefore, this method cannot collect diverse samples, resulting in an extreme lack of data labels for specific wear amounts with light wear during the training process.

[0004] Since the existing method of collecting slitting knife wear is carried out after the working cycle, experiments with different working conditions can be conducted (such as building a tool wear test bench to collect data through force, sound, multi-directional vibration sensors, current sensors, etc., or using open source data such as NASA and PMH competitions). However, in the industrial world, due to the limitations of high stability in manufacturing, equipment installation environment, complex production conditions, noise, high material costs, and production plan requirements, conducting experiments with different working conditions is very time-consuming, costly, and inefficient.

[0005] In related technology, the patent application document with publication number CN113868380A proposes segmenting intent recognition training data to convert the data into segment-based sequence data; performing the following operations N times on each sequence segment data to form N copies of data: performing one of the two operations, masking and random replacement, on each segment of the sequence segment data according to probability; concatenating the prompt template of the intent category for each of the N copies of data to obtain N times the training data; and achieving the goal of using a small amount of domain intent recognition training data to obtain a model with performance close to that obtained from a large amount of domain intent recognition training data in the absence of domain data. However, this scheme uses probabilistic random masking to replace data segments to obtain pre-training data. This scheme is applicable to the field of text processing because language has rich temporal information and a certain degree of information redundancy. However, because the values ​​of the random masking method are too random, it will introduce too much noise for numerical data in industry, which is not conducive to fitting parameter values. Summary of the Invention

[0006] The technical problem to be solved by the present invention is how to enhance the wear data of industrial on-site slitting knives.

[0007] The present invention solves the above technical problems through the following technical means:

[0008] The present invention proposes a method for enhancing data of industrial field slitting knife wear, the method comprising:

[0009] Collect the cutter wear cycle data from multiple industrial sites, aggregate the cutter wear cycle data from each industrial site in chronological order, and obtain the wear time series data corresponding to each industrial site;

[0010] The wear time series data corresponding to each industrial site are grouped according to the grouping variables to obtain multiple groups of grouped time series data;

[0011] Select all sampled data in each group of time series data, and replace part of the sampled data in the full sampled data with sampled data in other grouped time series data to obtain multiple groups of composite time series data as training data;

[0012] Replace the mask value of each single-dimensional data feature in the training data to obtain the masked data;

[0013] The masked data is used to perform regression training on the language model to generate a labeled dataset of cutter wear.

[0014] Furthermore, the cutter wear periodic data of each industrial site is aggregated in chronological order to obtain wear time series data corresponding to each industrial site, including:

[0015] The cutter wear cycle data of each industrial site is aggregated in chronological order, and the length of the aggregated data is truncated or padded with zeros to meet the set total length of the aggregated data.

[0016] Furthermore, the wear time series data corresponding to each industrial site is grouped according to the grouping variables to obtain multiple groups of grouped time series data, including:

[0017] The wear time series data corresponding to each industrial site are grouped according to [factory, material type] to obtain multiple groups of grouped time series data.

[0018] Furthermore, selecting all sampled data in each group of time series data and replacing part of the sampled data in the full sampled data with sampled data in other grouped time series data to obtain multiple groups of composite time series data as training data includes:

[0019] Each group of time series data is used as the central group, and the geometric distribution X~Geo (0.2) sampling is performed on the central group to select auxiliary groups;

[0020] Select all sampling data in the central group, and replace part of the sampling data in the full sampling data with the sampling data in the auxiliary group.

[0021] Furthermore, replacing part of the sampled data in the full sampled data with the sampled data in the auxiliary group includes:

[0022] The sampling data is searched from the auxiliary groups using the time of the central group as an index to replace part of the sampling data in the central group.

[0023] Furthermore, the mask value replacement of each single-dimensional data feature in the training data to obtain masked data includes:

[0024] Select m% of labal wear values ​​with wear values ​​and n% of non-grouped data features from the training data, where m% + n% = 100%;

[0025] The mask values ​​of m% of the labal wear values ​​with wear values ​​and n% of the non-grouped data features are replaced respectively to obtain the masked data.

[0026] Furthermore, the mask values ​​of m% of the labal wear values ​​with wear values ​​and n% of the non-grouping data features are replaced respectively to obtain the masked data, including:

[0027] Replace m% of the labal wear values ​​with wear values ​​or n% of the non-grouped data features with the [MASK] tag with a first percentage probability, replace them with the interpolated value with a second percentage probability, and keep the original word unchanged with a third percentage probability, where the sum of the first percentage, the second percentage, and the third percentage is 100%.

[0028] Furthermore, the non-grouping data features include sound, vibration and tension.

[0029] Furthermore, the interpolated value is calculated using a cubic spline interpolation method.

[0030] Furthermore, the loss function of the language model regression training is:

[0031]

[0032] in, is the predicted value, is the true value, is the total number of samples.

[0033] The advantages of the present invention are:

[0034] (1) The present invention uses the wear of slitting knives under different processing conditions in multiple industrial sites to migrate data, and then performs preprocessing such as time series aggregation, sampling and interpolation, and fills in coverage labels and features in a masked manner to reduce the data distribution differences introduced by factors such as variable working conditions and different test environments. The masked data is then used to train a language model, and the language model is used to calculate and interpolate the unknown wear amount based on the input data, ultimately generating more labeled data sets to solve the problem of difficult label collection.

[0035] (2) In the early stage of data processing, the present invention mixes the data of multiple factories, strengthens the correlation of the multi-factory data by constructing the data form and combining it with the language model to learn the cloze method, learns the underlying common logic of the multi-factory data, and the trained language model can fill in the missing factory life-wear amount or non-classified group feature data with values, thereby achieving the purpose of data enhancement.

[0036] (3) The present invention uses geometric distribution sampling when performing sample data replacement to ensure that each type of data is reasonably distributed.

[0037] (4) The present invention innovatively introduces an interpolation method to construct the mask value, which is more consistent with the actual distribution of the data.

[0038] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the slitting knife mentioned in the background technology part of the present invention;

[0040] Figure 2 The wear amount of the slitting knife mentioned in the background technology section of the present invention is obtained by placing the slitting knife under a microscope for observation;

[0041] Figure 3 This is a flow chart of a method for enhancing data on industrial field slitting blade wear, according to one embodiment of the present invention;

[0042] Figure 4 This is a process principle diagram of industrial field slitting knife wear data enhancement in one embodiment of the present invention;

[0043] Figure 5 is a time series diagram of slitting blade wear data in one embodiment of the present invention;

[0044] Figure 61 is a schematic diagram of data distribution after data sampling and replacement in one embodiment of the present invention;

[0045] Figure 7 Schematic diagram of the structure of the language model in one embodiment of the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0047] like Figures 3 and 4 As shown, an embodiment of the present invention provides a method for enhancing data of industrial field slitting knife wear, the method comprising the following steps:

[0048] S10, collecting cutter wear periodic data from multiple industrial sites, and aggregating the cutter wear periodic data from each industrial site in chronological order to obtain wear time series data corresponding to each industrial site;

[0049] S20, grouping the wear time series data corresponding to each industrial site according to the grouping variable to obtain multiple groups of grouped time series data;

[0050] S30, selecting all sampled data in each group of time series data, and replacing part of the sampled data in the all sampled data with sampled data in other grouped time series data, to obtain multiple groups of composite time series data as training data;

[0051] It should be noted that in this embodiment, the data is grouped into different working conditions and then mixed so that the training data contains not only single factory data but also composite data and wear amounts of multiple factories, ensuring that each type of data is reasonably distributed and maximizing the use of labeled data from multiple factories.

[0052] S40, replacing the mask value of each single-dimensional data feature in the training data to obtain masked data;

[0053] S50: Perform regression training on the language model using the masked data to generate a label data set of the cutter wear amount.

[0054] This embodiment uses the wear amount of slitting blades under different processing conditions at multiple industrial sites to migrate data, and then performs preprocessing such as time series aggregation, sampling, and interpolation. It fills in coverage labels and features in a masked manner to narrow the data distribution differences introduced by factors such as variable working conditions and different test environments. The masked data is then used to train a language model, and the language model is used to calculate and interpolate unknown wear amounts based on the input data, ultimately generating more labeled data sets to solve the problem of difficult label collection.

[0055] As a further preferred technical solution, in step S10, the cutter wear periodic data of each industrial site is aggregated in chronological order to obtain the wear time series data corresponding to each industrial site. Specifically, the cutter wear periodic data of each industrial site is aggregated in chronological order, and the length of the aggregated data is truncated or padded with zeros to meet the set total length of the aggregated data.

[0056] It should be noted that in this embodiment, the cutter wear data of multiple cycles in multiple factories are collected. Since the wear can only be collected when the tool is replaced, only the wear at the time of tool replacement can be collected. The wear data at other times are partially missing, so the wear at other times is unknown and recorded as None. The cutter wear data of each cycle in each factory are then aggregated in chronological order in the form of language generation. The cutter wear cycle data collected from multiple industrial sites in this embodiment are shown in Table 1 below. Other features except wear, such as tension, vibration, sound, etc., are recorded as non-grouped data features. Among them, the time series diagram of the slitting knife vibration data features is shown in the following table. Figure 5 As shown, the horizontal axis is time and the vertical axis is the vibration information of the knife when it is working:

[0057] Table 1 Cutting knife wear cycle data of each factory

[0058]

[0059] The cutter wear cycle data was aggregated in chronological order, mimicking the language-generated format. The total aggregated data length was set to T = 512, with the length of each sample being L, and N being the number of cycles. For example, N = 73 represents data from a maximum of 73 time points. The aggregated data length was L * N, with any data less than 512 padded with zeros and any data exceeding 512 truncated. The final wear time series data for Factory B resembled this:

[0060] 1 day, Factory B, Material B, xx meters, tension, vibration, sound, wear (None); 2 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (None); 3 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (None); 4 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (20um), 0000000 --- length is T = 512.

[0061] As a further preferred technical solution, step S20: grouping the wear time series data corresponding to each industrial site according to the grouping variable to obtain multiple groups of grouped time series data, specifically includes:

[0062] The wear time series data corresponding to each industrial site are grouped according to [factory, material type] to obtain multiple groups of grouped time series data.

[0063] Specifically, this embodiment groups the wear time series data corresponding to a factory into different working conditions. This ensures that the training data includes both single-factory data and composite data from multiple factories, along with wear measurements. This is accomplished by defining a grouping variable and specifying the number of groups. For example, if the grouping variable is [factory, material type], there are M possible grouping conditions: condition 1 = [factory A, material type A], condition 2 = [factory B, material type B], condition 3 ... condition - M.

[0064] As a further preferred technical solution, step S30: selecting all sampled data in each group of time series data, and replacing part of the sampled data in the full sampled data with sampled data in other grouped time series data, to obtain multiple groups of composite time series data as training data, specifically includes:

[0065] S31. Take each group of time series data as the central group, and perform geometric distribution X~Geo (0.2) sampling on the central group to select auxiliary groups;

[0066] S32: Select all sampling data in the central group, and replace part of the sampling data in the full sampling data with the sampling data in the auxiliary group.

[0067] Specifically, this embodiment performs full sampling on a single group as the central group: full sampling of condition1, full sampling of condition2, etc. Then, with a certain condition as the center, other conditions are used as auxiliary sampling to replace the data in the central group, until data replacement is performed on each group of time series data. For example, with condition1 as the center and other conditions as auxiliary:

[0068] First, select the fully sampled data in condition1. Examples of the data are:

[0069] 1 day, Factory A, Material A, xx meters, tension, vibration, sound, wear (None); 2 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (None); 3 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (None); 4 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (None); 5 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (50um)

[0070] Next, replace N pieces of data with data under multiple conditions.

[0071] Moreover, when selecting grouped data, this embodiment uses geometric distribution X~Geo(0.2) to sample and select groups. The central condition has the highest probability of being sampled, and other conditions gradually decrease. For example, with plant A as the central group, the mixed data after geometric distribution sampling is: plant A data (accounting for 20%), plant B accounted for (15%), plant C accounted for (15%), plant D accounted for (10%), etc., to ensure the rationality of the distribution of the sampled data.

[0072] As a further preferred technical solution, replacing part of the sampling data in the full sampling data with the sampling data in the auxiliary group includes: searching for sampling data from the auxiliary group using the time of the central group as an index to replace part of the sampling data in the central group.

[0073] Specifically, the data format after sampling replacement can be: 1 day, Factory A, Material A, xx meters, tension, vibration, sound, wear (None); 2 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (None); 3 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (None); 4 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (20um); 5 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (50um). The data distribution after data sampling replacement is as follows: Figure 6 shown.

[0074] The training data obtained at this time includes both single factory data and composite data of multiple factories and wear amount.

[0075] As a further preferred technical solution, step S40: replacing the data features of each single dimension in the training data with a mask value to obtain masked data, specifically includes the following steps:

[0076] S41, selecting m% of labal wear values ​​with wear values ​​and n% of non-grouped data features from the training data, where m%+n%=100%;

[0077] S42. Replace the mask values ​​of m% of the labal wear values ​​with wear values ​​and n% of the non-grouped data features to obtain masked data.

[0078] Furthermore, the step S42: performing mask value replacement on m% of the labal wear values ​​with wear values ​​and n% of the non-grouped data features to obtain masked data includes:

[0079] Replace m% of the labal wear values ​​with wear values ​​or n% of the non-grouped data features with the [MASK] tag with a first percentage probability, replace them with the interpolated value with a second percentage probability, and keep the original word unchanged with a third percentage probability, where the sum of the first percentage, the second percentage, and the third percentage is 100%.

[0080] Specifically, this embodiment retains 85% of the tokens without performing the "mask strategy" and performs the "mask strategy" on 15% of the tokens. During processing, the number of masked tokens remains unchanged (denoted as M). Among them, the token can be interpreted as a single-dimensional data feature, such as the tension value, the maximum value of the signal, etc.

[0081] When selecting the values ​​to be masked, 60% are selected as labal wear values ​​with wear values, and the remaining 40% are non-grouped features (such as sound, vibration, tension, etc.). This ensures that the model learns the correlation between features while learning the correlation between each label. Multi-class task learning enhances the model's generalization ability. For example, the sound features of the previous and next moments are known, and the goal is to predict the value of the sound features at the current moment.

[0082] Furthermore, the following processing is performed on the selected 60% of the labal wear values ​​with wear values:

[0083] Replace with [MASK] mark with 80% probability;

[0084] Replace with interpolated value with a probability of 10%;

[0085] With a probability of 10%, the original word is kept unchanged, that is, no replacement is made.

[0086] The remaining 40% of non-grouping features are processed as follows:

[0087] Replace with [MASK] mark with 80% probability;

[0088] Replace with interpolated value with a probability of 10%;

[0089] With a probability of 10%, the original word is kept unchanged, that is, no replacement is made.

[0090] As a further preferred technical solution, this embodiment innovatively introduces an interpolation method to construct the mask value, which is more consistent with the actual distribution of the data.

[0091] Specifically, this embodiment uses cubic spline interpolation, a curve interpolation method based on numerical characteristics, as it better matches the data distribution than linear interpolation. Cubic spline interpolation divides the known data into segments, constructing a cubic function for each segment. The connections between these cubic functions are guaranteed to be continuous at zeroth order, first-order derivatives, and second-order derivatives, resulting in smooth curve connections that are well-suited for life span interpolation.

[0092] For example: for the wear amount [1,3,5,x,10], x is the selected value to be masked, and the interpolation is output by its context. For the interpolation algorithm, x=cubic spline interpolation algorithm (1,3,5,10).

[0093] Specifically, the None value is marked with the symbol UNK, and the masked data is:

[0094] Label mask (wear mask data)

[0095] (1) 80% [MASK] mark

[0096] 1 day, Factory A, Material A, xx meters, tension, vibration, sound, wear (unk--um); 2 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (unk--um); 3 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (unk--um); 4 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (mask-um).

[0097] (2) 10% interpolation value

[0098] 1 day, Factory A, Material A, xx meters, tension, vibration, sound, wear (unk--um); 2 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (unk--um); 3 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (unk--um); 4 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (20um).

[0099] (3) 10% of the original words remain unchanged

[0100] 1 day, Factory A, Material A, xx meters, tension, vibration, sound, wear (unk--um); 2 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (unk--um); 3 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (unk--um); 4 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (20um).

[0101] Feature mask (data after feature mask)

[0102] (1) 80% [MASK] mark

[0103] 1 day, Factory A, Material A, xx meters, tension, vibration, sound (mask), wear (unk--um); 2 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (unk--um); 3 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (unk--um); 4 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (20um).

[0104] (2) 10% interpolation value

[0105] 1 day, Factory A, Material A, xx meters, tension, vibration, sound (43-interpolation), wear (unk--um); 2 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (unk--um); 3 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (unk--um); 4 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (20um).

[0106] (3) 10% of the original words remain unchanged

[0107] Factory A, material A, xx meters, tension, vibration, sound (45), wear (unk--um); 2 days, Factory B, material B, xx meters, tension, vibration, sound, wear (unk--um); 3 days, Factory A, material A, xx meters, tension, vibration, sound, wear (unk--um); 4 days, Factory A, material A, xx meters, tension, vibration, sound, wear (20um).

[0108] As a further preferred technical solution, in step S50, regression training is performed using a masked language model (LLM). Wear volume is modeled using the LLM approach, and the underlying network of the language model uses BERT. Masking is used to fill in labels and features, minimizing data distribution variations introduced by factors such as varying operating conditions and different test environments. Compared to time series LSTM and TCN models, these models can only train and update parameters targeting the beginning and end of a data segment's label values. However, this embodiment utilizes the masking approach of BERT, allowing training and parameter updates to be performed on any label value and feature value within a time period. This provides richer training data and a wider range of learning patterns.

[0109] Among them, such as Figure 7 As shown, the input of the language model is [N, T, F], and the output is [N, M], where M is the value to be predicted and the mask length. The output is not calculated by softmax for probability, and is defined as a regression task. The loss function uses MSE:

[0110]

[0111] in, is the predicted value, is the true value, is the total number of samples.

[0112] Furthermore, the input embedding of the language model is: the result after masking, and the input of the model is three types of vector embedding:

[0113] Token embedding: direct data input;

[0114] Segment embedding, paragraph encoding, represents a time series, different time points are encoded, and if time cannot be classified, the binning method is used: 1, 1, 1, 1, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4;

[0115] Position embedding, position encoding, is used to identify features of different categories: 1, 2, 3, 4, 5, 6, 7, 1, 2, 3, 4, 5, 6, 7, 1, 2, 3, 4, 5, 6, 7;

[0116] Among them, the binning method refers to a time coding strategy. The input of the binning is (a time series of a series of data), and the output is the grouped result (which can be classified into day shift and night shift, or divided by day of the week).

[0117] Furthermore, in this embodiment, the trained language model can be used to calculate and interpolate unknown wear amounts based on input data, ultimately generating more data sets to solve the problem of difficult label collection.

[0118] For example, input a value with a mask:

[0119] 1 day, Factory A, Material A, xx meters, tension, vibration, sound, wear (mask); 2 days, Factory B, Material B, xx meters, tension, vibration, sound, wear (mask); 3 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (mask); 4 days, Factory A, Material A, xx meters, tension, vibration, sound, wear (20um);

[0120] Output N (number of batches), the result of M mask values.

[0121] This embodiment collects the periodic data of slitting knife wear from various industrial sites and mixes the data from multiple factories during the preliminary data processing. It strengthens the relationship between the multi-factory data by constructing the data form and combining it with the model learning cloze method, and learns the underlying common logic of the multi-factory data. The trained model can fill in the missing values ​​of the life-wear or feature data of Factory A, thereby achieving the purpose of data enhancement.

[0122] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0123] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0124] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for enhancing data on industrial field slitting knife wear, characterized in that: The method comprises: Collect the cutter wear cycle data from multiple industrial sites, aggregate the cutter wear cycle data from each industrial site in chronological order, and obtain the wear time series data corresponding to each industrial site; The wear time series data corresponding to each industrial site are grouped according to the grouping variables to obtain multiple groups of grouped time series data; Select all sampled data in each group of time series data, and replace part of the sampled data in the full sampled data with sampled data in other grouped time series data to obtain multiple groups of composite time series data as training data; Replacing the data features of each single dimension in the training data with a mask value to obtain masked data includes: selecting m% of labal wear amounts with wear values ​​and n% of non-grouped data features from the training data, where m%+n%=100%; and replacing the m% of labal wear amounts with wear values ​​and n% of non-grouped data features with mask values ​​to obtain masked data, specifically replacing the m% of labal wear amounts with wear values ​​or the n% of non-grouped data features with a [MASK] tag with a first percentage probability, replacing them with an interpolated value with a second percentage probability, and keeping the original word unchanged with a third percentage probability, where the sum of the first percentage, the second percentage, and the third percentage is 100%; The masked data is used to perform regression training on the language model to generate a labeled dataset of cutter wear.

2. The industrial field slitting knife wear data enhancement method according to claim 1, characterized in that: The cutter wear periodic data of each industrial site is aggregated in chronological order to obtain the wear time series data corresponding to each industrial site, including: The cutter wear cycle data of each industrial site is aggregated in chronological order, and the length of the aggregated data is truncated or padded with zeros to meet the set total length of the aggregated data.

3. The industrial field slitting knife wear data enhancement method according to claim 1, characterized in that: The wear time series data corresponding to each industrial site is grouped according to the grouping variables to obtain multiple groups of grouped time series data, including: The wear time series data corresponding to each industrial site are grouped according to [factory, material type] to obtain multiple groups of grouped time series data.

4. The industrial on-site slitting blade wear data enhancement method according to claim 1, characterized in that: The method of selecting all sampled data from each group of time series data and replacing part of the sampled data in the all sampled data with sampled data from other grouped time series data to obtain multiple groups of composite time series data as training data includes: Each group of time series data is used as the central group, and the geometric distribution X~Geo (0.2) sampling is performed on the central group to select auxiliary groups; Select all sampling data in the central group, and replace part of the sampling data in the full sampling data with the sampling data in the auxiliary group.

5. The industrial on-site slitting blade wear data enhancement method according to claim 4, characterized in that: The replacing part of the sampled data in the full sampled data with the sampled data in the auxiliary group includes: The sampling data is searched from the auxiliary groups using the time of the central group as an index to replace part of the sampling data in the central group.

6. The industrial on-site slitting blade wear data enhancement method according to claim 1, characterized in that: The non-grouping data features include sound, vibration, and tension.

7. The industrial on-site slitting blade wear data enhancement method according to claim 1, characterized in that: The interpolated value is calculated using a cubic spline interpolation method.

8. The industrial on-site slitting blade wear data enhancement method according to claim 1, characterized in that: The loss function of the language model regression training is: in, is the predicted value, is the true value, is the total number of samples.

Images