Heat treatment processing technology stage segmentation method

Abstract

The invention relates to the field of heat treatment processing, in particular to a heat treatment processing process stage segmentation method based on analysis of a heat treatment time sequence. A piecewise linear representation algorithm based on key points is provided for the segmentation and detection problem of different process stages in the heat treatment machining process, a time sequence anomaly detection algorithm based on a sliding window is provided for data anomaly fluctuation in the production process, and real-time anomaly detection and processing are achieved. And the stability and accuracy of the segmentation algorithm are improved. And an interpolation algorithm based on segmentation errors is put forward for the problem that missing possibly exists in key point detection finally. Abnormal data in the heat treatment production process can be detected and processed, and the stability and accuracy of subsequent data are enhanced.

Description

Technical Field

[0001] The invention relates to the field of heat treatment processing, in particular to a method for implementing heat treatment processing process stage segmentation based on analysis of heat treatment time series. Background Art

[0002] Metal materials have a critical heating temperature. During the production process, when the temperature reaches the heating temperature corresponding to the material, the overall performance of the metal material will change. Heat treatment is the process of using this characteristic of metal materials to treat them accordingly and improve their performance. In heat treatment, common processes include annealing, normalizing, quenching, tempering, surface hardening, etc. Each of these processes corresponds to a different temperature and duration. In traditional manufacturing, workers rely on manual division of different process stages in the production process from historical production records. However, a heat treatment production usually lasts for several hours or even days, which will generate lengthy production data, and one worker is usually responsible for monitoring several machines at the same time. This puts a heavy burden on workers, and the manual processing speed is slow, which is easy to affect the advancement of subsequent manufacturing processes.

[0003] In the heat treatment manufacturing process, the data mainly consists of temperature-time series data. These data are collected at fixed time intervals by sensors located in the furnace and stored in the form of a series of (timestamp, temperature) pairs arranged in chronological order. As mentioned earlier, the temperature maintenance time at different levels during the heating process is a key indicator for determining the quality of the process. Therefore, in order to automate the detection and evaluation of each qualified process, we need to analyze the temperature-time series data and segment the different stages to extract the corresponding information of each stage. However, the original time series data is usually high-dimensional and contains various forms of noise. Direct analysis of the original time series data usually does not produce satisfactory results. So we need to align for feature extraction, denoising and dimensionality reduction expression.

[0004] At present, many methods have been proposed in the field of time series feature representation, such as methods that convert the feature domain of time series data into frequency domain representation, shapelets-based methods, singular value decomposition (SVD)-based methods, model-based methods, piecewise linear representation (PLR) methods, etc. Summary of the invention

[0005] The purpose of the present invention is to provide a method for dividing the heat treatment process stages, which automatically divides the heat treatment process stages, thereby reducing the workload of workers and accelerating the speed of dividing the heat treatment process stages and improving the accuracy of the division.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0007] The heat treatment process stage segmentation method comprises the following steps:

[0008] 1) Collect temperature data during processing through sensors;

[0009] 2) Use a time series anomaly detection algorithm based on a sliding window to perform real-time anomaly detection and judgment on temperature data;

[0010] 3) Using the angle change of the detected temperature data to determine and select the turning point, and generate candidate segments;

[0011] 4) Use the extreme trend segment algorithm to pre-screen the candidate segments;

[0012] 5) Use the BottomUp algorithm to fuse the extreme trend segments obtained by pre-screening;

[0013] 6) Traverse all the fused segments and determine the error of each segment. If the segment error exceeds the threshold, interpolation is performed on the segment to complete the segmentation of the processing technology stage.

[0014] The time series anomaly detection algorithm based on sliding window is specifically:

[0015] If the temperature difference between the two time points in the temperature data time series is greater than a certain threshold k, that is, |x t+1 -x t |>k, and x t+1 <x t , the left side of the window remains stationary, and the right side of the window begins to expand. When the temperature returns to the temperature x on the left side of the window left When the window length is within ±k neighborhood or exceeds the set length, stop the window expansion. At this time, detect the window length. If the window length is less than the set length, the value in the window is judged to be an abnormal value, and it is deleted or reset to x. left Corresponding temperature.

[0016] The step 3) comprises the following steps:

[0017] 3.1) The arctan value of the slope of the straight line connecting two adjacent points in the temperature data time series is used to calculate the value of the data point x i The angle θ i1 ,θ i2 :

[0018]

[0019] 3.2) Calculate the angle change σ:

[0020]

[0021] 3.3) If the angle change σ of a data point is greater than the angle threshold η, then add it to the candidate turning point sequence

[0022] Among them, s i is the index of the i-th turning point corresponding to the original time series;

[0023] 3.4) Use the alternative turning points to divide the original temperature data time series D into multiple segments, namely:

[0024] D=(d1,d2,…,d m )

[0025] Each segment d i From two adjacent candidate turning points x si ,x s(i+1) The points between them constitute

[0026] The step 4) comprises the following steps:

[0027] 4.1) Use the mean of all points in a segment as the representation of the segment

[0028]

[0029] 4.2) If a segment meets any of the following conditions, the segment will be retained:

[0030] Situation 1. Right now It is a maximum or minimum trend segment;

[0031] Situation 2. Right now and The segmentation standards are formed between them;

[0032] 4.3) For the segments that do not meet step 4.2), delete the turning points between the segments and merge the adjacent segments to obtain the filtered segments D′=(d′1,d′2,…,d′ u ), where u is the number of segments after screening.

[0033] The step 5) is specifically as follows:

[0034] If d′ i and d′ i+1 After fusion, if the sum of the errors between the obtained segmented linear expression and the original temperature data time series is less than the fusion error γ, fusion is performed to obtain the initial segmentation sequence Seg = (seg1, seg2, ..., seg p), where p is the number of initial segmentation sequences.

[0035] The step 6) is specifically as follows:

[0036] The initial segmentation sequence is scanned and the fitting error of each segment is calculated. If the fitting error of the segment is greater than the extreme trend segment judgment threshold ζ, the segment is interpolated to find the point with the largest error in the segment. The segment is split at this point and the segmented segment is tested again until it is decomposed into multiple sub-segments with an error of less than ζ for each sub-segment, and the final segmentation sequence Seg′=(seg′1,seg′1,…,seg′ q ), where q is the number of final segmentation sequences.

[0037] The fitting error e i for:

[0038] For the original temperature data time series X, X=(f1(v)+e1,f2(v)+e2,…,f p (v)+e p ), where f i (v) corresponds to segment seg i The linear fitting function of e i The fitting error corresponding to this segment, assuming that seg i The length of seg is j, that is, i contains j points but

[0039]

[0040] Among them, seg it Indicates that in the segment seg i At the tth time point, f i (v) indicates the segment used to fit the segment seg i A linear function of It means segmentation seg i The t-th time point in corresponds to the value in the original time series.

[0041] The present invention has the following beneficial effects and advantages:

[0042] 1. Able to detect and process abnormal data in the heat treatment production process, and enhance the stability and accuracy of subsequent data;

[0043] 2. The original time series can be segmented in a faster way and with a linear algorithm in time complexity, thereby realizing the segmentation of heat treatment process stages;

[0044] 3. Realize automated identification of process stages, reduce manual intervention, and help improve enterprise production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a partial flow chart of the core algorithm of the present invention;

[0046] Figure 2 This is a framework diagram for the overall use of the algorithm of the present invention. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0048] A piecewise linear representation algorithm based on key points is proposed to solve the segmentation and detection problems of different process stages in the heat treatment process. A time series anomaly detection algorithm based on sliding windows is proposed to solve the abnormal data fluctuations in the production process, which can realize real-time anomaly detection and processing and improve the stability and accuracy of the segmentation algorithm. In view of the problem that key point detection may be missed, an interpolation algorithm based on segmentation error is proposed to further improve the accuracy of stage division.

[0049] The specific steps are as follows:

[0050] Step 1: Set parameters

[0051] Step 2: Collect temperature data through sensors;

[0052] Step 3: Use the time series anomaly detection algorithm based on sliding windows to perform real-time anomaly detection, judgment and processing;

[0053] Step 4: Use the angle change to determine and select the turning point. If the angle is greater than the angle threshold, add the candidate turning point sequence to generate the candidate segment;

[0054] Step 5: Use the extreme trend segment algorithm to pre-screen the candidate segments;

[0055] Step 6: Use the BottomUp algorithm to fuse the extreme trend segments pre-screened in step 5. If the fusion error between two adjacent segments is less than the fusion error threshold, fusion is performed;

[0056] Step 7: Finally, traverse all the fused segments and determine the error of each segment. If the segment error exceeds the threshold, perform interpolation on the segment.

[0057] The parameters in step 1 include: angle threshold η, extreme trend segment determination threshold ζ, fusion error threshold γ, segmentation segment maximum error threshold τ;

[0058] In step 3, the time series anomaly detection algorithm based on sliding window is described as follows:

[0059] The root cause of this abnormal fluctuation in the heat treatment process is that it is necessary to open the furnace door during the production process to adapt to different processes and the current state of the remaining fuel, while also introducing the corresponding materials into the furnace. Therefore, this type of abnormal fluctuation is usually characterized by a rapid drop in temperature, followed by a return to the temperature before the furnace door was opened within a specified time frame. In order to exploit this pattern, we designed an algorithm combined with a sliding window technique. If the temperature difference between the two time points is greater than a certain threshold k, that is, |x t+1 -x t |>k, and x t+1 <x t , the window begins to expand until the temperature returns to the left side of the window x left ±k neighborhood or the window length exceeds the specified length, which is set by the time required for temporary furnace opening in the specific production process, usually 15 minutes.

[0060] If abnormal fluctuation data is detected, it will be deleted or reset to x left Corresponding temperature (determined by the operator).

[0061] In step 4, the definition of angle change is:

[0062] Angle definition: The angle definition is calculated using the arctan value of the slope of the straight line connecting two adjacent points:

[0063]

[0064] Angle change definition:

[0065]

[0066] If σ>η, add it to the candidate turning point sequence

[0067] Pre-screening strategy for step 5:

[0068] The candidate turning points selected in step 4 divide the original time series into multiple segments:

[0069] D=(d1,d2,…,d m )

[0070] Each d i From two adjacent candidate turning points x si ,x s(i+1) The points between At this time, the mean of all points in the segment is selected as the representation of the segment.

[0071]

[0072] Among them, s i is the index of the i-th turning point in the original time series. For example, if the first turning point is the 10th point in the original time series, then s1 = 10;

[0073] A segment is retained if:

[0074] Right now For extremely large or extremely small trend segments, select Keep; Right now and There are large differences between them, which constitute the segmentation criteria and are selected for retention;

[0075] Finally, the filtered segment D′=(d′1, d′2, …, d′ u ).

[0076] The fusion strategy of step 6 is:

[0077] For D′=(d′1,d′2,…,d′ u ) to fuse. If d′ i and d′ i+1 After fusion, if the sum of the errors between the linear expression of the obtained segment and the original time series is less than γ, fusion is performed to obtain the initial segmentation sequence Seg = (seg1, seg2, ..., seg p ).

[0078] The interpolation strategy of step 7 is:

[0079] In step 4, there may be a situation where the key turning point is not found in the candidate turning points selected. However, steps 5 and 6 cannot handle this situation no matter how the parameters are adjusted, so interpolation is introduced.

[0080] Scan the initial segmentation sequence obtained in step 6 and calculate the error of each segment. If the error of the segment is greater than ζ, interpolate the segment. Find the point with the largest error in the segment, split the segment at this point, and test the segmented segments again. Finally, decompose them into as few sub-segments as possible and each sub-segment has an error less than ζ, and update the final segmentation sequence Seg′=(seg′1,seg′1,…,seg′ q ).

[0081] Example

[0082] The present invention aims to identify and segment different process flows such as annealing, normalizing, quenching, tempering, etc.

[0083] After inputting the new temperature data into the process segmentor model and identifying different process flows in the heat treatment process, it also includes: judging process abnormalities according to the metal currently being processed and the corresponding standard processing parameters.

[0084] like Figure 1 As shown in the figure, the core algorithm of the whole system is divided into the following steps:

[0085] 1) Collect temperature data during processing through sensors;

[0086] 2) Sample sensor data at equal minute intervals;

[0087] 3) Use the angle change to preselect the turning point of the sampled data to obtain the initial segment sequence D = (d1, d2, ..., d m );

[0088] 4) The initial segmented sequence in step 3) is represented by the segmented mean, and the initial segmented sequence is preliminarily fused in combination with the segmented pre-screening mechanism to obtain the initial trend segment sequence D′=(d′1, d′2, …, d′ u );

[0089] 5) The initial trend segment sequence obtained in step 4) is merged from bottom to top according to the fitting error to obtain the initial segmentation sequence Seg = (seg1, seg2, ..., seg p );

[0090] 6) According to the initial segmentation sequence Seg = (seg1, seg2, ..., seg p ) in seg i The fitting error of the corresponding linear function is used to determine whether to perform interpolation. Finally, the segment exceeding the maximum error threshold τ of the segment is segmented, and the point with the largest error is selected for segmentation. The segmented subsequence is tested again until the segment is divided into as few sub-segments as possible with an error less than τ. Finally, Seg′=(seg′1,seg′1,…,seg′ q );

[0091] 7) respectively calculating the temperature, length, slope, fitting error and other statistical characteristics of the final segment obtained in step 6);

[0092] Finally, the statistical features obtained in step 7) and the metal being processed by the current heat treatment and its corresponding heat treatment processing standard parameters can be input into the classifier for classification to obtain the process type corresponding to the current stage, and the information can be combined to determine whether the corresponding process type processing is qualified.

[0093] The time series of the samples shown is defined as:

[0094] X=(x1,x2,…,x n )

[0095] Each point corresponds to the furnace temperature during processing sampled at fixed time intervals.

[0096] The turning point is defined as follows:

[0097] Angle definition: The angle definition is calculated using the arctan value of the slope of the straight line connecting two adjacent points:

[0098]

[0099] Angle change definition:

[0100]

[0101] If σ>η, add it to the candidate turning point sequence

[0102] The fitting error is defined as follows:

[0103] First, the original time series is divided into multiple segments: Seg = (seg1, seg2, ..., seg p ), for the original sequence: X=(f1(v)+e1,f2(v)+e2,…,f p (v)+e p ), where f i (v) corresponds to segment seg i The linear fitting function of i The fitting error corresponding to this segment is calculated as follows:

[0104] Assume seg i The length of is j, then

[0105]

[0106] like Figure 2 As shown, the overall structure is divided into:

[0107] 1) Query the previously collected and stored heat treatment temperature time series data to be analyzed from the database;

[0108] 2) Input the data into the corresponding process for processing, analysis and extraction to obtain the actual parameters and statistical characteristics of different processes in this processing;

[0109] 3) Input these data into the classifier for further analysis, and finally, the process status of this processing can be obtained.

[0110] The present invention also includes a system for identifying the processing state of a heat treatment process, comprising: a processor, a memory, and a temperature sensor arranged at the end of the heat treatment process processing equipment, the temperature sensor collects on-site temperature data, the memory stores a program, the processor loads the program to execute the steps of the method for identifying the processing state of a heat treatment process, and realizes the identification of the processing state of the heat treatment process.

[0111] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should be regarded as within the scope of protection of the present invention.

Claims

1. A method for segmenting heat treatment process stages, characterized in that: The following steps are involved: 1) Collect temperature data during processing through sensors; 2) Use a time series anomaly detection algorithm based on a sliding window to perform real-time anomaly detection and judgment on temperature data; 3) Using the angle change of the detected temperature data to determine and select the turning point, and generate candidate segments; 4) Use the extreme trend segment algorithm to pre-screen the candidate segments; 5) Use the BottomUp algorithm to fuse the extreme trend segments obtained by pre-screening; 6) Traverse all the fused segments and determine the error of each segment. If the segment error exceeds the threshold, interpolation is performed on the segment to complete the segmentation of the processing technology stage.

2. The heat treatment process stage segmentation method according to claim 1, characterized in that: The time series anomaly detection algorithm based on sliding window is specifically: If the temperature difference between the two time points in the temperature data time series is greater than a certain threshold k, that is, |x t+1 -x t |>k, and x t+1 <x t , the left side of the window remains stationary, and the right side of the window begins to expand. When the temperature returns to the temperature x on the left side of the window left When the window length is within ±k neighborhood or exceeds the set length, stop the window expansion. At this time, detect the window length. If the window length is less than the set length, the value in the window is judged to be an abnormal value, and it is deleted or reset to x. left Corresponding temperature.

3. The heat treatment process stage segmentation method according to claim 1, characterized in that: The step 3) comprises the following steps: 3.1) The arctan value of the slope of the straight line connecting two adjacent points in the temperature data time series is used to calculate the value of the data point x i The angle θ i1 ,θ i2 : 3.2) Calculate the angle change σ: 3.3) If the angle change σ of a data point is greater than the angle threshold η, then add it to the candidate turning point sequence Among them, s i is the index of the i-th turning point corresponding to the original time series; 3.4) Use the alternative turning points to divide the original temperature data time series D into multiple segments, namely: D=(d1,d2,…,d m ) Each segment d i From two adjacent candidate turning points x si ,x s(i+1) The points between them constitute 4. The heat treatment process stage segmentation method according to claim 1, characterized in that: The step 4) comprises the following steps: 4.1) Use the mean of all points in a segment as the representation of the segment 4.2) If a segment meets any of the following conditions, the segment will be retained: Situation 1. Right now It is a maximum or minimum trend segment; Situation 2. Right now and The segmentation standards are formed between them; 4.3) For the segments that do not meet step 4.2), delete the turning points between the segments and merge the adjacent segments to obtain the filtered segments D′=(d′1,d′2,…,d′ u ), where u is the number of segments after screening.

5. The heat treatment process stage segmentation method according to claim 1, characterized in that: The step 5) is specifically as follows: If d′ i and d′ i+1 After fusion, if the sum of the errors between the obtained segmented linear expression and the original temperature data time series is less than the fusion error γ, fusion is performed to obtain the initial segmentation sequence Seg = (seg1, seg2, ..., seg p ), where p is the number of initial segmentation sequences.

6. The heat treatment process stage segmentation method according to claim 1, characterized in that: The step 6) is specifically as follows: The initial segmentation sequence is scanned and the fitting error of each segment is calculated. If the fitting error of the segment is greater than the extreme trend segment judgment threshold ζ, the segment is interpolated to find the point with the largest error in the segment. The segment is split at this point and the segmented segment is tested again until it is decomposed into multiple sub-segments with an error of less than ζ for each sub-segment, and the final segmentation sequence Seg′=(seg′1,seg′1,…,seg′ q ), where q is the number of final segmentation sequences.

7. The heat treatment process stage segmentation method according to claim 6, characterized in that: The fitting error e i for: For the original temperature data time series X, X=(f1(v)+e1,f2(v)+e2,…,f p (v)+e p ), where f i (v) corresponds to segment seg i The linear fitting function of e i The fitting error corresponding to this segment, assuming that seg i The length of seg is j, that is, i contains j points but Among them, seg it Indicates that in the segment seg i At the tth time point, f i (v) indicates the segment used to fit the segment seg i A linear function of It means segmentation seg i The t-th time point in corresponds to the value in the original time series.

Images