AI quality inspection method and equipment based on time sequence perception and monotonicity constraint time sequence prediction, and medium

Abstract

The invention discloses an AI quality inspection method and device based on time sequence perception and monotonicity constraint time sequence prediction, and a medium, and the method comprises the steps: obtaining parameter data of a plurality of time sequence features under different time steps in a technological process, removing redundant data in the time sequence features based on prior knowledge, and obtaining first processing data as a training set; constructing a monotonicity constraint time sequence prediction model; aiming at the first processing data, using a small sample reinforcement learning method to expand a training set, and training the monotonicity constraint time sequence prediction model; and inputting the current time sequence data into the trained monotonicity constraint time sequence prediction model, and outputting a process product quality result to realize AI quality inspection. According to the method, the training set with prior knowledge is introduced, and the monotonicity constraint of the process mechanism is creatively added in the complex model, so that the interpretability of the model is enhanced, and the model is easier to train.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial data detection processing, and in particular to an AI quality inspection method based on time sequence perception and monotonicity constraint time sequence prediction, equipment and medium. BACKGROUND

[0002] AI quality inspection is a technology for indirectly estimating key variables that are difficult to measure directly in industrial processes through mathematical modeling. This technology uses auxiliary variables that are easy to measure to construct a mapping relationship between them, replacing the measurement function of "physical sensors" with "algorithm models". A typical case is an intelligent body weight scale, which estimates body fat rate through height, weight, and bioelectric resistance, which are easy to measure. The core elements include three parts: key variables, auxiliary variables, and mapping models between them. The core problems currently faced by AI quality inspection mainly include data scarcity and semi-supervised challenges, as well as the need for dynamic coupling and time sequence modeling. The label data of key variables (such as product quality) usually relies on offline methods such as chemical testing, which not only lacks data but also has high acquisition costs. Auxiliary variables can be collected in real time by instruments to generate a large amount of unlabeled data, but the lack of labels leads to the typical semi-supervised small sample problem in model training. Industrial data contains a large amount of meaningless noise data, and there is strong coupling and nonlinear correlation between variables, and the production process has continuous dynamic characteristics. The current state not only depends on the evolution of historical data, but also the time dependence scale of different auxiliary variables may differ, so the model not only needs to have time sequence modeling capabilities, but also needs a multi-scale attention mechanism to accurately capture dynamic correlations. SUMMARY

[0003] Technical purpose: In view of the defects in the prior art, the present application discloses an AI quality inspection method based on time sequence perception and monotonicity constraint time sequence prediction, equipment and medium, which introduces a training set with prior knowledge and innovatively adds process mechanism monotonicity constraints in a complex model. This method enhances the model's interpretability and makes the model easier to train.

[0004] Technical scheme: In order to achieve the above technical purpose, the present application adopts the following technical scheme.

[0005] An AI quality inspection method based on time sequence perception and monotonicity constraint time sequence prediction, the method comprising: Obtaining parameter data of a plurality of time sequence features at different time steps in a process, removing redundant data in the time sequence features based on prior knowledge to obtain first processed data as a training set; the first processed data includes parameter data of a plurality of time sequence features within a plurality of preset time steps; Constructing a monotonicity constraint time sequence prediction model; For the first processed data, a few-shot reinforcement learning method was used to expand the training set and train the monotonicity-constrained time series prediction model. The current time series data is input into the trained monotonicity-constrained time series prediction model, which outputs the process product quality results to achieve AI quality inspection.

[0006] An electronic device includes a memory for storing a computer program and a processor for executing the computer program to implement an AI quality inspection method based on time-awareness and monotonicity-constrained time-series prediction as described above.

[0007] A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the AI ​​quality inspection method based on time-series awareness and monotonicity constraint time-series prediction as described above.

[0008] Beneficial effects: This invention introduces a training set with prior knowledge and innovatively adds monotonicity constraints on process mechanisms to complex models. This approach enhances the interpretability of the model and makes it easier to train. This invention also uses dynamic time warping and inverse distance weighting to design a scheme for expanding labels for small-sample semi-supervised time series data, in order to solve the problem of few labels for key variables. Attached Figure Description

[0009] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the monotonicity-constrained time series prediction model structure of Embodiment 1 of the present invention. Detailed Implementation

[0010] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0011] Example 1 As attached Figure 1 As shown, an AI quality inspection method based on time-series awareness and monotonicity-constrained time-series prediction includes the following steps: S1: Obtain parameter data of several time-series features at different time steps in the process, remove redundant data in the time-series features based on prior knowledge, and obtain the first processed data as the training set; the first processed data includes parameter data of several time-series features in several preset time steps; The time sequence features refer to physical quantities with time sequence characteristics that have an impact on the process product quality, such as temperature, humidity, flow and the like. The redundant data is removed based on prior knowledge, including: combining process experience to determine the maximum influence time interval of each time sequence feature on the process product quality; for each time sequence feature, the parameter data obtained at other time steps outside the maximum influence time interval is removed to obtain the parameter data after removing the redundant data as the first processing data of each time sequence feature. In actual process, there are a large amount of useless redundant data in the collected industrial data, and the influence time scale of each feature on the process product quality also differs. Skilled technicians can determine the adjustment time of each feature according to the production process and prior experience knowledge, that is, the influence time interval dimension on the process product quality may not be consistent. The present application expects to introduce such prior knowledge into the pre-processing of data, mask according to the influence time interval for different time sequence features, that is, remove the parameter data obtained at some time steps that have no or little influence on the process product quality. For each feature, the irrelevant interval is masked to simulate the prior experience of technicians. The mask mode is to directly set 0, and only the time sequence interval that the current feature may influence the predicted target process product quality is left, which is convenient for subsequent model training.

[0012] S2: constructing a monotonicity-constrained time sequence prediction model; As shown in the accompanying Figure 2 The monotonicity-constrained time sequence prediction model structure includes: An input layer, inputting a time sequence of (T, N), that is, parameter data of a plurality of time sequence features within a plurality of preset time steps, wherein T is a time sequence step length determined according to a sampling time and an interval length, and N is a time sequence feature number, which is also an input layer dimension, that is, the input layer has N nodes.

[0013] An LSTM layer and a first fully connected layer, the LSTM layer is a long short-term memory network layer connected with the input layer, and the first fully connected layer, that is, a Dense layer, is connected with the LSTM layer, used for time sequence processing and prediction of the input time sequence, outputting first feature data, the dimension of the first feature data being consistent with the time sequence feature number N; A second fully connected layer having the same number of nodes as the input layer; the second fully connected layer is used as a monotonicity-constrained layer for respectively making monotonicity constraints on the time sequence features, outputting second feature data, and the second fully connected layer adopts a Dense layer; in the monotonicity constraints, if the relationship between the time sequence feature and the process product quality is monotonically increasing, the constraint is positive, if the relationship is monotonically decreasing, the constraint is negative, and otherwise no constraint is added, and the second fully connected layer is used for splicing the calculation results into a sequence with a dimension of N; The calculation formula of the second fully connected layer is: , wherein, are time series feature data respectively, b is a bias term, which is iteratively updated in the model training process, are monotonicity constraint weights respectively, which are valued according to the relationship between the time series feature and the process product quality, if the relationship between the time series feature and the process product quality is monotonically increasing, the constraint is positive, if the relationship between the time series feature and the process product quality is monotonically decreasing, the constraint is negative, otherwise, no constraint is added, and the specific value is confirmed according to the actual situation; an output layer, connected with the first full connection layer and the second full connection layer, used for multiplying the first feature data and the second feature data, and outputting the final result through the full connection layer, that is, the Dense layer, as the output of the monotonicity-constrained time series prediction model.

[0014] S3: using a small sample reinforcement learning method to expand the training set for the first processing data, and training the monotonicity-constrained time series prediction model; For the quality inspection task of the process, the process product quality is often obtained through chemical analysis, and this offline mode leads to high cost and small amount of label data, that is, product quality data, so the training process of the monotonicity-constrained time series prediction model in the present application is a typical small sample semi-supervised learning problem. S3 specifically includes the following steps: The process product quality data corresponding to the first processing data is used as the existing label data, it should be noted that in the actual production process, not every set of parameter data of the time series feature in the preset time step has its corresponding process product quality data, more often, the number of the first processing data is much larger than the number of the process product quality data, generally, the difference is two or more orders of magnitude, therefore, in the present application, a small sample reinforcement learning method is needed to expand the training set. For the existing label data, a new label is added by introducing a dynamic time warping distance, that is, dtw; the existing label data is combined with the inverse distance weighting method to estimate the value of the current unlabeled data, and the calculated result is used as the label of the current unlabeled time series data for training, for example, 100 existing labels can be used to produce 10000 new labels, in this embodiment, each set of parameter data of the time series feature in the preset time step in the first processing data can also obtain its corresponding label data, which is convenient for subsequent training. The calculation formula of the new label value is: , wherein, is the label value of the i-th existing label data, is the dynamic time warping distance of the current unlabeled time series data from the i-th existing label data, the dtw value can be realized according to the prior art, which is not described here, and V is the estimated label value of the current unlabeled time series data; The monotonicity-constrained time series prediction model is trained using gradient descent method with mean square error as loss function during training process.

[0015] S4: input the current time series data into the trained monotonicity-constrained time series prediction model, output the process product quality result, and realize AI quality inspection.

[0016] Embodiment 2 Taking the three-effect rectification process as an example, it can be known from process experience that the tower top temperature difference, tower top temperature, tower top pressure, tower bottom temperature difference, tower bottom temperature, tower plate temperature difference and filler layer temperature difference have the greatest impact on the process target, i.e. product quality, as the input of the soft measurement model, and the respective influence time range is based on the current time, the influence interval of the tower top temperature difference is from 90 minutes to 60 minutes before the current time, the tower top temperature is from 90 minutes to 60 minutes before the current time, the tower top pressure is from 30 minutes to 20 minutes before the current time, the tower bottom temperature difference is from 85 minutes to 55 minutes before the current time, the tower bottom temperature is from 85 minutes to 55 minutes before the current time, the filler layer temperature is from 40 minutes to 30 minutes before the current time, and the tower plate temperature difference is from 100 minutes to 90 minutes before the current time.

[0017] First, data processing is performed to construct a time series data set. It is known that the system data sampling frequency is 3s, and it is assumed that one day of data is used, and chemical detection is performed once every 2 hours, so there are 12 quality data as label values in one day, and 7 different features each have 28800 production data. It is known that the current quality data is only related to the production data in the previous 2 hours, and the 2 hours of data have a total of 7200s, with a sampling frequency of 3s per piece, a total of 7200 / 3=2400 pieces of data, and the time series step T=2400. Therefore, a time series data set is established, i.e. 2400 pieces of production data correspond to 1 piece of quality data, and 28800 pieces of production data can generate 26401 time series data, each with a length of 2400, of which 12 time series data have quality labels, and the rest have no labels. Using a small sample reinforcement learning method, 12 labeled time series data are used as the basis, and the dtw distance of each unlabeled time series data with the 12 quality data is calculated, and then the corresponding label value is calculated by the following formula:

[0018] wherein the i-th label value, is the DTW distance between the current time series data and the i-th existing labeled data, V is the estimated label value of the current unlabeled time series data, and finally the input data of (26401, 2400, 7) and the output label of (26401, 1) are obtained. Then, the attention mechanism mask is performed on each group of time series data. Specifically, for the top temperature difference, the influence interval is from the previous 90 minutes to the previous 60 minutes, and the interval in the time series data is 600 (2400 - 90 x 60 / 3) to 1200 (2400 - 60 x 60 / 3). Therefore, the remaining part outside the interval is set to 0, and the remaining features are also set to 0. Then, the data is standardized.

[0019] Then, a monotonicity-constrained time series prediction model is constructed. Specifically, the input data dimension is (None, 2400, 7). First, the data is passed through a layer of LSTM neural network with 120 neurons. Then, the output of the LSTM is passed through a layer of fully connected neural network with 7 output neurons, which is used as the time series feature extracted by the LSTM. Then, a monotonically constrained fully connected layer is used for each feature of the time series data. Taking the top temperature difference as an example, the time series data is 2400, which is passed through a monotonically constrained fully connected layer, that is,

[0020] wherein the weight The monotonicity constraint is increased, for example, the higher the temperature, the higher the quality, and the weight constraint is positive. If the pressure is higher, the quality is lower, and the weight constraint is negative. If the monotonicity is unknown, no monotonicity constraint is performed. The results of the 7 features are spliced together to obtain 7 neurons, which are used as prior constraint weights. The prior constraint weights are multiplied by the time series features extracted by the LSTM, and the data format is (None, 7). Then, a fully connected layer is used to output the predicted quality with a dimension of (None, 1). The mean square error is used as the loss function, and the gradient descent method is used to train the model.

[0021] The above is only an example and does not limit the application field of the present application.

[0022] The "first" and "second" in the names mentioned in the embodiments of the present application are only used for name identification, and do not represent the first and second in order.

[0023] In addition, the present application also discloses an electronic device comprising a memory for storing a computer program; a processor for executing the computer program to implement an AI quality inspection method based on time series perception and monotonicity-constrained time series prediction as described above.

[0024] Further, the application also discloses a computer readable storage medium, and the computer readable storage medium stores a computer program, wherein the computer program is executed by a processor to realize the steps of the AI quality inspection method based on timing perception and monotonicity constraint timing prediction. As can be clearly understood by those skilled in the art from the description of the above embodiments, all or part of the steps of the above-mentioned embodiment method can be realized by means of software and a general hardware platform. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product. The computer software product can be stored in a storage medium, and the storage medium can be various types of memories, such as random access memory (RAM), read-only memory (ROM), flash memory, etc., such as read-only memory (ROM) / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the method described in various embodiments or some parts of the embodiments of the present application.

[0025] The above-mentioned is only the preferred embodiment of the present application, and it should be pointed out that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. An AI quality inspection method based on time-series awareness and monotonicity-constrained time-series prediction, characterized in that the method... The method comprises the following steps: Obtaining parameter data of a plurality of time sequence characteristics at different time steps in a process, removing redundant data in the time sequence characteristics based on prior knowledge to obtain first processing data as a training set; the first processing data includes parameter data of a plurality of time sequence characteristics within a plurality of preset time steps; Constructing a monotonicity-constrained time sequence prediction model; Using a small sample reinforcement learning method to expand the training set for the first processing data, and training the monotonicity-constrained time sequence prediction model; Inputting current time sequence data into the trained monotonicity-constrained time sequence prediction model to output process product quality results, thereby realizing AI quality inspection.

2. The AI quality inspection method based on timing perception and monotonicity constraint timing prediction according to claim 1, characterized in that: Based on prior knowledge, redundant data is removed, including: combining process experience to determine the maximum influence time interval of each time sequence characteristic on process product quality; for each time sequence characteristic, the parameter data obtained at other time steps outside the maximum influence time interval is removed to obtain parameter data after removing redundant data as the first processing data of each time sequence characteristic.

3. The AI quality inspection method based on timing perception and monotonicity constraint timing prediction according to claim 1, characterized in that: The structure of the monotonicity-constrained time sequence prediction model comprises: The input layer inputs a time sequence of (T, N), that is, parameter data of a plurality of time sequence characteristics within a plurality of preset time steps, wherein T is the time sequence step length, and N is the number of time sequence characteristics, that is, the input layer has N nodes; The LSTM layer and the first full connection layer are connected with the input layer, the first full connection layer is connected with the LSTM layer, and are used for time sequence processing and prediction of the input time sequence, outputting first feature data, the dimension of the first feature data being consistent with the number of time sequence characteristics N; The second full connection layer has the same number of nodes as the input layer; the second full connection layer is used as a monotonicity constraint layer to respectively constrain the monotonicity of the time sequence characteristics, and outputs second feature data; The output layer is connected with the first full connection layer and the second full connection layer, and is used for multiplying the first feature data and the second feature data, and outputting the final result through the full connection layer as the output of the monotonicity-constrained time sequence prediction model.

4. The AI quality inspection method based on timing perception and monotonicity constraint timing prediction according to claim 3, characterized in that: The calculation formula of the second full connection layer is: ， wherein, are respectively timing characteristic data, b is a bias term, are respectively monotonicity constraint weights, which are determined according to the relationship between the timing characteristic and the process product quality, if the relationship between the timing characteristic and the process product quality is monotonically increasing, the constraint is positive, monotonically decreasing, the constraint is negative, otherwise no constraint is added.

5. The AI quality inspection method based on timing perception and monotonicity constraint timing prediction according to claim 3, characterized in that, The time sequence step length is determined according to the sampling time and the interval length.

6. The AI quality inspection method based on timing perception and monotonicity constraint timing prediction according to claim 1, characterized in that, The small sample reinforcement learning method is used to expand the training set, including: The process product quality data in the first processing data is used as existing label data, new labels are added by introducing a dynamic time warping distance for the existing label data; the existing label data is used to estimate the value of the current unlabeled data by using the inverse distance weighting method, and the calculated result is used as the label of the current unlabeled time sequence data for training.

7. The AI quality inspection method based on timing perception and monotonicity constraint timing prediction according to claim 6, characterized in that: The calculation process of the label value estimated by the current unlabeled time sequence data comprises: ， wherein, is a label value of the ith existing labeled data, is a dynamic time warping distance between the current unlabeled time series data and the ith existing labeled data, and V is an estimated label value of the current unlabeled time series data.

8. An electronic device, comprising: The computer readable storage medium stores a computer program, wherein the computer program is executed by the processor to implement the steps of the AI quality inspection method based on time sequence perception and monotonicity-constrained time sequence prediction.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, wherein the computer program is executed by the processor to implement the steps of the AI quality inspection method based on time sequence perception and monotonicity-constrained time sequence prediction.

Images