A powder pressing quality prediction method and system based on spatiotemporal feature fusion

Abstract

The application discloses a powder pressing quality prediction method and system based on space-time feature fusion, which comprises the following steps: pressing data cleaning, obtaining standard pressing data SD; pressing data space feature extraction, using pressing data space correlation analysis and convolution operation to complete space feature extraction of the standard pressing data SD, obtaining the pressing data SM after space feature extraction; pressing data time feature extraction, dividing SM into a time window sequence, using LSTM to complete time feature extraction of the pressing data in the sequence, and obtaining a pressing data time sequence prediction sequence PL; and pressing data quality prediction, training a pressing data quality prediction model PModel based on SD, SM and PL, and using the model to complete quality prediction of to-be-tested pressing data. The application can fully extract space and time features in the data under the condition that the pressing data features are not obvious, and improve the accuracy of the pressing data quality prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of computer science, and particularly relates to a powder pressing quality prediction method and system based on space-time feature fusion. BACKGROUND

[0002] The powder pressing process is one of the core links in the industrial fields of metallurgical production and fuel production, and has important practical significance for industrial development and environmental protection. With the development of science and technology, deep learning has been widely applied in the field of industry, and also provides a technical basis for powder pressing quality prediction. Domestic scholars have applied deep learning to the research of powder pressing quality prediction, but there are some deficiencies: (1) the traditional outlier detection algorithm is adopted, the average value of all data is used as the reference index, the time sequence of the pressing data cannot be reflected, and the outliers in the time sequence interval cannot be normally detected; (2) the traditional maximum and minimum normalization algorithm is adopted, and the maximum value and the minimum value are usually caused by abnormal points, and for the pressing data with time sequence, the data measurement equipment may be in different measurement conditions at different time periods, resulting in a large difference in the normalization amplitude of the same parameter, so the method is easy to cause the normalization distortion of the time sequence; (3) the traditional prediction model such as BP neural network model and CNN neural network model is adopted, which cannot extract the time and space features in the pressing data at the same time, and it is difficult to obtain high-precision prediction results. SUMMARY

[0003] The application solves the technical problems: overcoming the deficiencies of the prior art, providing a powder pressing quality prediction method and system based on space-time feature fusion, which can perform abnormal detection and normalization based on the characteristics of the pressing data, and effectively extract the space-time features in the pressing data, thereby improving the precision of the powder pressing quality prediction.

[0004] The application solves the technical problems by adopting the following technical solutions:

[0005] The powder pressing quality prediction method based on space-time feature fusion comprises the following steps:

[0006] Step one: cleaning the pressing data, performing abnormal detection and adaptive normalization of the pressing data, and obtaining the cleaned standard pressing data, which is specifically implemented as follows:

[0007] (11) pressing data abnormality processing: according to the abnormal value E of the to-be-detected data c t at time t is calculated, wherein H is a set of pressing data in the time range [t-5, t+5], and h is a piece of pressing data in H, H is the minimum difference between the maximum value, the minimum value and the average value of the suppressed data in H; if E exceeds the acceptable abnormal value 5, it is judged that the to-be-tested data c is abnormal, and the average value of the suppressed data in H is replaced, otherwise it is judged that the to-be-tested data c is normal; all time suppressed data are processed according to the above operation to obtain the abnormality processed suppressed data ED;

[0008] (12) Adaptive normalization of suppressed data: according to the data segmentation formula The ED of the calculation step (11) is segmented into k equal length suppressed data sequences NL={D1, D2,..., D k}, wherein k represents the data segmentation length, represents the floor function, and the remaining data outside the sequence is discarded. According to the adaptive factor formula, the adaptive factor r and f of each segment of the suppressed data sequence D are calculated, and then the normalized formula D'=(D-r) / f is used to calculate and obtain the normalized suppressed data sequence NL'={D'1, D'2, D'3,...};

[0009] (13) Standard suppressed data construction: based on the time sequence characteristics of the suppression parameters, the normalized suppressed data sequence NL' of step (12) is spliced according to the time sequence splicing rule to form a two-dimensional matrix as the standard suppressed data SD;

[0010] Step two: suppressed data space feature extraction, which is implemented as follows:

[0011] (21) Suppressed data space correlation calculation: according to the suppressed data space correlation formula The correlation ρ s，t between the space parameter s and the time parameter t in the standard suppressed data SD of step (13) is calculated j , t j respectively represent the values of the space parameter s and the time parameter t at the j time, respectively represent the average values of the space parameter s and the time parameter t in the whole suppression process, and T is the total time of the suppression process;

[0012] (22) Suppressed data space weight calculation: according to the suppressed data space weight formula The space correlation ρ s，t calculated in step (21) is converted into the space weight w s，t of the space parameter s corresponding to the time parameter t;

[0013] (23) Pressing data space feature matrix construction: Place the 8 space parameter values of each time of the standard pressing data SD in step (13) around the 5 time parameter values of this time, then multiply the space parameter values s around each time parameter t by the corresponding space weight w respectively s，t , to obtain the pressing data space feature matrix SFM, as shown in Figure 2 ;

[0014] (24) Pressing data space feature extraction: Calculate the extracted space feature of the pressing data SM according to the formula , where is the convolution operator, and KN is the pressing data space feature kernel;

[0015] Step three: Pressing data time feature extraction, the specific implementation is as follows:

[0016] (31) Pressing data time window slicing: According to the time window formula l = T-w, calculate the number of pressing data time windows l, where w is the size of the pressing data time window and where T is the total time of the pressing process; Place the pressing data of the first l time of SM in step (24) at the starting position of the l time window respectively, and take the remaining w-1 time data from the starting position of the pressing data according to the time of the pressing data, and finally obtain the pressing data time window sequence PL = {W1, W2,..., W l};

[0017] (32) Time feature extraction based on LSTM: Input the elements in the pressing data time window sequence PL in step (31) into the LSTM network in turn, use the unique cell structure and memory mechanism in the network to extract the time features of the pressing data time window sequence, and train the network parameters of the LSTM to output the time sequence prediction sequence of the pressing data PL = {y'1, y'2,..., y' l}, where y' i represents the predicted data of the next time after the i-th time window by the LSTM network;

[0018] Step four: Pressing data quality prediction, the specific implementation is as follows:

[0019] (41) Pressing data training sample set construction: The training sample set TS includes the pressing data input set IS and the pressing data label set LS, where IS is the standard pressing data SM in step (13), and LS is a set {y1, y2,..., y l} composed of the relative density data of the powder at the pressing time w+1, 2w+1,..., lw+1 in step (24), where y ipowder relative density data of the i-th time point in the SM;

[0020] (42) Pressing data prediction set construction: the pressing data input set IS of step (41) is sequentially subjected to step two spatial feature extraction and step three time feature extraction, and the PL in step (32) is obtained as the pressing data prediction set PS;

[0021] (43) Pressing data prediction loss value calculation: according to the loss formula The loss value loss of the pressing data prediction is calculated, wherein the discriminant factor λ takes different values in different pressing stages, generally 0.07 in the pressing stage, 0.03 in the holding stage, 0.02 in the unloading stage, and 0.08 in the ejection stage;

[0022] (44) Pressing data quality prediction parameter optimization: using the Adam optimizer to optimize the data segmentation length k of step (12), the pressing data spatial feature convolution kernel KN of step (24), and the network parameters of step (32) LSTM, and stopping training when the loss value is less than the threshold value 0.001, so as to obtain the optimal k, KN, LSTM network parameters when the pressing data prediction loss value loss reaches the minimum;

[0023] (45) Pressing data quality prediction: based on the optimal k, KN, LSTM network parameters obtained in step (44), a quality prediction model PModel of the pressing data is constructed. The to-be-predicted pressing data PD is taken as the input of PModel, and the quality prediction result RD is obtained. The training and prediction flow chart is shown in Figure 3

[0024] Further, the adaptive factor formula is f = f' - β * (f' - max(D) + min(D)), wherein r, f are the adaptive factors of the pressing sequence D, r', f' are the adaptive factors of the previous sequence of D, max(D), min(D) respectively represent the maximum value and the minimum value in the pressing sequence D, α, β are constant coefficients, generally α = 0.5, β = 0.1;

[0025] Further, the time sequence splicing rule is to arrange each piece of data in the pressing data sequence into a two-dimensional matrix according to the time sequence, each row of the matrix stores the data of one time point, and each column stores the data of all time points of one pressing parameter.

[0026] ​Further, the pressing parameters include 8 space parameters and 5 time parameters, the space parameters are environment temperature, environment humidity, environment air pressure, die friction coefficient, powder aspect ratio, powder Poisson ratio, powder elastic modulus and pressing machine service time, and the time parameters are powder relative density, upper die stamping pressure, upper die stamping speed, powder equivalent Mises stress and powder equivalent elastic strain.

[0027] Compared with the prior art, the present application has the following advantages:

[0028] (1) Based on the time sequence of the pressing data, the present application narrows the abnormal detection reference range to the vicinity of the to-be-detected value, realizes accurate detection of abnormal values in different time sequence intervals, solves the problem of large error in global detection when the pressing data changes unstably, and improves the accuracy of powder pressing quality prediction.

[0029] (2) The present application adopts a self-adaptive normalization method, divides the pressing data into equal sequences, considers the correlation of adjacent pressing data sequences, solves the problems of gradient explosion and sequence distortion in traditional methods, and can improve the training convergence speed and prediction accuracy of the quality prediction model PModel, thereby improving the accuracy of powder pressing quality prediction.

[0030] (3) The present application can fully extract the time and space features in the pressing data through the space correlation formula, space weight formula, convolution and LSTM algorithm, provide feature significant training data for the prediction model PModel, obtain high-precision pressing quality prediction results, and improve the accuracy of powder pressing quality prediction. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 The method flowchart of the present application is shown in the figure;

[0032] Figure 2 The pressing data space matrix diagram is shown in the figure;

[0033] Figure 3 The PModel model training and prediction flowchart is shown in the figure. DETAILED DESCRIPTION

[0034] The present application will be further described in detail below with reference to the accompanying drawings.

[0035] The present application relates to a powder pressing quality prediction method based on spatiotemporal feature fusion, which comprises a pressing data cleaning module design, a pressing data space feature extraction module design, a pressing data time feature extraction module design and a pressing data quality prediction module, can perform abnormal detection and normalization based on the characteristics of the pressing data, and effectively extract the spatiotemporal features in the pressing data, thereby improving the accuracy of powder pressing quality prediction.

[0036] As Figure 1As shown, the specific embodiments of the present invention are as follows:

[0037] (1) Compression data cleaning module 1 performs anomaly detection and adaptive normalization on the compression data to obtain cleaned standard compression data. The specific implementation is as follows:

[0038] (11) Suppression data anomaly handling: Suppression data collected at each time point is stored in a table Ptable in the database according to the chronological order. The i-th row of Ptable stores the suppression data collected at the i-th time point; according to... Calculate the measured data c at time t t The outlier E is defined as follows: H is the set of suppressed data within the time range [max(1, t-5), min(T, t+5)], T is the total number of suppressed data points, i.e., the total number of rows in the Ptable, and h is a single suppressed data point in H. Let ct be the minimum difference between the maximum, minimum, and average values ​​of the suppressed data in H. The data to be tested, ct, is the suppressed data in the t-th row of the suppressed data table. The count(cond) function is used to calculate the number of times the condition cond is satisfied, which is taken as the value of E. E is compared with the acceptable outlier value 5. If E exceeds the acceptable outlier value 5, the data to be tested, c, is judged to be abnormal and replaced with the average value of the suppressed data in H. Otherwise, the data to be tested, c, is judged to be normal. The suppressed data at all times are processed in the above way to obtain the suppressed data PD after the outlier processing.

[0039] (12) Suppression data adaptive normalization: based on the data segmentation formula The suppressed data after anomaly handling in step (11) is divided into k equal-length suppressed data sequences NL = {D1, D2, ..., D...}. k}, where k represents the data segmentation length, This indicates rounding down, discarding any remaining data outside the sequence. (Based on the adaptive factor formula...) The formula f = f′ - β * (f′ - max(D) + min(D)) calculates the adaptive factors r and f for each suppressed data sequence D, where r and f are the adaptive factors of the suppressed sequence D, r′ and f′ are the adaptive factors of the previous sequence of D, max(D) and min(D) represent the maximum and minimum values ​​in the suppressed sequence D, respectively, and α and β are constant coefficients, generally taken as α = 0.5 and β = 0.1. Then, the normalized suppressed data sequence NL′ = {D′1, D′2, D′3...} is calculated and obtained by using the normalization formula D′ = (Dr) / f.

[0040] (13) Construction of standard compression data: Based on the temporal characteristics of the compression parameters, the compression data sequence NL′ after normalization in step (12) is arranged into a two-dimensional matrix according to the time order of each data in NL′. Each row of the matrix stores the data at one time moment, and each column stores the data of all times for a compression parameter. A two-dimensional matrix is ​​constructed as the standard compression data SD. Among them, the compression parameters include 8 spatial parameters and 5 temporal parameters. The spatial parameters are ambient temperature, ambient humidity, ambient air pressure, negative mold friction coefficient, powder aspect ratio, powder Poisson's ratio, powder elastic modulus and compression machine usage time. The temporal parameters are powder relative density, upper mold punching pressure, upper mold punching speed, powder equivalent Mises stress and powder equivalent elastic strain.

[0041] (2) The implementation of the data spatial feature extraction module 2 is as follows:

[0042] (21) Calculation of spatial correlation of suppressed data: Based on the formula for spatial correlation of suppressed data In step (13), the correlation ρ between each time parameter t and all spatial parameters s in the standard suppression data SD is calculated. s，t Where j represents the j-th moment in the suppression process, s j , t j Let represent the values ​​of the spatial parameter s and the time parameter t at time j, respectively. These represent the average values ​​of the spatial parameter s and the time parameter t throughout the entire compression process, respectively.

[0043] (22) Calculation of suppression data space weight: According to the suppression data space weight formula The spatial correlation ρ calculated in step (21) s，t Spatial weight w converted into spatial parameter s corresponding to time parameter t s，t ;

[0044] (23) Construction of the spatial feature matrix of the suppressed data: Place the 8 spatial parameter values ​​of each time step in the standard suppressed data SD of step (13) around the 5 time parameter values ​​corresponding to that time step, so that each time parameter is surrounded by the 8 spatial parameters at the same time step. Then multiply the spatial parameter values ​​s around each time parameter t by the corresponding spatial weight w. s，t The suppressed data space feature matrix SFM is obtained, such as Figure 2As shown in the figure, each row of matrix M1 stores the compaction data at a certain time (the specific time is marked by brackets in the figure), the text (excluding brackets and the text in brackets) of each element in M1 represents the name of the time parameter it stores, and the element is composed of matrix M2, the text (excluding brackets and the text in brackets) of each element in M2 represents the calculation formula of the product of the space parameter it stores and the corresponding space weight, and the blank part in M2 stores the value of the corresponding time parameter, such as Figure 2 The blank part in M2 in the middle is the value of the powder equivalent Mises stress at time 1.

[0045] (24) Compaction data spatial feature extraction: according to the formula The compaction data SM after extracting the spatial features is calculated, where is a convolution operator, KN is a two-dimensional convolution kernel for extracting spatial features, which is represented by a two-dimensional matrix at the mathematical level, and the initial value of each element is obtained by sampling from a Gaussian distribution with a mean of 0 and a variance of 1; wherein, the dimension of KN is 3x3, and the dimension of SFM is 3Tx15, KN is scanned from the top left corner of SFM from left to right and from top to bottom, and the sum of the products of the corresponding position elements of KN and the covered part of SFM (both are 3x3 matrices) is calculated, which is an element of SM; the step of KN movement is (3, 3), which means that after each calculation, it moves three steps to the right, but if it cannot move to the right because it has reached the right end of SFM, it moves three steps down and moves to the left end of SFM to start calculation and movement.

[0046] (3) Compaction data time feature extraction module 3, which is implemented as follows:

[0047] (31) Compaction data time window slicing: according to the time window formula l = T-w, the number of compaction data time windows l is calculated, where w is the size of the compaction data time window and where T is the total time of the compaction process; construct l one-dimensional vectors of length w to represent the l time windows, and place the compaction data of the first l time in step (24) in the starting position of the l time windows, and each time window W takes the remaining w-1 time data from the starting position of the compaction data according to the time of the compaction data, and finally obtains the compaction data time window sequence PL = {W1, W2,..., W l};

[0048] (32) Temporal feature extraction based on LSTM: First, initialize the LSTM network, that is, set its hyperparameters. input_size is the dimension of the input features and is set to the size w of the time window calculated in step (31). hidden_size is the dimension of the hidden layer state and is set to 128. num_layers is the number of LSTM stacked layers and is set to 5. bias is set to true to indicate that the hidden layer state has a bias term. dropout is a parameter to prevent network overfitting and is set to 0.4. Input the elements in the time window sequence PL of the suppressed data in step (31) into the LSTM network in sequence. Use the unique cell structure and memory mechanism in the network to extract the temporal features of the time window sequence of the suppressed data. Train the weight matrix network parameters of the memory gate, forget gate and output gate of the LSTM to output the temporal prediction sequence of the suppressed data PL={y′1,y′2,...,y′ l}, where y′ i This represents the prediction data of the LSTM network for the next time step after the i-th time window. The initial values ​​of the weight matrix are all sampled from a Gaussian distribution with a mean of 0 and a variance of 1.

[0049] (4) Suppress data quality prediction pre-block 4, the specific implementation is as follows:

[0050] (41) Construction of the training sample set for compression data: The training sample set TS includes the compression data input set IS and the compression data label set LS, where IS is the standard compression data SM in step (13), and LS is a set {y1, y2, ..., y3} consisting of the powder relative density data of SM at compression times w+1, 2w+1, ..., lw+1 in step (24). l}, where y i This represents the relative density data of the powder at time i in SM;

[0051] (42) Construction of the suppression data prediction set: The suppression data input set IS in step (41) is subjected to spatial feature extraction in step two and temporal feature extraction in step three to obtain PL in step (32) as the suppression data prediction set PS.

[0052] (43) Calculation of loss value for data suppression prediction: Based on the loss formula Calculate the loss value of the compression data prediction, where the discriminant factor λ takes different values ​​in different compression stages. Generally, it takes a value of 0.07 in the pressurization stage, 0.03 in the pressure holding stage, 0.02 in the pressure release stage, and 0.08 in the mold removal stage.

[0053] (44) Compression data quality prediction parameter optimization: using Adam optimizer to optimize the data segmentation length k of step (12), the compression data space feature convolution kernel KN of step (24) and the network parameters of LSTM of step (32); before optimization, the parameters of Adam optimizer need to be set, including alpha=0.001, beta1=0.9, beta2=0.999 and epsilon=10E-8, wherein alpha is the learning rate, which is used to control the update rate, beta1 is the exponential decay rate of the first moment estimate, beta2 is the exponential decay rate of the second moment estimate, and the epsilon parameter is a very small number to prevent division by zero in implementation; during the training process, when the loss value is less than the threshold value 0.001, the training is stopped, so as to obtain the optimal k, KN, LSTM network parameters when the compression data prediction loss value loss reaches the minimum;

[0054] (45) Compression data quality prediction: based on the optimal k, KN, LSTM network parameters obtained in step (44), a compression data quality prediction model PModel is constructed. The to-be-predicted compression data PD is taken as the input of PModel, and the quality prediction result RD is obtained, and the training and prediction flowchart is as shown in Figure 3

[0055] In summary, the present application discloses a powder compression quality prediction method based on spatiotemporal feature fusion, including compression data cleaning module design, compression data space feature extraction module design, compression data time feature extraction module design and compression data quality prediction module, which can perform abnormal detection and normalization based on the characteristics of the compression data, and effectively extract the spatiotemporal features in the compression data, thereby improving the accuracy of powder compression quality prediction.

[0056] The contents not described in detail in the specification of the present application belong to the prior art known to those skilled in the art.

[0057] The above description is only the preferred embodiments of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.​

Claims

1. A powder pressing quality prediction method based on space-time feature fusion, characterized in that, Step 1: pressing data cleaning, abnormal detection and adaptive normalization of pressing data are performed to obtain cleaned standard pressing data, as follows: (11) According to ,calculate Data to be measured at time outliers , Time range The internal suppression data set, for One of the suppression data, for The minimum difference between the maximum, minimum, and average values ​​of the suppressed data; if If the outlier exceeds 5, then the data to be tested is considered invalid. Abnormal, and with If the average value of the suppressed data is used as the substitute, otherwise the test data is judged. Normal; Perform anomaly processing on all time-series suppression data using the above steps to obtain anomaly-processed suppression data. ; (12) According to the data segmentation formula , the calculation step (11) is segmented into the same length of the suppression data sequence , , which represents the data segmentation length, , represents the floor, the remaining data outside the sequence is discarded, and the adaptive factor of each segment of the suppression data sequence is calculated according to the adaptive factor formula and , the normalized suppression data sequence is calculated and obtained through the normalization formula ; (13) Based on the temporal characteristics of the suppression parameters, the suppression data sequence after normalization in step (12) is... The data is spliced ​​according to the time sequence splicing rules to construct a two-dimensional matrix as the standard suppressed data. The pressing parameters include 8 spatial parameters and 5 temporal parameters. The spatial parameters include ambient temperature, ambient humidity, ambient air pressure, coefficient of friction of the female mold, aspect ratio of the powder, Poisson's ratio of the powder, elastic modulus of the powder, and operating time of the pressing machine. The temporal parameters include relative density of the powder, pressing pressure of the upper mold, pressing speed of the upper mold, equivalent Mises stress of the powder, and equivalent elastic strain of the powder. Step 2: pressing data space feature extraction, as follows: (21) calculating the correlation between the space parameter and the time parameter in the standard pressing data of step (13) according to the pressing data space correlation formula ; denotes a certain moment in the pressing process, denotes the value of the space parameter and the time parameter at the moment denotes the average value of the space parameter and the time parameter in the entire pressing process, is the total time of the pressing process;​​​​​​​​​ (22) The spatial weight is converted into a spatial parameter according to the formula ;​​​​ (23) the standard pressing data of step (13) There are 5 time parameters at each moment, each of which is surrounded by 8 space parameters at the same moment, and each time parameter is multiplied by the corresponding space weight , respectively, to obtain the space feature matrix of the pressing data ;​​ (24) The suppressed data after extracting spatial features is calculated according to the formula , , is a convolution operator, is a suppressed data spatial feature kernel; Step 3: pressing data time feature extraction, as follows: (31) According to the time window formula Calculate the number of time windows for suppressing data. , To suppress the data time window size and , The total time for the suppression process; in step (24) The suppression data of the first t time steps are respectively placed in The starting position of each time window, and the starting position of each time window. Take the remaining data from the time corresponding to the suppression data at the starting position. The data at each moment ultimately yields the suppressed data time window sequence. ; (32) The time window sequence of the suppression data from step (31) Input the elements in sequence In the network, the unique cellular structure and memory mechanism of the network are used to extract the temporal features of the suppressed data time window sequence, and the network is trained. Network parameters are used to predict the time series of suppressed output data. , express Network for the first Predicted data for the next moment after a time window; Step 4: pressing data quality prediction, as follows: (41) training sample set including a set of suppressed data inputs and a set of suppressed data labels , for the standard suppressed data in step (13) , is a set of powder relative density data at the time points in step (24) , denotes the powder relative density data at the th time point in the set ;​ (42) inputting the pressed data of step (41) into a set The spatial feature extraction of step two and the temporal feature extraction of step three are performed in sequence to obtain the data in step (32) as a prediction set of pressed data ; (43) calculating a loss value for the suppressed data prediction according to a loss formula is a discriminant factor;​​ (44) using The optimizer optimizes the data segmentation length of step (12) , step (24) presses the data space feature convolution kernel And step (32) The network parameters are optimized, and the training is stopped when the loss value is less than the threshold value 0.001, so as to obtain the prediction loss value of the pressed data When the minimum is reached, the optimal 、 、 Network parameters; (45) based on step (44) to obtain the optimal 、 、 network parameters, construct a quality prediction model of the suppression data , input the to-be-predicted suppression data as input, and obtain a quality prediction result .

2. The powder compaction quality prediction method based on spatiotemporal feature fusion according to claim 1, characterized in that: The formula for calculating the adaptive factor in step (12) is , is the adaptive factor of the suppressed data sequence , is the adaptive factor of the previous sequence of , respectively represent the maximum value and the minimum value in the suppressed data sequence , is a constant coefficient, and the value is .

3. The powder compaction quality prediction method based on spatiotemporal feature fusion according to claim 1, characterized in that: In the step (13), the time sequence splicing rule is to arrange each data in the pressing data sequence into a two-dimensional matrix in time sequence, and each row of the matrix stores data at one time, and each column stores data of all times of one pressing parameter.

4. The powder compaction quality prediction method based on spatiotemporal feature fusion according to claim 1, characterized in that: In the step (43), The value is different in different pressing stages, 0.07 in the pressing stage, 0.03 in the pressure maintaining stage, 0.02 in the pressure releasing stage, and 0.08 in the die releasing stage.

5. A powder compaction quality prediction system based on spatio-temporal feature fusion, characterized in that, It comprises a pressing data cleaning module, a pressing data space feature extraction module, a pressing data time feature extraction module and a pressing data quality prediction module. The pressing data cleaning module performs abnormal detection and adaptive normalization of the pressing data to obtain cleaned standard pressing data, as follows: (11) A method according to any one of (1) to (10), wherein , for each time point, an outlier of the data to be tested is calculated , , is a set of suppressed data in a time range , is one of the suppressed data in , is the minimum difference between the maximum, minimum and average of the suppressed data in ; if is greater than the acceptable outlier 5, the data to be tested is determined to be abnormal and is replaced by the average of the suppressed data in , otherwise the data to be tested is determined to be normal; all the suppressed data at each time point is processed according to the above operation to obtain the processed suppressed data . (12) According to the data segmentation formula , the calculation step (11) is segmented into the same length of the suppression data sequence , , the data segmentation length is represented by , the remaining data outside the sequence is discarded, and the adaptive factor of each segment of the suppression data sequence is calculated according to the adaptive factor formula and , and the normalized suppression data sequence is calculated and obtained by the normalization formula ; (13) Based on the temporal characteristics of the suppression parameters, the suppression data sequence after normalization in step (12) is... The data is spliced ​​according to the time sequence splicing rules to construct a two-dimensional matrix as the standard suppressed data. The pressing parameters include 8 spatial parameters and 5 temporal parameters. The spatial parameters include ambient temperature, ambient humidity, ambient air pressure, coefficient of friction of the female mold, aspect ratio of the powder, Poisson's ratio of the powder, elastic modulus of the powder, and operating time of the pressing machine. The temporal parameters include relative density of the powder, pressing pressure of the upper mold, pressing speed of the upper mold, equivalent Mises stress of the powder, and equivalent elastic strain of the powder. The pressing data space feature extraction module, as follows: (21) calculating the correlation between the space parameter and the time parameter in the standard pressing data of step (13) according to the pressing data space correlation formula represents a certain moment in the pressing process, represents the value of the space parameter and the time parameter at the moment of represents the average value of the space parameter and the time parameter in the entire pressing process, is the total time of the pressing process;​​​​​​​​​​ (22) The spatial weight is converted into a spatial parameter according to the formula The spatial correlation calculated in step (21) is converted into a spatial parameter corresponding to the time parameter of the spatial weight ; (23) the standard pressing data of step (13) is multiplied by the spatial weight of the corresponding spatial parameter There are 5 time parameters at each moment, each of which is surrounded by 8 spatial parameters at the same moment, and each time parameter is multiplied by the spatial parameter value around it respectively corresponding spatial weight , to obtain the spatial feature matrix of the pressing data ; (24) The suppressed data after extracting spatial features is calculated according to the formula , , is a convolution operator, is a suppressed data spatial feature kernel; The pressing data time feature extraction module, as follows: (31) Calculate the number of time windows of the suppression data according to the time window formula (32) Calculate the size of the time window of the suppression data , (33) Calculate the total time of the suppression process , (34) Place the suppression data of the first t time points in step (24) at the start positions of the time windows, respectively, and each time window (35) Take the data of the remaining t-1 time points from the time point of the start position of the suppression data according to the time point of the start position of the suppression data, and finally obtain the time window sequence of the suppression data ​​​​ (32) The time window sequence of the suppression data from step (31) Input the elements in sequence In the network, the unique cellular structure and memory mechanism of the network are used to extract the temporal features of the suppressed data time window sequence, and the network is trained. Network parameters are used to predict the time series of suppressed output data. , express Network for the first Predicted data for the next moment after a time window; The pressing data quality prediction module, as follows: (41) training sample set including a set of pressed data inputs and a set of pressed data labels , for the standard pressed data in step (13) , is a set of powder relative density data at the time points in step (24) , , denotes the powder relative density data at the time point in step (24); and the time point in step (24). (42) inputting the pressed data of step (41) into a set The spatial feature extraction of step two and the temporal feature extraction of step three are performed in sequence to obtain the data in step (32) as a prediction set of pressed data ; (43) calculating a loss value for the suppressed data prediction according to a loss formula , is a discriminant factor;​ (44) using The optimizer optimizes the data segmentation length of step (12) , the convolution kernel of step (24) pressing data space features , and the network parameters of step (32) Optimize, stop training when the loss value is less than the threshold value 0.001, so as to obtain the prediction loss value of the pressing data When the minimum value is reached, the optimal 、 、 Network parameters; (45) based on step (44) to obtain the optimal 、 、 network parameters, construct a quality prediction model of the suppression data , input the to-be-predicted suppression data as input of the quality prediction model, and obtain a quality prediction result .

Images