Manufacturing data efficient dimension reduction processing device and method based on dimension reduction immune algorithm

Abstract

The invention discloses a manufacturing data efficient dimension reduction processing device and method based on a dimension reduction immune algorithm, and belongs to the field of manufacturing data processing. The device comprises a data acquisition module, a data processing module, a dimensionality reduction immune algorithm module and a result output module which are connected in sequence, the core is a dimensionality reduction immune algorithm module, and the device comprises an antibody antigen initialization unit, an affinity calculation unit with a dimensionality reduction penalty factor and the like. The method comprises the following steps: collecting manufacturing data, standardizing, initializing an antibody antigen, calculating affinity with a penalty factor, dynamically cloning a variant antibody, and screening an optimal feature subset after iterative judgment. According to the method, the prediction precision and the dimension reduction effect are balanced through the dimension reduction penalty factor, the efficiency is improved through dynamic clone variation, the manufacturing data dimension can be reduced, the precision is guaranteed, support is provided for optimal scheduling and intelligent decision making of a manufacturing system, and compatibility is high.

Description

Technical Field

[0001] This invention relates to the field of manufacturing data processing technology, specifically to a device and method for efficient dimensionality reduction processing of manufacturing data based on a dimensionality reduction immune algorithm. Background Technology

[0002] In modern manufacturing, raw enterprise data is a crucial foundation for building scheduling knowledge bases, optimizing production, and making intelligent decisions. However, raw data collected from actual manufacturing systems often suffers from numerous problems, such as missing data, data noise, excessively high data dimensionality, and complex data attributes. These issues not only increase the complexity of data processing but also severely impact the accuracy and efficiency of subsequent research processes such as feature selection, data prediction, and multi-objective optimization. Therefore, effective preprocessing of raw data is a key step in ensuring data quality and mining accuracy.

[0003] Data preprocessing mainly includes four parts: data cleaning, data integration, data transformation, and data reduction. Among them, data cleaning mainly addresses the problems of data noise and missing data; data integration focuses on merging datasets from different data sources; data transformation aims to convert sample data into a format that is easy for data mining algorithms to process; and data reduction, through feature selection or feature extraction, obtains smaller datasets that maintain integrity from large datasets, thereby reducing data redundancy and computational complexity.

[0004] The preprocessing techniques for complex manufacturing data mainly include three aspects: data attribute selection, data clustering, and data attribute discretization.

[0005] Data attribute selection aims to identify the key attributes that have the greatest impact on the target variable from a large number of attributes, thus addressing issues such as attribute redundancy, high computational cost, and long computation time. Existing methods include statistical methods, information theory-based methods, and machine learning-based methods. However, these methods often face the curse of dimensionality when dealing with high-dimensional data, making it difficult to effectively balance dimensionality reduction and prediction accuracy.

[0006] Data clustering: This involves classifying samples based on their similarity, grouping similar samples into the same category. Commonly used clustering algorithms include K-means, hierarchical clustering, and DBSCAN. However, these algorithms often face challenges such as high computational complexity and unstable clustering results when dealing with large-scale, high-dimensional data.

[0007] Data attribute discretization: Some algorithms and models (such as rough sets and decision trees) can only handle discrete data, thus requiring the transformation of continuous attribute values ​​into discrete attribute values. Existing discretization methods include equal-width discretization, equal-frequency discretization, and discretization based on information entropy. However, these methods often struggle to preserve the original information and distribution characteristics of high-dimensional continuous data when processing it.

[0008] Although existing methods can effectively reduce feature dimensionality and maintain high prediction accuracy to some extent, they still have the following significant drawbacks: Limited dimensionality reduction effect: The dimension of the feature attributes selected by existing methods is still relatively high, resulting in large data redundancy and making it difficult to meet the needs of subsequent data analysis and optimization.

[0009] High computational complexity: In high-dimensional data scenarios, the computational complexity of existing methods often increases exponentially, resulting in low processing efficiency and difficulty in meeting real-time requirements.

[0010] Balancing prediction accuracy and dimensionality reduction effectiveness is challenging: Existing methods often struggle to achieve high prediction accuracy while simultaneously improving dimensionality reduction, leading to data redundancy and increased computational complexity.

[0011] Therefore, it is necessary to explore an attribute selection method that can significantly reduce feature dimensionality while maintaining high prediction accuracy, in order to solve the bottleneck problem in the existing technology. Summary of the Invention

[0012] This invention aims to address the problems in existing technologies, such as high data dimensionality after feature selection, high computational complexity, and the difficulty in balancing prediction accuracy and dimensionality reduction effectiveness. By introducing a dimensionality reduction immune algorithm, this invention provides a device and method for efficient dimensionality reduction processing of manufacturing data based on this algorithm, further reducing feature dimensionality, improving data mining accuracy, reducing data redundancy, and providing strong support for the optimized scheduling and intelligent decision-making of manufacturing systems.

[0013] The technical solution of this invention: The present invention provides a high-efficiency dimensionality reduction processing device for manufacturing data based on a dimensionality reduction immune algorithm, comprising a data acquisition module, a data processing module, a dimensionality reduction immune algorithm module, and a result output module. The feature selection process based on the improved immune algorithm is as follows: Figure 1 As shown.

[0014] Immune algorithms are a class of algorithms that incorporate immunological principles into genetic algorithms. They borrow strategies from biological immune processes, such as maintaining antibody diversity and adding vaccines, to ensure the search for optimal solutions in the solution space. The core of immune algorithms is the clonal selection principle, proposed by immunologist Burner in 1957, which is widely used in artificial immune systems. This principle posits that under antigen stimulation, immune cells undergo mass clonal proliferation, subsequently dividing into antibodies and memory cells. During this process, only immune cells with high antigen binding affinity can be cloned and proliferated in large numbers, and mutations occur during proliferation, selecting immune cells with even higher affinity for the antigen, forming a positive feedback loop. In this paper, the attribute selection process based on the immune algorithm is as follows: Step 1: Antibody and antigen initialization. Here, antibodies are different combinations of the production line attribute set, i.e., the candidate set, where antibody encoding is as follows: Figure 2 As shown, "1" indicates that the production attribute is selected as a feature attribute, and "0" indicates that the production attribute is not selected as a feature attribute. Initial antibody population generate A number of antibodies were collected, and a portion of these antibodies were randomly selected and added to the antibody memory library; antigen populations were obtained from the sample data. Each antigen includes the antigen itself. (The performance metrics of interest) and antigenic determinants (Production line status attributes).

[0015] Step 2: Calculate the affinity between the antibody and the antigen (the strength of binding between the antibody and the antigenic determinant). The higher the value, the closer the candidate solution is to the true solution. Train a new antigen ontology using a BP neural network. : The BP neural network takes as input a set of production line performance attributes selected by the immune algorithm antibodies and outputs the performance indicators of interest. It is trained using a test set to generate the representation of these antibody groups. One test item can get one ,common The BP neural network here uses a three-layer structure, with one hidden layer containing 13 neurons.

[0016] and The similarity between them is calculated as follows: and The greater the similarity between the antibodies, the higher their affinity. With each The average affinity is taken as the final affinity of the antibody, as shown in the following formula: While immune algorithms outperform other common algorithms in prediction accuracy, they require a relatively large number of dimensions, which hinders subsequent optimization. Therefore, this paper proposes a dimensionality-reducible immune algorithm, called the dimensionality-reduced immune algorithm. Its basic process is similar to the immune algorithm, but a larger dimensionality reduction penalty factor is artificially added when calculating the affinity between antibodies and antigens, as shown in the following equation: in, yes Number of selected dimensions, This is the average dimensionality of the previous generation of antibodies. Through this operation, the dimensionality can be further reduced while improving the antigen-antibody fitness.

[0017] Step 3: Antibody Cloning. Based on antibody affinity, antibodies are cloned and mutated (the higher the antibody affinity, the more clones are produced and the lower the probability of mutation). Daughter antibodies with increased affinity are selectively used to replace parent antibodies, thereby improving the overall affinity of the antibody population.

[0018] Step 4: Determine if the maximum number of iterations has been reached. If the maximum number of iterations has been reached, proceed to step 7; otherwise, proceed to step 5.

[0019] Step 5: Calculate the average similarity of antibodies using the reciprocal of the Euclidean distance according to the following formula. The closer the distance between antibodies, the greater their similarity.

[0020] Step 6: If the average similarity of antibodies is less than the threshold, sort the antibodies in the antibody population by affinity and select the appropriate antibody group. Add the antibody with the highest affinity to the antibody library and replace an equal number of antibodies with the lowest affinity, then proceed to step 2; otherwise, proceed to step 7.

[0021] Step 7: Screen the antibodies in the antibody group and select the one with the highest affinity as the optimal solution.

[0022] This invention introduces a dimensionality reduction immune algorithm, which significantly reduces the dimensionality of the selected feature subset while maintaining high prediction accuracy, thus reducing data redundancy. Specifically, the beneficial effects of this invention include: Improving data mining efficiency: By reducing feature dimensions, the computational load and storage space requirements in the data mining process are reduced, thereby improving the efficiency of data mining.

[0023] Enhancing model generalization ability: The feature subset after dimensionality reduction is more compact and representative, which helps to improve the model's generalization ability and robustness.

[0024] Optimizing manufacturing system scheduling: The reduced-dimensional data provides more accurate and efficient data support for the optimized scheduling and intelligent decision-making of the manufacturing system, which helps to optimize the production process and make efficient use of resources. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the feature selection process based on the improved immune algorithm of the present invention.

[0026] Figure 2 This is a schematic diagram of the antibody encoding of the present invention. Detailed Implementation

[0027] See attached document Figure 1-2 This invention provides a device for efficient dimensionality reduction processing of manufacturing data based on a dimensionality reduction immune algorithm, comprising: The data acquisition module is used to collect raw, multi-dimensional data from the manufacturing system. A data processing module, connected to the data acquisition module, is used to perform data cleaning and standardization preprocessing on the raw multidimensional data. The dimension reduction immune algorithm module is connected to the data processing module and is used to run the dimension reduction immune algorithm to perform feature selection on the preprocessed data. The dimension reduction immune algorithm is based on the principle of immune clone selection and introduces a dimension reduction penalty factor related to the selected feature dimension in the antibody-antigen affinity evaluation function. The result output module is connected to the dimensionality reduction immune algorithm module and is used to output the optimal feature subset and the dimensionality-reduced data.

[0028] Furthermore, the dimensionality reduction immune algorithm module includes: The initialization unit is used to randomly generate the initial antibody population and initialize the antigen population. The antibodies are encoded in binary, with each bit representing whether a feature is selected. An affinity calculation unit is used to calculate the affinity between the antibody and the antigen. The affinity calculation employs a BP neural network to predict the performance of a selected feature subset, and uses the similarity between the predicted result and the true value as the primary evaluation criterion. Simultaneously, a dimensionality reduction penalty factor is introduced to correct the affinity; the correction formula is as follows: ; Cloning and mutation unit, used for cloning, proliferation and mutation operations based on antibody affinity; The memory bank update unit is used to update the antibody memory bank based on the similarity and affinity between antibodies to maintain population diversity; The iterative control unit is used to control the algorithm's iterative process until the termination condition is met and the optimal feature subset is output.

[0029] This invention provides a method for efficient dimensionality reduction processing of manufacturing data based on a dimensionality reduction immune algorithm, applied to the aforementioned device. The method includes the following steps: S1: Acquire raw multidimensional data of the manufacturing system through the data acquisition module; S2: The data processing module performs data cleaning and standardization preprocessing on the raw data; S3: Feature selection is performed through the dimensionality reduction immune algorithm module, the process of which includes: S31: Initialize the antibody population and antigen population; S32: Calculate the affinity of each antibody for the antigen, introducing a dimensionality reduction penalty factor during the calculation process; S33: Antibody cloning and mutation operations are performed based on affinity; S34: Calculate the average similarity of the antibody population and update the memory database; S35: Determine whether the termination condition has been met. If not, return to S32; if it has been met, proceed to S36. S36: Select the antibody with the highest affinity as the optimal feature subset; S4: Output the optimal feature subset and the dimensionality-reduced data through the result output module.

[0030] Furthermore, the dimensionality reduction penalty factor is used to penalize antibodies with a large number of selectable features in affinity calculation, and its intensity is controlled by the penalty coefficient.

[0031] The method used in this paper to obtain samples is to obtain scheduling-related simulation data by running a semiconductor production line simulation model. The simulation data results are complete and stored in a local database, so the sample data required in this paper only needs to undergo data transformation (digital processing and standardization).

[0032] Digitalization Computers can only perform numerical calculations, so textual content in the data must be represented in a numerical form that computers can recognize. In the actual simulation results presented in this paper, production attributes and performance indicators are concrete numbers, while scheduling policies are represented by English letters, such as FIFO. Therefore, the scheduling policy selected in the sample validation is converted into integers starting from 1.

[0033] standardization First, the sample data is standardized, and data mining is performed based on the standardized data. Commonly used data standardization methods include: Min-Max standardization, fractional scaling standardization, and Z-score standardization. Here, we choose Min-Max standardization: a linear transformation is applied to the original data, mapping the result to the range [0,1], as shown in the following equation: ; in, For standard data, This is the original data. The maximum value in the sample data. It is the minimum value in the sample data.

[0034] In the attribute selection phase, 67 dimensions of attributes related to the target of interest are given. Learning samples with different performance are obtained by providing different scheduling rule parameters. Combined with historical production data, the production line status is divided into three cases: light load, full load, and overload. Attribute selection operations are then performed on the simulation data for each case. A total of 1500 valid samples are selected for feature selection, with 500 samples for each load case. 100 samples are used as validation samples to verify the learned attribute set. Then, by comparing the prediction accuracy, the most suitable feature subset is selected. The feature subset results are shown in Tables 1-3.

[0035] Table 1. Subset of Production Attribute Features (Light Load)

[0036] Table 2. Subset of Production Attribute Features (Full Load)

[0037] Table 3. Subset of Production Attribute Features (Overload)

[0038] As can be seen from the three tables above, the immune algorithm outperforms other common algorithms in prediction accuracy. The dimensionality reduction immune algorithm not only reduces the dimensionality of the selected feature subset but also further improves prediction accuracy and reduces data redundancy. Therefore, the production line attribute set obtained using the dimensionality reduction immune algorithm is used for optimization operations.

Claims

1. A high-efficiency dimensionality reduction processing device for manufacturing data based on a dimensionality reduction immune algorithm, characterized in that, It includes a data acquisition module, a data processing module, a dimensionality reduction and immune algorithm module, and a result output module connected in sequence. The data acquisition module uses industrial-grade data acquisition sensors and data acquisition cards to collect production line status data, scheduling rule parameter data and performance index data in the manufacturing system in real time. The data processing module includes a digitization unit and a standardization unit. The digitization unit converts non-numerical information in the original data into integers, and the standardization unit uses the Min-Max standardization method to map the digitized data to the [0,1] interval. The standardization formula is: , where x ∗ The data are standardized, x is the original data, max is the maximum value of the sample data, and min is the minimum value of the sample data. The dimensionality reduction immune algorithm module is the core module, including an antibody antigen initialization unit, an affinity calculation unit, an antibody clonal mutation unit, an iterative judgment unit, an antibody similarity calculation unit, and an optimal solution screening unit; The result output module uses an industrial-grade display screen and USB and Ethernet data interfaces to display the optimal feature subset in tabular form and supports data export to storage devices or manufacturing system scheduling and decision-making platforms.

2. The efficient dimensionality reduction processing device for manufacturing data based on the dimensionality reduction immune algorithm according to claim 1, characterized in that, In the dimensionality reduction immune algorithm module, the antibody antigen initialization unit defines different combinations of manufacturing data attribute sets as antibodies, using binary encoding, where "1" represents that the production attribute is selected as a feature attribute, and "0" represents that the production attribute is not selected as a feature attribute; initial antibody population. generate A number of antibodies were collected, and a portion of these antibodies were randomly selected and added to the antibody memory library; antigen populations were obtained from the sample data. Each antigen includes performance metrics data of interest as its antigen ontology. Production line status attribute data as antigen determinant clusters .

3. The efficient dimensionality reduction processing device for manufacturing data based on the dimensionality reduction immune algorithm according to claim 1, characterized in that, The affinity calculation unit uses a BP neural network to train the antigen ontology. The BP neural network has a three-layer structure. The number of neurons in the input layer is the same as the number of feature attributes selected by the antibody. The number of neurons in the hidden layer is 13. The number of neurons in the output layer is the same as the number of performance indicators. The gradient descent method is used to optimize the parameters. The number of training iterations is 1000 and the learning rate is 0.

01. Training a new antigen ontology using a BP neural network : The BP neural network takes as input a set of production line performance attributes selected by the immune algorithm antibodies and outputs the performance indicators of interest. It is trained using a test set to generate the representation of these antibody groups. One test item can get one ,common indivual; and The similarity between them is calculated as follows: ； and The greater the similarity between the antibodies, the higher their affinity. With each The average affinity is taken as the final affinity of the antibody, as shown in the following formula: ； While immune algorithms outperform other common algorithms in prediction accuracy, they require a larger number of dimensions, which hinders subsequent optimization. Therefore, a dimensionality-reducing immune algorithm, called the dimensionality-reducing immune algorithm, is proposed here. Its basic process is the same as the immune algorithm, but a larger dimensionality reduction penalty factor is artificially added when calculating the affinity between antibodies and antigens, as shown in the following formula: ； in, yes Number of selected dimensions, This is the average dimensionality of the previous generation of antibodies. Through this operation, the dimensionality is further reduced while improving the antigen-antibody fitness.

4. The efficient dimensionality reduction processing device for manufacturing data based on the dimensionality reduction immune algorithm according to claim 1, characterized in that, The antibody cloning mutation unit determines the number of clones and the mutation probability based on the final affinity of the antibody. The number of clones ranges from 5 to 20, with higher affinity resulting in a larger number of clones. The mutation probability ranges from 0.01 to 0.05, with higher affinity resulting in a smaller mutation probability. Single-point mutation is performed on the cloned progeny antibodies, and progeny antibodies with higher affinity than the parent antibodies are selected to replace the parent antibodies.

5. The efficient dimensionality reduction processing device for manufacturing data based on the dimensionality reduction immune algorithm according to claim 1, characterized in that, The iteration judgment unit sets the maximum number of iterations to 100-200, and judges whether the current number of iterations has reached the maximum number of iterations. If it has, the antibody group is transferred to the optimal solution screening unit; otherwise, it is transferred to the antibody similarity calculation unit.

6. The efficient dimensionality reduction processing device for manufacturing data based on the dimensionality reduction immune algorithm according to claim 1, characterized in that, The antibody similarity calculation unit uses the reciprocal of the Euclidean distance to calculate the similarity between antibodies, using the formula: .

7. The efficient dimensionality reduction processing device for manufacturing data based on the dimensionality reduction immune algorithm according to claim 1, characterized in that, The optimal solution screening unit sorts the antibody groups in descending order of affinity, selects the antibody with the highest affinity as the optimal solution, and the corresponding feature attribute set is the optimal feature subset after dimensionality reduction, which is then transmitted to the result output module.

8. A method for efficient dimensionality reduction processing of manufacturing data based on a dimensionality reduction immune algorithm, characterized in that, Based on the apparatus according to any one of claims 1-7, the method includes the following steps: Step 1: Data Acquisition. The data acquisition module collects production line status data, scheduling rule parameter data, and performance index data in the manufacturing system in real time. The acquisition frequency is set to 5Hz, and the acquisition duration is determined according to the production cycle of the manufacturing system. Step 2: Data processing. Digital processing converts the non-numerical information in the original data into integers. Standardization processing uses the Min-Max standardization method to map the data to the [0,1] interval to obtain standardized sample data, which is then divided into training samples and test samples in an 8:2 ratio. Step 3: Antibody and antigen initialization. Different combinations of data attribute sets are defined as antibodies, encoded in binary. 50-100 antibodies are generated to form an initial antibody population. 20% of the antibodies are randomly selected and added to the antibody memory bank. Antigen populations are extracted from training samples. Step 4: Affinity calculation, construct and train the BP neural network to obtain prediction performance metrics. ， Calculate initial affinity ， Introducing a dimensionality reduction penalty factor to obtain the final affinity ； Step 5: Antibody cloning and mutation. Determine the number of clones and mutation probability based on the final affinity. Perform single-point mutation on the cloned progeny antibodies, select progeny antibodies with higher affinity than the parent to replace the parent, and update the antibody population. Step 6: Iteration judgment: Determine whether the current iteration count has reached the maximum iteration count. If it has, proceed to step 8; otherwise, proceed to step 7. Step 7: Antibody similarity calculation and antibody library update. Calculate the average similarity of the antibody population and compare it with a preset threshold. If the average similarity is less than the threshold, update the antibody memory library and return to step 4. If it is greater than or equal to the threshold, proceed to step 8. Step 8: Optimal solution screening. Sort the antibody groups in descending order of affinity, select the feature attribute set corresponding to the antibody with the highest affinity as the optimal feature subset, and output it through the result output module.

Images