Experimental data analysis method, device and electronic equipment based on OCR

Abstract

The present invention provides an experimental data analysis method, device and electronic device based on OCR, which obtains experimental data by performing OCR recognition on a target experimental report, and obtains a complete experimental detection data set corresponding to the target device; performs data pruning operation on the complete experimental detection data set through multiple data pruning strategies to obtain multiple partial experimental detection data sets; for each partial experimental detection data set, selects a target inference neural network in a neural network queue to process the partial experimental detection data set to obtain the inference analysis result corresponding to the partial experimental detection data set; the neural network queue includes multiple inference neural networks, and the multiple inference neural networks are obtained by debugging multiple debugging learning sample libraries under multiple debugging learning sample libraries; performs a fusion operation on the inference analysis results corresponding to the multiple partial experimental detection data sets to obtain the target inference analysis result. The present invention can increase the accuracy of the inference analysis result.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an experimental data analysis method, device and electronic equipment based on OCR. Background Art

[0002] In today's era of rapid technological development, performance testing and data analysis of various equipment play a vital role in many fields such as industrial production and scientific research experiments. Traditional experimental data analysis methods mainly rely on manual processing of data in experimental reports. This method is not only inefficient and prone to human errors, but also difficult to quickly and accurately mine valuable information when faced with a large amount of complex experimental data. With the emergence of optical character recognition (OCR) technology, although it can realize the conversion of text information in experimental reports to electronic text, there are still many deficiencies in the subsequent processing and analysis of experimental data. Most of the existing data analysis methods are directly based on all experimental data for processing, lacking effective data processing, resulting in the role of some high-value experimental detection data in the analysis link being reduced, and it is difficult to fully mine the characteristic information of each experimental detection data, which in turn affects the reliability and accuracy of the final analysis results. Summary of the invention

[0003] In view of this, an embodiment of the present invention provides an experimental data analysis method, device and electronic device based on OCR. The technical solution of the present invention is implemented as follows: On the one hand, the present invention provides an experimental data analysis method based on OCR, the method comprising: performing OCR recognition on a target experimental report to obtain experimental data, wherein the target experimental report is an experimental data report obtained by testing a target device; based on the experimental data, obtaining a complete experimental detection data set corresponding to the target device; the complete experimental detection data set includes multiple experimental detection data of the target device; through a variety of data pruning strategies, performing data pruning operations on the complete experimental detection data set to obtain multiple partial experimental detection data sets; different data pruning strategies are used to prune different experimental detection data in the complete experimental detection data set; the multiple data pruning strategies are used to indicate the pruning of different single experimental detection data in the complete experimental detection data set, or to indicate pruning. In addition to the experimental detection data corresponding to different data types in the complete experimental detection data set; for each of the partial experimental detection data sets, a target inference neural network for processing the partial experimental detection data set is selected from the neural network queue, and the partial experimental detection data set is processed based on the target inference neural network to obtain the inference analysis result corresponding to the partial experimental detection data set; the neural network queue includes multiple inference neural networks, and the multiple inference neural networks are obtained by debugging multiple debugging learning sample libraries under multiple debugging learning sample libraries, and the debugging learning samples in different debugging learning sample libraries under the same debugging learning sample library include different experimental detection training data; the inference analysis results corresponding to the multiple partial experimental detection data sets are fused to obtain the target inference analysis result corresponding to the target device.

[0004] On the other hand, the present invention provides an experimental data analysis device based on OCR, and the device includes: an OCR recognition module, which is used to perform OCR recognition on a target experimental report to obtain experimental data, wherein the target experimental report is an experimental data report obtained by testing a target device; a data acquisition module, which is used to obtain a complete experimental detection data set corresponding to the target device based on the experimental data; the complete experimental detection data set includes multiple experimental detection data of the target device; a data pruning module, which is used to perform data pruning operations on the complete experimental detection data set through multiple data pruning strategies to obtain multiple partial experimental detection data sets; different data pruning strategies are used to prune different experimental detection data in the complete experimental detection data set; the multiple data pruning strategies are used to indicate the pruning of different single experimental detection data in the complete experimental detection data set, or to indicate the pruning of The present invention relates to experimental detection data corresponding to different data types in the complete experimental detection data set; an inference analysis module, which is used to select a target inference neural network for processing the partial experimental detection data set from the neural network queue for each of the partial experimental detection data sets, and to process the partial experimental detection data set based on the target inference neural network to obtain the inference analysis result corresponding to the partial experimental detection data set; the neural network queue includes multiple inference neural networks, and the multiple inference neural networks are obtained by debugging multiple debugging learning sample libraries under multiple debugging learning sample libraries, respectively, and the debugging learning samples in different debugging learning sample libraries under the same debugging learning sample library include different experimental detection training data; a result fusion module, which is used to perform a fusion operation on the inference analysis results corresponding to the multiple partial experimental detection data sets to obtain the target inference analysis result corresponding to the target device.

[0005] On the other hand, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above method when executing the program.

[0006] The present invention adopts OCR technology to obtain experimental data from an experimental report, and obtains a complete experimental detection data set corresponding to a target device from the experimental data, wherein the complete experimental detection data set includes multiple experimental detection data of the target device; then, a data pruning operation is performed on the complete experimental detection data set through a variety of data pruning strategies to obtain multiple partial experimental detection data sets, and different data pruning strategies are used to prune different experimental detection data in the complete experimental detection data set; for each partial experimental detection data set, a target inference neural network for processing the partial experimental detection data set is selected from a neural network queue, and the partial experimental detection data set is processed based on the target inference neural network to obtain an inference analysis result corresponding to the partial experimental detection data set; and a fusion operation is performed on the inference analysis results respectively corresponding to the multiple partial experimental detection data sets to obtain a target inference analysis result corresponding to the target device. The present invention performs data pruning operations on a complete experimental detection data set through a variety of data pruning strategies, and will obtain multiple partial experimental detection data sets including different experimental detection data; then, for each partial experimental detection data set, the corresponding reasoning analysis result is determined based on its corresponding target reasoning neural network. Compared with determining the reasoning analysis result based on all the experimental detection data corresponding to the target device through only one neural network, for multiple partial experimental detection data sets including different experimental detection data, the corresponding reasoning analysis results are determined respectively, so that the feature information of each experimental detection data corresponding to the target device can be fully mined and analyzed, avoiding reducing the role of high-value experimental detection data in the reasoning link; then, the reasoning analysis results corresponding to the multiple partial experimental detection data sets are fused to obtain the target reasoning analysis result. On the premise that the reasoning analysis results corresponding to the multiple partial experimental detection data sets are accurate, the reliability and accuracy of the determined target reasoning analysis result are improved.

[0007] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The accompanying drawings herein are incorporated into and constitute a part of the specification. These drawings illustrate embodiments consistent with the present invention and, together with the specification, are used to explain the technical solutions of the present invention.

[0009] Figure 1 A schematic diagram of an implementation flow of an experimental data analysis method based on OCR provided in an embodiment of the present invention;

[0010] Figure 2 A schematic diagram of the composition structure of an experimental data analysis device based on OCR provided in an embodiment of the present invention;

[0011] Figure 3A schematic diagram of a hardware entity of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The embodiment of the present invention provides an experimental data analysis method based on OCR, which can be executed by a processor of an electronic device, wherein the electronic device can refer to a device with data processing capability, such as a server, a laptop, a tablet computer, and a desktop computer.

[0013] Figure 1 A schematic diagram of an implementation flow of an experimental data analysis method based on OCR provided in an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes: Step S100: performing OCR recognition on a target experiment report to obtain experiment data, where the target experiment report is an experiment data report obtained by testing a target device.

[0014] In step S100, the target device may be any device that needs to be tested for performance, function, etc., such as a smart phone in electronic equipment, a machine tool in industrial equipment, etc. The target experiment report is a document that records the detailed data information obtained by testing these target devices, which may be a paper document or an electronic document.

[0015] The purpose of OCR recognition by electronic devices is to convert the text information in the target lab report into editable electronic text data for subsequent data processing and analysis. OCR (Optical Character Recognition) technology is a technology that converts text in an image into text that can be recognized by a computer by optical means. When executing step S100, the electronic device first needs to capture an image of the target lab report. If the target lab report is a paper document, the electronic device can convert it into an image file through a scanner; if it is an electronic document, the electronic device can directly obtain its data in image format.

[0016] For example, for a target experiment report of a smartphone, the report may contain various performance test data of the phone, such as the processor performance test score, battery life, camera shooting parameters, etc. The electronic device scans this paper report into an image file in JPEG or PNG format through a scanner. Next, the electronic device performs preprocessing operations on the collected image. The purpose of preprocessing is to improve the quality of the image so that subsequent character recognition can be more accurate. Preprocessing operations usually include steps such as image grayscale, noise reduction, binarization, and tilt correction. Image grayscale is the process of converting a color image into a grayscale image, which can reduce the amount of image data while retaining the main information of the image. Commonly used grayscale methods include the average method and the weighted average method. Taking the average method as an example, for each pixel in the image, its grayscale value is equal to the average value of the pixel values ​​of the red, green, and blue channels of the pixel point. The formula is: Gray=(R+G+B) / 3, where Gray represents the grayscale value, and R, G, and B represent the pixel values ​​of the red, green, and blue channels respectively. Noise reduction is the process of removing noise interference in an image. Possible noises include salt and pepper noise, Gaussian noise, etc. For salt and pepper noise, median filtering can be used to remove it. The principle of median filtering is to sort the pixel values ​​in the neighborhood of each pixel in the image, and then take the middle value as the new value of the pixel. Tilt correction is to rotate the image so that its text is in a horizontal state for subsequent character recognition. Electronic devices can detect the tilt angle of the text in the image and then rotate the image accordingly.

[0017] After preprocessing, the electronic device can perform character recognition. Character recognition is the core step of OCR technology, and its purpose is to convert characters in an image into text that can be recognized by a computer. The feasible character recognition methods include template matching-based methods, feature extraction-based methods, and deep learning-based methods.

[0018] The method based on template matching is to compare the character to be recognized with the pre-stored character template to find the best matching template, thereby determining the category of the character. The advantage of this method is that it is simple to implement, but the disadvantage is that it has poor adaptability to the deformation and rotation of the character.

[0019] The feature extraction method is to extract the features of characters, such as strokes, contours, etc., and then classify and recognize them based on these features. The advantage of this method is that it has a certain adaptability to the deformation and rotation of characters, but the disadvantage is that the feature extraction process is relatively complicated.

[0020] The deep learning-based method uses deep learning models such as convolutional neural networks (CNN) to recognize characters. The advantages of this method are high recognition accuracy and good adaptability to character deformation and rotation. The disadvantage is that it requires a large amount of training data and computing resources.

[0021] For example, the electronic device uses a deep learning-based OCR engine to perform character recognition on a preprocessed smartphone target lab report image. The engine has been trained on a large amount of text image data and can accurately recognize various characters in the image. After the recognition is completed, the electronic device organizes and analyzes the obtained text data and extracts the lab data.

[0022] When collating and analyzing text data, the electronic device can use regular expressions and other techniques to extract the required experimental data according to the format and content of the target experimental report. For example, for a smartphone target experimental report, the electronic device can use regular expressions to match data such as processor running points and battery life.

[0023] Finally, the electronic device stores the extracted experimental data for use in subsequent steps. These experimental data will serve as the basis for further analysis of the performance and status of the target device.

[0024] In practical applications, the accuracy of OCR recognition by electronic devices may be affected by many factors, such as the quality of the target experimental report, the clarity of image acquisition, the font and size of characters, etc. In order to improve the accuracy of recognition, electronic devices can take the following measures: first, select high-quality image acquisition equipment to ensure that the acquired images are clear and complete; second, fully pre-process the images to remove noise and interference and improve the quality of the images; third, select appropriate OCR engines and recognition methods, and adjust and optimize them according to actual conditions; fourth, manually verify and correct the recognition results to ensure the accuracy of the experimental data.

[0025] In addition, electronic devices can also optimize and improve the OCR recognition process, such as using multi-threading technology to improve recognition efficiency, using adaptive pre-processing methods to process according to the characteristics of the image, etc. At the same time, electronic devices can also combine other technologies, such as natural language processing technology, to further analyze and understand the recognized text data and extract more valuable information.

[0026] Step S200: acquiring a complete experimental detection data set corresponding to the target device based on the experimental data; the complete experimental detection data set includes a plurality of experimental detection data of the target device.

[0027] When the electronic device obtains a complete experimental test data set, it first cleans the experimental data obtained through OCR recognition. The purpose of data cleaning is to remove noise, erroneous data and duplicate data in the experimental data to improve the quality of the data. For example, in a target experimental report on automobile performance testing, the experimental data obtained by OCR recognition may contain some erroneous values ​​due to scanning or recognition errors, or there may be duplicate records. Electronic devices can use statistical analysis methods, such as calculating the mean and standard deviation of the data, to determine whether the data is abnormal. For data points that deviate too much from the mean, they can be regarded as abnormal data and deleted or corrected. At the same time, electronic devices can identify and remove duplicate data by comparing the unique identifier or features of the data.

[0028] Next, the electronic device integrates the cleaned experimental data. Data integration is to merge and associate experimental data from different data sources to form a more complete data set. In actual situations, the experimental test data of the target device may come from multiple different test links or sensors, and these data may be stored in different files or databases. The electronic device can associate and merge these data based on information such as the timestamp of the data and the device identification. For example, for the experimental test data of a car, there may be engine performance test data, chassis test data, etc. stored in different files. The electronic device can integrate these data into one data set based on information such as the test time and the frame number of the car.

[0029] In the process of data integration, electronic devices may need to convert and unify data formats. Data formats from different data sources may differ, such as date format, numerical precision, etc. Electronic devices unify these data formats for subsequent processing and analysis. For example, date data in different formats are converted to a unified date format, and numerical data with different precisions are rounded or truncated.

[0030] In addition to cleaning and integrating the experimental data obtained by OCR recognition, the electronic device may also need to obtain supplementary data from other data sources to complete the complete experimental test data set. These supplementary data can come from historical experimental records, equipment design parameters, industry standards, etc. For example, for the experimental test of a car, the electronic device can obtain the design parameters of the car from the car manufacturer's database, such as the maximum power and torque of the engine, and add these data to the complete experimental test data set.

[0031] When acquiring supplementary data, the electronic device can communicate with other data sources through a network interface. For data stored in a database, the electronic device can use a database query language, such as SQL, to acquire the required data. For data stored in a file, the electronic device can use a file reading and parsing method to extract and process the data.

[0032] After acquiring and integrating all the data, the electronic device verifies and confirms the complete experimental test data set. The purpose of verification is to ensure the accuracy and completeness of the data and to confirm whether the data meets the expected requirements. The electronic device can use data verification rules and business logic for verification. For example, for the experimental test data of a car, the electronic device can verify whether the engine speed is within a reasonable range and whether the various performance indicators meet the design standards of the car.

[0033] In order to improve the efficiency and accuracy of obtaining complete experimental detection data sets, electronic devices can use some advanced technical means. For example, data mining technology can be used to discover potential patterns and associations in the data in order to better integrate and supplement the data. At the same time, electronic devices can establish a data quality monitoring mechanism to monitor the quality of data in real time and promptly discover and handle problems in the data.

[0034] Electronic devices can also use machine learning algorithms to preprocess and extract features from complete experimental detection data sets. For example, the principal component analysis (PCA) algorithm is used to reduce the dimension of the data, reduce the dimension of the data, and improve the efficiency of data processing. The principle of principal component analysis is to convert the original data into a set of principal components that are linearly independent of each dimension through linear transformation, where the first principal component has the largest variance, and the variances of subsequent principal components decrease in sequence. By selecting principal components with larger variances, the dimension of the data can be reduced while retaining the main information of the data.

[0035] In practical applications, the process of electronic devices obtaining a complete experimental test data set may be affected by many factors. For example, the reliability of the data source, the stability of data transmission, data security, etc. Electronic devices take corresponding measures to deal with these problems, such as evaluating and screening the data source, and using encryption technology to ensure the security of data transmission.

[0036] Step S300: Perform data pruning operations on the complete experimental test data set through multiple data pruning strategies to obtain multiple partial experimental test data sets; different data pruning strategies are used to prune different experimental test data in the complete experimental test data set. Multiple data pruning strategies are used to indicate the pruning of different single experimental test data in the complete experimental test data set, or to indicate the pruning of experimental test data corresponding to different data types in the complete experimental test data set.

[0037] The complete experimental test data set is obtained in step S200, which contains comprehensive and detailed experimental test information of the target device, while the partial experimental test data set is a subset obtained from the complete experimental test data set after pruning, and each subset contains a different combination of experimental test data. Before performing the data pruning operation, the electronic device first determines a variety of data pruning strategies. These data pruning strategies can be designed according to actual needs and data characteristics. The feasible data pruning strategies include indicating the pruning of different single experimental test data in the complete experimental test data set, or indicating the pruning of experimental test data corresponding to different data types in the complete experimental test data set. For example, for a complete experimental test data set for a CNC machine tool, it may include a variety of experimental test data such as machining accuracy, spindle speed, and feed speed. One data pruning strategy may be to prune the machining accuracy data at a specific time point, and another data pruning strategy may be to prune all feed speed data.

[0038] After determining the data pruning strategy, the electronic device can start to perform data pruning operations on the complete experimental detection data set. For the data pruning strategy that indicates the pruning of different individual experimental detection data, the electronic device can locate and delete specific experimental detection data according to the identifier or index of the data. For example, in a complete experimental detection data set containing data collected by multiple sensors, the data of each sensor has a unique identifier. The electronic device can select the individual experimental detection data to be pruned based on these identifiers.

[0039] For a data pruning strategy that indicates pruning experimental test data corresponding to different data types, the electronic device can filter and delete the data according to the data type label. For example, in a complete experimental test data set containing multiple data types such as temperature, pressure, and humidity, the electronic device can prune all temperature data according to the data type label.

[0040] When performing data pruning operations, the electronic device can process the complete experimental detection data set in a loop traversal manner. For each data pruning strategy, the electronic device sequentially checks each experimental detection data in the complete experimental detection data set to determine whether the data meets the pruning conditions, and if so, deletes it from the data set.

[0041] In practical applications, electronic devices can use rule engines to implement data pruning operations. A rule engine is a rule-based reasoning system that can filter and process data according to preset rules. Electronic devices can convert data pruning strategies into rules that can be recognized by the rule engine, and then input the complete experimental detection data set into the rule engine for processing.

[0042] In order to ensure the accuracy and reliability of the data pruning operation, the electronic device can back up the complete experimental test data set before performing the pruning operation. At the same time, the electronic device can verify the pruned part of the experimental test data set to check whether it meets the expected pruning effect. For example, the electronic device can count the number and type of experimental test data in the data set before and after pruning to compare whether it meets the requirements of the data pruning strategy.

[0043] When performing data pruning operations, electronic devices also need to consider the relevance and integrity of the data. Some experimental test data may have certain correlations, and pruning one of the data may affect the validity of other data. For example, in a complete experimental test data set containing voltage and current data, there is an Ohm's law relationship between voltage and current. If the voltage data is pruned, the meaning of the current data may be affected. Therefore, when performing data pruning operations, electronic devices need to analyze the relevance of the data to avoid destroying the integrity of the data due to pruning operations.

[0044] In addition, electronic devices can also design data pruning strategies based on the importance and value of data. For some data that have little impact on subsequent analysis and reasoning, they can be pruned first to reduce data redundancy and processing complexity. For example, in a complete experimental test data set containing a large amount of historical experimental test data, some early data may not be of much value to current analysis and reasoning. Electronic devices can design a data pruning strategy to prune these early data.

[0045] After completing the data pruning operation, the electronic device will obtain multiple partial experimental detection data sets. These partial experimental detection data sets will be used in the subsequent reasoning analysis steps. Each partial experimental detection data set contains a different combination of experimental detection data. By analyzing these different combinations of data, the characteristic information of the target device can be more comprehensively mined.

[0046] In practical applications, electronic devices can flexibly adjust data pruning strategies according to different application scenarios and requirements. For example, when resources are limited, a more aggressive data pruning strategy can be designed to reduce the amount of data processing; when data accuracy is required to be high, a more conservative data pruning strategy can be designed to retain more experimental detection data.

[0047] In the OCR-based experimental data analysis method, a variety of data pruning strategies play an important role, which are mainly used to indicate the pruning of different individual experimental test data in the complete experimental test data set, or to indicate the pruning of experimental test data corresponding to different data types in the complete experimental test data set. The complete experimental test data set is a collection of multiple experimental test data obtained by testing the target device, and the purpose of the data pruning strategy is to make the subsequent analysis more focused and efficient by removing some data.

[0048] For data pruning strategies that indicate the pruning of different individual experimental test data, the electronic device can select individual data to be pruned based on specific conditions or rules. For example, in the experimental test of a CNC machine tool, the complete experimental test data set contains multiple experimental test data such as the machining accuracy, spindle speed, and feed speed of the machine tool at different time points. If the machining accuracy data at a certain point in time is obviously abnormal, it may be caused by measurement errors or momentary equipment failures. The electronic device can judge the data as an outlier according to a pre-set threshold range, and then use the corresponding data pruning strategy to prune the machining accuracy data at this specific time point from the complete experimental test data set. The electronic device can use statistical analysis methods to calculate the mean and standard deviation of the machining accuracy data, and treat the data that deviates from the mean by more than a certain multiple of the standard deviation as outliers for pruning. Suppose the machining accuracy data is , mean , standard deviation , if a satisfy (k is a preset threshold), then Cut off.

[0049] For another example, in an experimental test of an intelligent sensor network, the complete experimental test data set includes the temperature, humidity, light intensity and other data collected by each sensor node. If a sensor node fails, causing the data collected by it to be unreliable, the electronic device can remove all the individual experimental test data collected by the node from the complete experimental test data set based on the identification information of the sensor node.

[0050] As for the data pruning strategy that indicates the pruning of experimental test data corresponding to different data types, the electronic device screens and deletes according to the type label of the data. For example, in the experimental test of an electric car, the complete experimental test data set contains multiple data types such as the vehicle's battery power, driving speed, motor temperature, and in-car air quality. If the electronic device is mainly concerned with the vehicle's power performance in this analysis, a data pruning strategy can be used to prune the data type of in-car air quality that has little correlation with power performance from the complete experimental test data set. The electronic device can delete all data of the type "in-car air quality" by checking the type label of the data.

[0051] Step S400: For each partial experimental detection data set, a target inference neural network for processing the partial experimental detection data set is selected from the neural network queue, and the partial experimental detection data set is processed based on the target inference neural network to obtain the inference analysis result corresponding to the partial experimental detection data set; the neural network queue includes multiple inference neural networks, and the multiple inference neural networks are obtained by debugging multiple debugging learning sample libraries under multiple debugging learning sample libraries, and the debugging learning samples in different debugging learning sample libraries under the same debugging learning sample library include different experimental detection training data.

[0052] The partial experimental detection data set is obtained by performing data pruning operation on the complete experimental detection data set in step S300, and each partial experimental detection data set contains different experimental detection data combinations. The neural network queue is composed of multiple inference neural networks, which are debugged by multiple debugging learning sample libraries under multiple debugging learning sample libraries, and the debugging learning samples in different debugging learning sample libraries under the same debugging learning sample library contain different experimental detection training data.

[0053] When selecting a target inference neural network, electronic devices need to consider the characteristics of some experimental detection data sets and the scope of application of the inference neural network. Different inference neural networks are debugged and trained for different data features and tasks. Therefore, for different experimental detection data sets, it is necessary to select a suitable inference neural network for processing to ensure accurate inference analysis results.

[0054] For example, for a partial experimental detection data set about power equipment, which contains data such as voltage, current, power, etc. of the equipment, the electronic device selects a target inference neural network from the neural network queue that can effectively process these electrical data and perform fault diagnosis or performance evaluation.

[0055] In order to select the target inference neural network, the electronic device can adopt a variety of methods. One method is to set the applicable data range and conditions for each inference neural network in advance according to the data type and characteristics of some experimental detection data sets. When selecting, the electronic device will check whether the data characteristics of some experimental detection data sets meet the applicable conditions of a certain inference neural network, and if so, it will be selected as the target inference neural network.

[0056] Another method is to select the target inference neural network by similarity calculation. The electronic device can calculate the similarity between part of the experimental detection data set and the debugging learning examples used by each inference neural network during the debugging learning process. Similarity can be calculated in many ways, such as using Euclidean distance, cosine similarity, etc.

[0057] After determining the target inference neural network, the electronic device inputs part of the experimental detection data set into the target inference neural network for processing. The inference neural network is an inference system based on a neural network model, which outputs the corresponding inference analysis results by extracting features and recognizing patterns on the input data.

[0058] The processing of inference neural network usually includes input layer, hidden layer and output layer. The input layer receives part of the experimental test data set as input, the hidden layer performs nonlinear transformation and feature extraction on the input data, and the output layer outputs the inference analysis results based on the features extracted by the hidden layer. Different types of inference neural networks may have different structures and processing methods, such as feedforward neural networks, recurrent neural networks, etc.

[0059] Taking a feedforward neural network as an example, the neurons in the input layer take each data point of a part of the experimental test data set as input value, then perform linear combination through weights and biases, and then perform nonlinear transformation through activation function, and pass the result to the neurons in the hidden layer. The neurons in the hidden layer repeat the same process and finally pass the result to the output layer, which outputs the reasoning analysis result according to the activation function.

[0060] When processing part of the experimental detection data set, the electronic device needs to ensure that the parameters of the inference neural network are properly debugged. The debugging process of the inference neural network is completed in steps S10-S40, and the inference neural network is trained and optimized by using the full set of debugging learning samples, the debugging learning sample library and the debugging learning sample library, so that the inference neural network can better adapt to different data features and tasks.

[0061] In order to improve the accuracy and reliability of the inference analysis results, the electronic device can post-process the output of the inference neural network. Post-processing methods include data smoothing, confidence assessment, etc. Data smoothing can reduce the noise and fluctuations in the inference analysis results through methods such as moving average and exponential smoothing. Confidence assessment is to evaluate the credibility of the inference analysis results. The electronic device can calculate the confidence of the inference analysis results based on the output of the inference neural network and the feedback of the debugging learning examples.

[0062] In practical applications, electronic devices may encounter situations where multiple target inference neural networks are applicable to a certain part of the experimental detection data set. In this case, the electronic device can use a method of fusing the output results of multiple inference neural networks to obtain more accurate inference analysis results. For example, the electronic device can perform a weighted average of the output results of multiple inference neural networks, and the weight can be determined based on the performance of each inference neural network during the debugging and learning process.

[0063] In addition, the electronic device can also adjust the selection and processing of the target inference neural network in real time according to the dynamic changes of some experimental detection data sets. If the characteristics of some experimental detection data sets change, the electronic device can re-evaluate and select a more appropriate target inference neural network for processing.

[0064] Step S500: performing a fusion operation on the inference analysis results corresponding to the multiple partial experimental detection data sets to obtain a target inference analysis result corresponding to the target device.

[0065] In step S400, the electronic device selects the target inference neural network for processing each partial experimental detection data set, and obtains the inference analysis results corresponding to each of the multiple partial experimental detection data sets. However, these individual inference analysis results may only reflect some characteristics of the target device from different angles. In order to more comprehensively and accurately evaluate the performance and status of the target device, these results need to be integrated.

[0066] Electronic devices can use different fusion methods to achieve this goal. Feasible methods include mean calculation and eccentricity adjustment fusion. Mean calculation is a relatively simple and direct fusion method. The electronic device will perform mean calculation on the reasoning and analysis results corresponding to multiple partial experimental detection data sets to obtain the target reasoning and analysis results. For example, suppose that the electronic device obtains three partial experimental detection data sets for the complete experimental detection data set of the target device through different data pruning strategies. After processing in step S400, the reasoning and analysis results corresponding to these three partial experimental detection data sets are A, B, and C respectively. When the electronic device adopts the fusion method of mean calculation, the target reasoning and analysis result R can be calculated by the formula R=(A+B+C) / 3. The advantage of this method is that the calculation is simple and the information of each reasoning and analysis result can be integrated to a certain extent, but it assumes that the importance of each reasoning and analysis result is the same, and does not take into account the differences in information contained in different partial experimental detection data sets.

[0067] Eccentricity adjustment fusion is a more flexible fusion method. It fuses the inference analysis results corresponding to multiple partial experimental detection data sets according to the eccentricity adjustment parameters (such as weight values) included in the target eccentricity adjustment parameter set. The target eccentricity adjustment parameter set reflects the importance of each inference analysis result in the fusion process. By adjusting these parameters, different weights can be given to different inference analysis results according to actual conditions. For example, for the above three inference analysis results A, B, and C, assuming that the target eccentricity adjustment parameter set is {w A , w B , w C}, where w A 、w B 、w C They are the eccentricity adjustment parameters corresponding to the reasoning analysis results A, B, and C, and they satisfy w A +w B +w C =1. Then, the target reasoning analysis result R obtained by eccentric adjustment fusion can be expressed by the formula R=w A ×A+w B ×B+w C ×C to calculate.

[0068] The process of the electronic device determining the target eccentricity adjustment parameter set is relatively complicated and needs to be implemented through a series of steps. First, the electronic device performs set selection processing and set conversion processing on the a-th eccentricity adjustment parameter set array to obtain the a+1-th eccentricity adjustment parameter set array. In the set selection processing, the electronic device determines the respective indexes of each eccentricity adjustment parameter set in the a-th eccentricity adjustment parameter set array based on the verification data. For example, for the a-th eccentricity adjustment parameter set, its index r eThe formula can be To calculate, where n is the number of eccentric adjustment parameters included in the eccentric adjustment parameter set, that is, the number of inference analysis results corresponding to each part of the experimental detection data set, h j is the jth reasoning analysis result, w ij is the eccentricity adjustment parameter corresponding to the jth inference analysis result in the eccentricity adjustment parameter set, h g To verify the annotated reasoning analysis results in the data, the electronic device selects the eccentricity adjustment parameter set to be converted from the ath eccentricity adjustment parameter set array according to these indicators.

[0069] In the set conversion process, the electronic device performs a conversion operation on the eccentricity adjustment parameters included in the eccentricity adjustment parameter set to be converted, and obtains an eccentricity adjustment parameter set corresponding to the a+1th eccentricity adjustment parameter set array. This process usually involves adjustment and optimization of the eccentricity adjustment parameters to make the final target reasoning analysis result more accurate. This operation will be repeated until a+1=s, where s is the preset maximum number of times. When the preset maximum number of times is reached, the electronic device selects the eccentricity adjustment parameter set with the highest index in the sth eccentricity adjustment parameter set array according to the respective indexes of each eccentricity adjustment parameter set in the sth eccentricity adjustment parameter set array, and uses it as the target eccentricity adjustment parameter set.

[0070] In practical applications, electronic devices need to consider a variety of factors when selecting fusion methods. If the reasoning analysis results corresponding to multiple partial experimental detection data sets are relatively stable and the importance of each result is relatively balanced, then the mean calculation fusion method may be a good choice because it is simple to calculate and efficient. However, if the information contained in different partial experimental detection data sets has a large difference in the degree of influence on the target reasoning analysis results, or some reasoning analysis results are more reliable, then the eccentricity adjustment fusion method can better reflect these differences, thereby obtaining more accurate target reasoning analysis results.

[0071] When performing fusion operations, electronic devices also need to evaluate and verify the fusion results. The target reasoning analysis results obtained by fusion can be compared with the known true results, or the accuracy and reliability of the fusion results can be evaluated through some evaluation indicators, such as mean square error, accuracy, etc. If the evaluation results are not ideal, the electronic device may need to readjust the fusion method or the target eccentricity adjustment parameter set to improve the quality of the fusion results.

[0072] As an implementation mode, in step S400, for each partial experimental detection data set, a target inference neural network for processing the partial experimental detection data set is selected from the neural network queue, and the partial experimental detection data set is processed based on the target inference neural network to obtain the inference analysis result corresponding to the partial experimental detection data set, including:

[0073] Step S410: for each partial experimental detection data set, in each neural network subarray included in the neural network queue, a target inference neural network for processing the partial experimental detection data set is selected; the neural network queue includes a plurality of neural network subarrays, the plurality of neural network subarrays are matched one-to-one with a plurality of debugging learning sample libraries obtained by arbitrary division, and each neural network subarray includes a plurality of inference neural networks, which are debugged according to a plurality of debugging learning sample libraries under the debugging learning sample library corresponding to the neural network subarray;

[0074] Step S420: based on multiple target inference neural networks, partially process the experimental detection data sets respectively to obtain multiple inference analysis results corresponding to the partial experimental detection data sets.

[0075] In step S410, the electronic device selects a target inference neural network for processing each partial experimental detection data set in each neural network subarray of the neural network queue. The key to this process is to match the characteristics of the partial experimental detection data set with the applicable scope of each inference neural network. Different neural network subarrays correspond to different debugging learning sample libraries, and the debugging learning samples in each debugging learning sample library have different characteristics, which makes the inference neural network in each neural network subarray also have different learning capabilities and applicable scenarios.

[0076] For example, for an experimental data analysis scenario about medical devices, assume that there are three neural network sub-arrays in the neural network queue, corresponding to three different debugging learning sample libraries. The debugging learning samples contained in the first debugging learning sample library are mainly about the performance data of medical devices in conventional clinical environments, the samples in the second debugging learning sample library focus on the performance data of medical devices in extreme temperature environments, and the samples in the third debugging learning sample library focus on the performance data of medical devices in high humidity environments. The inference neural network in each neural network sub-array is trained according to the characteristics of its respective debugging learning sample library.

[0077] When an electronic device faces a partial experimental detection data set, it needs to select a suitable target inference neural network in each neural network subarray. In order to achieve this selection process, the electronic device can use a similarity matching method. The electronic device will calculate the similarity between the partial experimental detection data set and the debugging learning examples used by each inference neural network in the debugging learning process. A variety of methods can be used to calculate the similarity, such as Euclidean distance, cosine similarity, etc. The electronic device will select the inference neural network with the highest similarity to the partial experimental detection data set as the target inference neural network.

[0078] The electronic device can also pre-set applicable data ranges and conditions for each inference neural network based on the data types and characteristics of some experimental detection data sets. When selecting a target inference neural network, the electronic device checks whether the data characteristics of some experimental detection data sets meet the applicable conditions of a certain inference neural network, and if so, selects it as the target inference neural network.

[0079] In step S420, the electronic device processes part of the experimental detection data set based on multiple target inference neural networks selected in each neural network subarray, thereby obtaining multiple inference analysis results corresponding to the part of the experimental detection data set. The inference neural network is an inference system based on a neural network model, which outputs corresponding inference analysis results by extracting features and recognizing patterns on input data.

[0080] The processing of inference neural network usually includes input layer, hidden layer and output layer. The input layer receives part of the experimental test data set as input, the hidden layer performs nonlinear transformation and feature extraction on the input data, and the output layer outputs the inference analysis results based on the features extracted by the hidden layer. Different types of inference neural networks may have different structures and processing methods, such as feedforward neural networks, recurrent neural networks, etc.

[0081] Taking a simple feedforward neural network as an example, the neurons in the input layer take each data point of a part of the experimental test data set as input value, then perform linear combination through weights and biases, and then perform nonlinear transformation through activation function, and pass the result to the neurons in the hidden layer. The neurons in the hidden layer repeat the same process and finally pass the result to the output layer, which outputs the reasoning analysis result according to the activation function.

[0082] In the example of medical devices, suppose that for a certain experimental test data set, the electronic device selects a target inference neural network from each of the three neural network subarrays. The first target inference neural network analyzes the partial experimental test data set based on the learning experience in the conventional clinical environment, and may obtain the performance evaluation results of the medical device in routine use; the second target inference neural network analyzes based on the learning samples in the extreme temperature environment, and may give the performance prediction of the medical device under extreme temperature conditions; the third target inference neural network analyzes based on the learning samples in the high humidity environment, and may obtain the failure risk assessment of the medical device in the high humidity environment.

[0083] When electronic devices process some experimental detection data sets, they need to ensure that the parameters of the inference neural network are properly debugged. The debugging process of the inference neural network is completed in the previous steps. The inference neural network is trained and optimized by using the full set of debugging learning samples, the debugging learning sample library, and the debugging learning sample library, so that the inference neural network can better adapt to different data features and tasks.

[0084] In order to improve the accuracy and reliability of the inference analysis results, the electronic device can post-process the output of the inference neural network. Post-processing methods include data smoothing, confidence assessment, etc. Data smoothing can reduce the noise and fluctuations in the inference analysis results through methods such as moving average and exponential smoothing. Confidence assessment is to evaluate the credibility of the inference analysis results. The electronic device can calculate the confidence of the inference analysis results based on the output of the inference neural network and the feedback of the debugging learning examples.

[0085] In practical applications, electronic devices may encounter situations where the output results of multiple target inference neural networks are quite different. At this time, the electronic device can further analyze the reasons for these differences, which may be caused by factors such as the complexity of some experimental detection data sets, the limitations of the inference neural network, or the incompleteness of the debugging learning examples. The electronic device can combine these analysis results to make a comprehensive judgment and adjustment on the inference analysis results.

[0086] In addition, the electronic device can also adjust the selection and processing of the target inference neural network in real time according to the dynamic changes of some experimental detection data sets. If the characteristics of some experimental detection data sets change, the electronic device can re-evaluate and select a more appropriate target inference neural network for processing.

[0087] After completing steps S410-S420, the electronic device obtains multiple reasoning analysis results corresponding to each partial experimental detection data set. These results reflect the characteristics and information of the partial experimental detection data set from different angles, providing a rich data basis for subsequent fusion operations. By selecting and processing the target reasoning neural network from multiple neural network subarrays, the electronic device can make full use of the information of different debugging learning sample libraries to improve the analysis ability and accuracy of the partial experimental detection data set. In different application scenarios, the electronic device can flexibly adjust the selection and processing strategy of the target reasoning neural network according to specific needs and data characteristics to achieve more efficient and accurate data analysis and reasoning. At the same time, the electronic device can also continuously optimize the debugging process and post-processing methods of the reasoning neural network to further improve the quality and reliability of the reasoning analysis results.

[0088] As another implementation, in step S400, for each partial experimental detection data set, a target inference neural network for processing the partial experimental detection data set is selected from the neural network queue, including:

[0089] Step S401: determining the target device type corresponding to the target device based on the complete experimental detection data set;

[0090] Step S402: For each partial experimental detection data set, based on the target device type, determine the target neural network subarray corresponding to the target device type in the neural network queue, and select a target inference neural network from the target neural network subarray to process the partial experimental detection data set; the neural network queue includes multiple neural network subarrays, and the multiple neural network subarrays are matched one-to-one with multiple debugging learning sample libraries corresponding to different device types, and each neural network subarray includes multiple inference neural networks, which are debugged according to multiple debugging learning sample libraries under the debugging learning sample library corresponding to the neural network subarray.

[0091] In step S401, the electronic device determines the target device type corresponding to the target device based on the complete experimental detection data set. The complete experimental detection data set is obtained in the previous step and contains comprehensive and detailed experimental detection information of the target device. The target device type is an identifier for classifying the target device. Different types of devices have different characteristics and operating rules.

[0092] In order to determine the target device type, the electronic device can use a variety of technical means. A commonly used method is a rule-based classification method, in which the electronic device judges the complete experimental detection data set according to preset rules. For example, for a complete experimental detection data set of an industrial device, if the data set contains information related to motor speed and torque, and the range and change law of this information conform to the characteristics of the motor, then the electronic device can judge that the target device type is a motor according to the preset rules.

[0093] Electronic devices can also use classification algorithms in machine learning to determine the target device type. Feasible classification algorithms include decision trees, support vector machines, and naive Bayes. Taking the decision tree algorithm as an example, electronic devices first need to use a large number of experimental test data sets of known device types to train the decision tree model. During the training process, the decision tree model will build decision rules based on the features in the data set and the corresponding device type. When determining the target device type, the electronic device inputs the complete experimental test data set into the trained decision tree model, and the model analyzes the data set according to the decision rules and finally outputs the target device type.

[0094] Assume that the electronic device processes a complete experimental test data set about a vehicle, which contains information such as speed, fuel consumption, and power output. If a decision tree algorithm is used, the trained decision tree model may determine the type of target device based on factors such as speed range and fuel consumption characteristics. If the speed is usually in a lower range and the fuel consumption is relatively low, it may be judged as an electric vehicle; if the speed is high and the power output is large, it may be judged as a car.

[0095] In step S402, for each partial experimental detection data set, the electronic device determines a target neural network subarray corresponding to the target device type in the neural network queue based on the target device type determined in step S401, and then selects a target inference neural network from the target neural network subarray for processing the partial experimental detection data set. The neural network queue includes multiple neural network subarrays, which are matched one-to-one with multiple debugging learning sample libraries corresponding to different device types, and each neural network subarray contains multiple inference neural networks, which are debugged based on multiple debugging learning sample libraries under the corresponding debugging learning sample library.

[0096] When determining the target neural network subarray, the electronic device can establish a mapping table between the device type and the neural network subarray. The mapping table records the identifiers of the neural network subarrays corresponding to different device types. The electronic device can determine the corresponding target neural network subarray by querying the mapping table according to the target device type. For example, the mapping table records that the motor type corresponds to the neural network subarray A, and the car type corresponds to the neural network subarray B. When the electronic device determines that the target device type is a motor, the corresponding target neural network subarray A can be found.

[0097] When selecting a target inference neural network in the target neural network subarray, the electronic device may adopt a strategy similar to the similarity matching method mentioned above. The electronic device calculates the similarity between the partial experimental detection data set and the debugging learning samples used by each inference neural network in the target neural network subarray during the debugging learning process. The electronic device selects the inference neural network with the highest similarity to the partial experimental detection data set as the target inference neural network.

[0098] The electronic device can also set priorities for the inference neural network in the target neural network subarray according to the specific characteristics and task requirements of some experimental detection data sets. For example, for some experimental detection data sets that have higher requirements for device stability, the electronic device can give priority to the inference neural network that has a better learning effect on stability-related data during the debugging and learning process as the target inference neural network.

[0099] Assuming that the target device type is a CNC machine tool, there are three inference neural networks in the corresponding target neural network subarray, which are debugged based on different debugging learning sample libraries. The first inference neural network focuses on learning about machining accuracy, the second focuses on learning about tool wear, and the third focuses on learning about spindle speed stability. When the electronic device faces a partial experimental detection data set about the machining accuracy of CNC machine tools, the first inference neural network will be selected as the target inference neural network by calculating similarity or considering priority.

[0100] After selecting the target inference neural network, the electronic device inputs part of the experimental detection data set into the target inference neural network for processing. The inference neural network outputs the corresponding inference analysis results by extracting features and recognizing patterns on the input data. The processing process of the inference neural network usually includes an input layer, a hidden layer, and an output layer. The input layer receives part of the experimental detection data set as input, the hidden layer performs nonlinear transformation and feature extraction on the input data, and the output layer outputs the inference analysis results based on the features extracted by the hidden layer.

[0101] In order to ensure the accuracy and reliability of the inference analysis results, the electronic device can post-process the output of the inference neural network. Post-processing methods include data smoothing, confidence assessment, etc. Data smoothing can reduce the noise and fluctuations in the inference analysis results through methods such as moving average and exponential smoothing. Confidence assessment is to evaluate the credibility of the inference analysis results. The electronic device can calculate the confidence of the inference analysis results based on the output of the inference neural network and the feedback of the debugging learning examples.

[0102] In practical applications, electronic devices may encounter situations where the target device type is difficult to accurately determine or there is no completely matching inference neural network in the target neural network subarray. For situations where the target device type is difficult to determine, the electronic device can use a multi-classifier fusion method to combine the results of multiple classification algorithms to determine the target device type. For situations where there is no completely matching inference neural network, the electronic device can select an inference neural network with relatively high similarity for processing, and can also be corrected in combination with other auxiliary information.

[0103] The electronic device can also adjust the selection of the target inference neural network in real time according to the dynamic changes of some experimental detection data sets. If the characteristics of some experimental detection data sets have changed significantly, the electronic device can re-evaluate the target device type and select a more appropriate target inference neural network for processing again.

[0104] As an implementation mode, step S500 performs a fusion operation on the reasoning analysis results corresponding to the multiple partial experimental detection data sets to obtain the target reasoning analysis result corresponding to the target device, including:

[0105] Step S510: Calculate the mean of the inference analysis results corresponding to the multiple partial experimental detection data sets to obtain the target inference analysis result; or;

[0106] Step S520: According to each eccentricity adjustment parameter included in the target eccentricity adjustment parameter set, the reasoning analysis results corresponding to the multiple partial experimental detection data sets are respectively fused to obtain the target reasoning analysis result.

[0107] In step S510, the electronic device calculates the mean of the reasoning and analysis results corresponding to the multiple partial experimental test data sets, so as to obtain the target reasoning and analysis results. Mean calculation is a basic and intuitive data fusion method. Its core idea is to assume that the reasoning and analysis results corresponding to each partial experimental test data set have the same importance, and to combine these results by taking the average value. For example, suppose that when the electronic device performs experimental data analysis on an industrial robot, it obtains four partial experimental test data sets through different data pruning strategies. After being processed by the inference neural network, the reasoning and analysis results corresponding to these four partial experimental test data sets are R1, R2, R3 and R4 respectively. When the electronic device executes step S510, it will calculate the reasoning and analysis results according to the formula R mean =(R1+R2+R3+R4) / 4 Calculate the target reasoning analysis result R mean .

[0108] In actual operation, the electronic device can first store these reasoning analysis results in an array or list, then traverse this data structure, accumulate the values ​​of each reasoning analysis result, and finally divide it by the number of results to obtain the mean. The advantage of this method is that the calculation is simple and efficient, and no additional parameters or complex algorithms are required. It can work well when the reliability and importance of each reasoning analysis result are relatively balanced, and can eliminate the random errors in individual results to a certain extent, and obtain a relatively stable comprehensive result. However, its disadvantages are also obvious. If different reasoning analysis results have different reliability or importance due to factors such as data pruning strategies and differences in reasoning neural networks, simple mean calculation may cover up these differences, resulting in the final target reasoning analysis result being inaccurate. For example, in the above-mentioned example of industrial robots, if one of the experimental detection data sets contains more accurate data on the robot's key performance indicators, its corresponding reasoning analysis result should have a higher weight, but the mean calculation cannot reflect this difference.

[0109] Step S520 requires the electronic device to perform eccentricity adjustment fusion on the reasoning analysis results corresponding to the multiple partial experimental detection data sets according to the eccentricity adjustment parameters included in the target eccentricity adjustment parameter set, so as to obtain the target reasoning analysis result. The target eccentricity adjustment parameter set reflects the importance of each reasoning analysis result in the fusion process. By adjusting these parameters, different weights can be given to different reasoning analysis results according to actual conditions, so that the fusion result is more in line with actual needs.

[0110] For example, taking the four inference analysis results R1, R2, R3 and R4 of the industrial robot as an example, assume that the target eccentricity adjustment parameter set is {w1, w2, w3, w4}, where w1, w2, w3, w4 are the eccentricity adjustment parameters corresponding to the inference analysis results R1, R2, R3, R4, respectively, and satisfy w1+w2+w3+w4=1. Then, the target inference analysis result R obtained by eccentricity adjustment fusion is weighted The formula R weighted =w1×R1+w2×R2+w3×R3+w4×R4 to calculate.

[0111] It is a complex process for electronic devices to determine the target eccentricity adjustment parameter set, which usually requires multiple iterative optimizations based on verification data. Verification data is pre-prepared data with known annotated reasoning analysis results, which is used to evaluate the effects of different eccentricity adjustment parameter sets. The electronic device first has an initial eccentricity adjustment parameter set array, and then performs set selection processing and set conversion processing on it to gradually obtain a better eccentricity adjustment parameter set.

[0112] In the set selection process, the electronic device determines the respective indicators of each eccentricity adjustment parameter set in the current eccentricity adjustment parameter set array based on the verification data. The calculation formula of the indicator can refer to the above content and will not be repeated here. The electronic device selects the eccentricity adjustment parameter set to be converted from the current eccentricity adjustment parameter set array based on these indicators. The smaller the indicator value, the closer the eccentricity adjustment parameter set is to the labeled result of the verification data when the fusion reasoning analysis result is, and the better it is.

[0113] In the set conversion process, the electronic device performs a conversion operation on the eccentricity adjustment parameters included in the eccentricity adjustment parameter set to be converted, and obtains an eccentricity adjustment parameter set corresponding to the next iteration. This conversion operation can be a fine-tuning of the parameters, such as increasing or decreasing the value of a certain parameter according to a certain step size, while ensuring that the sum of all parameters remains 1. The electronic device will continue to repeat the set selection process and the set conversion process until the preset maximum number of times is reached. When the preset maximum number of times is reached, the electronic device selects the eccentricity adjustment parameter set with the highest index (i.e., the closest to the verification data annotation result) in the array according to the respective indexes of each eccentricity adjustment parameter set in the last eccentricity adjustment parameter set array, and uses it as the target eccentricity adjustment parameter set.

[0114] The advantage of the eccentricity adjustment fusion method is that it can fully consider the importance differences of different reasoning analysis results. By reasonably adjusting the eccentricity adjustment parameters, more reliable and important reasoning analysis results can play a greater role in the fusion process, thereby improving the accuracy of the target reasoning analysis results. However, the computational complexity of this method is relatively high, requiring a large amount of verification data and multiple iterations to determine the optimal set of target eccentricity adjustment parameters. In addition, if the verification data itself is biased or unrepresentative, the obtained set of target eccentricity adjustment parameters may be inaccurate, thereby affecting the quality of the fusion result.

[0115] Electronic devices can evaluate and verify the results during the fusion process, for example, by comparing the fused target reasoning analysis results with known true results, or by using some evaluation indicators, such as mean square error, accuracy, etc., to judge the accuracy and reliability of the fusion results. If the evaluation results are not ideal, the electronic device may need to reselect the fusion method or adjust the target eccentricity adjustment parameter set to improve the quality of the fusion results. At the same time, the electronic device can use machine learning algorithms to optimize the fusion process, and further improve the performance and effect of eccentricity adjustment fusion through continuous learning and adjustment.

[0116] As an implementation mode, the target eccentricity adjustment parameter set is determined by the following steps:

[0117] Step S521: performing set selection processing and set conversion processing on the a-th eccentricity adjustment parameter set array to obtain the a+1-th eccentricity adjustment parameter set array; wherein the set selection processing includes: step S5211: determining the respective indexes of the eccentricity adjustment parameter sets in the a-th eccentricity adjustment parameter set array based on the verification data, and selecting the eccentricity adjustment parameter set to be converted from the a-th eccentricity adjustment parameter set array based on the respective indexes of the eccentricity adjustment parameter sets in the a-th eccentricity adjustment parameter set array; the set conversion processing includes: step S5212: performing conversion operation on the eccentricity adjustment parameters included in the eccentricity adjustment parameter set to be converted to obtain the eccentricity adjustment parameter set corresponding to the a+1-th eccentricity adjustment parameter set array; wherein 1≤a<s, s is the preset maximum number of times;

[0118] Step S522: when a+1=s, according to the respective indices of the eccentricity adjustment parameter sets in the s-th eccentricity adjustment parameter set array, the eccentricity adjustment parameter set with the highest index is selected in the s-th eccentricity adjustment parameter set array as the target eccentricity adjustment parameter set.

[0119] In step S521, the electronic device performs set selection processing and set conversion processing on the a-th eccentricity adjustment parameter set array to obtain the a+1-th eccentricity adjustment parameter set array, where a satisfies 1 ≤ a < s, and s is the preset maximum number of times. The set selection processing (step S5211) and the set conversion processing (step S5212) are steps for iteratively optimizing the eccentricity adjustment parameter set. Through continuous selection and conversion, the electronic device gradually finds a better eccentricity adjustment parameter set.

[0120] In step S5211, the electronic device determines the respective indexes of each eccentricity adjustment parameter set in the a-th eccentricity adjustment parameter set array based on the verification data. The verification data is pre-prepared data with known labeled reasoning analysis results, which is used to evaluate the effects of different eccentricity adjustment parameter sets when integrating the reasoning analysis results. The electronic device measures the pros and cons of each eccentricity adjustment parameter set based on these verification data.

[0121] For each eccentricity adjustment parameter set in the ath eccentricity adjustment parameter set array, the electronic device calculates its index value (refer to the aforementioned calculation method ).

[0122] By comparing the index values, the electronic device can determine which eccentricity adjustment parameter set is closer to the annotation result of the verification data when integrating the reasoning analysis results. Generally speaking, the closer the index value is to 0, the better the fusion effect of the eccentricity adjustment parameter set. In the above example, the index value of the first eccentricity adjustment parameter set is relatively closer to 0, so when selecting the set, the electronic device may give priority to the first set.

[0123] After calculating the index values ​​of each eccentricity adjustment parameter set in the a-th eccentricity adjustment parameter set array, the electronic device selects the eccentricity adjustment parameter set to be converted from the array according to these index values. The selection criteria can be set according to specific needs, for example, one or more eccentricity adjustment parameter sets with the smallest index value (closest to 0) can be selected as the set to be converted.

[0124] In step S5212, the electronic device performs a conversion operation on the eccentricity adjustment parameters included in the eccentricity adjustment parameter set to be converted selected in step S5211, and obtains an eccentricity adjustment parameter set corresponding to the a+1th eccentricity adjustment parameter set array. The purpose of the conversion operation is to fine-tune the eccentricity adjustment parameters in the hope of obtaining a better fusion effect.

[0125] The conversion operation can be performed in many ways, such as changing the value or crossing the values ​​in the set, or increasing or decreasing the eccentricity adjustment parameter according to a certain step size, while ensuring that the sum of all eccentricity adjustment parameters is still 1. For example, for the first eccentricity adjustment parameter set selected in the above example , the electronic device can set a step size of 0.05. The electronic device can try to Increase by 0.05. At the same time, in order to ensure that the sum of the parameters is 1, it is necessary to adjust accordingly. or Assume that the electronic device will Increase to 0.2+0.05=0.25, in order to , electronic devices can Reduced to 0.5-0.05=0.45, Keep it unchanged at 0.3, and get a new set of eccentricity adjustment parameters {0.25, 0.45, 0.3}.

[0126] The electronic device can generate multiple new eccentricity adjustment parameter sets by trying different adjustment methods multiple times, and these new sets constitute the a+1th eccentricity adjustment parameter set array. In actual operation, the electronic device can use random search, grid search and other methods to determine the appropriate adjustment method to increase the probability of finding a better eccentricity adjustment parameter set.

[0127] Step S522 requires the electronic device to select the eccentricity adjustment parameter set with the highest index (i.e., closest to 0) in the array of eccentricity adjustment parameter sets according to the respective indexes of each eccentricity adjustment parameter set in the s-time eccentricity adjustment parameter set array when a+1=s, that is, when the preset maximum number of times is reached, and use it as the target eccentricity adjustment parameter set. After multiple iterations of set selection and conversion processing, the electronic device gradually optimizes the eccentricity adjustment parameter set, so that the target eccentricity adjustment parameter set finally obtained can be as close as possible to the labeled reasoning analysis results of the verification data when the reasoning analysis results corresponding to multiple partial experimental detection data sets are integrated.

[0128] For example, assuming that the preset maximum number of times s=5, after the fifth iteration is completed, the electronic device obtains the fifth eccentricity adjustment parameter set array, which contains multiple eccentricity adjustment parameter sets that have been adjusted multiple times. The electronic device will calculate the index values ​​of these sets again, and then select the set with the index value closest to 0 as the target eccentricity adjustment parameter set. Assuming that there are three sets in the fifth eccentricity adjustment parameter set array, and their index values ​​are -0.005, -0.008, and -0.003, respectively, then the electronic device will select the set with the index value of -0.003 as the target eccentricity adjustment parameter set.

[0129] In the process of determining the entire target eccentricity adjustment parameter set, the electronic device needs to pay attention to some issues. First, the quality and representativeness of the verification data are crucial. If the verification data is biased or does not reflect the actual situation well, the index value calculated based on these verification data may be inaccurate, thus affecting the selection of the target eccentricity adjustment parameter set. Therefore, the electronic device ensures the accuracy and reliability of the verification data, and can improve the quality of the verification data by collecting more data, performing data cleaning and preprocessing, etc.

[0130] Secondly, the selection of the preset maximum number s also needs to be reasonable. If s is set too small, the eccentricity adjustment parameter set may not be fully optimized, resulting in a suboptimal target eccentricity adjustment parameter set; if s is set too large, the amount of calculation and time cost will increase. The electronic device can determine the appropriate maximum number s through experiments and experience.

[0131] In addition, the step size of the conversion operation will also affect the optimization effect. A step size that is too large may cause the optimal solution to be skipped, while a step size that is too small will increase the number of iterations and calculation time. Electronic devices can use an adaptive step size method, using a larger step size for rapid search at the beginning of the iteration, and gradually reducing the step size as the iteration proceeds to more accurately approach the optimal solution.

[0132] In practical applications, electronic devices can use parallel computing technology to improve the efficiency of the target eccentricity adjustment parameter set determination process. For example, the electronic device can simultaneously calculate the index values ​​of multiple eccentricity adjustment parameter sets, or perform multiple conversion operations in parallel, thereby reducing the overall calculation time. The electronic device can also combine the target eccentricity adjustment parameter set determination process with other data analysis and optimization algorithms to further improve the accuracy and reliability of the fusion result.

[0133] In summary, the process of determining the target eccentricity adjustment parameter set gradually optimizes the eccentricity adjustment parameter set through the set selection process and set conversion process of step S521 and the final selection of step S522, so as to obtain a target eccentricity adjustment parameter set that can more accurately fuse the corresponding reasoning analysis results of multiple partial experimental detection data sets. This process requires the electronic device to comprehensively consider factors such as the quality of the verification data, the preset maximum number of times, and the conversion step length, and can use technical means such as parallel computing to improve efficiency, thereby providing reliable parameter support for subsequent eccentricity adjustment fusion operations, and ultimately improving the accuracy and reliability of the target reasoning analysis results.

[0134] As an implementation method, the multiple inference neural networks included in the neural network queue are debugged using the following steps:

[0135] Step S10: Obtain a full set of debugging learning samples; the full set of debugging learning samples includes multiple debugging learning samples, each debugging learning sample includes a complete experimental detection training data set corresponding to the sample device, and a priori analysis results;

[0136] Step S20: planning multiple debugging learning samples in the full set of debugging learning samples into multiple debugging learning sample libraries;

[0137] Step S30: for each debugging learning sample library, based on multiple data pruning strategies, a data pruning operation is performed on the complete experimental detection training data set included in each debugging learning sample in the debugging learning sample library, so as to obtain a debugging learning sample library corresponding to multiple data pruning strategies under the debugging learning sample library;

[0138] Step S40: For the debugging learning sample library corresponding to each data pruning strategy under each debugging learning sample library, based on the debugging learning sample library, debug the inference neural network corresponding to the data pruning strategy; the inference neural network corresponding to the data pruning strategy is used to process the partial experimental detection data set obtained by performing data pruning operations on the complete experimental detection data set based on the data pruning strategy.

[0139] In step S10, the electronic device obtains a full set of debugging learning samples, which contains multiple debugging learning samples, and each debugging learning sample consists of a complete experimental detection training data set corresponding to the sample device and a priori analysis results. The full set of debugging learning samples is the basic data for debugging the inference neural network. It covers a rich variety of experimental detection data and corresponding analysis results, which enables the inference neural network to learn the characteristics and laws of different sample devices. For example, when performing experimental data analysis on smart home appliances, the full set of debugging learning samples may include a complete experimental detection training data set of sample devices such as refrigerators, air conditioners, and washing machines of different brands and models. These data sets contain information such as the operating parameters and performance indicators of the equipment, and the priori analysis results may be a judgment on whether the equipment is operating normally and whether the performance meets the standards. The electronic device can obtain the full set of debugging learning samples in a variety of ways, such as extracting from a historical experimental record database, cooperating with equipment manufacturers to obtain test data, etc.

[0140] Step S20 requires the electronic device to plan multiple debugging learning samples in the full set of debugging learning samples into multiple debugging learning sample libraries. The purpose of this is to classify and manage the debugging learning samples so that the samples in each debugging learning sample library have certain similarities, which is convenient for subsequent more targeted debugging for different sample libraries. For example, electronic devices can be divided according to factors such as the type of sample device and the use environment. For the full set of debugging learning samples of the above-mentioned smart home appliances, the electronic device can plan the debugging learning samples of the refrigerator to one debugging learning sample library, and plan the debugging learning samples of the air conditioner to another debugging learning sample library. The electronic device can use the cluster analysis method to cluster similar samples into the same debugging learning sample library based on the characteristics of the complete experimental detection training data set in the debugging learning samples. It can also be divided according to preset rules, such as according to information such as the brand and model of the device.

[0141] In step S30, for each debugging learning sample library, the electronic device performs data pruning operations on the complete experimental detection training data set included in each debugging learning sample in the debugging learning sample library based on multiple data pruning strategies, and obtains the debugging learning sample library corresponding to the multiple data pruning strategies under the debugging learning sample library. Different data pruning strategies are used to prune different experimental detection data in the complete experimental detection training data set, so that a debugging learning sample library containing different data combinations can be obtained, so that the inference neural network learns the rules under different data features. For example, for a debugging learning sample in the refrigerator debugging learning sample library, its complete experimental detection training data set includes the temperature, humidity, energy consumption and other data of the refrigerator. One data pruning strategy may be to prune the humidity data therein, and another data pruning strategy may be to prune the energy consumption data within a certain time period. When performing a data pruning operation, the electronic device can locate and delete specific experimental detection data according to the identification, index or type label of the data.

[0142] Step S40 requires the electronic device to debug the inference neural network corresponding to the data pruning strategy for each debugging learning sample library, based on the debugging learning sample library. The inference neural network corresponding to each data pruning strategy is used to process the partial experimental detection data set obtained by performing data pruning operations on the complete experimental detection data set based on the data pruning strategy. For example, for the debugging learning sample library corresponding to the data pruning strategy that prunes the humidity data, the electronic device will debug an inference neural network specifically used to process the partial experimental detection data set that does not contain humidity data. When debugging the inference neural network, the electronic device will use the complete experimental detection training data set in the debugging learning sample library as input and the prior analysis result as the expected output, and continuously adjust the parameters of the inference neural network so that the output of the inference neural network is as close as possible to the prior analysis result. Feasible debugging methods include the gradient descent method, the core idea of ​​which is to gradually reduce the value of the loss function by calculating the gradient of the loss function with respect to the neural network parameters, and then updating the parameters in the opposite direction of the gradient. The loss function can adopt the mean square error function, and the formula is , where n is the number of samples, is the result of a priori analysis, is the output of the inference neural network.

[0143] When acquiring the full set of debugging learning samples, the electronic device ensures the quality and integrity of the data. The electronic device can clean and preprocess the acquired data to remove noise data, erroneous data, and duplicate data, and normalize the data so that data with different characteristics have the same scale, which is convenient for learning inference neural networks. For example, for the operating parameter data of smart home appliances, the electronic device can use statistical analysis methods to determine whether the data is abnormal. For data points that deviate too much from the mean, they can be regarded as abnormal data and deleted or corrected. For operating parameters in different ranges, such as temperature, energy consumption, etc., the electronic device can use the minimum-maximum normalization method to map the data to the [0, 1] interval.

[0144] When planning debugging learning samples to the debugging learning sample library, the electronic device considers the similarities and differences between the samples. If the division is too rough, the samples in the same debugging learning sample library may be too different, affecting the learning effect of the inference neural network; if the division is too detailed, it may lead to too many debugging learning sample libraries, increasing the complexity of debugging. The electronic device can optimize the division of debugging learning samples by adjusting the parameters of cluster analysis or the conditions of preset rules. For example, when using cluster analysis, the electronic device can adjust the distance measurement method of clustering and the number of clusters to make the division results more reasonable.

[0145] When performing data pruning operations, electronic devices pay attention to the correlation and integrity of the data. There may be certain correlations between some experimental test data, and pruning one of the data may affect the validity of other data. For example, in the experimental test data of the refrigerator, there may be a correlation between the temperature and the operating status of the compressor. If the temperature data is pruned, it may cause the inference neural network to be unable to accurately learn the law of the compressor operating status. Therefore, when performing data pruning operations, electronic devices need to analyze the correlation of the data to avoid destroying the integrity of the data due to pruning operations. Electronic devices can determine the degree of correlation between data by calculating the correlation coefficient between data, such as the Pearson correlation coefficient. The formula is ,in and are the i-th data points of the two data features, and are the means of the two data features respectively.

[0146] When debugging the inference neural network, the electronic device selects the appropriate debugging method and parameters. Different debugging methods have different advantages and disadvantages. The electronic device selects the appropriate method according to the characteristics of the debugging learning sample library and the structure of the inference neural network. For example, for a small-scale debugging learning sample library, a simple gradient descent method may be able to achieve better results; for a large-scale debugging learning sample library, the stochastic gradient descent method or its variants, such as Adagrad, Adadelta, etc., may be more efficient. The electronic device also needs to adjust the debugging parameters, such as the learning rate, the number of iterations, etc. The learning rate controls the step size of the parameter update. If the learning rate is too large, the parameter update may be too fast and cannot converge to the optimal solution; if the learning rate is too small, the convergence speed will be too slow, increasing the debugging time. The electronic device can determine the appropriate learning rate and number of iterations through experiments and experience.

[0147] During the debugging process of the entire inference neural network, the electronic device also needs to evaluate and verify the debugging results. The electronic device can use the cross-validation method to divide the debugging learning sample library into a training set and a validation set, perform debugging on the training set, and evaluate the performance of the inference neural network on the validation set. Evaluation indicators may include accuracy, recall, mean square error, etc. For example, for the task of judging whether smart home appliances are operating normally, the accuracy can measure the proportion of correct judgments made by the inference neural network, and the recall rate can measure the proportion of normal operating devices correctly identified by the inference neural network. If the evaluation result is not ideal, the electronic device will readjust the debugging method, parameters, or data pruning strategy to improve the performance of the inference neural network.

[0148] In addition, electronic devices can use parallel computing technology to improve the debugging efficiency of inference neural networks. For example, electronic devices can use multiple processors or graphics processing units (GPUs) to parallelly calculate the gradient of the loss function and update parameters, thereby reducing the debugging time. Electronic devices can also perform incremental learning on the debugging learning sample library. When new debugging learning samples are added, electronic devices can fine-tune the original debugged inference neural network without having to redo the entire debugging process, which can improve the efficiency and flexibility of debugging.

[0149] As an implementation mode, step S20, planning multiple debugging learning examples in the full set of debugging learning examples into multiple debugging learning example libraries, includes:

[0150] Step S21: for each debugging learning example in the entire set of debugging learning examples, arbitrarily generate a first comparison value corresponding to the debugging learning example within a first value range; the first value range is determined based on the number of the debugging learning example library;

[0151] Step S22: Determine the sub-numerical range corresponding to the first control numerical value in the first numerical range, and plan the debugging learning sample to the debugging learning sample library corresponding to the sub-numerical range; the first numerical range includes sub-numerical ranges corresponding to multiple debugging learning sample libraries, and the sub-numerical ranges corresponding to the multiple debugging learning sample libraries are equally likely distributed in the first numerical range.

[0152] In step S21, the electronic device randomly generates a first control value corresponding to each debugging learning sample in the full set of debugging learning samples within a first value range, and the first value range is determined based on the number of debugging learning sample libraries. The core of this step is to randomly generate a control value for each debugging learning sample, which serves as the basis for subsequent planning to the debugging learning sample library. The setting of the first value range is related to the number of debugging learning sample libraries, and its purpose is to ensure that each debugging learning sample library has the opportunity to be assigned to the debugging learning sample.

[0153] For example, assuming that the electronic device has 5 debugging learning sample libraries, in order to determine the first numerical range, the electronic device can set the numerical range to [1, 5], and each integer in this range corresponds to a debugging learning sample library. When generating the first comparison numerical value, the electronic device can use a random number generation algorithm, such as the linear congruential method. The formula of the linear congruential method is ,in is the current random number, is the next random number, a is the multiplier, c is the increment, and m is the modulus. By continuously iterating this formula, the electronic device can generate a series of random numbers and then map them into the first numerical range as needed.

[0154] For each debugging learning sample in the full set of debugging learning samples, the electronic device generates a first comparison value for it using this random number generation method. Assume that there is a debugging learning sample about smartphone performance testing in the full set of debugging learning samples, and the electronic device generates a random number using the linear congruential method. After mapping, the first comparison value obtained is 3.

[0155] In step S22, the electronic device determines the sub-value range corresponding to the first comparison value in the first value range, and then plans the debugging learning sample to the debugging learning sample library corresponding to the sub-value range. The first value range includes sub-value ranges corresponding to multiple debugging learning sample libraries, and these sub-value ranges are equally distributed in the first value range. This means that each debugging learning sample has the same probability of being assigned to any debugging learning sample library.

[0156] Continuing with the above example, the first value range is [1, 5], and each integer corresponds to a sub-value range, that is, 1 corresponds to the first debugging learning sample library, 2 corresponds to the second debugging learning sample library, and so on. If the first comparison value generated by the electronic device is 3, then it will plan the debugging learning samples for smartphone performance testing into the third debugging learning sample library.

[0157] This random allocation method has certain advantages. First, it is simple and easy, and does not require complex analysis and calculation of the characteristics of the debugging learning samples. The allocation can be completed only through random number generation and range mapping. Secondly, since the sub-value ranges are distributed with equal probability in the first value range, each debugging learning sample library has the same chance to obtain debugging learning samples, which helps to ensure that the debugging learning samples in each debugging learning sample library have a certain degree of randomness and diversity. In practical applications, if the number of debugging learning samples in the full set of debugging learning samples is large enough, then through this random allocation method, the number of debugging learning samples in each debugging learning sample library will be roughly equal, and can cover various types of debugging learning samples.

[0158] Since it is a random allocation, the debugging learning samples in some debugging learning sample libraries may lack representativeness in certain features. For example, in the example of the smartphone performance test mentioned above, if after random allocation, most of the debugging learning sample libraries in a debugging learning sample library are debugging learning samples about a certain brand of smartphones, but lack samples of other brands, then when debugging the inference neural network based on this debugging learning sample library in the future, it may cause the inference neural network to make a more accurate performance judgment on the performance of the smartphone of this brand, but inaccurate performance judgment on smartphones of other brands. In order to improve the rationality of the allocation, the electronic device can make some adjustments based on the random allocation. For example, the electronic device can first make a preliminary classification of the debugging learning samples in the full set of debugging learning samples, and then randomly allocate them in each classification. Alternatively, the electronic device can make dynamic adjustments during the random allocation process according to the capacity limit of the debugging learning sample library to avoid allocating too many or too few debugging learning samples to a certain debugging learning sample library.

[0159] When the electronic device executes steps S21-S22, it is also necessary to consider the quality of random number generation. If the quality of random number generation is not high, it may cause deviations in the allocation results. The electronic device can adopt a more advanced random number generation algorithm, such as the Mersenne twister algorithm, which has the advantages of long cycle and good randomness. At the same time, the electronic device can perform statistical tests on the generated random numbers to ensure that they meet the requirements of uniform distribution. In addition, when determining the first numerical range and the sub-numerical range, the electronic device needs to make reasonable settings according to the actual situation of the debugging learning sample library. If the number of debugging learning sample libraries is large, the first numerical range can be expanded accordingly; if different weights need to be assigned to different debugging learning sample libraries, the electronic device can adjust the size and distribution of the sub-numerical range so that some debugging learning sample libraries have a higher probability of obtaining debugging learning samples.

[0160] After completing the planning of the debugging learning samples to the debugging learning sample library, the electronic device can also evaluate the allocation results. The electronic device can count the number and feature distribution of the debugging learning samples in each debugging learning sample library to check whether there is an obvious deviation. If it is found that the allocation result is not ideal, the electronic device can reallocate or adopt other allocation methods.

[0161] Steps S21-S22 randomly generate a first control value and plan the debugging learning samples into the debugging learning sample library according to its corresponding sub-value range. This method is simple and easy to implement, and can ensure a certain degree of randomness and diversity, but it also has some limitations. During the implementation process, electronic devices need to consider factors such as the quality of random number generation and the setting of value ranges, and can be adjusted and evaluated according to actual conditions to improve the rationality and effectiveness of the allocation of debugging learning samples and provide a better data basis for the subsequent debugging of the inference neural network.

[0162] As another implementation, step S20, planning multiple debugging learning samples in the full set of debugging learning samples into multiple debugging learning sample libraries, includes:

[0163] Step S201: performing cluster analysis based on the complete experimental detection training data set included in each of the multiple debugging learning samples in the full set of debugging learning samples, and determining the device types corresponding to the sample devices corresponding to the multiple debugging learning samples;

[0164] Step S202: for each debugging learning sample in the full set of debugging learning samples, determine that the device type corresponding to the corresponding sample device is the target device type, and plan the debugging learning sample into a debugging learning sample library corresponding to the target device type.

[0165] In step S201, the electronic device performs cluster analysis based on the complete experimental detection training data set included in each of the multiple debugging learning samples in the full set of debugging learning samples to determine the device types corresponding to the sample devices corresponding to the multiple debugging learning samples. Cluster analysis, also known as cluster analysis, is an analysis process that groups a collection of physical or abstract objects into multiple classes consisting of similar objects. The core idea is to make objects in the same class have a higher similarity, while objects in different classes have a greater difference.

[0166] Electronic devices can use a variety of clustering algorithms for cluster analysis, such as the K-means clustering algorithm. The K-means clustering algorithm is an iterative clustering analysis algorithm, and its steps are as follows: First, the electronic device determines the number of clusters K, that is, the number of device types to be divided. Then, the electronic device randomly selects K data points from the complete experimental detection training data set as the initial cluster centers. Next, for each complete experimental detection training data set, the electronic device calculates its distance from each cluster center, usually using the Euclidean distance. The electronic device assigns the data set to the class where the nearest cluster center is located. Afterwards, the electronic device recalculates the cluster center of each class, that is, the mean of all data sets in the class. Repeat the above steps of assigning and recalculating the cluster centers until the cluster center no longer changes significantly or reaches a preset number of iterations.

[0167] For example, suppose that the full set of debugging learning examples contains debugging learning examples about different types of industrial equipment, such as lathes, milling machines, grinders, etc. The electronic device determines K=3, that is, it considers that there are three main types of equipment. The electronic device randomly selects three complete experimental detection training data sets as the initial clustering centers, and then calculates the Euclidean distance between other data sets and these three centers, and assigns each data set to the class where the nearest center is located. After many iterations, the electronic device obtains three stable classes, corresponding to the three types of equipment: lathes, milling machines, and grinders.

[0168] In addition to the K-means clustering algorithm, electronic devices can also use hierarchical clustering algorithms. Hierarchical clustering algorithms calculate the similarity between data sets and gradually merge or split clusters to form a tree-like clustering structure. It can be divided into agglomerative hierarchical clustering and divisive hierarchical clustering. Agglomerative hierarchical clustering starts with each data set as a separate class, and then continuously merges the classes with the highest similarity until the preset number of clusters is reached or all data sets are merged into one class. Divisive hierarchical clustering, on the contrary, starts with all data sets as one class, and then continuously splits out classes with lower similarity.

[0169] In step S202, for each debugging learning sample in the full set of debugging learning samples, the electronic device determines that the device type corresponding to the corresponding sample device is the target device type, and then plans the debugging learning sample to the debugging learning sample library corresponding to the target device type. For example, in the above-mentioned industrial equipment example, if the sample device corresponding to a debugging learning sample is determined to be a lathe type through cluster analysis, then the electronic device will plan this debugging learning sample to the debugging learning sample library corresponding to the lathe type.

[0170] This equipment type-based planning method has obvious advantages. First, it enables the debugging learning samples in each debugging learning sample library to have similar equipment type characteristics, which helps the inference neural network to better learn the rules and characteristics of specific equipment types in the subsequent debugging process. For example, the samples in the debugging learning sample library of the lathe type are all related to lathes. The inference neural network can focus on learning the performance parameters, failure modes and other characteristics of the lathe, thereby improving the reasoning and analysis capabilities of lathe-related experimental data. Secondly, this method can improve the organization and manageability of debugging learning samples, and facilitate electronic equipment to classify, store and process different types of debugging learning samples.

[0171] However, this approach also has some challenges. The results of cluster analysis may be affected by many factors, such as the selection of clustering algorithms, the setting of clustering parameters, and the quality of data. If the clustering algorithm is not selected properly or the parameters are not set properly, the clustering results may be inaccurate, causing the debugging learning samples to be incorrectly assigned to inappropriate debugging learning sample libraries. For example, in the K-means clustering algorithm, if the K value is not selected properly, some device types may be merged or split, affecting the debugging effect of the subsequent inference neural network.

[0172] In order to improve the accuracy of cluster analysis, electronic devices can take some measures. Electronic devices can preprocess the complete experimental detection training data set before cluster analysis, such as data cleaning and normalization. Data cleaning can remove noise and outliers in the data and improve the quality of the data; normalization can map data of different dimensions to the same scale to avoid excessive influence of data of certain dimensions on the clustering results. Electronic devices can use multiple clustering algorithms for analysis, compare their results, and select the best clustering result. Electronic devices can also use methods such as cross-validation to evaluate the quality of clustering results and optimize clustering effects by adjusting clustering parameters.

[0173] After planning the debugging learning samples to the debugging learning sample library, the electronic device can also further manage and maintain the debugging learning sample library. The electronic device can regularly check the number and distribution of samples in the debugging learning sample library. If the number of samples in a debugging learning sample library is too small or unevenly distributed, the electronic device can consider transferring some samples from other libraries or collecting more related samples to supplement them. The electronic device can update the debugging learning sample library. When new debugging learning samples are added, the electronic device can re-perform cluster analysis and planning to ensure the effectiveness and accuracy of the debugging learning sample library.

[0174] As an implementation mode, step S30, based on multiple data pruning strategies, performs data pruning operations on the complete experimental detection training data set included in each debugging learning sample in the debugging learning sample library, and obtains the debugging learning sample library corresponding to multiple data pruning strategies under the debugging learning sample library, including:

[0175] Step S31: for each debugging learning example in the debugging learning example library, determining a target data pruning strategy corresponding to the debugging learning example from a plurality of data pruning strategies;

[0176] Step S32: Based on the target data pruning strategy, a data pruning operation is performed on the complete experimental detection training data set included in the debugging learning sample to obtain a partial experimental detection training data set corresponding to the debugging learning sample;

[0177] Step S33: Generate some debugging learning samples based on some experimental detection training data sets corresponding to the debugging learning samples and the prior analysis results in the debugging learning samples; integrate some debugging learning samples into the debugging learning sample library corresponding to the target data pruning strategy under the debugging learning sample library.

[0178] In step S31, for each debugging learning example in the debugging learning example library, the electronic device determines a target data pruning strategy corresponding to the debugging learning example from a plurality of data pruning strategies. Different data pruning strategies are used to prune different experimental detection data in the complete experimental detection data set. These strategies may be instructions for pruning different single experimental detection data in the complete experimental detection data set, or instructions for pruning experimental detection data corresponding to different data types in the complete experimental detection data set.

[0179] Electronic devices can use a variety of methods to determine the target data pruning strategy. One method is to select based on the characteristics of the debugging learning sample. For example, for a debugging learning sample about automobile performance testing, if the fluctuation of some sensor data in its complete experimental detection training data set is large and has little impact on the overall analysis, the electronic device can choose to prune the data pruning strategy of these sensor data as the target data pruning strategy. The electronic device can also use a rule engine to determine the target data pruning strategy according to preset rules. For example, if the number of a certain data type in the debugging learning sample exceeds a certain threshold, the electronic device selects a data pruning strategy to prune the data type.

[0180] Assume that there is a debugging learning example about smart home appliances in the debugging learning example library, and its complete experimental detection training data set contains data such as temperature, humidity, power, and running time of home appliances. Based on the characteristics of the debugging learning example, the electronic device believes that humidity data is less important for this analysis, so it selects the data pruning strategy of pruning humidity data as the target data pruning strategy.

[0181] In step S32, the electronic device performs a data pruning operation on the complete experimental detection training data set included in the debugging learning sample based on the target data pruning strategy determined in step S31, and obtains a partial experimental detection training data set corresponding to the debugging learning sample. The data pruning operation is to delete the corresponding experimental detection data from the complete experimental detection training data set according to the instruction of the target data pruning strategy.

[0182] If the target data pruning strategy is to prune a specific data type, the electronic device can traverse the complete experimental detection training data set and delete the data of that type according to the data type label. For example, for the above-mentioned smart home appliance debugging learning example, after determining that the target data pruning strategy is to prune humidity data, the electronic device will traverse the complete experimental detection training data set, find all humidity data and delete them, thereby obtaining a partial experimental detection training data set that does not contain humidity data.

[0183] If the target data pruning strategy is to prune specific single experimental detection data, the electronic device can locate and delete the data according to the data identifier or index. For example, in a debugging learning example about industrial equipment, the target data pruning strategy is to prune the vibration data at a certain moment. The electronic device can find the vibration data at that moment according to the timestamp identifier and delete it.

[0184] In step S33, the electronic device generates a partial debugging learning sample based on the partial experimental detection training data set corresponding to the debugging learning sample and the prior analysis result in the debugging learning sample, and integrates the partial debugging learning sample into the debugging learning sample library corresponding to the target data pruning strategy under the debugging learning sample library. The partial debugging learning sample consists of the partial experimental detection training data set and the prior analysis result. It is a debugging learning sample after data pruning and can reflect the data characteristics under the target data pruning strategy.

[0185] For example, for the debugging learning samples of the above-mentioned smart home appliances, after obtaining a partial experimental detection training data set that does not contain humidity data, the electronic device combines it with the prior analysis results in the original debugging learning sample to generate a partial debugging learning sample. Then, the electronic device puts this partial debugging learning sample into the debugging learning sample library corresponding to the data pruning strategy that prunes humidity data under the debugging learning sample library.

[0186] The electronic device can evaluate and compare some debugging learning samples in different debugging learning sample libraries. By evaluating the performance of the inference neural network under different data pruning strategies, the electronic device can understand which data pruning strategies are more effective, so as to give priority to these strategies in subsequent operations. The electronic device can use the cross-validation method to divide some debugging learning samples into training sets and validation sets, train the inference neural network on the training set, evaluate its performance on the validation set, and determine the optimal data pruning strategy by comparing the performance indicators of the inference neural network corresponding to different debugging learning sample libraries, such as accuracy, recall rate, etc.

[0187] When implementing steps S31-S33, the electronic device may also consider combining other technical means to improve the effect of data pruning. For example, the electronic device may use a machine learning algorithm to predict the impact of different data pruning strategies on the performance of the inference neural network, thereby selecting a more appropriate target data pruning strategy in advance. The electronic device may use a genetic algorithm to optimize the data pruning strategy, and by simulating the process of biological evolution, continuously iterate and optimize the data pruning strategy, so that the performance of the inference neural network is optimized.

[0188] As an implementation mode, step S31, for each debugging learning example in the debugging learning example library, determining a target data pruning strategy corresponding to the debugging learning example from a plurality of data pruning strategies, includes:

[0189] Step S311: for each debugging learning example in the debugging learning example library, arbitrarily generate a second comparison value corresponding to the debugging learning example within a second numerical range; the second numerical range is determined based on the number of data pruning strategies;

[0190] Step S312: Determine the sub-numerical range corresponding to the second control value within the second numerical range, and determine that the target data pruning strategy corresponding to the debugging learning sample is the data pruning strategy corresponding to the sub-numerical range; the second numerical range includes sub-numerical ranges corresponding to multiple data pruning strategies, and the sub-numerical ranges corresponding to the multiple data pruning strategies are equally likely distributed within the second numerical range.

[0191] In step S311, the electronic device generates a second comparison value corresponding to each debugging learning example in the debugging learning example library within a second numerical range, and the second numerical range is determined based on the number of data pruning strategies. The purpose of setting the second numerical range here is to assign a corresponding numerical interval to each data pruning strategy, and to determine which data pruning strategy to select by the randomly generated second comparison value.

[0192] For example, suppose that the electronic device has four different data pruning strategies, namely strategy A, strategy B, strategy C, and strategy D. In order to determine the second numerical range, the electronic device can set the numerical range to [1, 4], and each integer in this range corresponds to a data pruning strategy. When generating the second comparison value, the electronic device can use a random number generation algorithm. The feasible random number generation algorithm is the linear congruential method. For a debugging learning example about power equipment in the debugging learning example library, the electronic device uses the linear congruential method to generate a random number, and the second comparison value obtained after mapping is 3.

[0193] In step S312, the electronic device determines the sub-value range corresponding to the second control value within the second value range, and determines that the target data pruning strategy corresponding to the debugging learning example is the data pruning strategy corresponding to the sub-value range. The second value range includes sub-value ranges corresponding to a plurality of data pruning strategies, and these sub-value ranges are equally distributed within the second value range. This means that each data pruning strategy has the same probability of being selected as the target data pruning strategy.

[0194] Continuing with the above example, the second value range is [1, 4], 1 corresponds to data pruning strategy A, 2 corresponds to data pruning strategy B, 3 corresponds to data pruning strategy C, and 4 corresponds to data pruning strategy D. If the second comparison value generated by the electronic device is 3, it will determine data pruning strategy C as the target data pruning strategy corresponding to the power equipment debugging learning example.

[0195] This method of randomly selecting the target data pruning strategy has certain advantages. It is simple and easy to implement, and does not require complex analysis and judgment of the characteristics of the debugging learning examples. The target data pruning strategy can be determined only through random number generation and range mapping. Since the sub-value ranges are equally distributed in the second value range, each data pruning strategy has the same chance of being selected, which helps to ensure that all data pruning strategies can be applied as a whole, thereby providing diversified training data for subsequent inference neural networks. When the number of debugging learning examples is large enough, through this random selection method, the number of times each data pruning strategy is applied will be roughly equal, allowing the inference neural network to learn the data characteristics and rules under different data pruning strategies.

[0196] After the electronic device has completed the determination of the target data pruning strategy, it can also evaluate the selection results. The electronic device can count the number of times each data pruning strategy is selected and the characteristics of the corresponding debugging learning samples to check whether there is an obvious deviation. If it is found that the selection result is not ideal, the electronic device can reselect or use other selection methods.

[0197] In summary, steps S311-S312 determine the target data pruning strategy corresponding to the debugging learning sample by randomly generating a second control value and determining the target data pruning strategy according to its corresponding sub-value range. This method is simple and can ensure a certain degree of randomness and diversity, but there is also a risk of unreasonable selection. During the implementation process, electronic devices need to consider factors such as the quality of random number generation and the setting of value ranges, and can be adjusted and evaluated according to actual conditions to improve the rationality and effectiveness of the selection of target data pruning strategies, and provide better support for data pruning operations and debugging of inference neural networks.

[0198] Based on the foregoing embodiments, an embodiment of the present invention provides an experimental data analysis device based on OCR. The units included in the device and the modules included in each unit can be implemented by a processor in a computer device; of course, they can also be implemented by a specific logic circuit; in the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP) or a field programmable gate array (FPGA), etc.

[0199] Figure 2 A schematic diagram of the composition structure of an experimental data analysis device based on OCR provided by an embodiment of the present invention, such as Figure 2 As shown, the OCR-based experimental data analysis device 200 includes: an OCR recognition module 210, which is used to perform OCR recognition on a target experimental report to obtain experimental data, wherein the target experimental report is an experimental data report obtained by testing a target device; a data acquisition module 220, which is used to obtain a complete experimental detection data set corresponding to the target device based on the experimental data; the complete experimental detection data set includes multiple experimental detection data of the target device; a data pruning module 230, which is used to perform data pruning operations on the complete experimental detection data set through multiple data pruning strategies to obtain multiple partial experimental detection data sets; different data pruning strategies are used to prune different experimental detection data in the complete experimental detection data set; the multiple data pruning strategies are used to indicate the pruning of different single experimental detection data in the complete experimental detection data set, or to indicate the pruning of the complete experimental detection data set. The experimental detection data set corresponds to experimental detection data of different data types; the reasoning analysis module 240 is used to select a target reasoning neural network for processing the partial experimental detection data set in the neural network queue for each of the partial experimental detection data sets, and process the partial experimental detection data set based on the target reasoning neural network to obtain the reasoning analysis result corresponding to the partial experimental detection data set; the neural network queue includes multiple reasoning neural networks, and the multiple reasoning neural networks are obtained by debugging the multiple debugging learning sample libraries under multiple debugging learning sample libraries, and the debugging learning samples in different debugging learning sample libraries under the same debugging learning sample library include different experimental detection training data; the result fusion module 250 is used to fuse the reasoning analysis results corresponding to the multiple partial experimental detection data sets, and obtain the target reasoning analysis result corresponding to the target device.

[0200] The description of the above device embodiment is similar to the description of the above method embodiment, and has similar beneficial effects as the method embodiment. In some embodiments, the functions or modules included in the device provided in the embodiment of the present invention can be used to execute the method described in the above method embodiment. For technical details not disclosed in the device embodiment of the present invention, please refer to the description of the method embodiment of the present invention for understanding.

[0201] Figure 3 A schematic diagram of a hardware entity of an electronic device provided by an embodiment of the present invention, such as Figure 3 As shown, the hardware entity of the electronic device 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0202] The above description is only an implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An experimental data analysis method based on OCR, characterized in that: The method includes: performing OCR recognition on a target experimental report to obtain experimental data, wherein the target experimental report is an experimental data report obtained by testing a target device; obtaining a complete experimental detection data set corresponding to the target device based on the experimental data; the complete experimental detection data set includes multiple experimental detection data of the target device; performing data pruning operations on the complete experimental detection data set through multiple data pruning strategies to obtain multiple partial experimental detection data sets; different data pruning strategies are used to prune different experimental detection data in the complete experimental detection data set; the multiple data pruning strategies are used to indicate the pruning of different single experimental detection data in the complete experimental detection data set, or to indicate the pruning of experimental detection data corresponding to different data types in the complete experimental detection data set; for each of the partial experimental detection data sets, selecting a target inference neural network for processing the partial experimental detection data set from a neural network queue, and processing the partial experimental detection data set based on the target inference neural network to obtain an inference analysis result corresponding to the partial experimental detection data set; the neural network queue includes multiple inference neural networks, and the multiple inference neural networks are based on multiple debugging learning sample libraries under multiple debugging learning sample libraries. The learning sample library is debugged separately, and the debugging learning samples in different debugging learning sample libraries under the same debugging learning sample library include different experimental detection training data; the reasoning analysis results corresponding to the multiple partial experimental detection data sets are fused to obtain the target reasoning analysis results corresponding to the target device; wherein, for each of the partial experimental detection data sets, a target reasoning neural network for processing the partial experimental detection data set is selected from the neural network queue, including: based on the complete experimental detection data set, a target device type corresponding to the target device is determined; for each of the partial experimental detection data sets, based on the target device type, a target neural network subarray corresponding to the target device type is determined in the neural network queue, and a target reasoning neural network for processing the partial experimental detection data set is selected from the target neural network subarray; the neural network queue includes multiple neural network subarrays, and the multiple neural network subarrays are matched one by one with multiple debugging learning sample libraries corresponding to different device types, and each of the neural network subarrays includes multiple reasoning neural networks, which are debugged separately according to multiple debugging learning sample libraries under the debugging learning sample library corresponding to the neural network subarray.

2. The method according to claim 1, characterized in that For each of the partial experimental detection data sets, a target inference neural network for processing the partial experimental detection data set is selected from the neural network queue, and the partial experimental detection data set is processed based on the target inference neural network to obtain the inference analysis result corresponding to the partial experimental detection data set, including: for each of the partial experimental detection data sets, in each neural network subarray included in the neural network queue, a target inference neural network for processing the partial experimental detection data set is selected; the neural network queue includes multiple neural network subarrays, and the multiple neural network subarrays are matched one-to-one with the debugging learning sample libraries obtained by multiple arbitrary divisions, and each of the neural network subarrays includes multiple inference neural networks, which are debugged according to multiple debugging learning sample libraries under the debugging learning sample library corresponding to the neural network subarray; the partial experimental detection data sets are processed respectively based on the multiple target inference neural networks to obtain multiple inference analysis results corresponding to the partial experimental detection data sets.

3. The method according to claim 1 or 2, characterized in that: The fusing operation is performed on the reasoning and analysis results corresponding to the multiple partial experimental detection data sets to obtain the target reasoning and analysis result corresponding to the target device, including: performing mean calculation on the reasoning and analysis results corresponding to the multiple partial experimental detection data sets to obtain the target reasoning and analysis result; or: performing eccentricity adjustment fusion on the reasoning and analysis results corresponding to the multiple partial experimental detection data sets according to each eccentricity adjustment parameter included in the target eccentricity adjustment parameter set to obtain the target reasoning and analysis result.

4. The method according to claim 3, characterized in that The target eccentricity adjustment parameter set is determined by the following steps: performing set selection processing and set conversion processing on the a-th eccentricity adjustment parameter set array to obtain the a+1-th eccentricity adjustment parameter set array; wherein the set selection processing includes: determining the respective indexes of the eccentricity adjustment parameter sets in the a-th eccentricity adjustment parameter set array based on the verification data, and selecting the eccentricity adjustment parameter set to be converted from the a-th eccentricity adjustment parameter set array based on the respective indexes of the eccentricity adjustment parameter sets in the a-th eccentricity adjustment parameter set array; the set conversion processing includes: performing a conversion operation on the eccentricity adjustment parameters included in the eccentricity adjustment parameter set to be converted to obtain the eccentricity adjustment parameter set corresponding to the a+1-th eccentricity adjustment parameter set array; wherein 1≤a<s, s is the preset maximum number; when a+1=s, according to the respective indexes of the eccentricity adjustment parameter sets in the s-th eccentricity adjustment parameter set array, the eccentricity adjustment parameter set with the highest index in the s-th eccentricity adjustment parameter set array is selected as the target eccentricity adjustment parameter set.

5. The method according to claim 1, characterized in that: The multiple inference neural networks included in the neural network queue are debugged by the following steps: obtaining a full set of debugging learning samples; the full set of debugging learning samples includes multiple debugging learning samples, each of which includes a complete experimental detection training data set corresponding to the sample device, and a priori analysis results; the multiple debugging learning samples in the full set of debugging learning samples are planned to multiple debugging learning sample libraries; for each of the debugging learning sample libraries, based on the multiple data pruning strategies, data pruning operations are performed on the complete experimental detection training data set included in each debugging learning sample in the debugging learning sample library to obtain the debugging learning sample library corresponding to the multiple data pruning strategies under the debugging learning sample library; for the debugging learning sample library corresponding to each data pruning strategy under each of the debugging learning sample libraries, based on the debugging learning sample library, the inference neural network corresponding to the data pruning strategy is debugged; The inference neural network corresponding to the data pruning strategy is used to process the partial experimental detection data set obtained by performing a data pruning operation on the complete experimental detection data set based on the data pruning strategy.

6. The method according to claim 5, characterized in that The planning of the multiple debugging learning samples in the full set of debugging learning samples to multiple debugging learning sample libraries includes: for each debugging learning sample in the full set of debugging learning samples, arbitrarily generating a first comparison value corresponding to the debugging learning sample within a first numerical range; the first numerical range is determined based on the number of the debugging learning sample libraries; determining a sub-numerical range corresponding to the first comparison value in the first numerical range, and planning the debugging learning sample to the debugging learning sample library corresponding to the sub-numerical range; the first numerical range includes the sub-numerical ranges corresponding to the multiple debugging learning sample libraries respectively, and the multiple debugging learning sample libraries are divided into The sub-value ranges corresponding to the respective debugging learning samples are distributed with medium probability in the first value range; the planning of the multiple debugging learning samples in the full set of debugging learning samples into multiple debugging learning sample libraries includes: performing cluster analysis based on the complete experimental detection training data sets respectively included in the multiple debugging learning samples in the full set of debugging learning samples, and determining the device types corresponding to the sample devices corresponding to the multiple debugging learning samples respectively; for each of the debugging learning samples in the full set of debugging learning samples, determining the device type corresponding to the corresponding sample device as the target device type, and planning the debugging learning sample into the debugging learning sample library corresponding to the target device type.

7. The method according to claim 5, characterized in that The method of performing data pruning operations on the complete experimental detection training data sets included in each debugging learning sample in the debugging learning sample library based on the multiple data pruning strategies to obtain the debugging learning sample library corresponding to the multiple data pruning strategies respectively under the debugging learning sample library includes: for each debugging learning sample in the debugging learning sample library, determining a target data pruning strategy corresponding to the debugging learning sample from the multiple data pruning strategies; performing data pruning operations on the complete experimental detection training data sets included in the debugging learning sample based on the target data pruning strategy to obtain a partial experimental detection training data set corresponding to the debugging learning sample; generating partial debugging learning samples based on the partial experimental detection training data set corresponding to the debugging learning sample and the priori analysis results in the debugging learning samples; and integrating the partial debugging learning samples into the debugging learning sample library corresponding to the target data pruning strategy under the debugging learning sample library.

8. The method according to claim 7, characterized in that For each debugging learning sample in the debugging learning sample library, determining a target data pruning strategy corresponding to the debugging learning sample from among the multiple data pruning strategies includes: for each debugging learning sample in the debugging learning sample library, arbitrarily generating a second comparison value corresponding to the debugging learning sample within a second numerical range; the second numerical range is determined based on the number of the data pruning strategies; determining a sub-numerical range corresponding to the second comparison value within the second numerical range, and determining that the target data pruning strategy corresponding to the debugging learning sample is the data pruning strategy corresponding to the sub-numerical range; the second numerical range includes sub-numerical ranges corresponding to the multiple data pruning strategies, respectively, and the sub-numerical ranges corresponding to the multiple data pruning strategies are equally likely distributed within the second numerical range.

9. An experimental data analysis device based on OCR, characterized in that: The device includes: an OCR recognition module, which is used to perform OCR recognition on a target experimental report to obtain experimental data, wherein the target experimental report is an experimental data report obtained by testing a target device; a data acquisition module, which is used to obtain a complete experimental detection data set corresponding to the target device based on the experimental data; the complete experimental detection data set includes multiple experimental detection data of the target device; a data pruning module, which is used to perform data pruning operations on the complete experimental detection data set through multiple data pruning strategies to obtain multiple partial experimental detection data sets; different data pruning strategies are used to prune different experimental detection data in the complete experimental detection data set; the multiple data pruning strategies are used to indicate the pruning of different single experimental detection data in the complete experimental detection data set, or to indicate the pruning of experimental detection data corresponding to different data types in the complete experimental detection data set; an inference analysis module, which is used to select a target inference neural network for processing the partial experimental detection data set in a neural network queue for each of the partial experimental detection data sets, and to process the partial experimental detection data set based on the target inference neural network to obtain an inference analysis result corresponding to the partial experimental detection data set; the neural network queue includes multiple inference neural networks, and the multiple inference neural networks are based on A plurality of debugging learning sample libraries under a plurality of debugging learning sample libraries are debugged respectively, and the debugging learning samples in different debugging learning sample libraries under the same debugging learning sample library include different experimental detection training data; wherein, for each of the partial experimental detection data sets, a target inference neural network for processing the partial experimental detection data set is selected from the neural network queue, including: based on the complete experimental detection data set, a target device type corresponding to the target device is determined; for each of the partial experimental detection data sets, based on the target device type, a target neural network subarray corresponding to the target device type is determined in the neural network queue, and a target inference neural network for processing the partial experimental detection data set is selected from the target neural network subarray; the neural network queue includes a plurality of neural network subarrays, the plurality of neural network subarrays are matched one by one with a plurality of the debugging learning sample libraries corresponding to different device types, and each of the neural network subarrays includes a plurality of inference neural networks, which are debugged respectively according to a plurality of debugging learning sample libraries under the debugging learning sample library corresponding to the neural network subarray; a result fusion module is used to perform a fusion operation on the inference analysis results corresponding to the plurality of partial experimental detection data sets respectively, to obtain the target inference analysis result corresponding to the target device.

Images