Big data preprocessing method and system based on machine learning

Abstract

The present application discloses a machine learning-based big data preprocessing method and system, which includes: obtaining an RPA system's operation log data set, the operation log data set including multiple conversation interaction records; performing data cleaning on the operation log data set to obtain a preprocessed operation log data set; performing feature extraction on the preprocessed operation log data set based on a preset feature extraction algorithm to obtain text semantic features and operation sequence features of each conversation interaction record; calling a machine learning model to jointly model the text semantic features and operation sequence features, generating anomaly detection results for the conversation interaction records, and adjusting the RPA system's task execution strategy based on the anomaly detection results. This can improve the stability and efficiency of the RPA system and ensure the smooth progress of business processes.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of RPA data processing technology, and specifically to a big data preprocessing method and system based on machine learning. Background Art

[0002] In today's era of rapid digital business development, RPA systems have been widely used because they can automate repetitive tasks and greatly improve work efficiency. By simulating human operations, they automate various business processes, reduce human intervention, and lower the probability of errors.

[0003] However, existing RPA systems lack effective processing and analysis methods for the log data generated during system operation. This large amount of operational log data is often in a raw, disorganized state, making it difficult to extract valuable information and gain a comprehensive understanding of the system's operational status. Furthermore, existing technologies are unable to accurately detect anomalies within the system, making it even more difficult to rationally adjust task execution strategies based on these anomalies. This can lead to issues such as task execution confusion and reduced efficiency when the system encounters anomalies, impacting the normal operation of business processes.

[0004] In summary, existing RPA system technology has shortcomings in processing operation log data, detecting anomalies, and adjusting task execution strategies. There is an urgent need for a more effective technical solution to solve these problems and improve the stability and efficiency of the RPA system. Summary of the Invention

[0005] The embodiments of the present application provide a big data preprocessing method and system based on machine learning, which are used to improve the stability and efficiency of the RPA system and ensure the smooth progress of business processes.

[0006] In a first aspect, an embodiment of the present application provides a big data preprocessing method based on machine learning, which is applied to a big data preprocessing system, the method comprising: obtaining an operation log data set of an RPA system, the operation log data set comprising multiple session interaction records; performing data cleaning processing on the operation log data set to obtain a preprocessed operation log data set; performing feature extraction processing on the preprocessed operation log data set based on a preset feature extraction algorithm to obtain text semantic features and operation sequence features of each session interaction record; calling a machine learning model to jointly model the text semantic features and the operation sequence features, generating anomaly detection results for the session interaction records, and adjusting the task execution strategy of the RPA system based on the anomaly detection results.

[0007] In a second aspect, an embodiment of the present application provides a big data preprocessing system, comprising:

[0008] processor;

[0009] a storage device having a computer program stored thereon,

[0010] When the computer program is executed by the processor, the processor implements any of the machine learning-based big data preprocessing methods.

[0011] An embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the big data preprocessing method based on machine learning are implemented.

[0012] The embodiment of the present application first obtains the operation log data set of the RPA system and merges it for data cleaning processing, effectively removing noise and incomplete data, improving data quality, and making the pre-processed operation log data set more accurate and reliable; secondly, based on the preset feature extraction algorithm, the text semantic features and operation sequence features of each conversation interaction record are extracted, which can accurately characterize the characteristics of the interaction record from different dimensions, providing a comprehensive and detailed basis for anomaly detection; then, the machine learning model is called to jointly model the text semantic features and operation sequence features, which can fully explore the potential relationship between the features and generate accurate anomaly detection results. Finally, the task execution strategy of the RPA system is adjusted according to the anomaly detection results, which can dynamically adapt to the system operation status, optimize the task execution process, improve the stability and efficiency of the RPA system, and ensure the smooth progress of the business process. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flowchart of a big data preprocessing method based on machine learning provided in an embodiment of the present application.

[0014] Figure 2 A schematic diagram of the basic structure of a big data preprocessing system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the embodiments of the present application are further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0016] See also Figure 1 As shown in FIG, this figure is a flow chart of a method for preprocessing big data based on machine learning provided in an embodiment of the present application, which can be applied to a big data preprocessing system. Figure 1 As shown, the method includes steps 110 to 140.

[0017] It should be noted that in the embodiments of this application, the RPA system's operational log data collection strictly adheres to the principles of privacy protection and user authorization, ensuring the legitimacy and transparency of data processing. Upon first using the system, a clear privacy policy and user agreement inform users of the scope, purpose, and usage of data collection, including recording session interaction logs for system optimization and anomaly detection. Users are required to actively select authorization before enabling relevant features. The system adheres to the principle of minimum necessary, collecting only operational events and response data directly related to business processes, avoiding the inclusion of sensitive personal information. Data fields that may contain user identifiers (such as user IDs) are anonymized using de-identification technology to ensure that log records cannot be linked to specific individuals. A strict data access control mechanism is established within the enterprise, authorizing only specific operations and risk management personnel to access raw logs, and all data processing activities are tracked through audit logs. Users can review the type of data collected at any time through system settings and have the right to request the export or deletion of their personal operation records. Furthermore, the system regularly sends data usage reports to users, explaining the purpose and results of log analysis, to continuously safeguard users' right to be informed and their right to choose.

[0018] Step 110: Obtain an operation log data set of the RPA system, where the operation log data set includes multiple session interaction records.

[0019] In the embodiment of the present application, an RPA system used within an enterprise is taken as an example. The RPA system is responsible for processing various business processes, such as order processing, data entry, etc. Over a period of time, the system will record the interactions between many users and the system, and these records constitute an operation log data set. For example, in the order processing process, user A submitted an order, and the system recorded user A's operation event and the corresponding system response event; user B modified an order, which was also recorded. These various operation events and system response events of different users at different times together constitute multiple session interaction records in the operation log data set. The session interaction records store all key information during the operation of the RPA system and provide raw materials for subsequent data processing and analysis. Then, the system will integrate the data scattered in each session interaction record to form a complete operation log data set that can be processed later.

[0020] Step 120: performing data cleaning processing on the operation log data set to obtain a pre-processed operation log data set.

[0021] In this example, each session interaction record consists of at least one user operation event and a corresponding system response event. The data cleaning process includes deleting redundant operation events, correcting inconsistent system response events, and filling in missing session interaction records. Based on this, step 120 includes:

[0022] Step 121: traverse each session interaction record in the operation log data set, identify logical conflict events between user operation events and system response events in the session interaction records, and the logical conflict events include events in which user operation events are not correctly reflected by system response events; perform context association analysis on the logical conflict events, determine the target system response events corresponding to the logical conflict events, and perform logic correction processing on the target system response events based on a preset rule base to obtain a corrected system response event.

[0023] For example, in the order processing process, within the operational log data set, there may be a session interaction record showing a user submitting an order for a product. However, the system response event indicates that the order failed with an error message unrelated to the product. This is a logical conflict event, where the user action is not correctly reflected by the system response event. In this case, the system performs contextual analysis on this logical conflict event, examining other related actions and system responses before and after the order submission. For example, the analysis may reveal that the product was in sufficient stock before the order was submitted, but then suddenly ran out of stock. Based on a pre-set rule base that explicitly states that when a product is out of stock, the system should correctly indicate that the order failed due to insufficient stock. Based on this rule base, the target system response event is modified to explicitly indicate that the product is out of stock, resulting in a modified system response event. This ensures the correct logical relationship between the system response and the user action, providing an accurate data foundation for subsequent analysis and processing.

[0024] In a preferred embodiment, the logic correction processing of the target system response event based on a preset rule base to obtain a corrected system response event includes:

[0025] Step 1211: Obtain the semantic description text of the user operation event corresponding to the target system response event, perform intent analysis on the semantic description text, and obtain the target operation intention of the user operation event.

[0026] In the order processing example above, the semantic description text of the user action event corresponding to the target system response event is obtained, such as "User submits an order to purchase Product X." This semantic description text is then subjected to intent parsing, using natural language processing techniques to analyze the keywords and grammatical structure within the text. For example, by identifying the core verb "purchase" and the action object "Product X," the system can determine that the target action intent of the user action event is to purchase Product X. This target action intent is an important basis for subsequent actions, clarifying the user's purpose and enabling the system to better understand the user's needs and respond accordingly.

[0027] Step 1212: Match a corresponding expected response event template in the rule base based on the target operation intention, where the expected response event template is pre-configured with a standard response event format corresponding to the target operation intention.

[0028] Optionally, a match is performed against the determined target action intent of purchasing product X within the rule library. The rule library pre-configured expected response event templates corresponding to various target action intents. For a purchase action intent, these expected response event templates may include response formats for successful purchases, such as "Order submitted successfully. Product X will be shipped within XX timeframe," as well as response formats for failed purchases, such as "The order for product X failed due to XX reason." By searching the rule library, the expected response event template corresponding to the target action intent of purchasing product X is found. This template provides a standard format for subsequent judgment and correction of system response events.

[0029] Step 1213: Compare the formats of the target system response event and the expected response event template. If the response format of the target system response event is inconsistent with the expected response event template, convert the format of the target system response event according to the format of the expected response event template.

[0030] Optionally, the target system response event "Order processing failed and the error message is not related to the product" is formatted and compared with the expected response event template. It is discovered that the format of the target system response event is inconsistent with the format of the purchase failure in the expected response event template. The expected response event template requires that the failure cause be clearly stated to be related to the product, but the current target system response event does not meet this format requirement. Therefore, the target system response event is formatted according to the expected response event template, converting it to "The order for product X failed due to insufficient inventory of product X" to conform to the standard format requirements, facilitating subsequent logical consistency verification with user operation events.

[0031] Step 1214: Perform a logical consistency check between the target system response event after format conversion and the user operation event. If the check passes, use the target system response event after format conversion as the corrected system response event.

[0032] Optionally, a logical consistency check is performed between the converted target system response event ("The order for product X failed due to insufficient product X inventory") and the user action event ("User submitted an order for product X"). By analyzing the logical relationship between the two, it is determined whether the reason for the order failure (inadequate product inventory) and the user's purchase action are reasonably associated. If the logical relationship is found to be reasonable and the check passes, the converted target system response event is used as the corrected system response event, completing the correction of the system response event and ensuring data accuracy and logical consistency.

[0033] Step 122: traverse each session interaction record in the operation log data set, and detect whether there are continuously repeated redundant operation events in the session interaction record. If so, deduplicate the redundant operation events, retain the first user operation event and delete subsequent repeated redundant operation events.

[0034] In the order processing log data set, each session interaction record is traversed. For example, if a user submits the same order for product Y multiple times in a short period of time in a session interaction record, these repeated order submissions are considered redundant. The system deduplicates these redundant operation events, retaining the first user's order for product Y and deleting subsequent duplicate order submissions. This reduces redundant information in the data, improves data quality, and avoids interference caused by duplicate data during subsequent analysis and processing.

[0035] Step 123: traverse each session interaction record in the operation log data set, and detect whether there is a missing timestamp between the user operation event and the system response event in the session interaction record. If so, interpolate and fill in the missing timestamp based on the timestamp sequence of adjacent session interaction records.

[0036] Continuing with the order processing example, while traversing the run log data set, the system discovered a missing timestamp for a user action event (order submission) in a session interaction record. In this case, the system examines the timestamp sequence of adjacent session interaction records. For example, the timestamp for a user submitting another order in the previous session interaction record is 10:00, while the timestamp for the system's response to that order in the subsequent session interaction record is 10:05. By analyzing this time series and considering the typical time intervals and logical sequence of order processing, an appropriate interpolation method, such as linear interpolation, is used to estimate the missing timestamp for the user submitting the order to be 10:02. Based on this method, the missing timestamp is interpolated to fill in the missing timestamp, ensuring the complete time information in the session interaction record, consistent and accurate data, and providing reliable data for subsequent time series-based analysis and processing.

[0037] Step 124: Based on the session interaction records that have completed the system response event correction, completed the redundant operation event deduplication, and completed the interpolation and filling processing, the pre-processed operation log data set is obtained.

[0038] Optionally, after processing each session interaction record in the operation log data set for system response event correction, redundant operation event deduplication, and interpolation, these processed session interaction records are integrated to obtain a preprocessed operation log data set. This preprocessed operation log data set removes erroneous and redundant information from the original data and supplements missing information, significantly improving data quality and providing a high-quality data foundation for subsequent feature extraction and model processing. For example, after such processing, the session interaction records related to order processing are more accurate and complete, better reflecting the operations and responses in the actual business process.

[0039] Step 130: performing feature extraction processing on the pre-processed operation log data set based on a preset feature extraction algorithm to obtain text semantic features and operation sequence features of each conversation interaction record.

[0040] As an implementation method, the feature extraction process is performed on the pre-processed operation log data set based on a preset feature extraction algorithm to obtain text semantic features and operation sequence features of each session interaction record, including:

[0041] Step 131: performing text segmentation processing on the user operation event in the conversation interaction record to obtain multiple operation word units, and performing part-of-speech tagging processing on the multiple operation word units to identify the core operation verbs and auxiliary operation objects in the user operation event.

[0042] For example, consider a conversational interaction record for order processing, where the user action event is "submit an order to purchase a computer." This text is segmented to produce multiple action word units, such as "submit," "one," "purchase," "computer," and "order." These action word units are then tagged with part-of-speech, with "submit" and "purchase" marked as verbs, "one" as a quantifier, and "computer" and "order" as nouns. By analyzing these parts of speech, we identify "submit" and "purchase" as core action verbs, and "computer" and "order" as auxiliary action objects. These core action verbs and auxiliary action objects accurately reflect the key information of the user action event, providing a foundation for the subsequent generation of text semantic features.

[0043] Step 132: Call a pre-trained language encoding model to perform semantic encoding processing on the core operation verbs and the auxiliary operation objects, generate a text vector representation of the user operation event, and use the text vector representation as the text semantic feature.

[0044] In this process, a language encoding model that has been pre-trained on a large amount of text data, such as the BERT model, is called. The identified core operation verbs "submit" and "purchase" and the auxiliary operation objects "computer" and "order" are input into the model. The model will perform semantic encoding processing on these words, and generate corresponding vector representations for each word based on the contextual information and semantic relationships of the words in a large amount of text. For example, "submit" may be encoded as a vector containing multiple numerical values, and "purchase", "computer" and "order" also have their own corresponding vectors. These vectors are then combined or further processed to generate a text vector representation of the user operation event. This text vector representation can comprehensively reflect the semantic information of the user operation event, and it is used as a text semantic feature to provide semantic input for subsequent machine learning models.

[0045] Step 133: Perform operation sequence parsing processing on the system response event in the conversation interaction record, identify the atomic operation instruction set contained in the system response event, and perform sequence pattern encoding processing on the atomic operation instruction set to generate the operation sequence code of the system response event.

[0046] In order processing, a system response event may be "verify order information, check inventory, generate a delivery note, and arrange logistics." The system response event is subjected to operation sequence parsing to identify the atomic operation instruction set contained therein, such as "verify order information," "check inventory," "generate a delivery note," and "arrange logistics." Each atomic operation instruction represents a basic operation unit. As an optional embodiment, the sequence pattern encoding process is performed on the atomic operation instruction set to generate the operation sequence code of the system response event, including:

[0047] Step 1331: Traverse each atomic operation instruction in the atomic operation instruction set to determine the operation type identifier and operation parameter set of the atomic operation instruction.

[0048] Furthermore, the aforementioned atomic operation instruction set is traversed. For the atomic operation instruction "Verify Order Information," its operation type identifier is determined to be "Order Verification," and its operation parameter set may include the order number, customer information, etc. For the atomic operation instruction "Check Inventory," the operation type identifier is "Inventory Check," and the operation parameter set includes the product name, inventory quantity, etc. Through the above method, the operation type identifier and operation parameter set are determined for each atomic operation instruction for subsequent encoding processing.

[0049] Step 1332: Map the atomic operation instruction to a preset operation type space based on the operation type identifier, and generate a type encoding vector for the atomic operation instruction.

[0050] In an embodiment of the present application, an operation type space is preset. For example, "order verification" is mapped to a region in the operation type space. Based on the encoding rules of the region, a type encoding vector for the "order verification" atomic operation instruction is generated. The type encoding vector is a multi-dimensional vector whose dimensions and values ​​are determined according to the definition and encoding rules of the operation type space. Similarly, other atomic operation instructions such as "inventory check" are also mapped to the operation type space, generating their own type encoding vectors. In this way, the operation type information is digitally represented.

[0051] Step 1333: Perform parameter type parsing on each operation parameter in the operation parameter set, determine the data type identifier and parameter value range of the operation parameter, and generate a parameter encoding vector of the operation parameter based on the data type identifier and the parameter value range.

[0052] For the order number in the "Verify Order Information" operation parameter set, its data type identifier is parsed as "string type." The parameter value range can be a string of a certain length consisting of numbers and letters. Based on the data type identifier and parameter value range, a parameter encoding vector for the order number operation parameter is generated. The same process is performed for other operation parameters, such as customer information, to generate their own parameter encoding vectors. These parameter encoding vectors accurately represent the type and value range of the operation parameter.

[0053] Step 1334: Concatenate the type coding vector and the parameter coding vector to obtain the instruction coding vector of the atomic operation instruction, serialize the instruction coding vector based on the execution order of the atomic operation instruction in the system response event, and generate the operation sequence code.

[0054] Furthermore, the type encoding vector and parameter encoding vector of each atomic operation instruction are concatenated. For example, the type encoding vector of "Verify Order Information" and the parameter encoding vectors such as the order number and customer information are concatenated together to obtain the instruction encoding vector of the atomic operation instruction "Verify Order Information." These instruction encoding vectors are then serialized according to the execution order of the atomic operation instructions in the system response event, such as "Verify Order Information," "Check Inventory," "Generate Invoice," and "Arrange Logistics." These vectors are arranged and combined in sequence to generate the operation sequence code of the system response event. This operation sequence code can fully reflect the operation sequence and specific operation content information of the system response event.

[0055] Step 134: perform time dimension alignment processing on the operation sequence encoding and the text vector representation to obtain the operation sequence features of the conversation interaction record.

[0056] In the order processing session interaction records, the generated operation sequence codes and text vector representations are aligned in time. Because user operation events and system response events follow a temporal sequence, time alignment allows the operation sequence codes and text vector representations to correspond to each other. For example, the text vector representation corresponding to the user's order submission operation is aligned with the operation sequence code of the earliest "Verify Order Information" operation in the system response. Subsequent operation sequence codes and text vector representations are also aligned in chronological order. This alignment process yields the operation sequence features of the session interaction records, which comprehensively reflect the temporal relationship and operation sequence between user operations and system responses.

[0057] Step 140: Call a machine learning model to jointly model the text semantic features and the operation sequence features, generate anomaly detection results of the conversation interaction records, and adjust the task execution strategy of the RPA system based on the anomaly detection results.

[0058] In a preferred embodiment, the calling of a machine learning model to jointly model the text semantic features and the operation sequence features to generate an anomaly detection result of the conversation interaction record includes:

[0059] Step 141: Input the text semantic features into the text encoding branch of the machine learning model, and generate a text high-order feature vector through multi-layer nonlinear transformation.

[0060] The machine learning model of the embodiment of the present application is a deep neural network model with multiple branches and layers. The generated text semantic features are input into the text encoding branch of the model, which contains multiple nonlinear transformation layers, such as a fully connected layer and an activation function layer. After entering these layers, the text semantic feature vector first passes through the fully connected layer, which performs a linear transformation on the input vector according to a preset weight matrix, and then performs a nonlinear transformation through the activation function. For example, using the ReLU activation function, the result after the linear transformation is nonlinearly transformed to enhance the model's ability to express semantic information. After multiple layers of the above-mentioned nonlinear transformations, the text semantic features are gradually converted into text high-order feature vectors, which can more deeply represent the complex information in the text semantics.

[0061] Step 142: Input the operation sequence features into the sequence encoding branch of the machine learning model, and generate a sequence high-order feature vector through a temporal convolution operation.

[0062] Optionally, the operation sequence features are input into the sequence encoding branch of the machine learning model. This branch employs a temporal convolution operation, which performs sliding convolution on the operation sequence feature vector along the time dimension using a convolution kernel. For example, the convolution kernel size can be 3, which convolves three consecutive vector values ​​along the time dimension of the operation sequence encoding vector. This convolution extracts local features and patterns of the operation sequence along the time dimension. After multiple temporal convolution operations, the operation sequence features are converted into a high-order feature vector that better reflects the temporal variation and characteristics of the operation sequence.

[0063] Step 143: Perform cross-modal attention interaction processing on the text high-order feature vector and the sequence high-order feature vector to determine an association weight matrix between the text high-order feature vector and the sequence high-order feature vector.

[0064] Furthermore, step 143 includes:

[0065] Step 1431: Take the text high-order feature vector as the first query vector, and the sequence high-order feature vector as the first key vector and the first value vector, and calculate the first attention score between the first query vector and the first key vector; normalize the first attention score to obtain the first attention weight distribution between the text high-order feature vector and the sequence high-order feature vector; perform weighted summation on the first value vector based on the first attention weight distribution to obtain the first cross-modal context vector corresponding to the text high-order feature vector; perform residual connection processing on the first cross-modal context vector and the text high-order feature vector to obtain an updated text high-order feature vector.

[0066] In this step, the text high-order feature vector is used as the first query vector, and the sequence high-order feature vector is used as the first key vector and the first value vector. The first attention score is obtained by calculating the dot product between the query vector and the key vector. For example, the dot product formula is used to calculate the sum of the products of each element in the text high-order feature vector and the corresponding element in the sequence high-order feature vector to obtain the first attention score. The score is then normalized, such as using a softmax function for normalization, to obtain the first attention weight distribution between the text high-order feature vector and the sequence high-order feature vector. The weight distribution represents the degree of attention of the text high-order feature vector to different parts of the sequence high-order feature vector.

[0067] Based on the first attention weight distribution, a weighted summation is performed on the first value vector (i.e., the sequence high-order feature vector). This weight distribution is multiplied by the corresponding elements of the sequence high-order feature vector and then summed to obtain the first cross-modal context vector corresponding to the text high-order feature vector. This cross-modal context vector contains information related to the text high-order feature vector obtained from the sequence high-order feature vector. Next, a residual connection is performed on the first cross-modal context vector and the text high-order feature vector. This residual connection adds the corresponding elements of the two to obtain an updated text high-order feature vector. This residual connection helps the model better learn and retain the information in the original text high-order feature vector while integrating relevant information from the sequence high-order feature vector.

[0068] Step 1432: Use the sequence high-order feature vector as the second query vector, and the text high-order feature vector as the second key vector and the second value vector, and calculate the second attention score between the second query vector and the second key vector; normalize the second attention score to obtain a second attention weight distribution between the sequence high-order feature vector and the text high-order feature vector; perform weighted summation on the second value vector based on the second attention weight distribution to obtain a second cross-modal context vector corresponding to the sequence high-order feature vector; perform residual connection processing on the second cross-modal context vector and the sequence high-order feature vector to obtain an updated sequence high-order feature vector.

[0069] Similarly, the sequence high-order feature vector is used as the second query vector, and the text high-order feature vector is used as the second key vector and second value vector. Using a similar calculation method, the dot product between the second query vector and the second key vector is first calculated to obtain a second attention score. The second attention score is then normalized, for example, using the softmax function again, to obtain a second attention weight distribution between the sequence high-order feature vector and the text high-order feature vector. This weight distribution reflects the degree of attention paid by the sequence high-order feature vector to different parts of the text high-order feature vector.

[0070] Based on the second attention weight distribution, a weighted summation is performed on the second value vector (i.e., the text high-order feature vector) to obtain the second cross-modal context vector corresponding to the sequence high-order feature vector. Finally, a residual connection is performed on the second cross-modal context vector and the sequence high-order feature vector, that is, corresponding elements are added together to obtain the updated sequence high-order feature vector. In this way, the sequence high-order feature vector also incorporates relevant information from the text high-order feature vector, further enriching its own feature representation.

[0071] Step 1433: Generate a third query vector based on the updated text high-order feature vector, and generate a third key vector based on the updated sequence high-order feature vector; calculate the bidirectional cross-attention score matrix between the third query vector and the third key vector; perform row-column bidirectional normalization on the bidirectional cross-attention score matrix to generate an association weight matrix between the text high-order feature vector and the sequence high-order feature vector.

[0072] Furthermore, the updated text high-order feature vector is used to generate a third query vector, and the updated sequence high-order feature vector is used to generate a third key vector. A bidirectional cross-attention score matrix is ​​calculated between the third query vector and the third key vector. This means that not only the attention score in the text-to-sequence direction but also the attention score in the sequence-to-text direction is calculated, and these scores are combined into a matrix. For example, each element in the matrix represents the attention score between a certain dimension of the text high-order feature vector and a certain dimension of the sequence high-order feature vector.

[0073] Next, the bidirectional cross-attention score matrix is ​​subjected to row-column bidirectional normalization. In the row direction, the elements of each row are normalized so that the sum of the elements in each row is 1; in the column direction, the elements of each column are also normalized so that the sum of the elements in each column is also 1. After the above-mentioned row-column bidirectional normalization, the association weight matrix between the text high-order feature vector and the sequence high-order feature vector is generated. This association weight matrix can accurately reflect the degree of mutual correlation between the two modal feature vectors, providing an important basis for subsequent weighted fusion.

[0074] Step 144: performing weighted fusion processing on the text high-order feature vector and the sequence high-order feature vector based on the association weight matrix to generate a joint feature representation vector;

[0075] In an exemplary embodiment, performing weighted fusion processing on the text high-order feature vector and the sequence high-order feature vector based on the association weight matrix to generate a joint feature representation vector includes:

[0076] Step 1440: decompose the association weight matrix into a text attention weight vector and a sequence attention weight vector; perform weighted scaling processing on the text high-order feature vector based on the text attention weight vector to obtain a scaled text feature vector; perform weighted scaling processing on the sequence high-order feature vector based on the sequence attention weight vector to obtain a scaled sequence feature vector; perform element-by-element addition processing on the scaled text feature vector and the scaled sequence feature vector to obtain an initial fused feature vector; perform layer normalization processing on the initial fused feature vector to eliminate feature scale differences to obtain a normalized fused feature vector; input the normalized fused feature vector into the fully connected layer for feature compression processing to generate the joint feature representation vector.

[0077] Specifically, the association weight matrix is ​​decomposed into a text attention weight vector and a sequence attention weight vector according to certain rules. For example, the matrix is ​​divided according to the row or column information, and the weight information related to the text high-order feature vector is extracted to form the text attention weight vector, and the weight information related to the sequence high-order feature vector is extracted to form the sequence attention weight vector.

[0078] Then, the text high-order feature vector is weighted and scaled using the text attention weight vector. That is, each element of the text attention weight vector is multiplied by the corresponding element of the text high-order feature vector to obtain the scaled text feature vector. Similarly, the sequence high-order feature vector is weighted and scaled using the sequence attention weight vector to obtain the scaled sequence feature vector.

[0079] Next, the scaled text feature vector and the scaled sequence feature vector are added element-by-element, adding the elements at corresponding positions in the two vectors to obtain the initial fused feature vector. Because the scales of different feature vectors may vary, which can affect model training and performance, the initial fused feature vector is subjected to layer normalization. Layer normalization normalizes the vector along each dimension, eliminating feature scale differences through calculations such as subtracting the mean and dividing by the standard deviation, resulting in a normalized fused feature vector.

[0080] Finally, the normalized fused feature vector is fed into a fully connected layer for feature compression. The fully connected layer performs matrix multiplication on the input vector using a weight matrix, compressing the high-dimensional fused feature vector into a lower-dimensional space. This generates a joint feature representation vector that incorporates the semantics of the text and the feature information of the operation sequence, providing a more comprehensive and effective feature representation for subsequent anomaly detection.

[0081] Step 145: Input the joint feature representation vector into the classifier layer of the machine learning model, and output the anomaly detection result of the session interaction record, where the anomaly detection result is used to indicate whether there is an operational logic anomaly or a system response anomaly in the session interaction record.

[0082] The generated joint feature representation vector is then input into the classifier layer of the machine learning model. This classifier layer can be a simple fully connected layer or a more complex classification model, such as a softmax classifier. The classifier layer calculates and judges the joint feature representation vector based on pre-trained weight parameters. For example, the weight matrix in the classifier layer is multiplied by the joint feature representation vector. The result is then converted into a probability distribution using an activation function (such as the softmax function). This probability distribution represents the probability of the session interaction record belonging to different categories (normal or abnormal). Based on the probability distribution, the model outputs a decision result, namely, the anomaly detection result for the session interaction record. If the output indicates that the probability of belonging to the abnormal category exceeds a certain threshold (for example, 0.5), the session interaction record is determined to contain an operational logic anomaly or an abnormal system response. Conversely, if the probability of belonging to the normal category is higher, the session interaction record is considered normal. This anomaly detection result can help the RPA system promptly identify operational problems and provide a basis for adjusting task execution strategies.

[0083] As an implementation manner, adjusting the task execution strategy of the RPA system based on the anomaly detection result includes:

[0084] Step 146: Count the target operation event distribution and the abnormal type distribution of the abnormal session interaction records in the abnormal detection results to generate an abnormal event statistical report; and determine the frequent abnormal operation event set and the risk abnormal type set in the RPA system based on the abnormal event statistical report.

[0085] In the order processing scenario, anomaly detection results are statistically analyzed. First, all session interaction records identified as abnormal are traversed to count the frequency and distribution of target operation events. For example, the "submit order" operation is found to have a high frequency of anomalies in abnormal session interaction records. The frequency and distribution of different types of anomalies, such as "order failure due to insufficient inventory but not properly notified" and "order information verification failure with unknown reasons," are also recorded. Based on this statistical information, an abnormal event statistical report is generated. The report details the frequency of anomalies for various target operation events and the distribution of different anomaly types. Based on this abnormal event statistical report, a set of frequently abnormal operation events and a set of risky anomaly types in the RPA system are further identified. The frequently abnormal operation event set may include operations such as "submit order" and "modify order," which appear frequently in abnormal session interaction records. The risky anomaly type set may include anomalies with a high business impact and a relatively high frequency of occurrence, such as "order failure due to insufficient inventory but not properly notified," which may lead to customer churn.

[0086] Step 147: For each frequent abnormal operation event in the set of frequent abnormal operation events, a pre-verification rule is inserted into the task execution process of the RPA system, and the pre-verification rule is used to activate the secondary confirmation process when the frequent abnormal operation event is detected.

[0087] As you can understand, for the "Submit Order" operation, which is part of the frequently abnormal operation event set, a pre-verification rule is inserted into the RPA system's order processing task execution process. When the system detects a user initiating the "Submit Order" operation, the pre-verification rule is triggered first. For example, the pre-verification rule checks the integrity of product inventory information and user order information. If a situation that may cause an abnormality is detected, such as product inventory approaching a critical value or partial missing order information, a secondary confirmation process is activated. During this secondary confirmation process, the system will pop up a prompt to the user, informing them of the possible problem and asking them to reconfirm whether they want to submit the order. This can avoid frequent abnormalities caused by user errors or incomplete information, thereby improving the accuracy and success rate of order processing.

[0088] Step 148: For each risk exception type in the risk exception type set, configure an exception fuse mechanism in the task execution process of the RPA system, and the exception fuse mechanism is used to suspend the current task process and activate the manual intervention process when a preset number of risk exception types are detected continuously.

[0089] As you can understand, for the risk anomaly type "Insufficient inventory leading to order failure without proper notification" in the risk anomaly type set, an exception circuit breaker mechanism is configured within the RPA system's order processing task execution process. A preset number of times, for example, if this risk anomaly type is detected three times in a row, triggers the exception circuit breaker mechanism. When the exception circuit breaker mechanism is triggered, the currently processing order task process is paused to prevent further orders that could result in errors or adverse effects. Simultaneously, the system activates the manual intervention process, notifying relevant staff to intervene. Staff can then check the inventory management system to confirm actual inventory levels and correct system notifications to resolve the issue that caused the risk anomaly, ensuring that subsequent order processing can proceed normally.

[0090] In a non-limiting embodiment, after adjusting the task execution strategy of the RPA system based on the anomaly detection result, it also includes: obtaining a target conversation interaction record set corresponding to the anomaly detection result, extracting target text semantic features and target operation sequence features of the target conversation interaction record, and generating a derived training sample set; optimizing and training the machine learning model based on the derived training sample set, and calculating the gradient difference vector between the current model parameters and the model parameters after derivative training; adjusting the weight distribution of the machine learning model according to the gradient difference vector to generate an optimized machine learning model; deploying the optimized machine learning model to the real-time detection module of the RPA system, performing anomaly detection on the newly added conversation interaction records, and updating the derived training sample set based on the detection results.

[0091] In the order processing scenario, based on the anomaly detection results, a corresponding set of target session interaction records is obtained. This includes those identified as anomalous and a representative set of normal session interaction records. Target text semantic features and target operation sequence features are extracted from these target session interaction records. The same feature extraction methods as before are used, such as performing text segmentation and part-of-speech tagging on user operation events to generate text semantic features, and parsing and encoding the operation sequence of system response events to generate operation sequence features. These newly extracted features are combined to generate a derived training sample set. This derived training sample set is then used to optimize and train the machine learning model.

[0092] During training, the model calculates the gradient difference vector between the current model parameters and the model parameters trained using the derived training sample set. This gradient difference vector reflects the direction and degree of parameter change after training with the new sample set. Based on the gradient difference vector, an optimization algorithm (such as stochastic gradient descent) is used to adjust the weight distribution of the machine learning model. This adjusted weight distribution enables the model to better adapt to new sample data, thereby generating an optimized machine learning model. The optimized machine learning model is deployed to the real-time detection module of the RPA system. When new session interaction records enter the system, the real-time detection module uses the optimized model for anomaly detection. If new anomalies or new representative normal situations are detected, these new session interaction records are added to the derived training sample set, and the sample set is further updated to continuously optimize the model and improve its accuracy and adaptability.

[0093] In another non-limiting embodiment, after adjusting the task execution strategy of the RPA system based on the anomaly detection result, the method further includes: collecting operation log data of the RPA system after executing the adjusted task execution strategy in real time, and extracting a set of conversation interaction records after the strategy adjustment; performing anomaly rechecking on the set of conversation interaction records after the strategy adjustment, and generating a recheck anomaly rate and anomaly type distribution change matrix; calculating the difference coefficient between the recheck anomaly rate and the historical anomaly rate, and analyzing the frequency change gradient of each type of anomaly in the anomaly type distribution change matrix; if the difference coefficient exceeds a preset threshold or the frequency change gradient indicates that the anomaly type is spreading, adjusting the activation condition of the pre-check rule and the pause threshold of the anomaly fuse mechanism; generating a strategy optimization instruction based on the adjusted activation condition and pause threshold, driving the RPA system to perform strategy iterative updates and output a strategy adjustment performance report.

[0094] In the order processing system, real-time log data is collected after the RPA system implements the adjusted task execution policy. For example, after inserting pre-verification rules and configuring the exception circuit breaker mechanism, the system continuously records information about user operations and system responses. A set of session interaction records after the policy adjustment is extracted from this log data. This set is then rechecked for anomalies, using a machine learning model to detect anomalies in these session interaction records. The rechecked anomaly rate is calculated, which represents the proportion of abnormal session interaction records in the total session interaction records.

[0095] At the same time, an anomaly type distribution change matrix is ​​generated, which records the change in the frequency of different anomaly types before and after the policy adjustment. The difference coefficient between the rechecked anomaly rate and the historical anomaly rate is calculated. For example, the difference between the two is divided by the historical anomaly rate to obtain the difference coefficient. The frequency change gradient of each anomaly type in the anomaly type distribution change matrix is ​​analyzed to observe which anomaly types are increasing or decreasing in frequency. If the difference coefficient exceeds a preset threshold, such as 0.1, or the frequency change gradient shows a rapid increase in the frequency of certain anomaly types, indicating anomaly type proliferation, the current policy adjustment may be unsatisfactory and requires further adjustment. In this case, the activation conditions of the pre-check rules can be adjusted, such as loosening or tightening the inventory check threshold, and the suspension threshold of the anomaly circuit breaker mechanism can be adjusted, such as increasing the number of consecutive detections of risky anomaly types from three to four. Based on the adjusted activation conditions and suspension thresholds, a policy optimization instruction is generated and sent to the RPA system. The RPA system performs an iterative policy update based on this instruction, updating the relevant rules and mechanisms in the task execution process.

[0096] Finally, the system outputs a policy adjustment effectiveness report, which records in detail the anomaly detection situation before and after the policy adjustment, changes in task execution efficiency, and other information, so as to evaluate the effect of the policy adjustment and further optimize the system.

[0097] It is understood that when implementing the above-mentioned technical solutions of the embodiments of this application, those skilled in the art can optimize and supplement the processing based on the existing natural language processing, sequence pattern encoding, and cross-modal attention mechanisms. Specifically, for the construction of the rule base in data cleaning, an expected response template matching mechanism can be established by combining the domain knowledge graph and the historical abnormal event library. The standard response format is dynamically generated through entity relationship extraction and logical reasoning to avoid correction failure caused by the static nature of the rule base.

[0098] When encoding the operation sequence, the operation type space can be defined by referring to the instruction classification system of the industrial standard OPCUA, a lightweight encoder can be used to normalize and map the parameter types, and a hash function can be used to process string parameters to generate a fixed-length parameter encoding vector to ensure the uniformity of the dimensions of different data types.

[0099] For cross-modal attention interaction, a multi-head attention mechanism can be introduced and a dynamic dimensional projection layer can be set. The dimensional difference between text vectors and sequence encodings can be eliminated through adaptive weight distribution, and the gated residual network can be combined to control the feature fusion ratio.

[0100] For the implementation of the classifier layer, a hierarchical softmax structure can be used and the Focal Loss loss function can be introduced to optimize the sample imbalance problem, and the sliding window mechanism can be combined to dynamically adjust the anomaly judgment threshold.

[0101] In the policy adjustment phase, a policy optimizer can be built based on the reinforcement learning framework. The Q-learning algorithm can be used to quantify the triggering conditions of the pre-verification rules and the circuit breaker mechanism. Combined with online learning, the policy parameters can be updated in real time to ensure dynamic adaptation of the exception handling logic to the business scenario.

[0102] By introducing the above-mentioned existing technologies to optimize and improve the implementation of the technical solution, it is possible to fully realize the processing from log cleaning, feature fusion to policy iteration, so that technical personnel in the relevant field can clearly and completely implement the embodiments of the present application.

[0103] The embodiment of the present application first obtains the operation log data set of the RPA system and merges it for data cleaning processing, effectively removing noise and incomplete data, improving data quality, and making the pre-processed operation log data set more accurate and reliable; secondly, based on the preset feature extraction algorithm, the text semantic features and operation sequence features of each conversation interaction record are extracted, which can accurately characterize the characteristics of the interaction record from different dimensions, providing a comprehensive and detailed basis for anomaly detection; then, the machine learning model is called to jointly model the text semantic features and operation sequence features, which can fully explore the potential relationship between the features and generate accurate anomaly detection results. Finally, the task execution strategy of the RPA system is adjusted according to the anomaly detection results, which can dynamically adapt to the system operation status, optimize the task execution process, improve the stability and efficiency of the RPA system, and ensure the smooth progress of the business process.

[0104] See also Figure 2 As shown, this figure is a schematic diagram of the basic structure of a big data preprocessing system 200 provided in an embodiment of the present application. The big data preprocessing system 200 includes:

[0105] Processor 201;

[0106] a storage device 202 having a computer program 2020 stored thereon;

[0107] When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the machine learning-based big data preprocessing methods.

[0108] Based on the above, a readable storage medium is provided, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0109] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems or devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

Claims

1. A big data preprocessing method based on machine learning, characterized in that: include: Obtaining an RPA system operation log data set, wherein the operation log data set includes multiple session interaction records; Performing data cleaning processing on the operation log data set to obtain a pre-processed operation log data set; Performing feature extraction on the preprocessed operation log data set based on a preset feature extraction algorithm to obtain text semantic features and operation sequence features of each session interaction record: performing text segmentation on the user operation events in the session interaction records to obtain multiple operation word units, and performing part-of-speech tagging on the multiple operation word units to identify core operation verbs and auxiliary operation objects in the user operation events; calling a pre-trained language encoding model to perform semantic encoding on the core operation verbs and the auxiliary operation objects to generate a text vector representation of the user operation event, and using the text vector representation as the text semantic feature; Performing operation sequence parsing processing on the system response event in the session interaction record, identifying an atomic operation instruction set contained in the system response event, and performing sequence pattern encoding processing on the atomic operation instruction set to generate an operation sequence code for the system response event; Performing time dimension alignment processing on the operation sequence encoding and the text vector representation to obtain operation sequence features of the conversation interaction record; The sequence pattern encoding processing is performed on the atomic operation instruction set to generate the operation sequence code of the system response event, including: traversing each atomic operation instruction in the atomic operation instruction set to determine the operation type identifier and operation parameter set of the atomic operation instruction; mapping the atomic operation instruction to a preset operation type space based on the operation type identifier to generate a type coding vector of the atomic operation instruction; performing parameter type parsing processing on each operation parameter in the operation parameter set to determine the data type identifier and parameter value range of the operation parameter, and generating a parameter coding vector of the operation parameter based on the data type identifier and the parameter value range; splicing the type coding vector with the parameter coding vector to obtain the instruction coding vector of the atomic operation instruction, and serializing the instruction coding vector based on the execution order of the atomic operation instruction in the system response event to generate the operation sequence code; A machine learning model is called to jointly model the text semantic features and the operation sequence features, generate an anomaly detection result of the conversation interaction record, and adjust the task execution strategy of the RPA system based on the anomaly detection result.

2. The method according to claim 1, characterized in that Each session interaction record consists of at least one user operation event and a corresponding system response event. The data cleaning process includes deleting redundant operation events, correcting inconsistent system response events, and filling in missing session interaction records. The data cleaning process is performed on the operation log data set to obtain a pre-processed operation log data set, including: Traversing each session interaction record in the operation log data set, identifying logical conflict events between user operation events and system response events in the session interaction records, wherein the logical conflict events include events in which the user operation events are not correctly reflected by the system response events; performing context association analysis on the logical conflict events, determining a target system response event corresponding to the logical conflict event, and performing logical correction processing on the target system response event based on a preset rule base to obtain a corrected system response event; Traversing each session interaction record in the operation log data set, detecting whether there are consecutively repeated redundant operation events in the session interaction record, and if so, deduplicating the redundant operation events, retaining the first user operation event and deleting subsequent repeated redundant operation events; Traversing each session interaction record in the operation log data set, detecting whether there is a missing timestamp between a user operation event and a system response event in the session interaction record, and if so, interpolating and filling the missing timestamp based on a timestamp sequence of adjacent session interaction records; The pre-processed operation log data set is obtained based on the session interaction records that have completed the system response event correction, completed the redundant operation event deduplication, and completed the interpolation and filling processing.

3. The method according to claim 2, characterized in that The logic correction processing of the target system response event based on the preset rule base to obtain a corrected system response event includes: Obtaining a semantic description text of a user operation event corresponding to the target system response event, performing intent analysis processing on the semantic description text, and obtaining a target operation intent of the user operation event; Matching a corresponding expected response event template in the rule base based on the target operation intention, wherein the expected response event template is pre-configured with a standard response event format corresponding to the target operation intention; Comparing the target system response event with the expected response event template in terms of format; if the response format of the target system response event is inconsistent with the expected response event template, converting the target system response event in terms of format according to the expected response event template; A logical consistency check is performed on the target system response event after format conversion and the user operation event. If the check passes, the target system response event after format conversion is used as the corrected system response event.

4. The method according to claim 1, wherein The calling of the machine learning model to perform joint modeling processing on the text semantic features and the operation sequence features to generate an anomaly detection result of the conversation interaction record includes: Inputting the text semantic features into the text encoding branch of the machine learning model, and generating a text high-order feature vector through multi-layer nonlinear transformation; Inputting the operation sequence features into the sequence encoding branch of the machine learning model, and generating a sequence high-order feature vector through a temporal convolution operation; Performing cross-modal attention interaction processing on the text high-order feature vector and the sequence high-order feature vector to determine an association weight matrix between the text high-order feature vector and the sequence high-order feature vector; Performing weighted fusion processing on the text high-order feature vector and the sequence high-order feature vector based on the association weight matrix to generate a joint feature representation vector; The joint feature representation vector is input into the classifier layer of the machine learning model, and an anomaly detection result of the session interaction record is output, where the anomaly detection result is used to indicate whether there is an operation logic anomaly or a system response anomaly in the session interaction record.

5. The method according to claim 4, characterized in that The performing cross-modal attention interaction processing on the text high-order feature vector and the sequence high-order feature vector to determine an association weight matrix between the text high-order feature vector and the sequence high-order feature vector includes: Taking the text high-order feature vector as a first query vector and the sequence high-order feature vector as a first key vector and a first value vector, calculating a first attention score between the first query vector and the first key vector; normalizing the first attention score to obtain a first attention weight distribution between the text high-order feature vector and the sequence high-order feature vector; performing weighted summation on the first value vector based on the first attention weight distribution to obtain a first cross-modal context vector corresponding to the text high-order feature vector; performing residual connection processing on the first cross-modal context vector and the text high-order feature vector to obtain an updated text high-order feature vector; Taking the sequence high-order feature vector as a second query vector, taking the text high-order feature vector as a second key vector and a second value vector, calculating a second attention score between the second query vector and the second key vector; normalizing the second attention score to obtain a second attention weight distribution between the sequence high-order feature vector and the text high-order feature vector; performing weighted summation on the second value vector based on the second attention weight distribution to obtain a second cross-modal context vector corresponding to the sequence high-order feature vector; performing residual connection processing on the second cross-modal context vector and the sequence high-order feature vector to obtain an updated sequence high-order feature vector; Generate a third query vector based on the updated text high-order feature vector, and generate a third key vector based on the updated sequence high-order feature vector; calculate a bidirectional cross-attention score matrix between the third query vector and the third key vector; perform row-column bidirectional normalization on the bidirectional cross-attention score matrix to generate an association weight matrix between the text high-order feature vector and the sequence high-order feature vector.

6. The method according to claim 4, characterized in that The step of performing weighted fusion processing on the text high-order feature vector and the sequence high-order feature vector based on the association weight matrix to generate a joint feature representation vector includes: Decomposing the association weight matrix into a text attention weight vector and a sequence attention weight vector; Performing weighted scaling processing on the text high-order feature vector based on the text attention weight vector to obtain a scaled text feature vector; Performing weighted scaling processing on the sequence high-order feature vector based on the sequence attention weight vector to obtain a scaled sequence feature vector; Performing element-by-element addition processing on the scaled text feature vector and the scaled sequence feature vector to obtain an initial fused feature vector; Performing layer normalization processing on the initial fused feature vector to eliminate feature scale differences and obtain a normalized fused feature vector; The normalized fused feature vector is input into a fully connected layer for feature compression processing to generate the joint feature representation vector.

7. The method according to claim 1, characterized in that The adjusting the task execution strategy of the RPA system based on the anomaly detection result includes: Counting the target operation event distribution and the anomaly type distribution of the abnormal conversation interaction records in the anomaly detection results, and generating an abnormal event statistical report; Determine a set of frequent abnormal operation events and a set of risk abnormality types in the RPA system based on the abnormal event statistical report; For each frequent abnormal operation event in the set of frequent abnormal operation events, a pre-verification rule is inserted into the task execution process of the RPA system, wherein the pre-verification rule is used to activate a secondary confirmation process when the frequent abnormal operation event is detected; For each risk exception type in the risk exception type set, an exception fuse mechanism is configured in the task execution process of the RPA system, and the exception fuse mechanism is used to suspend the current task process and activate the manual intervention process when a preset number of risk exception types are detected continuously.

8. A big data preprocessing system, characterized in that: include: processor; A storage device having a computer program stored thereon, wherein when the computer program is executed by the processor, the processor implements the big data preprocessing method based on machine learning as described in any one of claims 1 to 7.

Images