A Data Security Traceability Method and System Based on Artificial Intelligence

Abstract

This invention relates to the field of data security technology and provides an artificial intelligence-based data security tracing method and system. The method collects multi-dimensional operational data of industrial robots during continuous production, including continuous physical parameter sequences and discrete action command sequences; establishes a dynamic sliding window baseline model to track slow parameter drift and constructs a normal action pattern library to record high-frequency action sub-sequences; monitors baseline deviation events and rare pattern events in real time, cross-validates these two types of events on the timeline, and marks overlapping time windows that point to the same robot as suspicious contamination periods, tracing the data injection source in reverse; excludes data within suspicious contamination periods from the baseline update queue, and stores rare action sub-sequences in a blacklist pattern library. This invention can adapt to changes in equipment status, integrate multi-dimensional data cross-validation, accurately identify and trace data contamination attacks, and ensure data security in continuous industrial production.

Description

Technical Field

[0001] This invention belongs to the field of data security technology, and in particular relates to a data security tracing method and system based on artificial intelligence. Background Technology

[0002] With the rapid development of industrial automation and intelligent manufacturing, industrial robots play a central role in continuous production processes. As industrial production systems become increasingly open and networked, data security risks have increased significantly. Data contamination attacks targeting industrial robots—that is, disrupting normal production processes by injecting altered physical parameters or forged control commands—have become a key threat affecting production safety and product quality. Existing security measures are mostly focused on static measures such as network boundary defense, access control, and data encryption, which are insufficient to address the real-time detection and traceability requirements after malicious tampering of internal data flows.

[0003] Most existing technical solutions cannot effectively distinguish between the normal slow drift of parameters caused by mechanical wear and environmental changes during long-term operation of industrial robots and sudden malicious data injection. Static thresholds and fixed whitelists are prone to generating a large number of false alarms or false negatives. Detection based solely on a single dimension of physical parameters or a single dimension of instruction sequence is easily bypassed by attackers by coordinating the forgery of data in another dimension, resulting in low detection reliability. When anomalies are detected, it is difficult to accurately determine the actual time period of data contamination and reverse the source of data injection, which cannot provide a reliable basis for subsequent data cleaning and system recovery. Summary of the Invention

[0004] The purpose of this invention is to provide a data security traceability method and system based on artificial intelligence, which aims to solve the technical problems existing in the prior art as identified in the background art.

[0005] This invention is implemented as follows: a data security tracing method based on artificial intelligence, the method comprising:

[0006] Collect multi-dimensional operational data and workshop production status labels generated by industrial robots during continuous production. The multi-dimensional operational data includes a continuous physical parameter sequence and a discrete action command sequence.

[0007] For each industrial robot, a dynamic sliding window baseline model is established based on the continuous physical parameter sequence. The dynamic sliding window baseline model maintains the mean and tolerance range of each continuous physical parameter within the sliding window and tracks the slow drift of each continuous physical parameter over time.

[0008] For each industrial robot, a normal motion pattern library is constructed based on the discrete motion instruction sequence, and the normal motion pattern library records frequently occurring motion sub-sequences;

[0009] Real-time monitoring of whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, and determination of whether the deviation belongs to the recorded slow drift; if not, a baseline deviation event is generated.

[0010] Real-time monitoring is conducted to check whether rare action subsequences not recorded in the normal action pattern library appear in newly acquired discrete action instruction sequences. If they do appear, a rare pattern event is generated.

[0011] Cross-validate the baseline deviation event and the rare pattern event on the timeline. If the occurrence time windows of the two events overlap and point to the same industrial robot, then mark the time window as a suspected contamination period and trace back to the data injection source when the suspected contamination period first appeared.

[0012] Based on the data contamination results confirmed by cross-validation, the multi-dimensional operational data during the suspected contamination period are excluded from the update queue of the dynamic sliding window baseline model, and the rare action subsequences that appear during the suspected contamination period are stored in the blacklist pattern library.

[0013] As a further aspect of the present invention, the collection of multi-dimensional operational data and workshop production status tags generated by the industrial robot during continuous production specifically includes:

[0014] A continuous sequence of physical parameters of an industrial robot is collected at a fixed sampling frequency. The continuous sequence of physical parameters includes torque values, speed values, and temperature values.

[0015] The robot controller's instruction bus is monitored, and the opcode and corresponding timestamp of each motion control instruction are intercepted and arranged in chronological order to form a discrete motion instruction sequence.

[0016] Read the current production order type, the current shift handover identifier, and the current equipment maintenance mode identifier, and combine and package the three pieces of information to generate a workshop production status label.

[0017] As a further aspect of the present invention, the establishment of the dynamic sliding window baseline model specifically includes:

[0018] For each continuous physical parameter of each industrial robot, an independent sliding window of length L is created, and the arithmetic mean of all sampled values ​​within the sliding window is calculated as the mean.

[0019] Calculate the sample standard deviation of all sampled values ​​within the sliding window, and use the mean plus or minus k times the sample standard deviation as the tolerance interval;

[0020] Every sampling interval, the sliding window is moved forward by one sampling point, the mean and tolerance interval are recalculated, and the trend of mean change in the same direction and single change amount less than 0.1 times the initial sample standard deviation in three consecutive updates is recorded as slow drift.

[0021] As a further aspect of the present invention, the construction of the normal action mode library specifically includes:

[0022] The collected discrete action command sequence is divided into several sliding subsequence windows of length N according to the time sequence. Each sliding subsequence window contains N consecutive opcodes and their order.

[0023] For each sliding subsequence window, construct a frequent pattern tree and set a minimum support threshold;

[0024] Extract all subsequences whose support is not less than the minimum support threshold from the frequent pattern tree and store them in the normal action pattern library.

[0025] As a further aspect of the present invention, the real-time monitoring of whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, and determining whether the deviation belongs to the recorded slow drift, specifically includes:

[0026] When a newly acquired continuous physical parameter value is reached, the current mean and current tolerance range of the continuous physical parameter in the dynamic sliding window baseline model are searched, and it is determined whether the continuous physical parameter value exceeds the upper and lower boundaries of the current tolerance range.

[0027] If the value exceeds the upper or lower boundary of the current tolerance range, it is determined whether the continuous physical parameter value is within the next prediction range predicted by the recorded slow drift.

[0028] If the continuous physical parameter value is within the prediction range, the slowly drifting record is updated without generating any event; if the continuous physical parameter value is not within the prediction range, the current timestamp and the name of the continuous physical parameter are recorded as a baseline deviation event.

[0029] As a further aspect of the present invention, the real-time monitoring of whether rare action subsequences not recorded in the normal action pattern library appear in the newly acquired discrete action command sequence specifically includes:

[0030] The opcodes in the newly acquired discrete action instruction sequence are read one by one into a queue of length M in chronological order. The queue is initially empty. Each time a new opcode is read, the new opcode is added to the tail of the queue and the old opcode at the head of the queue is removed, and the queue is updated.

[0031] After each queue update, the opcode sequence currently contained in the queue is compared with all the patterns in the normal action pattern library one by one for complete matching.

[0032] If no match is found, the opcode subsequence is marked as a rare action subsequence, and the timestamp of the first opcode in the rare action subsequence is recorded as the occurrence time of the rare mode event. The opcode order of the rare action subsequence is also recorded.

[0033] As a further aspect of the present invention, the step of cross-validating the baseline deviation event with the rare pattern event on the time axis specifically includes:

[0034] Using the timestamp recorded in the baseline deviation event as the time center, extend forward and backward by a fixed duration T to form the first event time window;

[0035] Using the timestamps recorded in the rare pattern events as the time center, the fixed duration T is extended forward and backward to form a second event time window;

[0036] Determine whether there is a time intersection between the first event time window and the second event time window, and simultaneously determine whether the baseline deviation event and the rare pattern event originate from the same data acquisition channel of an industrial robot. If both conditions are met, then the time intersection interval is marked as the suspected contamination period.

[0037] Retrieve the sender identifier record of the first abnormal data within the suspected contamination period, and use the sender identifier as the data injection source.

[0038] As a further aspect of the present invention, the step of excluding the multi-dimensional operational data during the suspected contamination period from the update queue of the dynamic sliding window baseline model, and simultaneously storing the rare action subsequences appearing during the suspected contamination period into a blacklist pattern library, specifically includes:

[0039] Based on the suspected contamination periods marked after cross-validation, delete all continuous physical parameter values ​​and all discrete action command values ​​corresponding to the suspected contamination periods from the collected multi-dimensional operational data.

[0040] The rare action subsequences that appear during the suspected pollution period are stored in the blacklist pattern library in the form of hash values. When monitoring newly collected discrete action command sequences in real time, the hash value of the current opcode subsequence is compared with the hash value in the blacklist pattern library first. If the comparison is consistent, the step of matching the normal action pattern library is skipped and rare pattern events are generated directly.

[0041] Another objective of this invention is to provide a data security traceability system based on artificial intelligence, the system comprising:

[0042] The data acquisition module is used to collect multi-dimensional operational data and workshop production status labels generated by industrial robots during continuous production. The multi-dimensional operational data includes a continuous physical parameter sequence and a discrete action command sequence.

[0043] The dynamic baseline model building module is used to build a dynamic sliding window baseline model for each industrial robot based on the continuous physical parameter sequence. The dynamic sliding window baseline model maintains the mean and tolerance range of each continuous physical parameter within the sliding window and tracks the slow drift of each continuous physical parameter over time.

[0044] The motion pattern library construction module is used to construct a normal motion pattern library for each industrial robot based on the discrete motion instruction sequence. The normal motion pattern library records frequently occurring motion sub-sequences.

[0045] The baseline deviation monitoring module is used to monitor in real time whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, and to determine whether the deviation belongs to the recorded slow drift. If it does not belong to the slow drift, a baseline deviation event is generated.

[0046] The rare pattern monitoring module is used to monitor in real time whether rare action subsequences not recorded in the normal action pattern library appear in newly acquired discrete action instruction sequences. If they appear, a rare pattern event is generated.

[0047] The event cross-validation module is used to cross-validate the baseline deviation event and the rare pattern event on the time axis. If the occurrence time windows of the two overlap and point to the same industrial robot, the time window is marked as a suspected contamination period, and the data injection source when the suspected contamination period first appears is traced back.

[0048] The data cleaning and blacklist update module is used to exclude multi-dimensional operational data during the suspected contamination period from the update queue of the dynamic sliding window baseline model based on the data contamination results confirmed by cross-validation, and at the same time store rare action subsequences that appear during the suspected contamination period into the blacklist pattern library.

[0049] The beneficial effects of this invention are:

[0050] This invention achieves adaptive response to natural state changes of equipment by collecting continuous physical parameter sequences and discrete action command sequences, constructing a dynamic sliding window baseline model capable of tracking slow parameter drift, and a normal pattern library of statistically high-frequency action subsequences. Based on this, it monitors baseline deviation events and rare pattern events in real time and cross-validates these two types of events on the timeline. Data contamination is only determined when physical parameter anomalies and control command anomalies overlap within the time window and point to the same robot, significantly reducing the interference of single-dimensional false alarms and improving the accuracy of attack detection. Simultaneously, by tracing the sender identifier of the first abnormal data within a suspected contamination period, the source of data injection can be accurately located, and contaminated data can be excluded from the baseline update queue to prevent model poisoning. Confirmed rare malicious subsequences are also stored in a blacklist to accelerate subsequent responses. This invention enables data security traceability in continuous industrial production environments to possess adaptive, multimodal collaborative, and proactive defense capabilities, effectively ensuring the stability and reliability of the production process. Attached Figure Description

[0051] Figure 1 A flowchart illustrating an artificial intelligence-based data security tracing method provided in this embodiment of the invention;

[0052] Figure 2 A flowchart for collecting multi-dimensional operational data and workshop production status labels generated by industrial robots during continuous production, as provided in an embodiment of the present invention;

[0053] Figure 3 This is a flowchart for establishing a dynamic sliding window baseline model provided in an embodiment of the present invention;

[0054] Figure 4 A flowchart for constructing a normal action mode library provided in an embodiment of the present invention;

[0055] Figure 5 A flowchart for real-time monitoring of whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, provided for embodiments of the present invention;

[0056] Figure 6 A flowchart for real-time monitoring of whether rare action subsequences not recorded in the normal action pattern library appear in newly acquired discrete action instruction sequences provided in an embodiment of the present invention;

[0057] Figure 7 A flowchart for cross-validating the baseline deviation event and the rare pattern event on a time axis, provided as an embodiment of the present invention;

[0058] Figure 8A flowchart for excluding multi-dimensional operational data during the suspected pollution period from the update queue of the dynamic sliding window baseline model, provided for embodiments of the present invention;

[0059] Figure 9 This is a structural block diagram of a data security traceability system based on artificial intelligence, provided as an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] Figure 1 A flowchart of a data security tracing method based on artificial intelligence provided in an embodiment of the present invention is shown below. Figure 1 As shown, the method includes:

[0062] S100, collects multi-dimensional operational data and workshop production status labels generated by industrial robots during continuous production. The multi-dimensional operational data includes a continuous physical parameter sequence and a discrete action instruction sequence.

[0063] Multi-dimensional operational data includes continuous physical parameter sequences and discrete motion instruction sequences. These two data types reflect the robot's actual physical state and the behavioral intent at the control level, respectively. Continuous physical parameter sequences include values ​​such as torque, speed, and temperature that change continuously over time, reflecting the robot's physical characteristics during task execution, such as load, energy consumption, and thermal balance. Discrete motion instruction sequences record the opcodes and time sequence of each motion control instruction sent by the controller to the robot, directly reflecting the logical flow and operational steps of the production task. Simultaneously, workshop production status tags are collected, including current production order type, shift handover identifiers, and equipment maintenance mode identifiers. These tags provide contextual information about the operating environment for subsequent baseline models and pattern libraries. This collaborative collection of multi-source heterogeneous data allows subsequent detection not only to discover anomalies in a single dimension but also to determine whether data has been maliciously tampered with through the spatiotemporal correlation between continuous parameters and discrete instructions. If only continuous physical parameters are collected, attackers can bypass static threshold detection by slowly injecting false parameters; if only discrete instructions are collected, attackers can disguise a normal instruction sequence while actually modifying the execution parameters. By aligning the two types of data and collecting them on the same timeline, and attaching production status labels, we can effectively support subsequent cross-validation mechanisms and ensure the reliability of anomaly detection and the accuracy of tracing.

[0064] S200, for each industrial robot, a dynamic sliding window baseline model is established based on the continuous physical parameter sequence. The dynamic sliding window baseline model maintains the mean and tolerance range of each continuous physical parameter within the sliding window and tracks the slow drift of each continuous physical parameter over time.

[0065] The torque, speed, and temperature parameters of industrial robots are not constant. Normal factors such as mechanical wear, lubricant performance degradation, and ambient temperature fluctuations can cause these parameters to exhibit a slow, unidirectional drift trend. If a fixed threshold or static baseline is used, the system will generate numerous false alarms in the early stages of operation, and will miss genuine anomalies in the later stages. The dynamic sliding window baseline model maintains the mean and tolerance range of each parameter within the most recent time window, allowing the baseline to automatically update as the equipment's normal state changes. It also specifically tracks slow drift trends, distinguishes between normal aging and sudden anomalies, ensuring that only deviations exceeding the predicted range are marked as baseline deviation events. This design significantly reduces sensitivity to changes in the equipment's natural state, avoids frequent alarms caused by normal drift, and maintains rapid response capabilities to sudden anomalies such as data injection attacks. The model's adaptive update mechanism requires no manual intervention and can adapt to robots under different operating conditions and at different aging stages, improving the robustness and availability of the entire data security traceability system in real-world production environments.

[0066] S300: For each industrial robot, a normal motion pattern library is constructed based on the discrete motion instruction sequence. The normal motion pattern library records frequently occurring motion sub-sequences.

[0067] While industrial robots generate a vast number of control commands, the number of repeatedly executed action subsequences in actual production is limited, such as the operation code combinations corresponding to standard processes like grasping, placing, and welding. By segmenting the command sequence into fixed-length sliding subsequence windows and extracting subsequences with support above a threshold using a frequent pattern tree, the system can automatically learn the statistical characteristics of normal action patterns. This frequency-based statistical method eliminates the need for manual labeling of normal behavior, relying entirely on production data and adapting to differences in action habits across different product types, shifts, and maintenance modes. The normal action pattern library provides a matching benchmark for subsequent rare pattern monitoring; any action subsequence not recorded in the library is considered a potential anomaly. When used in conjunction with a physical parameter baseline model, rare patterns at the discrete command level can cross-validate with baseline deviations at the continuous parameter level, significantly enhancing the detection capability against data injection attacks. Even if an attacker can forge normal physical parameters, it is difficult to simultaneously mimic high-frequency action subsequences that perfectly match the current production state, thereby improving the accuracy of traceability and resistance to deception.

[0068] S400: Real-time monitoring of whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, and determination of whether the deviation belongs to the recorded slow drift; if not, a baseline deviation event is generated.

[0069] The newly acquired parameter values ​​are first compared with the tolerance range in the current sliding window baseline model. If they do not exceed the range, they are considered normal, and the baseline continues to be updated. If they exceed the tolerance range, it is further determined whether the deviation falls within the next prediction range predicted by the recorded slow drift. Slow drift reflects long-term normal trends such as mechanical wear and temperature changes, and its direction and magnitude are continuous and predictable. When the deviation value conforms to this predicted trend, the system determines that the equipment is aging normally, does not generate any events, and only updates the drift record, so that the baseline model can smoothly evolve with the equipment state. Only when the deviation value exceeds the tolerance range and is inconsistent with the prediction range of slow drift is a baseline deviation event generated, recording the timestamp and parameter name. This two-level judgment mechanism significantly reduces the false alarm rate, ensuring that only anomalies caused by data pollution or sudden failures will trigger alarms. At the same time, continuous updating of slow drift records allows the model to adapt to the state changes of the equipment throughout its entire life cycle, without the need for manual threshold reset, improving the reliability of the system in long-term unattended production environments. The baseline deviation event, as one of the inputs for subsequent cross-validation, is used together with rare pattern events at the discrete instruction level to determine suspicious pollution periods.

[0070] S500: Real-time monitoring of whether rare action subsequences not recorded in the normal action pattern library appear in newly acquired discrete action instruction sequences; if so, a rare pattern event is generated.

[0071] By maintaining a fixed-length first-in-first-out (FIFO) queue, the system can capture the latest received continuous opcode subsequences in real time using a sliding window. After each queue update, the current subsequence is compared against all patterns in the normal action pattern library. If a match fails, it indicates that the subsequence rarely or never occurred during normal production, violating normal action execution logic. Such rare action subsequences are likely illegal instructions injected by attackers or abnormal call sequences. The system immediately records the timestamp of the first opcode in the subsequence as the occurrence time of the rare pattern event and saves the specific opcode order. Using a queue-based sliding matching method avoids storing all historical instruction sequences, has a fixed memory footprint, and meets the real-time requirements of industrial production. Prioritizing comparison with the normal pattern library ensures that only truly rare patterns are marked, reducing the false positive rate. Rare pattern events and baseline deviation events together form a dual chain of evidence for cross-validation. False positives in one dimension can be filtered out by the normal state in the other dimension. Only when both dimensions are abnormal simultaneously is the judgment of a suspicious contamination period triggered, thus significantly improving the accuracy and anti-interference capability of data security tracing.

[0072] S600, cross-validate the baseline deviation event and the rare pattern event on the time axis. If the occurrence time windows of the two overlap and point to the same industrial robot, mark the time window as a suspected contamination period and trace back to the data injection source when the suspected contamination period first appeared.

[0073] By cross-verifying continuous physical parameters and discrete action commands, the authenticity of data contamination is confirmed and the source of contamination is located. Relying solely on physical parameters or command sequences can lead to false alarms. For example, equipment malfunctions may cause parameter deviations while commands remain normal, or rare commands may appear during the debugging phase without abnormal parameters. By extending a fixed duration around the timestamp of each event to form a time window, minor delays in physical response after malicious injection can be tolerated, while filtering out isolated events irrelevant in time. When the time windows of two events overlap and originate from the same industrial robot's data acquisition channel, it indicates that both physical and control behaviors are abnormal within the same time period, ruling out single-dimensional sporadic factors, thus determining with high confidence that data contamination exists in that time period. After marking the time intersection interval as a suspected contamination period, the sender identifier record of the first abnormal data within that period is further retrieved and identified as the data injection source. This reverse tracing mechanism directly locates the starting point of the attack, providing a clear target for subsequent security response and forensics. Cross-validation avoids the interference of a large number of false alarms to operation and maintenance personnel, ensuring that only security events that truly need to be handled are marked. At the same time, it achieves accurate matching of asynchronous events through overlapping time windows, improving the practicality and response efficiency of the tracing system.

[0074] S700, based on the data contamination results confirmed by cross-validation, exclude the multi-dimensional operational data during the suspected contamination period from the update queue of the dynamic sliding window baseline model, and store the rare action subsequences that appear during the suspected contamination period into the blacklist pattern library.

[0075] The purpose of excluding contaminated data is to prevent attack data from poisoning the baseline model and to avoid subsequent detection failures due to baseline contamination. If malicious data is included, the sliding window mean will shift towards outliers, the tolerance range will be unreasonably widened, and real attacks will fail to be detected. By deleting all continuous physical parameter values ​​and discrete action command values ​​within suspicious time periods, the baseline model is ensured to be updated only based on clean, normal data, maintaining the reliability of the detection benchmark. Rare action subsequences appearing within suspicious contaminated time periods are stored in a blacklist pattern library in the form of hash values. In subsequent monitoring, the blacklist is prioritized for comparison. If a match is found, the matching in the normal pattern library is skipped, and rare pattern events are generated directly. The blacklist mechanism enables the system to instantly identify confirmed attack patterns without repeatedly performing frequent pattern tree matching, reducing computational overhead, and preventing the re-activation of attacks from the same source. This closed-loop design of data cleaning and blacklist updates enables the tracing system to have self-purification and adversarial memory capabilities. Each time a contaminated attack is detected, the defense library is simultaneously strengthened, realizing the evolution from passive detection to active defense, significantly improving the long-term security of the system in continuous production environments.

[0076] like Figure 2 As shown, the collection of multi-dimensional operational data and workshop production status tags generated by the industrial robot during continuous production specifically includes:

[0077] S110, acquires a continuous sequence of physical parameters of an industrial robot at a fixed sampling frequency, the continuous sequence of physical parameters including torque values, speed values ​​and temperature values;

[0078] S120 monitors the robot controller's instruction bus, intercepts the operation code and corresponding timestamp of each motion control instruction, and arranges them in chronological order to form a discrete motion instruction sequence.

[0079] S130 reads the current production order type, the current shift handover identifier, and the current equipment maintenance mode identifier, and merges and packages the three pieces of information to generate a workshop production status label.

[0080] like Figure 3 As shown, establishing the dynamic sliding window baseline model specifically includes:

[0081] S210, create a sliding window of length L independently for each continuous physical parameter of each industrial robot, and calculate the arithmetic mean of all sampled values ​​within the sliding window as the mean;

[0082] ;

[0083] in:

[0084] For the first At each sampling moment, the arithmetic mean of a certain continuous physical parameter of an industrial robot within the current sliding window;

[0085] This is the fixed length of the sliding window, i.e., the number of sampling points;

[0086] For the first The actual sampled value of the continuous physical parameter at each sampling time;

[0087] This is the time index of the latest sampling time.

[0088] S220, calculate the sample standard deviation of all sampled values ​​within the sliding window, and use the mean plus or minus k times the sample standard deviation as the tolerance interval;

[0089] ;

[0090] in:

[0091] For the first At each sampling moment, the sample standard deviation of a certain continuous physical parameter of an industrial robot within the current sliding window;

[0092] For the first The arithmetic mean of the continuous physical parameter within the current sliding window at each sampling time.

[0093] S230, every sampling interval, the sliding window is moved forward by one sampling point, the mean and tolerance interval are recalculated, and the trend of the mean changing in the same direction and the single change amount being less than 0.1 times the initial sample standard deviation in three consecutive updates is recorded as slow drift.

[0094] During long-term operation, industrial robots experience slow, unidirectional, and small-amplitude changes in their physical parameters (such as motor temperature and joint torque) due to normal factors like mechanical wear, lubricant consumption, and gradual changes in ambient temperature. If this normal drift is not recognized, the baseline model may misinterpret it as an abnormal deviation, resulting in numerous false alarms. By separately recording these slow drifts and allowing them to update the baseline, the false alarm rate of the system can be significantly reduced while maintaining sensitivity to sudden, large-amplitude anomalies.

[0095] The tolerance range of continuous physical parameters is expressed as:

[0096] ;

[0097] in:

[0098] For the first The lower boundary of the tolerance range of the continuous physical parameter at each sampling time;

[0099] For the first The upper boundary of the tolerance range of the continuous physical parameter at each sampling time;

[0100] The confidence coefficient is a pre-set value, usually 2 or 3, corresponding to a confidence level of 95% or 99.7%.

[0101] The criteria for determining slow drift are:

[0102] ;

[0103] in:

[0104] This is a sign function; it returns 1 when the input is positive, -1 when the input is negative, and 0 when the input is zero.

[0105] , , These are the sliding window averages for the first 1, 2, and 3 sampling times, respectively.

[0106] This refers to the initial sample standard deviation of this continuous physical parameter calculated during the system initialization phase.

[0107] When all of the above conditions are met simultaneously, it is determined that the continuous physical parameter has a slow drift trend.

[0108] like Figure 4 As shown, the construction of the normal action mode library specifically includes:

[0109] S310, the collected discrete action command sequence is divided into several sliding subsequence windows of length N according to the time sequence, and each sliding subsequence window contains N consecutive opcodes and their order;

[0110] S320: Construct a frequent pattern tree for each sliding subsequence window and set a minimum support threshold;

[0111] S330, extract all subsequences with support not less than the minimum support threshold from the frequent pattern tree and store them in the normal action pattern library.

[0112] ;

[0113] in:

[0114] For action subsequences Support level;

[0115] For all lengths of In the sliding subsequence window, action subsequence Total number of occurrences;

[0116] For all lengths segmented from a discrete action instruction sequence The total number of sliding subsequence windows;

[0117] For any one of continuous A sequence of action subsequences consisting of opcodes.

[0118] like Figure 5 As shown, the real-time monitoring of whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, and determining whether the deviation belongs to the recorded slow drift, specifically includes:

[0119] S410, when a newly acquired continuous physical parameter value arrives, find the current mean and current tolerance range of the continuous physical parameter in the dynamic sliding window baseline model, and determine whether the continuous physical parameter value exceeds the upper and lower boundaries of the current tolerance range;

[0120] S420, if the value exceeds the upper or lower boundary of the current tolerance range, determine whether the continuous physical parameter value is within the next prediction range predicted by the recorded slow drift.

[0121] The next prediction range for slow drift is represented as:

[0122]

[0123] in:

[0124] This represents the lower boundary of the slow drift prediction range;

[0125] This represents the upper boundary of the slow drift prediction range;

[0126] It is the arithmetic mean of the three consecutive mean changes in a recorded slow drift trend.

[0127] S430, if the continuous physical parameter value is within the prediction range, then update the slowly drifting record without generating any event; if the continuous physical parameter value is not within the prediction range, then record the current timestamp and the name of the continuous physical parameter as a baseline deviation event.

[0128] No events are generated when the newly acquired parameter values ​​are within the slowly drifting prediction range. The purpose is to allow the baseline model to adapt to normal changes in device status and avoid false alarms.

[0129] Slow drift is a predictable trend caused by normal equipment aging or environmental changes, and its direction and magnitude are known. If this is still classified as an anomaly and generated as an event, it will lead to a large number of meaningless alarms, interfering with maintenance personnel's assessment of the real security threats. Meanwhile, updating slow drift records allows the baseline model to continuously track changes in the equipment's normal state, maintaining the accuracy and timeliness of the detection model.

[0130] like Figure 6 As shown, the real-time monitoring of whether rare action subsequences not recorded in the normal action pattern library appear in the newly acquired discrete action command sequence specifically includes:

[0131] S510: The opcodes in the newly acquired discrete action instruction sequence are read one by one into a queue of length M in chronological order. The queue is initially empty. Each time a new opcode is read, the new opcode is added to the tail of the queue and the old opcode at the head of the queue is removed, and the queue is updated.

[0132] A queue of fixed length M is used to process discrete action instruction sequences. The core purpose is to achieve real-time subsequence matching using a sliding window approach.

[0133] The motion commands of industrial robots are continuous streams of data, which cannot all be loaded into memory for processing at once. By maintaining a first-in, first-out (FIFO) queue of length M, a new action subsequence of length M can be automatically generated each time a new opcode is received. This ensures the temporal order and continuity of the subsequence, and the memory usage is fixed and does not increase with the running time.

[0134] S520, after each queue update, the current sequence of opcodes in the queue is compared with all the patterns in the normal action mode library for a complete match.

[0135] S530, if no match is found, the opcode subsequence is marked as a rare action subsequence, and the timestamp of the first opcode in the rare action subsequence is recorded as the occurrence time of the rare mode event. At the same time, the opcode order of the rare action subsequence is recorded.

[0136] like Figure 7 As shown, the step of cross-validating the baseline deviation event with the rare pattern event on the timeline specifically includes:

[0137] S610, using the timestamp recorded in the baseline deviation event as the time center, extend forward and backward by a fixed duration T to form the first event time window;

[0138] S620, using the timestamp recorded in the rare pattern event as the time center, extend the fixed duration T forward and backward to form a second event time window;

[0139] A time window is formed by extending a fixed duration T around the event timestamp before and after it. The purpose is to solve the problem of time asynchrony between baseline deviation events and rare pattern events.

[0140] In real-world data contamination attacks, the injection of malicious commands and the abnormal changes in physical parameters often do not occur simultaneously: the malicious command is executed first, and then after a brief delay, the physical parameters deviate from the baseline. By extending the time window, these two correlated events with slight temporal discrepancies can be correctly matched, improving the accuracy of cross-validation and reducing false negatives.

[0141] At the same time, setting a fixed duration T can also filter out irrelevant events that are too far apart in time, avoiding false matching.

[0142] S630, determine whether the first event time window and the second event time window have a time intersection, and at the same time determine whether the baseline deviation event and the rare pattern event come from the same data acquisition channel of the industrial robot. If both conditions are met, mark the time intersection interval as the suspected pollution period.

[0143] S640, retrieve the sender identifier record of the first abnormal data within the suspected contamination period, and use the sender identifier as the data injection source.

[0144] like Figure 8As shown, the step of excluding multi-dimensional operational data during the suspected contamination period from the update queue of the dynamic sliding window baseline model, and simultaneously storing rare action subsequences occurring during the suspected contamination period into a blacklist pattern library, specifically includes:

[0145] S710, based on the suspected contamination period marked after cross-validation, delete all continuous physical parameter values ​​and all discrete action command values ​​corresponding to the suspected contamination period from the collected multi-dimensional operation data;

[0146] S720, the rare action subsequence that appears during the suspected pollution period is stored in the blacklist pattern library in the form of a hash value. When monitoring newly collected discrete action command sequences in real time, the hash value of the current opcode subsequence is compared with the hash value in the blacklist pattern library first. If the comparison is consistent, the step of matching the normal action pattern library is skipped and a rare pattern event is generated directly.

[0147] In subsequent monitoring, the blacklist is compared first, and if a match is found, an event is generated directly. The core purpose is to improve the detection speed of known threats and the overall performance of the system.

[0148] The blacklist pattern library stores malicious action sequences that have been cross-validated. If these sequences reappear, they inevitably constitute data contamination attacks. Prioritizing comparisons with the blacklist enables "zero-latency" detection of known threats, eliminating the need for time-consuming one-by-one comparisons with the normal pattern library. This significantly reduces the number of matches required by the normal pattern library, lowers system CPU and memory usage, and improves the system's processing capabilities in high-concurrency data acquisition scenarios, ensuring the real-time requirements of industrial production.

[0149] Figure 9 A structural block diagram of an artificial intelligence-based data security traceability system provided in an embodiment of the present invention is shown below. Figure 9 As shown, the system includes

[0150] The data acquisition module 100 is used to collect multi-dimensional operating data and workshop production status labels generated by industrial robots during continuous production. The multi-dimensional operating data includes a continuous physical parameter sequence and a discrete action command sequence.

[0151] The dynamic baseline model establishment module 200 is used to establish a dynamic sliding window baseline model for each industrial robot based on the continuous physical parameter sequence. The dynamic sliding window baseline model maintains the mean and tolerance range of each continuous physical parameter within the sliding window and tracks the slow drift of each continuous physical parameter over time.

[0152] The motion pattern library construction module 300 is used to construct a normal motion pattern library for each industrial robot based on the discrete motion instruction sequence. The normal motion pattern library records frequently occurring motion sub-sequences.

[0153] The baseline deviation monitoring module 400 is used to monitor in real time whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, and to determine whether the deviation belongs to the recorded slow drift. If it does not belong to the slow drift, a baseline deviation event is generated.

[0154] The rare pattern monitoring module 500 is used to monitor in real time whether a rare action subsequence not recorded in the normal action pattern library appears in the newly acquired discrete action instruction sequence. If it appears, a rare pattern event is generated.

[0155] The event cross-validation module 600 is used to cross-validate the baseline deviation event and the rare pattern event on the time axis. If the occurrence time windows of the two overlap and point to the same industrial robot, the time window is marked as a suspected contamination period, and the data injection source when the suspected contamination period first appears is traced back.

[0156] The data cleaning and blacklist update module 700 is used to exclude multi-dimensional operational data during the suspected pollution period from the update queue of the dynamic sliding window baseline model based on the data pollution results confirmed by cross-validation, and at the same time store rare action subsequences that appear during the suspected pollution period into the blacklist pattern library.

[0157] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0158] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0159] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data security traceability method based on artificial intelligence, characterized in that, The method includes: Collect multi-dimensional operational data and workshop production status labels generated by industrial robots during continuous production. The multi-dimensional operational data includes a continuous physical parameter sequence and a discrete action command sequence. For each industrial robot, a dynamic sliding window baseline model is established based on the continuous physical parameter sequence. The dynamic sliding window baseline model maintains the mean and tolerance range of each continuous physical parameter within the sliding window and tracks the slow drift of each continuous physical parameter over time. For each industrial robot, a normal motion pattern library is constructed based on the discrete motion instruction sequence, and the normal motion pattern library records frequently occurring motion sub-sequences; Real-time monitoring of whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, and determination of whether the deviation belongs to the recorded slow drift; if not, a baseline deviation event is generated. Real-time monitoring is conducted to check whether rare action subsequences not recorded in the normal action pattern library appear in newly acquired discrete action instruction sequences. If they do appear, a rare pattern event is generated. Cross-validate the baseline deviation event and the rare pattern event on the timeline. If the occurrence time windows of the two events overlap and point to the same industrial robot, then mark the time window as a suspected contamination period and trace back to the data injection source when the suspected contamination period first appeared. Based on the data contamination results confirmed by cross-validation, the multi-dimensional operational data during the suspected contamination period are excluded from the update queue of the dynamic sliding window baseline model, and the rare action subsequences that appear during the suspected contamination period are stored in the blacklist pattern library.

2. The method according to claim 1, characterized in that, The collection of multi-dimensional operational data and workshop production status tags generated by industrial robots during continuous production specifically includes: A continuous sequence of physical parameters of an industrial robot is collected at a fixed sampling frequency. The continuous sequence of physical parameters includes torque values, speed values, and temperature values. The robot controller's instruction bus is monitored, and the opcode and corresponding timestamp of each motion control instruction are intercepted and arranged in chronological order to form a discrete motion instruction sequence. Read the current production order type, the current shift handover identifier, and the current equipment maintenance mode identifier, and combine and package the three pieces of information to generate a workshop production status label.

3. The method according to claim 2, characterized in that, The establishment of the dynamic sliding window baseline model specifically includes: For each continuous physical parameter of each industrial robot, an independent sliding window of length L is created, and the arithmetic mean of all sampled values ​​within the sliding window is calculated as the mean. Calculate the sample standard deviation of all sampled values ​​within the sliding window, and use the mean plus or minus k times the sample standard deviation as the tolerance interval; Every sampling interval, the sliding window is moved forward by one sampling point, the mean and tolerance interval are recalculated, and the trend of mean change in the same direction and single change amount less than 0.1 times the initial sample standard deviation in three consecutive updates is recorded as slow drift.

4. The method according to claim 3, characterized in that, The construction of the normal action mode library specifically includes: The collected discrete action command sequence is divided into several sliding subsequence windows of length N according to the time sequence. Each sliding subsequence window contains N consecutive opcodes and their order. For each sliding subsequence window, construct a frequent pattern tree and set a minimum support threshold; Extract all subsequences whose support is not less than the minimum support threshold from the frequent pattern tree and store them in the normal action pattern library.

5. The method according to claim 4, characterized in that, The real-time monitoring of whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, and the determination of whether the deviation belongs to the recorded slow drift, specifically includes: When a newly acquired continuous physical parameter value is reached, the current mean and current tolerance range of the continuous physical parameter in the dynamic sliding window baseline model are searched, and it is determined whether the continuous physical parameter value exceeds the upper and lower boundaries of the current tolerance range. If the value exceeds the upper or lower boundary of the current tolerance range, it is determined whether the continuous physical parameter value is within the next prediction range predicted by the recorded slow drift. If the continuous physical parameter value is within the prediction range, the slowly drifting record is updated without generating any event; if the continuous physical parameter value is not within the prediction range, the current timestamp and the name of the continuous physical parameter are recorded as a baseline deviation event.

6. The method according to claim 5, characterized in that, The real-time monitoring of whether rare action subsequences not recorded in the normal action pattern library appear in newly acquired discrete action command sequences specifically includes: The opcodes in the newly acquired discrete action instruction sequence are read one by one into a queue of length M in chronological order. The queue is initially empty. Each time a new opcode is read, the new opcode is added to the tail of the queue and the old opcode at the head of the queue is removed, and the queue is updated. After each queue update, the opcode sequence currently contained in the queue is compared with all the patterns in the normal action pattern library one by one for complete matching. If no match is found, the opcode subsequence is marked as a rare action subsequence, and the timestamp of the first opcode in the rare action subsequence is recorded as the occurrence time of the rare mode event. The opcode order of the rare action subsequence is also recorded.

7. The method according to claim 6, characterized in that, The step of cross-validating the baseline deviation event with the rare pattern event on the time axis specifically includes: Using the timestamp recorded in the baseline deviation event as the time center, extend forward and backward by a fixed duration T to form the first event time window; Using the timestamps recorded in the rare pattern events as the time center, the fixed duration T is extended forward and backward to form a second event time window; Determine whether there is a time intersection between the first event time window and the second event time window, and simultaneously determine whether the baseline deviation event and the rare pattern event originate from the same data acquisition channel of an industrial robot. If both conditions are met, then the time intersection interval is marked as the suspected contamination period. Retrieve the sender identifier record of the first abnormal data within the suspected contamination period, and use the sender identifier as the data injection source.

8. The method according to claim 7, characterized in that, The step of excluding multi-dimensional operational data during the suspected contamination period from the update queue of the dynamic sliding window baseline model, and simultaneously storing rare action subsequences occurring during the suspected contamination period into a blacklist pattern library, specifically includes: Based on the suspected contamination periods marked after cross-validation, delete all continuous physical parameter values ​​and all discrete action command values ​​corresponding to the suspected contamination periods from the collected multi-dimensional operational data. The rare action subsequences that appear during the suspected pollution period are stored in the blacklist pattern library in the form of hash values. When monitoring newly collected discrete action command sequences in real time, the hash value of the current opcode subsequence is compared with the hash value in the blacklist pattern library first. If the comparison is consistent, the step of matching the normal action pattern library is skipped and rare pattern events are generated directly.

9. A data security traceability system based on artificial intelligence, characterized in that, The system includes The data acquisition module is used to collect multi-dimensional operational data and workshop production status labels generated by industrial robots during continuous production. The multi-dimensional operational data includes a continuous physical parameter sequence and a discrete action command sequence. The dynamic baseline model building module is used to build a dynamic sliding window baseline model for each industrial robot based on the continuous physical parameter sequence. The dynamic sliding window baseline model maintains the mean and tolerance range of each continuous physical parameter within the sliding window and tracks the slow drift of each continuous physical parameter over time. The motion pattern library construction module is used to construct a normal motion pattern library for each industrial robot based on the discrete motion instruction sequence. The normal motion pattern library records frequently occurring motion sub-sequences. The baseline deviation monitoring module is used to monitor in real time whether newly acquired continuous physical parameter values ​​deviate from the tolerance range in the dynamic sliding window baseline model, and to determine whether the deviation belongs to the recorded slow drift. If it does not belong to the slow drift, a baseline deviation event is generated. The rare pattern monitoring module is used to monitor in real time whether rare action subsequences not recorded in the normal action pattern library appear in newly acquired discrete action instruction sequences. If they appear, a rare pattern event is generated. The event cross-validation module is used to cross-validate the baseline deviation event and the rare pattern event on the time axis. If the occurrence time windows of the two overlap and point to the same industrial robot, the time window is marked as a suspected contamination period, and the data injection source when the suspected contamination period first appears is traced back. The data cleaning and blacklist update module is used to exclude multi-dimensional operational data during the suspected contamination period from the update queue of the dynamic sliding window baseline model based on the data contamination results confirmed by cross-validation, and at the same time store rare action subsequences that appear during the suspected contamination period into the blacklist pattern library.

Images