Computer-automated parameter tuning method and system for behavior pattern recognition

Abstract

The application relates to the technical field of machine learning, and discloses a computer-automated parameter adjustment method and system for behavior pattern recognition, which comprises the following steps: performing phase point detection on a verification set precision sequence to divide the sequence into three subsequences, namely an initial learning subsequence, a middle adjustment subsequence and a late stabilization subsequence. Then, the long-term trend of the middle adjustment subsequence is determined to obtain a trend identifier, and the extreme points and oscillation characteristics of the initial learning subsequence and the late stabilization subsequence are analyzed to generate stable or jitter feature codes. Next, the feature codes are encapsulated into a multi-dimensional performance image with the trend identifier as an index. Risk assessment is performed on an endogenous trajectory chain formed based on historical images to generate a discrimination instruction. Finally, the current parameters are subjected to directional disturbance or the historical optimal parameter combination is output according to the instruction, so that an adjustment result set is obtained. The application can improve the accuracy of computer-automated parameter adjustment for behavior pattern recognition.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a computer-automated parameter tuning method and system for behavioral pattern recognition. Background Technology

[0002] In the current parameter tuning process of behavior pattern recognition models, parameter adjustments often rely on expert experience or pre-set fixed rules. This approach is not only inefficient and unable to cope with complex and ever-changing training dynamics, but also fails to perform fine-grained stage division and real-time state analysis of the learning process. This results in a lack of targeted parameter tuning decisions, a long overall optimization cycle, and unstable results.

[0003] Furthermore, while some automated methods can achieve parameter tuning, they typically only focus on the final performance metrics or a simple convergence state, lacking continuous monitoring and deep feature mining throughout the entire training process, from early to late stages. This results in the system's inability to promptly identify risk trends during training and its difficulty in providing historically validated optimal parameter rollback schemes when making termination decisions, thus affecting the reliability, adaptability, and final performance of parameter tuning. Summary of the Invention

[0004] This invention provides a computer-automated parameter tuning method and system for behavior pattern recognition, the main purpose of which is to solve the problem of low accuracy in computer-automated parameter tuning of behavior pattern recognition.

[0005] To achieve the above objectives, the present invention provides a computer-automated parameter tuning method for behavior pattern recognition, comprising: Phase transition points are detected on the validation set accuracy sequence of the target process to obtain the initial learning sequence, the intermediate adjustment sequence, and the later stable sequence of the target process. The long-term trend component of the intermediate adjustment sequence is oriented to obtain the trend identifier of the target process; Based on the distribution of extreme points in the initial learning sequence and the oscillation characteristics of the later stable sequence, the stable convergence category and the high-frequency jitter category in the target process are identified, and the stable convergence category and the high-frequency jitter category are converted into corresponding stable feature codes and jitter feature codes. Using the trend identifier as a structural index, the stable feature encoding and jitter feature encoding are encapsulated in a targeted manner to obtain a multi-dimensional performance profile of the target process; A risk assessment is performed on the endogenous trajectory chain of the multi-dimensional performance profile to obtain the discrimination instructions for the target process; The state parameter stream of the target process is updated based on the logical branch of the discrimination instruction to obtain the parameter tuning result set of the target process.

[0006] In a preferred embodiment, the step of performing phase transition point detection on the validation set accuracy sequence of the target process to obtain the initial learning sequence, the intermediate adjustment sequence, and the later stable sequence of the target process includes: The first-order difference calculation is performed on the accuracy sequence of the validation set to obtain the difference sequence of the target process; The first position point in the difference sequence where the sign changes from positive to zero or negative is defined as the first phase transition point of the target process; Starting from the first phase transition point, the difference sequence is iteratively scanned, and the position point in the difference sequence where the sign changes from negative to zero or positive is defined as the second phase transition point of the target process; The validation set accuracy sequence is segmented based on the first phase transition point and the second phase transition point. The subsequence before the first phase transition point is used as the initial learning sequence of the target process, the subsequence between the first phase transition point and the second phase transition point is used as the intermediate adjustment sequence of the target process, and the subsequence after the second phase transition point is used as the later stable sequence of the target process.

[0007] In a preferred embodiment, the step of determining the direction of the long-term trend component of the intermediate-term adjustment sequence to obtain the trend identifier of the target process includes: The intermediate adjustment sequence is subjected to sliding smoothing to obtain the smoothed sequence of the target process; Connect the start and end points of the smooth sequence to obtain the baseline segment of the target process; Calculate the vertical distance from the data points on the smoothed sequence to the baseline segment, and calculate the algebraic sum of the vertical distances and the sign of the corresponding algebraic sum; If the algebraic sum is positive, it is determined to be an upward trend, and a first-type trend identifier for the target process is obtained; if the algebraic sum is negative, it is determined to be a downward trend, and a second-type trend identifier for the target process is obtained; if the algebraic sum is zero, it is determined to be a stationary trend, and a third-type trend identifier for the target process is obtained.

[0008] In a preferred embodiment, the step of identifying the stable convergence category and high-frequency jitter category in the target process based on the extreme point distribution of the initial learning sequence and the oscillation characteristics of the later stable sequence, and converting the stable convergence category and high-frequency jitter category into corresponding stable feature codes and jitter feature codes, includes: Local extrema are detected in the later stable sequence to obtain the local maxima and local minima of the target process; Calculate the amplitude difference between the local maxima and the local minima to obtain the oscillation amplitude sequence of the target process; Analyze the decay trend of the oscillation amplitude sequence: if the oscillation amplitude sequence shows a monotonically decreasing trend, it is determined to be the stable convergence category; if the oscillation amplitude sequence has no significant decay and the values ​​fluctuate violently, it is determined to be the high-frequency jitter category. Based on the initial learning sequence, the stable convergence category and the high-frequency jitter category are fused to generate the stable feature code or jitter feature code of the target process.

[0009] In a preferred embodiment, based on the initial learning sequence, the stable convergence category and the high-frequency jitter category are fused to generate a stable feature code or jitter feature code for the target process, including: Based on the oscillation amplitudes corresponding to the stable convergence category and the high-frequency jitter category, the first-class and second-class coding values ​​of the target process are determined; The first-class and second-class encoding values ​​are concatenated with the number of extreme points in the initial learning sequence to generate stable feature codes and jitter feature codes for the target process.

[0010] In a preferred embodiment, the step of using the trend identifier as a structural index to perform targeted encapsulation of the stable feature code and the jitter feature code to obtain a multi-dimensional performance profile of the target process includes: Create an association mapping table with the trend identifier as the key; The stable feature code and the jitter feature code are stored as data items in the value range associated with the trend identifier in the association mapping table; By attaching timestamp information to the association mapping table, a multi-dimensional performance profile of the target process is obtained.

[0011] In a preferred embodiment, the step of performing a risk assessment on the endogenous trajectory chain of the multi-dimensional performance profile to obtain the discrimination instruction for the target process includes: The risk assessment value of the target process is calculated based on the pattern symbol sequence and time series correlation strength in the endogenous trajectory chain. Linear interpolation is performed on the risk assessment value to obtain the value corresponding to the specified quantile in the target process, and the value is set as the dynamic discrimination threshold of the target process; The discrimination instruction for the target process is generated based on the comparison result between the risk assessment value and the dynamic discrimination threshold.

[0012] In a preferred embodiment, the formula for calculating the risk assessment value includes:

[0013] in, The risk assessment value is... It is a nonlinear saturation function. This is the scale scaling factor. For feature fusion function, For trend dynamics identifier vectors, For steady-state convergent encoding vectors, This is a high-frequency perturbation coding vector. It is a positive damping coefficient. This represents the time interval between key pattern symbols.

[0014] In a preferred embodiment, updating the state parameter stream of the target process based on the logical branch of the discrimination instruction to obtain the parameter tuning result set of the target process includes: If the discrimination instruction is a task continuation instruction, then based on the encoding features of the multi-dimensional performance profile, the current hyperparameter vector in the state parameter stream is subjected to directional perturbation to obtain the hyperparameters to be adjusted for the target process. If the discrimination instruction is a task termination instruction, then the optimal value record is retrieved from the historical multidimensional performance profile of the target process, and the hyperparameter combination associated with the optimal value record is extracted to obtain the parameter tuning result set of the target process.

[0015] To address the above problems, the present invention also provides a computer-automated parameter tuning system for behavior pattern recognition, the system comprising: The data analysis module performs phase transition point detection on the validation set accuracy sequence of the target process to obtain the initial learning sequence, the intermediate adjustment sequence, and the later stable sequence of the target process. The trend identifier module determines the direction of the long-term trend component of the intermediate adjustment sequence to obtain the trend identifier of the target process. The state judgment module identifies the stable convergence category and high-frequency jitter category in the target process based on the distribution of extreme points in the initial learning sequence and the oscillation characteristics of the later stable sequence, and converts the stable convergence category and high-frequency jitter category into corresponding stable feature codes and jitter feature codes. The multi-dimensional performance profiling module uses the trend identifier as a structural index to perform targeted encapsulation of the stable feature encoding and jitter feature encoding to obtain a multi-dimensional performance profile of the target process. The discrimination instruction module performs a risk assessment on the endogenous trajectory chain of the multi-dimensional performance profile to obtain the discrimination instruction for the target process; The parameter tuning result set module updates the state parameter stream of the target process based on the logical branch of the discrimination instruction to obtain the parameter tuning result set of the target process.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves real-time, refined monitoring and structured description of the parameter tuning process through fully automated phase transition point detection, trend discrimination, feature encoding generation and encapsulation, replacing the inefficient traditional method that relies on manual observation and stage division. The system can automatically identify the early, middle, and late stages of the learning process and accurately extract key features such as trends and stability at each stage, integrating them into a multi-dimensional performance profile. This significantly improves the automation level and processing efficiency of behavior pattern recognition and parameter adjustment decisions.

[0017] 2. Based on the constructed multi-dimensional performance profile sequence (endogenous trajectory chain) and dynamic risk assessment mechanism, the system can adaptively generate "continue" or "terminate" judgment commands, triggering intelligent directional perturbation or historical optimal parameter backtracking mechanisms respectively. This makes the entire parameter tuning process form a complete intelligent closed loop from perception, analysis, decision-making to execution, which not only significantly improves the accuracy and success rate of single parameter tuning, but also enables the system to continuously evolve through the accumulation of historical data, greatly improving the overall efficiency and reliability of computer-automated parameter tuning. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a computer-automated parameter tuning method for behavior pattern recognition provided in an embodiment of the present invention; Figure 2 This is a functional block diagram of a computer-automated parameter tuning system for behavior pattern recognition provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] This application provides a computer-automated parameter tuning method for behavior pattern recognition. The executing entity of this computer-automated parameter tuning method for behavior pattern recognition includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the computer-automated parameter tuning method for behavior pattern recognition can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0021] Reference Figure 1 The diagram shown is a flowchart illustrating a computer-automated parameter tuning method for behavior pattern recognition according to an embodiment of the present invention. In this embodiment, the computer-automated parameter tuning method for behavior pattern recognition includes: In this embodiment, when performing phase transition point detection on the validation set accuracy sequence of the target process to obtain the initial learning sequence, intermediate adjustment sequence, and late stable sequence of the target process, it is specifically used for: The first-order difference calculation is performed on the accuracy sequence of the validation set to obtain the difference sequence of the target process; The first position point in the difference sequence where the sign changes from positive to zero or negative is defined as the first phase transition point of the target process; Starting from the first phase transition point, the difference sequence is iteratively scanned, and the position point in the difference sequence where the sign changes from negative to zero or positive is defined as the second phase transition point of the target process; The validation set accuracy sequence is segmented based on the first phase transition point and the second phase transition point. The subsequence before the first phase transition point is used as the initial learning sequence of the target process, the subsequence between the first phase transition point and the second phase transition point is used as the intermediate adjustment sequence of the target process, and the subsequence after the second phase transition point is used as the later stable sequence of the target process.

[0022] Specifically, starting from the second data point in the sequence, the value of each subsequent data point is subtracted from the value of the preceding data point to calculate the difference between every two adjacent data points. These calculated differences are arranged in the order corresponding to the original data points, thus forming a new sequence, which is the difference sequence.

[0023] Specifically, each value in the difference sequence is sequentially traversed, and its sign attribute is determined. The condition for the sign to change from positive to zero or negative is that when traversing to a certain position, the sign of the value at that position in the difference sequence has changed relative to the sign of the value at its immediate preceding position, and the change conforms to a specific pattern.

[0024] Specifically, starting from the determined position of the first phase transition point in the difference sequence, the remaining data points in the difference sequence are traversed sequentially. During the traversal, the sign of each data point is determined, and the first point that satisfies the condition of the sign changing from negative to zero or positive is found.

[0025] Specifically, the validation set accuracy sequence is segmented based on the first phase transition point and the second phase transition point. The subsequence before the first phase transition point is used as the initial learning sequence of the target process, the subsequence between the first and second phase transition points is used as the intermediate adjustment sequence of the target process, and the subsequence after the second phase transition point is used as the later stable sequence of the target process. The segmentation operation is performed based on the corresponding positions of the first and second phase transition points in the original validation set accuracy sequence. Specifically, the continuous data extracted from the starting data point of the validation set accuracy sequence to the data point preceding the first phase transition point constitutes the initial learning sequence.

[0026] Furthermore, each value in the difference sequence represents the amount of change in the accuracy of the original validation set between two adjacent time points or iteration steps, with positive values ​​indicating an increase in accuracy, negative values ​​indicating a decrease in accuracy, and zero values ​​indicating no change in accuracy.

[0027] Furthermore, it is necessary to find the first position point that satisfies the following conditions: the difference value corresponding to this position point is zero or negative, and the difference value corresponding to the immediately preceding position point is positive. The index of this first found position point that satisfies the above sign change condition in the difference sequence is defined as the first phase transition point.

[0028] Furthermore, the criterion for the sign to change from negative to zero or positive is: the difference value corresponding to the currently examined position is zero or positive, while the difference value corresponding to the immediately preceding position is negative. When the scanning process first finds a position that meets this condition, the scanning stops, and the index of that position in the difference sequence is defined as the second phase transition point. This point marks the end of the performance degradation or adjustment phase of the target process, and it may enter a new state.

[0029] Furthermore, the continuous data segment extracted from the data point corresponding to the first phase transition point to the data point preceding the second phase transition point constitutes the intermediate adjustment sequence. Finally, the continuous data segment extracted from the data point corresponding to the second phase transition point to the last data point of the validation set precision sequence constitutes the late-stabilized sequence. Through this segmentation, the original continuous precision sequence is divided into three stage-specific subsequences with different variation characteristics.

[0030] In summary, by stripping away the absolute level of accuracy, the dynamic characteristics of each stage of the learning process are directly highlighted, providing a clear and well-organized data foundation for the subsequent automatic and accurate detection of performance inflection points. This avoids the subjectivity and instability that may be introduced by directly dividing the stages from the original accuracy sequence with large fluctuations.

[0031] In summary, this method can automatically and unambiguously identify the end of the rapid ascent phase in the early stages of the learning process, providing precise boundaries for subsequent segmentation into pure early learning sequences. This replaces coarse-grained methods that rely on experience or fixed iteration counts for segmentation, making the stage segmentation more closely reflect the actual dynamics of model training.

[0032] In summary, the system automatically captures signals indicating the end of the adjustment period and the beginning of the later stage, thus distinguishing the intermediate stage, which may involve repeated adjustments and significant fluctuations, from the subsequent relatively stable stage. This achieves precise definition of the lower bound of the intermediate adjustment sequence and the upper bound of the later stable sequence.

[0033] In summary, this approach achieves the structured deconstruction of complex learning processes, transforming continuous monitoring data into subsequences with different semantics and features. This structured division lays a solid foundation for subsequent feature extraction and encoding at different stages, serving as a key preprocessing step in the entire automated and refined parameter tuning process.

[0034] In this embodiment, when determining the direction of the long-term trend component of the intermediate-term adjustment sequence to obtain the trend identifier of the target process, it is specifically used for: The intermediate adjustment sequence is subjected to sliding smoothing to obtain the smoothed sequence of the target process; Connect the start and end points of the smooth sequence to obtain the baseline segment of the target process; Calculate the vertical distance from the data points on the smoothed sequence to the baseline segment, and calculate the algebraic sum of the vertical distances and the sign of the corresponding algebraic sum; If the algebraic sum is positive, it is determined to be an upward trend, and a first-type trend identifier for the target process is obtained; if the algebraic sum is negative, it is determined to be a downward trend, and a second-type trend identifier for the target process is obtained; if the algebraic sum is zero, it is determined to be a stationary trend, and a third-type trend identifier for the target process is obtained.

[0035] Specifically, a fixed-width processing window is defined, covering a series of consecutive data points in the sequence. Starting from the beginning of the sequence, the center of the window is aligned with the first data point, and the average value of all data points within the window is calculated. This average value is then used as the data value for the corresponding position in the new sequence. Subsequently, the window is slid forward by one data point, and the average value of all data points within the new window is calculated again, becoming the data value for the next position in the new sequence.

[0036] Specifically, the starting point of the smoothed sequence refers to the position of the first data point in the sequence in a two-dimensional coordinate system. The horizontal axis of this position represents its temporal or indexal order, and the vertical axis represents its value after smoothing. The ending point of the smoothed sequence refers to the position of the last data point in the sequence in a two-dimensional coordinate system.

[0037] Specifically, for any data point in the sequence, its specific position in the two-dimensional coordinate system is determined based on its horizontal and vertical coordinate values. The perpendicular distance from this point to the baseline segment is calculated. Geometrically, this means that the length between the point and the intersection of the perpendicular line with the baseline segment is the perpendicular distance.

[0038] Specifically, when the sign is positive, it indicates that most data points are above the baseline segment, meaning the overall trend of the actual smoothed sequence is more skewed upwards than a simple line connecting the start and end points. This is interpreted as an inherent upward trend, and the system generates a specific identifier or code based on this, called the first type of trend identifier. When the sign is negative, it indicates that most data points are below the baseline segment, and the overall trend is more skewed downwards, interpreted as a downward trend, and the system generates the second type of trend identifier.

[0039] Furthermore, this sliding and averaging operation is repeated until the processing window covers the last data point of the sequence. For data points at the ends of the sequence where the window center cannot be aligned, the average is calculated by reducing the window width to include only the existing data points. Finally, the averages obtained from this series of calculations are arranged in their original order to form a new sequence that eliminates short-term fluctuations, namely the smoothed sequence.

[0040] Furthermore, in a two-dimensional coordinate system, a straight line segment is drawn with the starting point as one end and the ending point as the other end. This line segment is defined as the baseline line segment, which has a definite slope and intercept in the coordinate system. The sign of its slope reflects the overall direction of numerical change from the starting point to the ending point.

[0041] Furthermore, if a data point is above the baseline segment (i.e., its ordinate value is greater than the ordinate value of the baseline segment at that point), the distance value is positive; if it is below the baseline segment, the distance value is negative; and if it falls exactly on the baseline segment, the distance value is zero. After calculating the signed vertical distance values ​​corresponding to all data points in the smoothed sequence, these distance values ​​are algebraically added together to obtain a total value, which is the algebraic sum of the vertical distances. This algebraic sum is itself a specific numerical value, and the sign of this value is the sign of the algebraic sum.

[0042] Furthermore, when the algebraic sum is zero, it indicates that the data points are distributed in a balanced manner above and below the baseline segment, and the overall trend is consistent with the trend expressed by the line connecting the starting point to the end point. This is interpreted as a stable trend, and the system generates the third type of trend identifier.

[0043] In summary, without altering the main trend of the sequence, a smooth curve is obtained that better reflects the potential long-term trend during the medium-term adjustment phase. This provides a clear and stable analytical object for subsequent macroeconomic trend identification, avoids direct interference from noisy data on trend judgment, and improves the robustness and reliability of trend analysis.

[0044] In summary, this baseline segment intuitively expresses the result of a uniform linear change if the adjustment process is the result of such a change. It establishes a clear geometric reference for quantifying the overall deviation between the actual nonlinear fluctuation process and this linear expectation, thus providing a stable and calculable comparison standard for trend judgment.

[0045] In summary, by summing the deviations of all points, a comprehensive assessment was made as to whether the actual sequence as a whole lies above or below the baseline. This quantitative method based on the algebraic sum of geometric distances can comprehensively and holistically capture the overall morphological tendency of the sequence, transforming abstract trend judgments into concrete mathematical calculations, thus achieving objectivity and automation in trend identification.

[0046] In summary, the complex sequence patterns are ultimately reduced to three clear semantic labels: rising, falling, or stable, and corresponding trend identifiers are generated. This identifier is a concise summary of the macroscopic performance evolution direction in the mid-term stage, providing a key basis for subsequent steps to use the trend as a structural index to organize other feature encodings, and realizing high-level semantic induction and indexing of the trajectories of different parameter tuning processes.

[0047] In this embodiment, the step of identifying the stable convergence category and high-frequency jitter category in the target process based on the extreme point distribution of the initial learning sequence and the oscillation characteristics of the later stable sequence, and converting the stable convergence category and high-frequency jitter category into corresponding stable feature codes and jitter feature codes, is specifically used for: Local extrema are detected in the later stable sequence to obtain the local maxima and local minima of the target process; Calculate the amplitude difference between the local maxima and the local minima to obtain the oscillation amplitude sequence of the target process; Analyze the decay trend of the oscillation amplitude sequence: if the oscillation amplitude sequence shows a monotonically decreasing trend, it is determined to be the stable convergence category; if the oscillation amplitude sequence has no significant decay and the values ​​fluctuate violently, it is determined to be the high-frequency jitter category. Based on the initial learning sequence, the stable convergence category and the high-frequency jitter category are fused to generate the stable feature code or jitter feature code of the target process.

[0048] Specifically, each data point in the sequence, except for the first and last points, is sequentially traversed, and its value is compared with the value of its immediately preceding and immediately following data points. If the value of a data point is greater than both the values ​​of its preceding and immediately following data points, then that data point is identified as a local maximum.

[0049] Specifically, according to the alternating order of the extreme points in the original sequence, an adjacent local maximum point is paired with a local minimum point.

[0050] Specifically, starting from the first amplitude value of the sequence, each subsequent amplitude value is less than or equal to its immediate preceding amplitude value. The values ​​of the entire sequence exhibit a state of continuous change or continuous decrease, with no subsequent value being greater than its preceding value. If this condition is met, the sequence trend is determined to be monotonically decreasing, and this stage characteristic of the target process is classified into the stable convergence category.

[0051] Specifically, the initial learning sequence obtained in the previous steps is used to count the number of extreme points. That is, using the same method as for detecting local extreme points in the later stable sequence, the total number of local maxima and local minima contained in the initial learning sequence is counted. Then, encoding is generated by combining the determined category information.

[0052] Furthermore, if the value of a data point is simultaneously less than the values ​​of both the preceding and following data points, then that data point is identified as a local minimum. After the traversal is complete, all identified local maxima are recorded and collected in the order they appear in the sequence, forming a set of local maxima; all identified local minima are recorded and collected in the order they appear in the sequence, forming a set of local minima.

[0053] Furthermore, for each pair of extreme points, the value of the local maximum point is taken, and the value of the local minimum point in the same pair is subtracted to calculate the absolute difference between the two. This difference is the amplitude of the oscillation at that point. All amplitude differences calculated in the order in which the extreme points appear are arranged sequentially to form the above oscillation amplitude sequence.

[0054] Furthermore, the numerical values ​​of the sequence did not show a continuous decreasing trend, and by calculating the mean and standard deviation of all amplitude values ​​in the sequence, it was found that the ratio of the standard deviation to the mean exceeded a pre-set fixed threshold, indicating that the amplitude values ​​fluctuated greatly around their mean. If this condition is met, it is determined to be the high-frequency jitter category. This classification result is a qualitative description of the dynamic characteristics of the later stable stage of the target process.

[0055] Further, if the target process is determined to belong to the stable convergence category, a stable feature code is generated. This code is a string consisting of characters and numbers, containing a category identifier prefix representing stable convergence and the total number of extreme points connected to the initial learning sequence. If the target process is determined to belong to the high-frequency jitter category, a jitter feature code is generated. This code is a string consisting of characters and numbers, containing a category identifier prefix representing "high-frequency jitter" and the total number of extreme points connected to the initial learning sequence. Finally, based on the actual determined category, a corresponding feature code is output, namely the stable feature code or the jitter feature code.

[0056] In summary, it meticulously captures all peaks and troughs within the sequence within a local range. This surpasses the crude approach of merely observing the overall mean or final value of the sequence, accurately quantifying any minute fluctuations or oscillation patterns present within the stable phase. By identifying these extreme points, it provides precise input pairs for subsequent calculations of oscillation amplitudes, thus ensuring that the assessment of stability is not a subjective impression but an objective measure based on the fluctuation pattern of the sequence itself.

[0057] In summary, it transforms spatial fluctuations into a new sequence that evolves over time, characterizing the intensity of the fluctuations. This oscillation amplitude sequence directly reflects the magnitude and changes in the model's performance swing during the stable phase. It is the core quantitative basis for judging whether the performance has truly converged or is merely fluctuating at a high level, making the characterization of the dynamics in the later stages more profound and accurate.

[0058] In summary, by using the stringent criterion of monotonically decreasing performance, it can reliably identify a healthy process where performance is continuously converging and tending towards true stability. Simultaneously, by judging whether fluctuations are severe, it can effectively filter out states that, although having high average values, are extremely unstable and pose a risk of oscillation. This dichotomy provides a clear decision-making basis for automated systems, enabling them to distinguish between two distinct situations: healthy convergence and dangerous jitter.

[0059] In summary, the generated encoding is not a snapshot of a single stage, but a composite feature descriptor spanning the early and later stages of the process. This allows the final feature encoding to simultaneously contain activity information from the early stages of learning and convergence quality information from the later stages, forming a more discriminative and informative process fingerprint. This provides rich and structured input for subsequent construction of multi-dimensional performance profiles and risk assessments.

[0060] In this embodiment, when generating the stable feature code or jitter feature code of the target process based on the initial learning sequence by fusing the stable convergence category and the high-frequency jitter category, it is specifically used for: Based on the oscillation amplitudes corresponding to the stable convergence category and the high-frequency jitter category, the first-class and second-class coding values ​​of the target process are determined; The first-class and second-class encoding values ​​are concatenated with the number of extreme points in the initial learning sequence to generate stable feature codes and jitter feature codes for the target process.

[0061] Specifically, the first value in the oscillation amplitude sequence is extracted, namely the amplitude difference at the beginning of the sequence. This value reflects the oscillation intensity at the initial moment of the later stable phase of the target process, and this value is directly used as the first type of encoded value.

[0062] Specifically, the number of extreme points in the initial learning sequence, which has been counted in the previous steps, is obtained. This number is an integer value obtained by detecting the total number of all local maxima and local minima in the initial learning sequence. Next, the encoded value is converted into a string, and this string is concatenated with the string representation of the number of extreme points in a predetermined order to form a new combined string. This new string is the stable feature code.

[0063] Furthermore, the last value in the oscillation amplitude sequence, i.e., the amplitude difference at the end of the sequence, is extracted. This value reflects the oscillation intensity at the end of the later stable phase of the target process, and is directly used as the second-class encoded value. In this way, the first-class encoded value and the second-class encoded value are defined as the beginning and end values ​​of the oscillation amplitude sequence, respectively, capturing the start and end states of the oscillation behavior in the time dimension in a simple numerical form.

[0064] Further, the two types of encoded values ​​are converted into string form, and this string is concatenated with the string representation of the same number of extreme points in the same predetermined order to form another new combined string. This new string is the jitter feature code. The predetermined order can be encoded value first, followed by the number of extreme points, or vice versa, as long as it remains consistent throughout the system. Finally, the system outputs two feature codes: a stable feature code and a jitter feature code. These encapsulate the indicators based on the initial state of the oscillation amplitude and the indicators based on the initial learning dynamics, as well as the indicators based on the final state of the oscillation amplitude and the indicators based on the initial learning dynamics, respectively.

[0065] In summary, a highly efficient and physically meaningful method is employed to quantify the initial and final states of oscillation behavior. By directly extracting these two feature points, one type of encoded value captures the initial oscillation intensity at the start of the later stable phase, while the other type captures the oscillation intensity at its end. Together, they constitute a concise yet crucial feature pair describing the temporal evolution of the oscillation amplitude. This avoids the computational overhead and information redundancy associated with introducing complex statistical indicators, preserving the core comparative information for assessing whether the oscillation has decayed with minimal data dimensionality.

[0066] In summary, the generated final feature codes not only include qualitative categories from later stages of the process but also incorporate the activity or complexity of the early learning process by integrating the number of extreme points from earlier stages. This makes each generated feature code a compressed signature of key characteristics throughout the early, middle, and late stages of the learning process, greatly enhancing the discriminative and representational capabilities of the codes and providing a solid and rich data foundation for subsequent accurate risk assessment and decision-making based on these codes.

[0067] In this embodiment, when the trend identifier is used as a structural index to perform targeted encapsulation of the stable feature code and the jitter feature code to obtain a multi-dimensional performance profile of the target process, it is specifically used for: Create an association mapping table with the trend identifier as the key; The stable feature code and the jitter feature code are stored as data items in the value range associated with the trend identifier in the association mapping table; By attaching timestamp information to the association mapping table, a multi-dimensional performance profile of the target process is obtained.

[0068] Specifically, an empty data structure supporting key-value pair storage is initialized in computer memory, allowing a unique key to be associated with and stored with a corresponding value. The trend identifier is then set as the key in this data structure.

[0069] Specifically, the entry with the trend identifier as the key in the created association mapping table is located. Then, the value range space allocated to that key is obtained. Next, the two data items, the stable feature code and the jitter feature code, are stored as a data set within this value range space.

[0070] Specifically, in addition to the association mapping table data structure that stores the feature codes, a new, independent data field or attribute is created. This field stores the current standard time obtained from the system clock, which is typically accurate to the second or millisecond.

[0071] Furthermore, a space is pre-allocated for storing the associated values ​​for this key; this space is called the value range of the key. At this point, an association mapping table with the trend identifier as the key and its value range empty is created. The structure of this table establishes a framework for organizing subsequent feature data with the trend direction as the core index.

[0072] Furthermore, in the association mapping table, the trend identifier key is associated with a set containing two feature codes, meaning that by querying the trend identifier, the stable feature code and jitter feature code bound to it can be directly obtained.

[0073] Furthermore, this timestamp field is encapsulated or bound to the previously existing association mapping table containing the trend identifier key and its associated feature code value range, so that they together constitute a single, composite data object. This final, complete data object, integrating the trend index, feature code set, and timestamp, constitutes the multi-dimensional performance profile of the target process.

[0074] In summary, scattered feature data was systematically categorized and stored according to the trend context in which it occurred. This structured organization allows subsequent systems to quickly and accurately retrieve all detailed state features related to a trend by querying the high-level semantic label of the trend, greatly improving the efficiency of data management and laying the data structure foundation for differentiated processing and analysis based on different trend scenarios.

[0075] In summary, this storage method fully associates each trend category with the specific state details that may occur under that trend, ensuring data integrity and contextual consistency. This avoids the problem of feature data being disconnected from the background of the process stage, enabling subsequent risk assessment and decision-making to be carried out in a unified and information-rich context.

[0076] In summary, the addition of timestamps allows each performance evaluation to be precisely positioned on the timeline. This not only supports real-time analysis of individual processes but also enables all historical profiles to be arranged in a time series, forming an endogenous trajectory chain. This provides the possibility for analyzing the evolution of the process over time, calculating the strength of temporal correlations, and conducting dynamic risk assessments. This represents a leap from static cross-sectional analysis to dynamic spatiotemporal analysis of the parameter tuning process.

[0077] In this embodiment, when performing risk assessment on the endogenous trajectory chain of the multi-dimensional performance profile to obtain the discrimination instruction for the target process, it is specifically used for: The risk assessment value of the target process is calculated based on the pattern symbol sequence and time series correlation strength in the endogenous trajectory chain. Linear interpolation is performed on the risk assessment value to obtain the value corresponding to the specified quantile in the target process, and the value is set as the dynamic discrimination threshold of the target process; The discrimination instruction for the target process is generated based on the comparison result between the risk assessment value and the dynamic discrimination threshold.

[0078] Specifically, each historical sequence is traversed, the number of occurrences of each pair of adjacent symbols is recorded, the adjacent symbol pairs that appear in all historical sequences are summarized, the total number of occurrences of each unique symbol pair is calculated and divided by the total number of occurrences of all symbol pairs to obtain the transition probability of each symbol pair. These probability values ​​constitute the temporal correlation strength table.

[0079] Specifically, all risk assessment values ​​calculated from historical multi-dimensional performance profiles are collected and sorted in ascending order to form an ordered sequence. A quantile is a predefined percentile, representing the percentage of data points in the ordered sequence that are below the corresponding numerical value. The position index of the quantile in the ordered sequence is calculated; the position index equals the quantile percentage multiplied by the total number of values ​​in the ordered sequence.

[0080] Specifically, if the risk assessment value is greater than or equal to the dynamic discrimination threshold, the system determines that the target process is currently in a low-risk state, which means that the process behavior pattern is consistent with the historical stable pattern. Therefore, a task continuation instruction is generated, which instructs the automated parameter tuning system to continue executing the current operation process.

[0081] Furthermore, when calculating the risk assessment value, the latest pattern symbol sequence of the current target process is extracted from the endogenous trajectory chain. For each pair of adjacent symbols in this latest sequence, the corresponding transition probability value is found in the temporal correlation strength table. All the probability values ​​found are added together to obtain the sum. Then, the sum is divided by the number of adjacent symbol pairs in the latest sequence to obtain the average probability value. This average probability value is the risk assessment value.

[0082] Furthermore, if the position index is an integer, the value at that index position in the ordered sequence is directly taken; if the position index is not an integer, the two values ​​corresponding to the integer part of the position index and the next integer part are taken, and the interpolation result is obtained by adding the fractional part to the value corresponding to the integer part and multiplying it by the difference between the two values. This value calculated through linear interpolation is set as the dynamic discrimination threshold.

[0083] Furthermore, if the risk assessment value is less than the dynamic discrimination threshold, the system determines that the target process is currently in a high-risk state. This means that the process's behavior pattern deviates from its historical norm, and therefore generates a task termination instruction. This instruction instructs the system to stop the current parameter tuning process and trigger a rollback mechanism, such as retrieving the best-performing multi-dimensional performance profile from historical records and extracting the associated hyperparameter combinations. The specific form of the discrimination instruction is a predefined code or string, and the system switches to the corresponding logical branch based on this instruction.

[0084] In summary, this approach not only considers the characteristics of the current state but also incorporates the temporal patterns implicit in historical behavior, thus giving risk assessment context-awareness and historical basis. This avoids the one-sidedness of judging risk based on static rules or a single current state, improves the accuracy and reliability of risk identification, and enables the system to more keenly detect abnormal behavior that deviates from historical normal patterns.

[0085] In summary, this dynamic threshold setting method fully considers the normal fluctuation range of process risks in different tasks or stages. It can raise the threshold to encourage continued exploration when process performance is generally good, and lower the threshold to avoid risks in advance when performance is average. This enhances the system's adaptability and robustness, reduces premature termination or ineffective continuation due to improper threshold settings, and optimizes resource allocation and parameter tuning efficiency.

[0086] In summary, when the risk assessment value indicates that the current process status is consistent with historical successful patterns, the system instructions continue to execute, promoting parameter tuning exploration; when the risk assessment value indicates that the process may have deviated from its course or fallen into an inefficient state, the system instructions terminate and promptly trigger a rollback mechanism. This decision-making mechanism based on explicit quantification and dynamic thresholds enables the entire parameter tuning process to have self-monitoring and self-regulating capabilities, significantly improving the efficiency and success rate of automated parameter tuning while reducing the need for manual intervention.

[0087] In this embodiment, the formula for calculating the risk assessment value is specifically used for:

[0088] in, The risk assessment value is... It is a nonlinear saturation function. This is the scale scaling factor. For feature fusion function, For trend dynamics identifier vectors, For steady-state convergent encoding vectors, This is a high-frequency perturbation coding vector. It is a positive damping coefficient. This represents the time interval between key pattern symbols.

[0089] Specifically, each component in the formula originates from specific data objects and preset constants generated during the target process analysis. The trend dynamics identifier vector originates from the first, second, or third type trend identifiers obtained after smoothing and direction determination of the mid-term adjustment sequence of the target process. These identifiers are converted into numerical form and arranged into a vector. The steady-state convergence encoding vector originates from the stable feature encoding generated when the later stable sequence is determined to be of the stable convergence category after oscillation analysis. This encoding string is parsed and converted into a numerical vector. The high-frequency disturbance encoding vector originates from the jitter feature encoding generated when the later stable sequence is determined to be of the high-frequency jitter category after oscillation analysis. This encoding string is also parsed and converted into a numerical vector. The feature fusion function is a fixed procedure that takes these three vectors as input and merges them into a new composite vector through concatenation or weighted summation. The Euclidean norm is calculated for this composite vector, measuring its overall length or strength. The scaling factor is a pre-defined fixed numerical constant used to amplify or reduce the norm calculation results. The time interval between key mode symbols is a specific time length value calculated from the timestamp difference between two consecutive key mode symbols in the endogenous trajectory chain of the target process. The positive damping coefficient is another pre-set fixed positive constant used to adjust the impact of the time interval on the overall calculation. The nonlinear saturation function is a mathematical function mapping that compresses the input values ​​to a fixed output range, such as between zero and one.

[0090] Furthermore, the significance of this calculation formula lies in quantifying the risk level of the target process in its current state. By fusing the trend dynamics identifier vector reflecting the trend direction, the steady-state convergence encoding vector reflecting stability, and the high-frequency disturbance encoding vector reflecting disturbances, the feature fusion function generates a feature vector representing the comprehensive state of the process. Calculating the Euclidean norm of this feature vector yields a scalar value representing the overall strength of the state feature. After adjusting this strength value with a scaling factor, it is scaled by dividing it by the square root of a damping factor related to the time interval. This damping factor consists of a positive damping coefficient multiplied by the time interval, and its function is to reduce the contribution of the feature strength as the time interval increases. Finally, a nonlinear saturation function processes the scaled result to ensure that the final output value is limited to a predefined, bounded numerical range, thereby obtaining a standardized and easily interpretable risk assessment value, which is directly used for subsequent discrimination instruction generation.

[0091] In summary, the formula reflects the trend of risk assessment values ​​changing with input features and time intervals. When the feature strength represented by the trend dynamics identifier vector, steady-state convergence encoding vector, or high-frequency perturbation encoding vector increases, the Euclidean norm of the feature fusion vector increases. Under the influence of the scaling factor, the numerator increases, leading to a tendency for the risk assessment value to rise. When the time interval between key mode symbols increases, the damping factor in the denominator of the calculation formula increases, causing the denominator to increase, thus leading to a tendency for the risk assessment value to decrease. The nonlinear saturation function stabilizes the output value near the upper or lower limit of the saturation range when the numerator is too large or too small, preventing the assessment value from growing indefinitely or becoming overly sensitive. Overall, the formula trend indicates that the risk assessment value is positively correlated with the intensity of abnormal features of the process state and negatively correlated with the time interval of key mode changes, and the saturation mechanism ensures the stability and robustness of the assessment results.

[0092] In this embodiment, when the logical branch based on the discrimination instruction updates the state parameter stream of the target process to obtain the parameter tuning result set of the target process, it is specifically used for: If the discrimination instruction is a task continuation instruction, then based on the encoding features of the multi-dimensional performance profile, the current hyperparameter vector in the state parameter stream is subjected to directional perturbation to obtain the hyperparameters to be adjusted for the target process. If the discrimination instruction is a task termination instruction, then the optimal value record is retrieved from the historical multidimensional performance profile of the target process, and the hyperparameter combination associated with the optimal value record is extracted to obtain the parameter tuning result set of the target process.

[0093] Specifically, the current hyperparameter vector is a set of adjustable parameters used by the target process during its current runtime, and these parameters are managed in vector form. The encoded features of the multi-dimensional performance profile specifically refer to the string information contained in the stable feature encoding or jitter feature encoding stored therein.

[0094] Specifically, the historical multidimensional performance profile is a collection of all multidimensional performance profiles generated and saved in all previous runs of the target process. The specific operation for retrieving the record with the highest value is as follows: traverse the entire collection of historical multidimensional performance profiles; for each profile, extract its timestamp information; and based on that timestamp, find the final validation set precision value reached by the target process at the end of that run. Compare all these final validation set precision values ​​and find the one with the largest value. The historical multidimensional performance profile corresponding to this largest precision value is then determined as the record with the highest value.

[0095] Further, the encoded feature string is parsed to extract the portion containing the number of extreme points in the initial learning sequence, which is then converted into an integer value. This integer value is then used as the base perturbation amplitude. Next, each parameter value of the current hyperparameter vector in the state parameter stream is read, and for each parameter value, a random perturbation value is generated. The sign of this perturbation value is randomly determined, but its absolute value is equal to the base perturbation amplitude multiplied by a fixed, extremely small coefficient to ensure that the perturbation is minute. The generated random perturbation value is added to the corresponding original parameter value to obtain a new parameter value. After performing this operation on all parameters, the vector composed of these new parameter values ​​is the hyperparameter to be adjusted after targeted perturbation. This perturbation method correlates the magnitude of parameter adjustment with the dynamic complexity of the early learning phase.

[0096] Furthermore, in the storage system or database, each historical multidimensional performance profile is persistently associated with the set of hyperparameters used during its generation. By querying this association, using the identified optimal record as an index or key, the complete set of hyperparameters used when generating the profile can be directly retrieved. Finally, this retrieved set of hyperparameters is output as the tuning result set, which represents the parameter configuration that achieved optimal validation performance during historical runs.

[0097] In summary, by analyzing the encoded features encapsulated in the multi-dimensional performance profile and quantifying this information as the basis for perturbation amplitude, a correlation between parameter adjustment and the historical dynamic characteristics of the process is achieved. This targeted perturbation allows the magnitude of small parameter changes to adaptively reflect the learning complexity or activity level in the early stages of the process, thereby guiding the search in a direction that better matches the current learning state. While continuing the exploration, this improves the targeting and potential efficiency of parameter updates, and helps to more intelligently escape local optima or fine-grained optimization.

[0098] In summary, by automatically retrieving the record with the highest validation accuracy from the accumulated historical profiles and extracting the complete hyperparameter combination associated with it as the final output, the system effectively performs an intelligent rollback and optimization. This fully utilizes the best experience gained from historical exploration, ensuring the quality and reproducibility of the hyperparameter tuning result set, while also promptly stopping potentially ineffective resource consumption, significantly improving the practicality and reliability of the entire automated process.

[0099] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves real-time, refined monitoring and structured description of the parameter tuning process through fully automated phase transition point detection, trend discrimination, feature encoding generation and encapsulation, replacing the inefficient traditional method that relies on manual observation and stage division. The system can automatically identify the early, middle, and late stages of the learning process and accurately extract key features such as trends and stability at each stage, integrating them into a multi-dimensional performance profile. This significantly improves the automation level and processing efficiency of behavior pattern recognition and parameter adjustment decisions.

[0100] 2. Based on the constructed multi-dimensional performance profile sequence (endogenous trajectory chain) and dynamic risk assessment mechanism, the system can adaptively generate "continue" or "terminate" judgment commands, triggering intelligent directional perturbation or historical optimal parameter backtracking mechanisms respectively. This makes the entire parameter tuning process form a complete intelligent closed loop from perception, analysis, decision-making to execution, which not only significantly improves the accuracy and success rate of single parameter tuning, but also enables the system to continuously evolve through the accumulation of historical data, greatly improving the overall efficiency and reliability of computer-automated parameter tuning.

[0101] like Figure 2 The diagram shown is a functional block diagram of a computer-automated parameter tuning system for behavior pattern recognition provided in an embodiment of the present invention.

[0102] The computer-automated parameter tuning system 100 for behavior pattern recognition described in this invention can be installed in an electronic device. Depending on the functions implemented, the computer-automated parameter tuning system 100 for behavior pattern recognition may include a data analysis module 101, a trend identifier module 102, a state judgment module 103, a multi-dimensional performance profiling module 104, a discrimination instruction module 105, and a parameter tuning result set module 106. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and which are stored in the memory of the electronic device.

[0103] In this embodiment, the functions of each module / unit are as follows: The data analysis module performs phase transition point detection on the validation set accuracy sequence of the target process to obtain the initial learning sequence, the intermediate adjustment sequence, and the later stable sequence of the target process. The trend identifier module determines the direction of the long-term trend component of the intermediate adjustment sequence to obtain the trend identifier of the target process. The state judgment module identifies the stable convergence category and high-frequency jitter category in the target process based on the distribution of extreme points in the initial learning sequence and the oscillation characteristics of the later stable sequence, and converts the stable convergence category and high-frequency jitter category into corresponding stable feature codes and jitter feature codes. The multi-dimensional performance profiling module uses the trend identifier as a structural index to perform targeted encapsulation of the stable feature encoding and jitter feature encoding to obtain a multi-dimensional performance profile of the target process. The discrimination instruction module performs a risk assessment on the endogenous trajectory chain of the multi-dimensional performance profile to obtain the discrimination instruction for the target process; The parameter tuning result set module updates the state parameter stream of the target process based on the logical branch of the discrimination instruction to obtain the parameter tuning result set of the target process.

[0104] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0105] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0106] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0107] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0108] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A computer-automated parameter tuning method for behavioral pattern recognition, characterized in that, The method includes: Phase transition points are detected on the validation set accuracy sequence of the target process to obtain the initial learning sequence, the intermediate adjustment sequence, and the later stable sequence of the target process. The long-term trend component of the intermediate adjustment sequence is oriented to obtain the trend identifier of the target process; Based on the distribution of extreme points in the initial learning sequence and the oscillation characteristics of the later stable sequence, the stable convergence category and the high-frequency jitter category in the target process are identified, and the stable convergence category and the high-frequency jitter category are converted into corresponding stable feature codes and jitter feature codes. Using the trend identifier as a structural index, the stable feature encoding and jitter feature encoding are encapsulated in a targeted manner to obtain a multi-dimensional performance profile of the target process; A risk assessment is performed on the endogenous trajectory chain of the multi-dimensional performance profile to obtain the discrimination instructions for the target process; The state parameter stream of the target process is updated based on the logical branch of the discrimination instruction to obtain the parameter tuning result set of the target process.

2. The computer-automated parameter tuning method for behavior pattern recognition as described in claim 1, characterized in that, The phase transition point detection is performed on the validation set accuracy sequence of the target process to obtain the initial learning sequence, the intermediate adjustment sequence, and the later stable sequence of the target process, including: The first-order difference calculation is performed on the accuracy sequence of the validation set to obtain the difference sequence of the target process; The first position point in the difference sequence where the sign changes from positive to zero or negative is defined as the first phase transition point of the target process; Starting from the first phase transition point, the difference sequence is iteratively scanned, and the position point in the difference sequence where the sign changes from negative to zero or positive is defined as the second phase transition point of the target process; The validation set accuracy sequence is segmented based on the first phase transition point and the second phase transition point. The subsequence before the first phase transition point is used as the initial learning sequence of the target process, the subsequence between the first phase transition point and the second phase transition point is used as the intermediate adjustment sequence of the target process, and the subsequence after the second phase transition point is used as the later stable sequence of the target process.

3. The computer-automated parameter tuning method for behavior pattern recognition as described in claim 1, characterized in that, The step of determining the direction of the long-term trend component of the intermediate-term adjustment sequence to obtain the trend identifier of the target process includes: The intermediate adjustment sequence is subjected to sliding smoothing to obtain the smoothed sequence of the target process; Connect the start and end points of the smooth sequence to obtain the baseline segment of the target process; Calculate the vertical distance from the data points on the smoothed sequence to the baseline segment, and calculate the algebraic sum of the vertical distances and the sign of the corresponding algebraic sum; If the algebraic sum is positive, it is determined to be an upward trend, and a first-type trend identifier for the target process is obtained; if the algebraic sum is negative, it is determined to be a downward trend, and a second-type trend identifier for the target process is obtained; if the algebraic sum is zero, it is determined to be a stationary trend, and a third-type trend identifier for the target process is obtained.

4. The computer-automated parameter tuning method for behavior pattern recognition as described in claim 1, characterized in that, The step of identifying the stable convergence category and high-frequency jitter category in the target process based on the extreme point distribution of the initial learning sequence and the oscillation characteristics of the later stable sequence, and converting the stable convergence category and high-frequency jitter category into corresponding stable feature codes and jitter feature codes, includes: Local extrema are detected in the later stable sequence to obtain the local maxima and local minima of the target process; Calculate the amplitude difference between the local maxima and the local minima to obtain the oscillation amplitude sequence of the target process; Analyze the decay trend of the oscillation amplitude sequence: if the oscillation amplitude sequence shows a monotonically decreasing trend, it is determined to be the stable convergence category; if the oscillation amplitude sequence has no significant decay and the values ​​fluctuate violently, it is determined to be the high-frequency jitter category. Based on the initial learning sequence, the stable convergence category and the high-frequency jitter category are fused to generate the stable feature code or jitter feature code of the target process.

5. The computer-automated parameter tuning method for behavior pattern recognition as described in claim 4, characterized in that, Based on the initial learning sequence, the stable convergence category and the high-frequency jitter category are fused to generate the stable feature code or jitter feature code of the target process, including: Based on the oscillation amplitudes corresponding to the stable convergence category and the high-frequency jitter category, the first-class and second-class coding values ​​of the target process are determined; The first-class and second-class encoding values ​​are concatenated with the number of extreme points in the initial learning sequence to generate stable feature codes and jitter feature codes for the target process.

6. The computer-automated parameter tuning method for behavior pattern recognition as described in claim 1, characterized in that, The step of using the trend identifier as a structural index to perform targeted encapsulation of the stable feature encoding and jitter feature encoding to obtain a multi-dimensional performance profile of the target process includes: Create an association mapping table with the trend identifier as the key; The stable feature code and the jitter feature code are stored as data items in the value range associated with the trend identifier in the association mapping table; By attaching timestamp information to the association mapping table, a multi-dimensional performance profile of the target process is obtained.

7. The computer-automated parameter tuning method for behavior pattern recognition as described in claim 1, characterized in that, The risk assessment of the endogenous trajectory chain of the multi-dimensional performance profile to obtain the discrimination instruction for the target process includes: The risk assessment value of the target process is calculated based on the pattern symbol sequence and time series correlation strength in the endogenous trajectory chain. Linear interpolation is performed on the risk assessment value to obtain the value corresponding to the specified quantile in the target process, and the value is set as the dynamic discrimination threshold of the target process; The discrimination instruction for the target process is generated based on the comparison result between the risk assessment value and the dynamic discrimination threshold.

8. The computer-automated parameter tuning method for behavior pattern recognition as described in claim 7, characterized in that, The formula for calculating the risk assessment value is as follows: ； in, The risk assessment value is... It is a nonlinear saturation function. This is the scale scaling factor. For feature fusion function, For trend dynamics identifier vectors, For steady-state convergent encoding vectors, This is a high-frequency perturbation coding vector. It is a positive damping coefficient. This represents the time interval between key pattern symbols.

9. The computer-automated parameter tuning method for behavior pattern recognition as described in claim 1, characterized in that, The logical branch based on the discrimination instruction updates the state parameter stream of the target process to obtain the parameter tuning result set of the target process, including: If the discrimination instruction is a task continuation instruction, then based on the encoding features of the multi-dimensional performance profile, the current hyperparameter vector in the state parameter stream is subjected to directional perturbation to obtain the hyperparameters to be adjusted for the target process. If the discrimination instruction is a task termination instruction, then the optimal value record is retrieved from the historical multidimensional performance profile of the target process, and the hyperparameter combination associated with the optimal value record is extracted to obtain the parameter tuning result set of the target process.

10. A computer-automated parameter tuning system for behavior pattern recognition, used to implement the computer-automated parameter tuning method for behavior pattern recognition as described in any one of claims 1-9, characterized in that, The system includes: The data analysis module performs phase transition point detection on the validation set accuracy sequence of the target process to obtain the initial learning sequence, the intermediate adjustment sequence, and the later stable sequence of the target process. The trend identifier module determines the direction of the long-term trend component of the intermediate adjustment sequence to obtain the trend identifier of the target process. The state judgment module identifies the stable convergence category and high-frequency jitter category in the target process based on the distribution of extreme points in the initial learning sequence and the oscillation characteristics of the later stable sequence, and converts the stable convergence category and high-frequency jitter category into corresponding stable feature codes and jitter feature codes. The multi-dimensional performance profiling module uses the trend identifier as a structural index to perform targeted encapsulation of the stable feature encoding and jitter feature encoding to obtain a multi-dimensional performance profile of the target process. The discrimination instruction module performs a risk assessment on the endogenous trajectory chain of the multi-dimensional performance profile to obtain the discrimination instruction for the target process; The parameter tuning result set module updates the state parameter stream of the target process based on the logical branch of the discrimination instruction to obtain the parameter tuning result set of the target process.

Images