Assembly connection load intelligent control method based on machine learning

Abstract

The invention relates to the technical field of assembly load control, and discloses an assembly connection load intelligent control method based on machine learning. The method comprises the following steps: acquiring load data and corresponding real-time regulation and control parameters of an assembly connection point at each moment, and constructing a complete time sequence data set; and taking the load data at the current moment as a target load, and accurately deducing ideal regulation and control parameters in a historical sequence according to a distribution mode of the target load in the historical time sequence. And the regulation and control precision of each target load is quantified by analyzing the deviation between ideal and instant regulation and control parameters in a historical sequence. And finally, calculating estimated regulation and control parameters of each target load at the next moment by means of a machine learning algorithm by integrating regulation and control precision, historical instant regulation and control parameters and ideal and instant regulation and control parameters at the current moment. According to the method, the inherent rule of data is deeply excavated, dynamic intelligent adaptation of the assembly connection load is achieved, and the intelligent development requirement of precision assembly in the modern manufacturing industry is met.

Description

Technical Field

[0001] This invention relates to the field of assembly load control technology, specifically to an intelligent control method for assembly connection loads based on machine learning. Background Technology

[0002] In the assembly process of manufacturing, the reliability of assembly connections directly determines the overall performance and service life of the product, and the control quality of assembly connection loads is a core element in ensuring connection reliability. With the advancement of industry, the manufacturing industry is gradually transforming towards intelligent and flexible production, and the assembly precision requirements of various complex products are constantly increasing, posing more stringent challenges to the real-time performance, adaptability, and accuracy of assembly connection load control.

[0003] Traditional load control methods for assembly connections primarily rely on operator experience or pre-set fixed control rules. Operators set control parameters based on past experience, which is not only inefficient but also prone to mismatches between control parameters and actual load requirements due to subjective judgment biases. Furthermore, manual parameter adjustments are slow to respond to instantaneous fluctuations in load data during assembly, often resulting in control lags that affect assembly accuracy and may even lead to product assembly defects. Fixed control rules, on the other hand, cannot adapt to changes in assembly conditions. When factors such as product model, assembly environment, and component characteristics change, pre-set parameters struggle to quickly adapt to the new load state, leading to a significant decrease in control effectiveness.

[0004] In recent years, with the development of sensor and automation technologies, some assembly systems have begun to introduce automated control schemes. However, existing technologies still have many shortcomings. Most automated control methods adjust parameters based solely on load data at a single moment, neglecting the continuity and correlation of load data over time and failing to fully explore the load variation patterns inherent in historical data. Furthermore, existing methods lack effective evaluation mechanisms for control effectiveness, making it impossible to accurately quantify the gap between current control parameters and the ideal state. This results in a lack of clear direction in the parameter optimization process, making it difficult to effectively guarantee control accuracy. These problems make existing assembly connection load control methods unable to meet the needs of modern precision assembly, becoming a significant bottleneck restricting the improvement of manufacturing production efficiency and product quality. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based intelligent control method for assembly connection loads to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a machine learning-based intelligent control method for assembly connection loads, the method comprising: Collect load data of assembly connection points at different times and their corresponding real-time control parameters; Each load data point at the current moment is considered as the target load, and based on its distribution pattern in the historical time series, its ideal control parameters in the historical series are derived. The control precision of each target load is quantified by analyzing the deviation between the ideal control parameters and the instantaneous control parameters in the historical sequence. By combining the control precision, the real-time control parameters in the historical sequence, and the ideal control parameters and real-time control parameters at the current moment, the estimated control parameters for each target load at the next moment are calculated.

[0007] Preferably, the derivation process of the ideal control parameter is as follows: Identify load data that appears frequently in historical time series as core load data; For each core load data point, its significance is assessed based on its numerical fluctuations, temporal distribution, and load size in the historical sequence; For non-core load data in the target load, assess its significance based on its load size and the strength of its correlation with core load data; Each target load is assigned an ideal control parameter from the historical sequence based on its significance.

[0008] Preferably, the significance of core load data is assessed through the following steps: For any core load data and any point in the historical sequence, use machine learning feature extraction techniques to generate feature vectors and their dimensional information from the load value at that point. By combining the parameter coordinates of the core load data in the historical sequence control model, the location information at that moment, and the system state, the control parameters and their values ​​at that moment are derived. The difference between the number of feature vectors and the number of control parameters is calculated as the feature bias; The reference index is obtained by adding the characteristic deviation to a preset constant and taking the reciprocal. By using a similarity matching method, the feature vectors are compared with the control parameters, and the number of successfully matched feature vectors is counted as the match count. Calculate the load percentage of the core load data at that moment, and use it as a reference coefficient; The product of the match count, reference index, and reference coefficient is normalized to represent the significance of the core load data at that moment.

[0009] Preferably, the significance of non-core workload data is calculated in the following way: For any non-core load data and the time when it appears in the historical sequence, calculate the minimum correlation metric between the non-core load data and each core load data at that time, and use it as the correlation distance. The minimum correlation distance is added to a preset constant, and the reciprocal is taken as the target index; Obtain the percentage of the load value of this non-core load data at that moment, and use it as the target coefficient; The product of the target index and the target coefficient is normalized to obtain the reference significance level; The difference between the total number of control parameters and another preset constant is calculated and used as the correction value; Divide the correction value by the total number of levels of the control parameters to obtain the correction weight; The adjusted weight is multiplied by the reference significance to obtain the significance of the non-core load data at that moment.

[0010] The preferred allocation rule for the ideal control parameters is as follows: The numerical range of significance is divided into multiple intervals, with the number of intervals being the same as the total number of levels of the control parameter. Each interval corresponds to a control parameter level, and the interval with the highest significance corresponds to the lowest control parameter level; For any target load data and the time when it appears in the historical sequence, if the significance of the target load data at that time falls into a certain interval, then the control parameter level corresponding to that interval is taken as the ideal control parameter at that time.

[0011] The preferred method for calculating the control precision is as follows: For any target load data, calculate the difference between the ideal control parameter and the instantaneous control parameter at each time point in the historical sequence, and use it as the control deviation; The sum of all control deviations is inversely converted and normalized to obtain the control accuracy of the target load data.

[0012] Preferably, the method for generating the predicted control parameters is adjusted according to the control precision: If the control accuracy of the target load data exceeds the preset threshold, the estimated control parameters for the next moment are determined based on the real-time control parameters of the target load data at the current moment and its real-time control parameters in the historical sequence. If the control accuracy does not exceed the preset threshold, the estimated control parameters for the next moment are determined based on the ideal control parameters and the instantaneous control parameters at the current moment.

[0013] Preferably, when the control precision exceeds the threshold, the steps for determining the predicted control parameters include: Arrange the historical sequence of the target load data in chronological order to form a time series; Arrange the instantaneous control parameters at each moment in the historical sequence in chronological order to obtain the control parameter sequence; A trend line is obtained by linearly fitting the sequence of control parameters; If the slope of the trend line is positive, then the change in the control parameter is set to increase by one unit. If the slope of the trend line is zero, then the change is set to zero; If the slope of the trend line is negative, the change is set to decrease by one unit; The current instantaneous control parameter is added to the change, and the result is used as the estimated control parameter for the next instant.

[0014] Preferably, when the control precision does not exceed the threshold, the rule for determining the estimated control parameters is as follows: If the ideal control parameter is not greater than the instantaneous control parameter at the current moment, then the ideal control parameter will be used as the estimated control parameter for the next moment. If the ideal control parameter is greater than the instantaneous control parameter, then the instantaneous control parameter will be used as the predicted control parameter for the next moment.

[0015] Preferably, the method further includes an iterative control step, comprising: The estimated control parameters for the next moment are updated to the current control parameters, and the process of claim 1 is executed cyclically to achieve continuous load control.

[0016] Compared with the prior art, the beneficial effects of the present invention are: In terms of data utilization, this method comprehensively collects load data and corresponding real-time control parameters at different moments in the assembly connection points, constructing a complete time-series dataset and breaking the limitations of traditional methods that rely solely on data from a single moment. By deeply mining the distribution patterns of load data in historical time series, it can accurately capture the inherent laws and trends of load changes, allowing the setting of control parameters to go beyond superficial numerical matching and be based on a profound understanding of the essential relationships between the data. This full activation of historical data enables control strategies to better adapt to the dynamic changes in the assembly process, forming targeted parameter setting logic for both routine operating conditions and complex and ever-changing assembly scenarios.

[0017] In the derivation of ideal control parameters, this method uses the current load data as the target load and combines it with its distribution characteristics in historical sequences to deduce parameters, ensuring a high degree of fit between ideal parameters and actual load conditions. Compared to traditional parameter setting methods that rely on manual experience or fixed rules, this data-based derivation model is more objective and scientific, effectively avoiding deviations caused by human factors, making the control parameters more consistent with the physical characteristics and actual needs of the assembly connection, and reducing assembly risks caused by parameter mismatch.

[0018] The quantitative design of control precision provides a clear feedback path for assembly load control. By analyzing the deviation between ideal control parameters and real-time control parameters in historical sequences, weak links in the control process of each target load can be accurately located, and the fluctuation of control effect can be clearly presented. This quantitative evaluation method makes problems in the control process perceptible and knowable, provides a clear direction for subsequent parameter optimization, avoids the accumulation and expansion of deviations, helps maintain the stability and consistency of assembly connection loads, and ensures the smooth progress of assembly processes.

[0019] In the calculation of the predicted control parameters for the next moment, this method comprehensively integrates information from multiple aspects, including control accuracy, real-time control parameters from historical sequences, ideal control parameters at the current moment, and real-time control parameters. Leveraging the powerful data fusion and analysis capabilities of machine learning algorithms, it achieves the organic synergy of multi-source information. This multi-factor comprehensive decision-making model allows the predicted control parameters to fully incorporate effective information from historical experience and the current state, exhibiting stronger foresight and adaptability. When assembly conditions undergo sudden changes or load data shows abnormal fluctuations, this method can quickly adjust the predicted parameters, ensuring the timeliness and accuracy of control actions and effectively addressing the challenges of various complex assembly scenarios.

[0020] This method, driven by machine learning, enables the autonomous evolution and continuous optimization of control strategies. As the assembly process progresses and data samples accumulate, the machine learning model continuously uncovers new patterns, further improving the accuracy of control parameter calculations and creating a virtuous cycle. This intelligent control mode is widely adaptable to different types of assembly tasks, reducing reliance on operators' professional skills and minimizing uncertainties in the assembly process. It promotes a more efficient and stable assembly process, providing strong technical support for the intelligent upgrading of the manufacturing industry. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the working principle of the machine learning-based intelligent control method for assembly connection loads described in this invention. Figure 2 Flowchart for deriving the ideal control parameters; Figure 3 A flowchart for the allocation rules of ideal control parameters. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 This invention provides a machine learning-based intelligent load control method for assembly connection points. The method includes continuously recording load data and corresponding real-time control parameters of assembly connection points at consecutive timestamps, forming a historical time series for analysis. Each load data point collected at the current moment is set as a target load, and each target load is examined within its historical context. Based on its recurring patterns in the historical time series and its correlation with other load data, the ideal control parameters that should have been applied under similar past conditions are deduced. This derivation process relies on in-depth data mining and pattern recognition of historical data, aiming to find the inherent laws between load states and optimal control actions.

[0024] After obtaining the ideal control parameters at each moment in the historical sequence, the method enters the quantification stage of control accuracy. By comparing the ideal control parameters at each moment in the history with the actual instantaneous control parameters, the deviation between the two is calculated. The cumulative effect of this deviation reflects the accuracy and reliability of the historical control strategy for that specific target load. Control accuracy is an important feedback indicator, quantifying the system's control capability for that load. Finally, the system needs to integrate multi-source information to calculate the estimated control parameters for the next moment. This information includes the calculated control accuracy, a series of instantaneous control parameters recorded in the historical sequence, and the newly derived ideal control parameters and the actual instantaneous control parameters applied at the current moment.

[0025] Example 1: See Figure 2The derivation of ideal control parameters for the machine learning-driven intelligent load control method for assembly connections is based on deep pattern recognition of historical load data, which originates from time series records of assembly connection points during continuous operation. Identification of core load data is the starting point of this process. The system performs frequency analysis on the historical time series, identifying load data points that repeat more than a preset frequency threshold within a statistical period as core load data. Core load data constitutes the main framework of the system's load state, representing stable and common operating conditions. Each core load data point needs to undergo a significance evaluation, with evaluation dimensions covering the numerical fluctuation range of the load data in the historical series, the distribution density of the load data on the time axis, and the absolute value of the load data itself. The numerical fluctuation range reflects the stability of the load value, the distribution density reveals the periodicity or regularity of the load occurrence, and the load value establishes its relative position in the global load spectrum. These three characteristics together constitute a multi-dimensional coordinate system for evaluating the significance of core load data. For any core load data point and any specific moment in the historical series, machine learning feature extraction technology is applied to the global load information recorded at that moment. The feature extraction process extracts multi-dimensional feature vectors from the raw load values ​​to characterize their inherent patterns. These feature vectors contain information such as the load's spectral characteristics, gradient changes, and correlations between neighboring points. The dimensionality of the feature vectors is recorded synchronously; the number of dimensions depends on the complexity of the selected feature extraction algorithm. Simultaneously, the system accesses a pre-constructed historical sequence control model, which stores the mapping relationship between core load data and system parameters. By combining the parameter coordinates of the core load data in the model, the absolute and relative positions of that specific moment in the time series, and the overall operating state parameters of the system at that time, the system derives the theoretical values ​​of the control parameters applied to the core load data in this specific context.

[0026] The system performs a comparative analysis between the feature vector set and the derived control parameter set. The arithmetic difference between the total number of generated feature vectors and the total number of derived control parameters is calculated; this difference is defined as the feature bias. The magnitude of the feature bias reflects the matching relationship between the richness of feature extraction and the refinement of model derivation. A reference index is obtained by adding a preset constant to the feature bias and then taking the reciprocal of the sum. The reference index is designed so that the smaller the feature bias, the larger its value, indicating a better matching degree. The system's built-in similarity matching algorithm begins to work, comparing each feature vector with each control parameter. The comparison criteria are based on distance metrics or cosine similarity in the vector space. Successfully matched feature vectors are counted, and the total number is recorded as the match count. The match count directly reflects the degree of agreement between features and parameters at the specific numerical level. The proportion of the core load data's load value at a specific moment to the sum of all load values ​​at that moment is calculated; this proportion serves as a reference coefficient. The reference coefficient reflects the weight of the core load data in the instantaneous load composition. The match count, reference index, and reference coefficient are multiplied to obtain an unnormalized significance score. This product comprehensively considers the balance between the matching degree between data features and model parameters, feature richness and parameter derivation, and instantaneous importance. Normalization maps the product to a standard interval; the normalization method uses min-max normalization or Z-score standardization. The normalized result is ultimately determined as the significance of the core payload data at that specific historical moment. Significance is a quantitative value between 0 and 1; a higher value indicates that the core payload data is more prominent and critical at that moment.

[0027] The significance assessment logic for non-core load data in historical sequences differs from that for core load data. Non-core load data refers to load points with a frequency below a preset threshold, constituting the edge or special operating conditions of the load state. The significance assessment of non-core load data depends on its correlation with the core load data cluster. The correlation strength is quantified by calculating the distance between non-core load data and each core load data point in the feature space; the distance metric can be Euclidean or Mahalanobis distance. The minimum value among all distances is taken as the correlation distance; the smaller the correlation distance, the closer the non-core load data is to a certain core load pattern. The correlation distance is converted into a target exponent after being added to a preset constant and taking its reciprocal. The load value proportion of non-core load data at that moment is calculated as the target coefficient. The product of the target exponent and the target coefficient is normalized to form a reference significance level. The system predefines a total number of levels for a control parameter, which represents the control precision. The difference between the total number of levels and another preset constant is calculated; this difference is called the correction value. The correction value is divided by the total number of levels to obtain the correction weight. The adjustment weight is a coefficient between 0 and 1 used to adjust the reference significance. Finally, the adjustment weight is multiplied by the reference significance, and the product is the final significance of the non-core load data at that historical moment. This method ensures that the significance assessment of non-core load data considers both its similarity to the core pattern and its own proportion in the instantaneous load, while also being influenced by the granularity of system regulation, resulting in a more comprehensive and objective assessment. The entire significance calculation process assigns a quantitative importance index to each load data point in the historical sequence, which is the direct basis for subsequent allocation of ideal regulation parameters. The higher the significance of the load data point, the greater its influence on regulation decisions, and the assigned ideal regulation parameter level usually tends to be in a more conservative or stable range. This data-driven importance assessment mechanism enables the formulation of regulation strategies to focus on key load states, optimize resource allocation, and improve overall regulation efficiency and system robustness.

[0028] Example 2: Machine Learning-Driven Intelligent Load Control Method for Assembly Connections. The processing of non-core load data requires a saliency calculation paradigm independent of that for core load data. Non-core load data refers to load data points whose frequency of occurrence in historical time-series statistical analysis is below a preset threshold. Although these data points do not constitute the main state of system operation, their characteristics and occurrence patterns have a significant impact on the completeness and robustness of the overall control strategy. The core of calculating the saliency of non-core load data lies in quantifying its correlation with known core load data clusters and comprehensively evaluating it in conjunction with its own instantaneous characteristics. The correlation measurement is based on feature space distance calculation. The system executes the correlation metric calculation process for each non-core load data and each specific timestamp in its historical sequence. For any non-core load data and its specific time of occurrence in the historical sequence, the system needs to calculate the correlation metric between the non-core load data and all data points identified as core load data at that time. The correlation metric is calculated using the Euclidean distance formula in multidimensional space, constructing a feature vector for each load data point composed of its load value, timestamp index, and spectral features extracted from the original load signal. The feature vectors of non-core load data are used to calculate the Euclidean distance with the feature vectors of each core load data point. The Euclidean distance directly reflects the relative positional relationship between the two load data points in the feature space; the smaller the distance value, the more similar the two are in terms of pattern. The system iterates through the feature vectors corresponding to all core load data at that time, obtaining a set of distance values ​​with the non-core load data.

[0029] From the calculated set of Euclidean distance values ​​with all core load data, the smallest distance is selected. This smallest distance is defined as the correlation distance of the non-core load data at this moment. The concept of correlation distance captures the proximity between the non-core load data and the closest member of the core load data cluster. A smaller correlation distance means that the non-core load data is close to a known, common core state in terms of behavior. A preset constant is added to this correlation distance. The preset constant is used to prevent calculation anomalies when the correlation distance is zero and to give greater weight to extremely small distances. The reciprocal of the sum is then taken, and the resulting value is called the target exponent. The target exponent is designed so that the smaller the correlation distance, the larger its value, indicating a stronger correlation between the non-core load data and the core cluster. The system calculates the proportion of the non-core load data's load value to the total system load value at that specific moment. This proportion is defined as the target coefficient. The target coefficient reflects the contribution of the non-core load data to the overall system load state at a specific moment. A higher target coefficient means that even if the load data itself is not common, it has a relatively important influence at the time of its occurrence. The calculated target index is multiplied by the target coefficient to obtain a preliminary product value. This product takes into account the similarity between non-core load data and core patterns, as well as its own instantaneous weight.

[0030] The system normalizes this initial product value using a linear transformation projecting it into the interval between zero and one. The normalized result is labeled as the reference significance. The reference significance is a preliminary assessment of significance, but it needs further refinement by incorporating the granularity of the system's global control parameters. The system predefines the total number of control parameters, representing the fineness of the control action. The arithmetic difference between the total number of levels and another preset constant is calculated; this difference is called the correction value. Dividing the correction value by the total number of levels yields the correction weight, a coefficient between zero and one. The introduction of the correction weight ensures that the final significance value is adjusted within the system's available control precision. The finer the control granularity, the closer the correction weight is to zero, and the smaller the adjustment to the reference significance. Multiplying the correction weight by the reference significance, the result of this multiplication is taken as the final significance of the non-core load data at that specific historical moment. This final significance level is a standardized value between zero and one, simultaneously encompassing the correlation between non-core load data and the core cluster, its instantaneous importance, and the system's control capability. This calculation method ensures that even low-frequency non-core load data receives a significance value commensurate with its actual impact. This value guarantees fairness and completeness in the subsequent allocation of ideal control parameters, preventing control strategies from favoring high-frequency core data while ignoring low-frequency but potentially critical special operating conditions. The precise calculation of the significance level of non-core load data ensures that the intelligent control method can cover a more comprehensive system state space, enhancing its adaptability and decision-making rationality in the face of rare load fluctuations or special operating conditions.

[0031] Example 3: See Figure 3After calculating the significance of each load data point in the historical sequence, the machine learning-driven intelligent load control method for assembly connections enters the stages of allocating ideal control parameters and quantifying control precision. The allocation of ideal control parameters is based on the interval division and mapping relationship of significance values. The significance values ​​are continuously distributed from zero to one. The system divides the entire zero-to-one value interval equally, with the number of intervals exactly matching the pre-set total number of control parameter levels. The total number of control parameter levels represents the level of precision of the control actions that the system can execute. Each equally divided interval has the same width, and the interval boundaries gradually increase from zero to one. For example, when the total number of control parameter levels is ten, the interval division includes ten consecutive intervals: [0, 0.1), [0.1, 0.2), up to [0.9, 1.0]. Each interval is uniquely mapped to a control parameter level. The mapping rule is that the interval with the highest significance value corresponds to the lowest value among the control parameter levels, and the interval with the lowest significance value corresponds to the highest value among the control parameter levels. This inverse mapping relationship reflects that high-significance load states often require more stable or conservative control strategies. For any target load data point in the historical sequence and any specific time it appears, the system retrieves the calculated significance value of that target load data at that time and determines which equal interval it falls into. Once the interval is determined, the system assigns the control parameter level mapped to that interval as the ideal control parameter for that target load data at that historical time. Through this mechanism, each load data point in the historical sequence is assigned an ideal control parameter value that matches its significance, and the formation of the ideal control parameter sequence provides a benchmark for evaluating historical control performance. The calculation of control accuracy is performed for each target load data point, which can be core load data or non-core load data. Control accuracy reflects the historical control effect of the system on that specific load data. For any target load data point, the system iterates through all times it appears in the historical sequence. At each time point, the system calculates the absolute difference between the ideal control parameter assigned at that time and the actual instantaneous control parameter recorded at that time. This absolute difference is defined as the control deviation at that time. The control deviation quantifies the degree of dispersion between the ideal control action and the actual control action. The system sums the control deviations at all times to obtain the cumulative control deviation, which represents the overall deviation of historical control behavior from the ideal state. The larger the cumulative control deviation, the less the historical control strategy matches the ideal demand of the load data.

[0032] The cumulative control deviation needs to undergo an inverse proportional transformation. This transformation involves taking the reciprocal of the cumulative control deviation to obtain an inverse proportional value. The introduction of this inverse proportional value means that the larger the cumulative control deviation, the smaller the inverse proportional value, and vice versa. The inverse proportional value reflects the raw measure of control accuracy, but to bring it into a standardized range for easier comparison, normalization is required. Normalization uses a minimum-maximum scaling method, mapping the inverse proportional value to the interval between zero and one. The normalization process requires determining the minimum and maximum values ​​based on the range of inverse proportional values ​​for all target load data in historical data. The formula for calculating control accuracy is as follows:

[0033] in: It represents the control precision and is a dimensionless scalar with a value between zero and one. This represents the total number of times the target load data appears in the historical sequence; It is an index variable that iterates from 1 to... Every moment; This indicates the target load data in the historical sequence. The ideal control parameter at any given time is a value on the order of an integer. This indicates the target load data in the historical sequence. The instantaneous adjustment parameter at any given moment is a value on the order of an integer. Indicates the first The control deviation at any given moment is the absolute value of the difference between the ideal control parameter and the instantaneous control parameter. This represents the cumulative control deviation, which is the sum of the control deviations at all points in time. This represents the value after inverse proportional conversion; This represents the set of inversely proportional transformation values ​​of all target load data in historical data; Represents a set The minimum value in; Represents a set The maximum value in the molecule; Shift the inverse ratio to a point starting from the minimum value; denominator part It is a set The range of values ​​is used for scaling; the entire fraction scales the translated inverse ratio to the interval between zero and one, obtaining the final control precision. .

[0034] All symbols in the formula maintain consistent dimensions; ideal control parameters. Real-time adjustment parameters , control deviation and cumulative control deviation Both have dimensions of the control parameter level. After inverse proportional conversion, the dimensions become the reciprocal of the level, but due to the normalization of the denominator... It also has the dimension of the reciprocal of the order; after dividing the dimensions of the numerator and denominator, the control precision is... It becomes a dimensionless ratio. Control precision. The calculation provides a unified performance index for each target load data point; a higher accuracy value indicates that historical control actions are closer to the ideal state. The allocation of ideal control parameters and the calculation of control accuracy together constitute a key link in historical data analysis, providing a data-driven decision-making basis for subsequent dynamic adjustments to the predicted control parameters. The significance of load data is transformed into specific control levels through interval mapping, while control accuracy quantifies historical execution effects. These two parts enable intelligent control methods to optimize based on objective data rather than subjective experience, enhancing the system's adaptability and reliability.

[0035] Example 4: The predictive control parameter generation stage of the machine learning-driven intelligent load control method for assembly connections exhibits adaptive logic. Its core lies in dynamically selecting a calculation strategy based on the control accuracy level of the target load data. Control accuracy is a quantitative indicator calculated from historical data, reflecting the degree of agreement between the system's past control effects on the load and the ideal state. The system internally presets an accuracy threshold, which serves as a decision boundary, dividing the control behavior into high-confidence and low-confidence intervals. When the calculated control accuracy of the target load data exceeds the preset threshold, it indicates that the historical control actions for this load have high consistency and reliability, and their changing trends contain valuable dynamic information. In this case, the generation of predictive control parameters will fully utilize the trend information contained in the historical real-time control parameter sequence, using time series analysis methods for extrapolation prediction. The generation process begins with the organization of historical data, strictly arranging all moments of the target load data in the historical sequence according to chronological order to form a continuous time index sequence. Corresponding to this time index sequence, the real-time control parameter values ​​actually recorded and executed at each moment in the historical sequence are extracted in the same chronological order to form a real-time control parameter sequence. This real-time control parameter sequence serves as the fundamental data source for trend analysis, recording the specific path of actual control actions over time for that load. Trend analysis employs a linear fitting algorithm, with the least squares method used to find a straight line that best represents the overall direction of change in the real-time control parameter sequence; this line is called the trend line. The mathematical expression for the trend line is y = kx + b, where the slope k characterizes the direction and rate of change of the control parameters, and the intercept b represents the initial level.

[0036] The sign and magnitude of the trend line slope *k* are crucial in determining the change in the control parameter at the next moment. The calculated slope *k* is evaluated as follows: If the slope *k* is greater than zero, it indicates a clear upward trend in the historical real-time control parameter sequence, with the control intensity gradually increasing over the historical period. In this case, the system sets the change in the control parameter at the next moment to increase by one unit, with the change being positive. If the slope *k* is equal to zero or fluctuates within a very small tolerance range around zero, it indicates that the historical real-time control parameter sequence remains stable without a clear monotonic trend. In this case, the system sets the change in the control parameter at the next moment to zero, meaning maintaining the current control intensity. If the slope *k* is less than zero, it indicates a clear downward trend in the historical real-time control parameter sequence, with the control intensity gradually weakening over the historical period. In this case, the system sets the change in the control parameter at the next moment to decrease by one unit, with the change being negative.

[0037] After determining the change, the system reads the current instantaneous control parameter value at the latest timestamp. The instantaneous control parameter at the current moment is algebraically added to the determined change, and the result is used to determine the estimated control parameter for the next moment. This trend-based extrapolation prediction method relies on the continuity and regularity of historical data and is suitable for high-precision scenarios where the system operates smoothly and the control strategy is consistent. It allows control decisions to continue historically effective strategies and reasonably extend their future trends. To more clearly illustrate this decision-making logic, refer to Table 1, which lists calculation examples of how to determine the estimated control parameter for the next moment based on the current instantaneous control parameter under different trend line slopes. The trend slope judgment in the table is based on a near-zero tolerance threshold ε.

[0038] Table 1: Rules for Determining Predicted Control Parameters When High Control Accuracy is Required

[0039] The linear fitting process smooths out random fluctuations in historical data, helping to capture macro trends rather than being affected by short-term noise. The goodness of fit of the trend line can serve as an auxiliary reference to assess the reliability of trend prediction; however, in this implementation, the decision is based solely on the sign of the slope k. The instantaneous control parameters at the current moment act as a benchmark in the calculation. Trend extrapolation involves adding or subtracting a fixed unit level from this benchmark. This design avoids overly aggressive adjustments and maintains the stability of the system's control behavior. The entire process of generating the predicted control parameters embodies the adaptive characteristics of data-driven decision-making. Within a high-precision confidence interval, the system trusts historical behavior patterns and makes future decisions accordingly, giving the control strategy a certain degree of foresight and continuity. This method is particularly suitable for assembly and connection systems with strong inertia or periodicity in load changes, effectively utilizing historical patterns to improve the accuracy of future control.

[0040] Example 5: Machine Learning-Driven Intelligent Load Control Method for Assembly Connections. When the control accuracy of the target load data does not exceed a preset threshold, a conservative decision rule based on the current state is adopted. The fact that the control accuracy does not exceed the preset threshold means that historical control records deviate significantly from the ideal state or lack a clear pattern. In this scenario, continuing to rely on historical trends for extrapolation is risky. The generation logic of the estimated control parameters shifts to directly comparing the numerical relationship between the ideal control parameter and the instantaneous control parameter at the current moment. The ideal control parameter at the current moment is the theoretically optimal value recently derived from the significance of the load data and interval mapping rules. The instantaneous control parameter at the current moment is the actual control command being executed. The comparison between the two directly reflects the gap between the current control action and the current ideal state. The core of the decision rule is to determine whether the ideal control parameter at the current moment is greater than the instantaneous control parameter at the current moment. The comparison operation uses direct numerical comparison. If the ideal control parameter at the current moment is not greater than the instantaneous control parameter at the current moment, the mathematical relationship includes two cases: less than or equal to. This case indicates that the theoretically optimal control intensity calculated based on the current system state is lower than or equal to the actual control intensity. The system decision is to directly set the estimated control parameter for the next moment as the ideal control parameter for the current moment. This decision prompts the control action to approach the theoretical optimal value, reducing the control intensity to match the ideal demand of the current load. For example, suppose the system control parameter level ranges from 1 to 10, where level 1 represents the lowest control intensity and level 10 represents the highest. At a specific moment, for a target load data, the calculated ideal control parameter for the current moment is level 4, while the actual applied instantaneous control parameter is level 6. The ideal control parameter (level 4) is not greater than the instantaneous control parameter (level 6). According to the rule, the estimated control parameter for the next moment is set to level 4. This adjustment means that the system considers the current control intensity of level 6 to be too high, and the next cycle should reduce it to level 4 to approach the ideal state.

[0041] If the ideal control parameter at the current moment is greater than the immediate control parameter at the current moment, this indicates that theoretical analysis requires stronger control intervention than the actual implementation suggests. However, due to the low accuracy of historical control, the system lacks confidence in its own judgment, and hastily increasing the control intensity may introduce instability risks. The system decision is to set the estimated control parameter for the next moment to the same value as the immediate control parameter at the current moment. The decision logic tends to maintain the status quo, avoiding potentially overly aggressive adjustments when reliability is insufficient. Continuing with the above example, at another moment, for the same target load data, the calculated ideal control parameter for the current moment is level 7, while the immediate control parameter at the current moment is still level 6. At this time, the ideal control parameter (level 7) is greater than the immediate control parameter (level 6). According to the rule, the estimated control parameter for the next moment is set to level 6, consistent with the current immediate control parameter, and the control intensity remains unchanged. This decision-making mechanism reflects the principle of robustness, prioritizing the continuity and stability of system behavior in areas of insufficient control confidence, and avoiding the amplification of erroneous decisions caused by poor historical data. The rule essentially balances "correction" and "stability maintenance." The system only decisively executes corrective action when the ideal state indicates a need for relaxed control (ideal value not exceeding the current value); conversely, when the ideal state indicates a need for strengthened control (ideal value exceeding the current value), the system chooses a conservative stability maintenance strategy. This asymmetric response mechanism helps prevent the system from falling into over-adjustment oscillations when information is uncertain. The machine learning-driven intelligent load control method for assembly connections includes a closed-loop iterative control mechanism, enabling the above decision-making process to run continuously. Once the estimated control parameters for the next moment are calculated, these parameters are output to the actuators at the assembly connection points. As time progresses, this "next moment" becomes the new "current moment," and the previously calculated estimated control parameters are updated to the real-time control parameters for this new moment. The system synchronously collects the load data of the assembly connection points at this new moment, and this new data is incorporated into the considerations. The entire process is then restarted: the new load data is regarded as the target load, the system re-derives the historical ideal control parameters based on the expanded historical time series (which includes the newly added data points), recalculates the control accuracy, and selects the corresponding rules again based on the latest control accuracy level to calculate the estimated control parameters for the next time step.

[0042] The iterative control loop constitutes a dynamic feedback system, where decisions at each time step are based on all available information up to the current time. System status, load conditions, and control effects are continuously monitored and used as input for the next decision. This cyclical mechanism enables the intelligent control method to adapt to long-term changes in load conditions and learn from historical experience. For example, load data may initially be in a low-precision range due to sparse or highly volatile historical data, and the control strategy may adopt conservative rules. As the system continues to run, a large amount of data about this load is accumulated, and its control accuracy may improve as historical records gradually match the ideal state, exceeding a preset threshold. Once the threshold is exceeded, the system automatically switches the decision logic from the conservative rules of this embodiment to the trend extrapolation rules of embodiment 4, achieving adaptive upgrading of the control strategy. The machine learning-driven intelligent control method for assembly and connection loads, through the combination of conservative decision rules and iterative control loops, ensures that the system's control behavior remains basically stable even in stages with insufficient information and poor historical performance, accumulating high-quality data for subsequent learning.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based intelligent control method for assembly connection loads, characterized in that, The method includes: Collect load data of assembly connection points at different times and their corresponding real-time control parameters; Each load data point at the current moment is considered as the target load, and based on its distribution pattern in the historical time series, its ideal control parameters in the historical series are derived. The control precision of each target load is quantified by analyzing the deviation between the ideal control parameters and the instantaneous control parameters in the historical sequence. By combining the control precision, the real-time control parameters in the historical sequence, and the ideal control parameters and real-time control parameters at the current moment, the estimated control parameters for each target load at the next moment are calculated.

2. The intelligent control method for assembly connection load based on machine learning as described in claim 1, characterized in that, The derivation process of the ideal control parameter is as follows: Identify load data that appears frequently in historical time series as core load data; For each core load data point, its significance is assessed based on its numerical fluctuations, temporal distribution, and load size in the historical sequence; For non-core load data in the target load, assess its significance based on its load size and the strength of its correlation with core load data; Each target load is assigned an ideal control parameter from the historical sequence based on its significance.

3. The intelligent control method for assembly connection load based on machine learning as described in claim 2, characterized in that, The significance of core load data is assessed using the following steps: For any core load data and any point in the historical sequence, use machine learning feature extraction techniques to generate feature vectors and their dimensional information from the load value at that point. By combining the parameter coordinates of the core load data in the historical sequence control model, the location information at that moment, and the system state, the control parameters and their values ​​at that moment are derived. The difference between the number of feature vectors and the number of control parameters is calculated as the feature bias; The reference index is obtained by adding the characteristic deviation to a preset constant and taking the reciprocal. By using a similarity matching method, the feature vectors are compared with the control parameters, and the number of successfully matched feature vectors is counted as the match count. Calculate the load percentage of the core load data at that moment, and use it as a reference coefficient; The product of the match count, reference index, and reference coefficient is normalized to represent the significance of the core load data at that moment.

4. The intelligent control method for assembly connection load based on machine learning as described in claim 2, characterized in that, The significance of non-core workload data is calculated as follows: For any non-core load data and the time when it appears in the historical sequence, calculate the minimum correlation metric between the non-core load data and each core load data at that time, and use it as the correlation distance. The minimum correlation distance is added to a preset constant, and the reciprocal is taken as the target index; Obtain the percentage of the load value of this non-core load data at that moment, and use it as the target coefficient; The product of the target index and the target coefficient is normalized to obtain the reference significance level; The difference between the total number of control parameters and another preset constant is calculated and used as the correction value; Divide the correction value by the total number of levels of the control parameters to obtain the correction weight; The adjusted weight is multiplied by the reference significance to obtain the significance of the non-core load data at that moment.

5. The intelligent control method for assembly connection load based on machine learning as described in claim 2, characterized in that, The allocation rules for the ideal control parameters are as follows: The numerical range of significance is divided into multiple intervals, with the number of intervals being the same as the total number of levels of the control parameter. Each interval corresponds to a control parameter level, and the interval with the highest significance corresponds to the lowest control parameter level; For any target load data and the time when it appears in the historical sequence, if the significance of the target load data at that time falls into a certain interval, then the control parameter level corresponding to that interval is taken as the ideal control parameter at that time.

6. The intelligent control method for assembly connection load based on machine learning as described in claim 1, characterized in that, The calculation method for control precision is as follows: For any target load data, calculate the difference between the ideal control parameter and the instantaneous control parameter at each time point in the historical sequence, and use it as the control deviation; The sum of all control deviations is inversely converted and normalized to obtain the control accuracy of the target load data.

7. The intelligent control method for assembly connection load based on machine learning as described in claim 1, characterized in that, The method for generating the predicted control parameters is adjusted according to the control precision: If the control accuracy of the target load data exceeds the preset threshold, the estimated control parameters for the next moment are determined based on the real-time control parameters of the target load data at the current moment and its real-time control parameters in the historical sequence. If the control accuracy does not exceed the preset threshold, the estimated control parameters for the next moment are determined based on the ideal control parameters and the instantaneous control parameters at the current moment.

8. The intelligent control method for assembly connection load based on machine learning as described in claim 7, characterized in that, When the control precision exceeds the threshold, the steps for determining the predicted control parameters include: Arrange the historical sequence of the target load data in chronological order to form a time series; Arrange the instantaneous control parameters at each moment in the historical sequence in chronological order to obtain the control parameter sequence; A trend line is obtained by linearly fitting the sequence of control parameters; If the slope of the trend line is positive, then the change in the control parameter is set to increase by one unit. If the slope of the trend line is zero, then the change is set to zero; If the slope of the trend line is negative, the change is set to decrease by one unit; The current instantaneous control parameter is added to the change, and the result is used as the estimated control parameter for the next instant.

9. The intelligent control method for assembly connection load based on machine learning as described in claim 7, characterized in that, When the control precision does not exceed the threshold, the rule for determining the predicted control parameters is as follows: If the ideal control parameter is not greater than the instantaneous control parameter at the current moment, then the ideal control parameter will be used as the estimated control parameter for the next moment. If the ideal control parameter is greater than the instantaneous control parameter, then the instantaneous control parameter will be used as the predicted control parameter for the next moment.

10. The intelligent control method for assembly connection load based on machine learning as described in claim 1, characterized in that, It also includes iterative control, including: The estimated control parameters for the next moment are updated to the current control parameters, and the process of claim 1 is executed cyclically to achieve continuous load control.

Images