A method of correlating spring defect data traceability of production lots

Abstract

The present application relates to the field of industrial production quality management and big data analysis technology, and particularly relates to a spring defect data traceability method for associated production batches, which comprises: obtaining a spring defect data sequence to be traced and a plurality of process parameter sequences corresponding to candidate production batches, calculating an effective payload cumulative value according to the numerical drift of the process parameters, and determining a drift directionality coefficient; fusing the two to obtain a physical cause weighting coefficient for representing the physical damage of parameter fluctuation; based on the principle of Granger causality test, correcting the prediction error by using the physical cause weighting coefficient, determining the Granger causality degree through the corrected prediction error, and determining the responsible batch and the cause parameter according to the sorting result of the Granger causality degree to realize defect traceability. The method solves the problem of false cause misjudgment and improves the accuracy of spring defect traceability in a multi-batch parallel operation environment.

Description

Technical Field

[0001] This invention relates to the field of industrial production quality management and big data analysis technology. Specifically, it relates to a method for tracing the source of spring defect data related to a production batch. Background Technology

[0002] The production of mechanical springs involves multiple continuous and highly coupled processes, including coiling, stress-relief annealing, end-face grinding, and high-pressure treatment. In actual production, due to the parallel operation of multiple batches, when the final quality inspection finds specific defects such as large stiffness dispersion, insufficient fatigue life, or excessive free height tolerance, it is necessary to trace the source of the defect. Accurately identifying the responsible production batch that caused the defect is key to reducing scrap rate and recall risk.

[0003] In existing technologies, the Granger causality test is a classic statistical algorithm widely used to analyze causal relationships between multivariate time series. The basic principle of this algorithm is: if the prediction effect of variable Y, including information about variable X, is better than the prediction effect of Y using only past information about variable Y, that is, if variable X helps reduce the prediction error of variable Y, then variable X is considered a Granger cause of variable Y. In industrial data traceability, process parameters for each batch are typically used as X, and product quality indicators as Y. The F-statistic is calculated to determine whether parameter anomalies led to quality defects.

[0004] However, when applied to tracing defects in spring production, the traditional Granger causality test algorithm suffers from significant physical lag and energy inequivalence defects. The Granger test mainly determines causality based on the statistical reduction in prediction error, assuming that all reductions in prediction error are equivalent. The process parameters of a normal production batch may simply be due to random fluctuations, and in time they happen to show a strong statistical correlation with the actual physical defects of another batch, causing the algorithm to incorrectly identify it as the causative batch. In the spring manufacturing process, especially in the heat treatment and high-pressure processes, the influence of process parameters on product performance has a significant energy accumulation effect and thermal inertia. For example, although the small high-frequency fluctuations in the heat treatment furnace temperature over a short period of time are statistically mathematically correlated with the small fluctuations in certain quality data and show significant results in the Granger test, from the perspective of physical metallurgy, such short-term fluctuations, due to insufficient energy, cannot cause any change in the metallographic structure of the spring. They belong to the category of fluctuation noise data and interfere with the traceability results. On the contrary, a certain temperature drift with a long duration but small amplitude, due to the accumulation of a large amount of heat energy, is the fundamental cause of the decline in fatigue performance, that is, the physical cause of spring defects, which is of great significance to the traceability results.

[0005] Traditional Granger causality test algorithms cannot distinguish between noisy data and physical causes. This makes it easy to produce false causal misjudgments when faced with electromagnetic interference or mechanical vibrations that are common in industrial sites. Normal batches that only have data fluctuations but no actual physical damage are misjudged as responsible batches, which affects the accuracy of the traceability results. Summary of the Invention

[0006] To address the problem that existing Granger causality tests in spring defect tracing often result in false causal errors and low accuracy due to neglecting the energy accumulation effect and thermal inertia of process parameters and failing to distinguish between statistical correlation noise and physical causal bias, this invention proposes a spring defect data tracing method that correlates production batches, comprising: Obtain all candidate production batches corresponding to the spring defect data sequence to be traced, and collect multiple process parameter sequences for each candidate production batch; For each type of process parameter sequence, the cumulative effective load value of the process parameter sequence is calculated based on the numerical drift of the process parameter sequence, and the drift directionality coefficient of the process parameter sequence is determined based on the cumulative effective load value. By combining the cumulative effective load value and the drift directionality coefficient, the physical causal weighting coefficient of this type of process parameter sequence is determined; Based on the Granger causality test principle, each type of process parameter sequence is introduced to predict the data sequence of spring defects to be traced, and the prediction error of the process parameter sequence is obtained. The prediction error is corrected by the physical cause weighting coefficient of the process parameter sequence. The Granger causality between the process parameter sequence and the spring defect data sequence to be traced is calculated based on the corrected prediction error. The traceability result of the spring defect data sequence to be traced is determined based on the Granger causality between the process parameter sequence and the spring defect data sequence to be traced for all candidate production batches.

[0007] This technical solution first acquires the data sequence of spring defects to be traced and the corresponding candidate production batches of multiple process parameter sequences. For each sequence, the effective load accumulation value is calculated and the drift directionality coefficient is determined in combination with the numerical drift situation. This deeply simulates the physical energy accumulation process of heat treatment and processing stress in spring production. On this basis, the physical causation weighting coefficient determined by the above indicators is integrated to inject rigorous physical mechanism constraints into the subsequent statistical inference. By introducing the process parameter sequence for prediction based on the Granger causality test principle, the prediction error is nonlinearly corrected using the physical causation weighting coefficient, which effectively filters out common interferences such as electromagnetic interference or data fluctuations in industrial sites. Based on the corrected prediction error, the Granger causality is calculated and the traceability results are output. This ensures that under the complex working conditions of multiple batches operating in parallel, the causal parameters and responsible batches that cause product performance defects can be accurately identified, thus improving the accuracy of traceability.

[0008] Preferably, the cumulative effective payload value satisfies the following relationship:

[0009] in, For the first The first candidate production batch The first of the process parameter sequence A number of effective payload values, For the first The first candidate production batch The first of the process parameter sequence A number, , and These are the first and second batches of the pre-obtained defect-free production batches. The mean, standard deviation, and time decay constant of the process parameter series. For the first The first candidate production batch The first of the process parameter sequence The cumulative drift time at each value, It is the absolute value symbol. For the natural exponential function; let the first... The first candidate production batch The effective loads of all values ​​in the sequence of process parameters are summed to obtain the first value. The first candidate production batch Cumulative effective load value of a series of process parameters This is the preset parameter for preventing zero.

[0010] This technical solution assesses the energy accumulation effect of process deviations at the physical level by analyzing the degree of numerical deviation from the mean and considering the cumulative drift time of time decay. By summing up the effective load of all values ​​to obtain the cumulative effective load value, it can more accurately identify those physical causes that have a long duration and have a substantial impact on the spring.

[0011] Preferably, the time decay constant and the cumulative drift duration are determined based on the following methods: Calculate the first defective production batch obtained in advance. The autocorrelation coefficients of process parameter sequences at different lag times are decayed from 1 to... The corresponding lag time length determines the first The time decay constant of the process parameter sequence, where, For the natural constant; from the perspective of the first The first candidate production batch A series of process parameters, from Start tracing back, according to Each previous value and The sign of the difference indicates the sum of each value. The deviation polarity is statistically consistent for a length of time, and the cumulative drift duration is taken as the length of time the deviation polarity remains consistent.

[0012] Preferably, the drift directionality coefficient satisfies the following relationship:

[0013] in, and The first The first candidate production batch The drift directionality coefficient and cumulative effective load of the process parameter sequence. and They are the first The first candidate production batch The index and length of the process parameter sequence. For symbolic functions, and The first The first candidate production batch The first of the process parameter sequence The number and the first A number of effective payload values, For the first defect-free production batch obtained in advance The average value of the process parameter sequence. To prevent positive numbers with a denominator of 0, For the first The reciprocal of the mean of the cumulative effective load values ​​of all types of process parameter sequences for each candidate production batch. It is the hyperbolic tangent function. It is the absolute value symbol.

[0014] The drift directionality coefficient determined by this technical solution injects spatial polarity dimensional constraints into the source tracing analysis. It uses a sign function to capture the offset direction of process parameters relative to the ideal value, performs global mapping with effective load as weight, and finally performs nonlinear smoothing through a hyperbolic tangent function. This can accurately distinguish between symmetrical random fluctuations and trend-based unidirectional drifts. In spring physics, stress or temperature shifts in a specific direction often correspond to specific types of defects. Directionality makes the source tracing results more directional in terms of failure mechanism.

[0015] Preferably, the physical cause weighting coefficients satisfy the following relationship:

[0016] in, For the first The first candidate production batch Physical cause weighting coefficients for process parameter sequences. For the first The first candidate production batch Cumulative effective load value of a series of process parameters The standard deviation of the drift directionality coefficients for all types of process parameter sequences in a pre-obtained defect-free production batch is given. It is the standard deviation of the cumulative effective values ​​of the sequence of process parameters of all types in a pre-acquired defect-free production batch.

[0017] The physical causal weighting coefficient determined by this technical solution achieves deep coupling between energy accumulation and directional characteristics. By introducing a pre-acquired defect-free standard deviation as a benchmark, the complex physical evolution process is effectively transformed into a quantifiable weighting factor. This modeling logic essentially adds a physical filter to the traditional statistical correlation to ensure that the corresponding causal weight is activated only when the process deviation reaches a significant physical threshold in both amplitude accumulation and directional shift.

[0018] Preferably, based on the Granger causality test principle, each type of process parameter sequence is introduced to predict the data sequence of spring defects to be traced, and the prediction error of this type of process parameter sequence is obtained, including: For each value in the traceable spring defect data sequence, according to the index of the value, all values ​​before the index are divided in both the traceable spring defect data sequence and the process parameter sequence of this type, forming two sequences. The two sequences are used to perform binary regression fitting on the value to obtain the first predicted value of the value. The difference between the value and the first predicted value is calculated to obtain the first prediction residual. The variance of the first prediction residuals of all values ​​is used as the prediction error of this type of process parameter sequence.

[0019] Preferably, the method for correcting the prediction error is to multiply the prediction error of each type of process parameter sequence by the reciprocal of the physical cause weighting coefficient of that type of process parameter sequence to obtain the corrected prediction error of that type of process parameter sequence.

[0020] This technical solution embeds a physical penalty mechanism within the purely statistical Granger causality framework by multiplying the prediction error of the process parameter sequence by the reciprocal of the physical causation weighting coefficient. The core logic is that if a parameter fluctuation is mathematically highly correlated with defect data (i.e., the prediction error is small), but its physical causation weighting coefficient is low (i.e., insufficient energy accumulation or misalignment), its reciprocal will act as a large penalty factor, amplifying the corrected prediction error and thus suppressing spurious causal interference. Conversely, only parameters with strong statistical correlation and significant physical causation can maintain a low level of corrected prediction error. This correction logic endows the algorithm with the ability to identify the essence of physical damage, ensuring that the final source tracing result is based on the true cause confirmed by data correlation and physical mechanism, thereby improving the accuracy of the source tracing result.

[0021] Preferably, Granger causality is determined as follows: For each value in the traceable spring defect data sequence, according to the index of that value, all values ​​before that index are divided in the traceable spring defect data sequence to form a sequence. The sequence is used to perform autoregressive fitting on the value to obtain a second predicted value. The difference between the value and the second predicted value is calculated to obtain a second prediction residual. The variance of the second prediction residuals of all values ​​is taken as the prediction error of the traceable spring defect data sequence. The ratio of the prediction error of the traceable spring defect data sequence to the corrected prediction error of the process parameter sequence is taken as the logarithmic function value of the ratio as the Granger causality between the process parameter sequence and the traceable spring defect data sequence.

[0022] This technical solution transforms the reduction in statistically predicted error variance into an indicator at the causal level, constituting the final criterion for source tracing analysis. It determines the fluctuation benchmark of the spring defect data to be traced through autoregressive fitting, and then calculates the ratio of its predicted error variance to the predicted error variance after process parameter correction and takes the logarithm. This approach essentially evaluates the marginal contribution of specific process parameters to eliminating the uncertainty of defect data. Since the denominator has been nonlinearly corrected by the physical cause weighting coefficient, the obtained Granger causality not only reflects the correlation at the data level, but also deeply couples the load accumulation effect at the physical level, enabling the precise extraction of root cause parameters with actual physical destructive force from interfering factors.

[0023] Preferably, the tracing results of the spring defect data sequence to be traced are determined based on the following method: The Granger causality of all process parameter sequences in all candidate production batches is uniformly sorted. The process parameter category corresponding to the process parameter sequence with the largest Granger causality value is determined as the causative parameter causing the spring defect. The batch corresponding to the process parameter sequence with the largest Granger causality value is determined as the responsible batch causing the spring defect.

[0024] Preferably, the multiple categories of process parameter sequences are determined based on the following method: obtaining the time series of spring quality defect indicators detected in the terminal quality inspection process as the spring defect data sequence to be traced, wherein the defect indicators include stiffness, free height, and fatigue life; according to the time period corresponding to the spring defect data sequence to be traced, the production process is traced back, and all production batches that overlap with the time period are identified as candidate production batches; the various process parameters of each candidate production batch throughout the entire time period are arranged in time sequence to form multiple categories of process parameter sequences, and the process parameter sequences of each category are aligned with the spring defect data sequence to be traced through time sequence alignment and resampling operations; wherein, the various process parameters include tempering furnace temperature, spring winding speed, cutting force, and pressure.

[0025] The present invention has the following effects: This invention constructs a tracing model that deeply integrates physical causal mechanisms and statistical causal testing. It uses the cumulative effective load value and drift directionality coefficient to accurately reflect the energy threshold of process parameter fluctuations transforming into physical damage to products. This value is then used as a weight to nonlinearly correct prediction errors, injecting physical constraints into the statistical model. This allows the model to automatically filter out noise fluctuations caused by insufficient energy in industrial settings and focus on persistent shifts with substantial destructive power. This solves the problem of false causal misjudgment caused by neglecting the physical cumulative effect in the complex working conditions of spring production using traditional Granger testing. It enables precise location of the responsible batch and causal parameters that lead to specific defects such as stiffness and fatigue life, improving the accuracy of defect tracing in multi-batch parallel operation environments. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the various process parameter sequences and the defect data sequence to be traced in this invention; Figure 3 This is a schematic diagram of the cumulative effective load value calculated based on numerical drift according to the present invention; Figure 4 This is a schematic diagram comparing the Granger causality of the present invention with the causality calculated by the traditional Granger causality test algorithm. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0028] Reference Figure 1 A method for tracing spring defect data related to a production batch, specifically including the following steps: S1: Obtain all candidate production batches corresponding to the spring defect data sequence to be traced, and collect multiple process parameter sequences for each candidate production batch.

[0029] In modern spring production processes, data sources are scattered and sampling frequencies are inconsistent. Defect data typically originates from downstream quality inspection stages, such as AOI visual inspection or pressure testing, presenting as discrete sequences. In contrast, process parameters such as tempering furnace temperature, spring winding speed, cutting force, and pressure come from sensors in upstream equipment, usually representing high-frequency, continuous time series. In the traceability process of spring defect data, this data reflects the performance degradation exhibited by the finished spring during the final quality inspection stage, including insufficient stiffness, excessive free height, or fatigue life below design values.

[0030] This step aims to build a unified spatiotemporal benchmark, mapping data from different dimensions onto the same timeline. This is a prerequisite for causal tracing and ensures that each defect data point can be accurately mapped to the process state at its production moment, thus avoiding causal tracing errors caused by time misalignment.

[0031] First, obtain the defect data sequence of the springs to be traced. Utilize the quality defect indicators collected in the factory's terminal quality inspection process (such as visual inspection, fully automatic pressure testing instruments, etc.) to construct a data sequence that changes over time. Specifically, this includes key performance indicators such as spring stiffness, free height, and fatigue life. These indicators are recorded in time series form, which can intuitively reflect the impact of production fluctuations on product quality.

[0032] Secondly, the system identifies candidate production batches. Based on the time period corresponding to the defect sequence to be traced, and combined with the production scheduling records of the factory's MES (Manufacturing Execution System), the system traces back to the upstream production process. By comparing the time overlap, all production batches that are in the processing state within that time period are identified as candidate production batches. This mechanism ensures the completeness of the traceability scope, covering all potentially affected processes such as coiling, tempering, and grinding.

[0033] Subsequently, multi-dimensional process parameter acquisition was performed. For each locked candidate production batch, data from multiple process sensors throughout the entire time period were retrieved simultaneously, including but not limited to tempering furnace temperature, spring speed, cutting force, and pressure. These parameters were constructed into multiple categories of process parameter sequences according to the original sampling time sequence, serving as a set of independent variables to be analyzed.

[0034] Finally, timing alignment and resampling are performed. Since the sampling frequencies of quality inspection data and equipment-side sensor data often differ by orders of magnitude—for example, the tempering furnace temperature is recorded once per minute, while the spring speed is recorded multiple times per second—timing alignment and resampling operations must be performed. Using a unified timestamp reference, high-frequency process sequences are downsampled or low-frequency sequences are interpolated to ensure that each process parameter sequence and defect data sequence completely overlaps on the time axis and has the same sequence length. This high-precision alignment operation eliminates time misalignment caused by physical transmission delays or equipment clock asynchrony, which is the foundation for ensuring the accuracy of subsequent analysis.

[0035] A specific example is: When quality inspection revealed that a batch of springs had a problem with inconsistent stiffness: (1) Obtain the data sequence of spring defects to be traced: The automatic pressure testing machine detects one spring per second and records the stiffness deviation of 300 consecutive springs to generate the data sequence of spring defects to be traced. The recording time period is from 09:00 to 09:16 in the morning.

[0036] (2) Locking out candidate production batches: Based on the quality inspection time of 09:00, deduct the 2-hour delay time required for grinding and cooling, trace back to the production process, and lock out batch A and batch B, which were operating in the spring coiling machine and tempering furnace at around 07:00, as candidate production batches.

[0037] (3) Collect multiple process parameter sequences for each candidate production batch: Collect the process parameter sequences for batch A and batch B respectively, including: tempering furnace temperature, the sensor records the temperature once every 10 seconds to form the tempering furnace temperature sequence; spring speed, the servo motor records the speed once every 0.1 seconds to form the spring speed sequence.

[0038] (4) Resampling and alignment: The timestamp is mapped to the product serial number through the production counter. If the 100th spring is produced at 07:08:30, the tempering furnace temperature value at that time is established in correspondence with the stiffness deviation of the spring. In this way, it is ensured that the process parameter sequence of batch A and batch B is the same as the length of the defect data sequence of the spring to be traced, and the index is consistent.

[0039] like Figure 2As shown, the evolution of various process parameter sequences and the defect data sequence to be traced after time-series alignment of batch A is compared. The shared horizontal axis in the figure is a unified time index. The upper curve shows that the tempering furnace temperature exhibits a slow downward trend after the 80th value. The middle curve shows that the coiling speed exhibits high-frequency and violent oscillations around the baseline value, especially with extremely large fluctuations in the range of the 120th to 220th index values. The lower curve shows that the defect index (stiffness deviation) shows a gradually increasing deterioration trend over time. If only the statistical characteristics of the waveform amplitude are observed, the coiling speed in the middle layer, due to its large variance and violent fluctuations, is easily misjudged as the main cause of system instability in traditional correlation analysis. Although the upper tempering furnace temperature exhibits a physical drift that is negatively correlated with the defect trend, its small single-step changes are easily drowned out by background noise. Therefore, it is necessary to look beyond the appearance of high-frequency fluctuations and capture the weak signals that truly cause defects from the physical dimension of energy accumulation.

[0040] S2: For each type of process parameter sequence, calculate the cumulative effective load value of that type of process parameter sequence based on the numerical drift of that type of process parameter sequence.

[0041] After completing the time-series alignment of the data, considering that spring manufacturing is a thermomechanical coupling process involving metal plastic deformation and recrystallization, its quality defects (such as elasticity decay and fracture) are often not caused by instantaneous parameter jumps, but by the accumulation of continuous energy input deviations. For example, in the stress-relief annealing process, the furnace temperature sensor may generate instantaneous spikes due to electrical interference. Such fluctuations, which are extremely short in duration, are limited by the thermal conductivity and specific heat capacity of the metal material and cannot penetrate into the spring core to cause microstructural transformation. Conversely, if the furnace temperature control module ages, causing the temperature to remain below the set lower limit for a long time, even if the deviation is not large, insufficient heat accumulation will lead to incomplete stress relief and subsequent fatigue fracture.

[0042] Therefore, this step aims to quantify the cumulative pressure on the spring mass caused by the process parameters of each candidate production batch from the perspective of physical energy, transform discrete parameter fluctuations into continuous energy potential, calculate the cumulative effective load value of each type of process parameter sequence for each candidate batch, eliminate invalid fluctuation signals that are large in amplitude but extremely short in duration from the source, and construct a physical benchmark that reflects the actual thermodynamic effects.

[0043] In one embodiment, the cumulative payload value of each type of process parameter sequence in each candidate batch satisfies the following relationship: First, calculate the effective load for each value in each type of process parameter sequence. For the first... The first candidate production batch Class of process parameter sequence, the first The first candidate production batch The first of the process parameter sequence A number of effective payload values:

[0044] in, For the first The first candidate production batch The first of the process parameter sequence A number of effective payload values, For the first The first candidate production batch The first of the process parameter sequence A number, , and These are the first and second batches of the pre-obtained defect-free production batches. The mean, standard deviation, and time decay constant of the process parameter series. For the first The first candidate production batch The first of the process parameter sequence The cumulative drift time at each value, It is the absolute value symbol. It is a natural exponential function. This is a preset zero-prevention parameter used to prevent calculation errors caused by a denominator of 0. It is usually set to a very small positive number. .

[0045] The core of this relationship is to simulate the nonlinear damage accumulation process of spring materials under the influence of process deviations during heat treatment or forming processes. The first term... middle, This represents the absolute magnitude of the deviation of the process parameters from the ideal state, i.e., the instantaneous load. By amplifying the absolute magnitude by a power of two, the energy magnitude of the deviation of this type of process parameter is simulated. This is done using signal processing theory, because in signal processing, the square of the amplitude is usually used to reflect the energy magnitude. The larger the energy deviation of this type of process parameter, the more severe the deviation, and the more likely it is to cause defects. The second term... It utilizes the time decay constant A filtering mechanism based on physical inertia was established. For short-term, occasional deviations in process parameters, considering that such occasional fluctuations do not affect the physical properties of the spring, they are filtered out as noise interference. These occasional deviations, due to their accumulated drift time... If the length is too short, the second term approaches 0, effectively suppressing the misleading effect of random noise on the tracing results. However, for continuous deviations in process parameters, considering that such continuous parameter deviations affect the physical properties of the spring, they are more likely to be the target of defect tracing. Therefore, the second term approaches 1, reflecting the impact of such continuous parameter deviations on the physical properties of the spring. Through deep coupling of amplitude and time, the formula is the effective load. The calculation constructs an integrator with low-frequency triggering characteristics, which can automatically eliminate harmless process fluctuations and accurately extract effective physical loads with substantial destructive energy, laying the foundation for distinguishing between statistical correlation and real-cause bias.

[0046] This relationship is based on the response curve of a first-order linear time-invariant system in control engineering and thermodynamics (such as the charging and discharging of RC circuits or Newton's law of cooling). The quadratic operation in the first half eliminates the positive and negative polarities and amplifies the amplitude energy. The second half is based on the time ratio and is converted into a dimensionless attenuation coefficient. The overall output is a dimensionless value, which conforms to the physical objective law that the cumulative potential energy is nonlinearly related to time.

[0047] Then, the effective loads of all values ​​in each type of process parameter sequence are summed to obtain the cumulative effective load value for that type of process parameter sequence. For example, the first... The first candidate production batch The effective loads of all values ​​in the sequence of process parameters are summed to obtain the first value. The first candidate production batch Cumulative effective load value of a series of process parameters.

[0048] like Figure 3 As shown, the cumulative effective load of the tempering furnace temperature sequence exhibits a clear increasing trend, while the effective load of the coil spring speed sequence remains at a low background noise level and does not show an increasing trend. The calculation of the effective load in this invention incorporates the cumulative drift duration and time decay constant. For the coil spring speed sequence, although its original waveform fluctuates violently, it exhibits an alternating positive and negative random walk, lacking a sustained deviation polarity, leading to the following in the formula: The term approaches 0, thus suppressing the effective load value. Conversely, although the tempering furnace temperature sequence has a small instantaneous deviation, its energy is effectively accumulated by the algorithm due to its continuous unidirectional drift. This step successfully simulates the physical process of the thermal inertia and energy accumulation of the spring, effectively distinguishing between invalid high-frequency noise and causal deviations with substantial physical destructive force.

[0049] In one embodiment, the time decay constant and the cumulative drift duration are determined based on the following: For the time decay constant: calculate the first time of the pre-obtained defect-free production batch. The autocorrelation coefficients of process parameter sequences at different lag times are decayed from 1 to... The corresponding lag time length determines the first The time decay constant of the process parameter sequence, where, It is a natural constant.

[0050] Specifically, the time decay constant is essentially the memory duration (i.e., physical inertia) of a measured process parameter. First, defect-free batch data of this type of process parameter is selected as a standard sample, and the autocorrelation coefficient of the standard sample at different lag times is calculated. When the lag time is... When the standard sample is perfectly correlated with itself, the autocorrelation coefficient is . As the lag time increases, the autocorrelation coefficient of the standard sample will increase from... Gradually decreasing trend This means that as the lag time increases, the correlation between data weakens. Therefore, a standard physical decay threshold, i.e., the natural constant, should be set. The reciprocal of, approximately equal to In physics, this value typically represents a critical point where energy or signal decays to a significant change, when the autocorrelation coefficient first drops to a certain level. When the autocorrelation coefficient is used, the lag time length corresponding to it is taken as the time decay constant of this type of process parameter.

[0051] For tempering furnace temperatures, thermal inertia is high and memory is long; the current temperature is very close to the temperature 10 seconds ago, showing a high correlation. This correlation typically decays only after a slow period, such as 600 seconds. Then the time decay constant of the furnace temperature is This indicates that it is a slow variable. For spring speed, it exhibits rapid fluctuations and short memory; the current speed and the speed from one second ago both show random jumps. For example, if only 0.5 seconds have passed, the autocorrelation decays to zero. The following is the time decay constant of the coil spring speed. This indicates that it is a fast variable.

[0052] Regarding cumulative drift duration: from the perspective of the first... The first candidate production batch A series of process parameters, from Start tracing back, according to Each previous value and The sign of the difference indicates the sum of each value. The deviation polarity is statistically consistent for a length of time, and the cumulative drift duration is taken as the length of time the deviation polarity remains consistent.

[0053] Specifically, the cumulative drift duration is essentially a deviation polarity counter. It is used to count how long the process parameters have deviated from the standard average. If the temperature is too high at the current moment, it will go back to see how many consecutive moments it has been too high. Once it finds that the temperature was too low at a previous moment (the polarity has changed), the counting will stop.

[0054] First, determine the polarity of the deviation at the current moment, using... Subtract the average value of the defect-free batches If the result is positive, it indicates that the current state is relatively high; if the result is negative, it indicates that the current state is relatively low. Then, by backtracking along the time axis, for... , Then, check the deviation polarity of each data point one by one. Finally, count the consecutive length. Count how many consecutive values ​​before the current time have the same deviation polarity as the current time. As long as the deviation signs are the same, they are all positive or all negative, the counter is incremented by 1. Once a value with a deviation sign opposite to the current one or zero is encountered, the count is immediately terminated. The final accumulated number of values ​​is the accumulated drift time.

[0055] For example, the standard average temperature of a batch of tempering furnaces The tempering furnace temperature sequence is used to calculate the cumulative drift time of the 5th temperature, subtracting the following values ​​sequentially. The first temperature The temperature is too low, the deviation is negative, and it is the second temperature. The temperature is too low, the deviation is negative, and it is the third temperature. The temperature is too high, and the deviation polarity is positive. This is the polarity reversal point, the fourth temperature. The temperature is too high, with a positive deviation polarity, and is the 5th temperature. The value is too high, and the polarity of the deviation is positive.

[0056] Ultimately, the cumulative drift time of the 5th temperature is 3 (the length of 3 sampling times), which includes the time length corresponding to the three consecutively higher temperatures of the 5th, 4th, and 3rd.

[0057] S3: Determine the drift directionality coefficient of each type of process parameter sequence based on the cumulative effective load value of each type of process parameter sequence.

[0058] After determining the cumulative load of the process parameters, it is also necessary to identify the pattern of their changes. Physical mechanisms show that equipment failures usually lead to monotonic trend drift of parameters, while environmental disturbances often cause bidirectional random oscillations. Therefore, this step aims to construct a trend discriminator by analyzing the gradient direction and rate stability of parameter changes. The goal is to identify, from the perspective of time-domain morphology, whether the parameter sequence has a structural trend that leads to continuous quality deterioration, thereby eliminating random disturbances that, although fluctuating greatly, have no clear direction.

[0059] In one embodiment, the drift directionality coefficient satisfies the following relationship:

[0060] in, and The first The first candidate production batch The drift directionality coefficient and cumulative effective load of the process parameter sequence. and They are the first The first candidate production batch The index and length of the process parameter sequence. For symbolic functions, and The first The first candidate production batch The first of the process parameter sequence The number and the first A number of effective payload values, For the first defect-free production batch obtained in advance The average value of the process parameter sequence, To prevent positive numbers with a denominator of 0, For the first The reciprocal of the mean of the cumulative effective load values ​​of all types of process parameter sequences for each candidate production batch. It is the hyperbolic tangent function. It is the absolute value symbol.

[0061] In this relationship, the drift directionality coefficient The introduction of this technology aims to accurately capture the unidirectional offset trend of process anomalies, for part, It uses a sign function to extract the direction of parameter offset, in order to determine the effective load. Weighted averaging is applied to the weights. When process parameters show a trend of deviation, the effective loads in the same direction will have a superposition effect, causing the molecules to increase. If the process is only a disordered bidirectional oscillation, the effective loads in different directions will cancel each other out during the accumulation process, thereby revealing the unidirectional physical cause that truly leads to the defect. For One part provides a nonlinear modulation function, utilizing Adjusting the saturation of the directional weights means that the directional features will only be given the highest weight when the parameter deviation accumulates to a sufficiently high level, ensuring the robustness of the source tracing logic in complex industrial noise environments.

[0062] This relationship is constructed based on the AC / DC power ratio in signal processing and the hyperbolic tangent activation function in artificial neural networks. Equipment failures typically cause monotonic DC drift, while environmental disturbances exhibit positive and negative symmetrical AC oscillations. A dimensionless sign function is used to extract polarity, and... The extreme value convergence property of a function mathematically prevents infinity and physically conforms to the law that the trend of deviation has a saturation upper limit. Weighted averaging extracts the trend polarity, while introducing... The activation function maps the infinitely divergent cumulative error to a standard range, thus avoiding individual extreme outliers.

[0063] S4: Combine the cumulative effective load and drift directionality coefficient of each type of process parameter sequence to determine the physical cause weighting coefficient of that type of process parameter sequence.

[0064] Through the first two steps, the features of the energy dimension and the morphology dimension were obtained respectively. This step aims to construct a dual-constraint fusion model, which requires that both energy and trend satisfy the significance condition before it is considered that the parameter is more likely to cause defects.

[0065] In one embodiment, the physical causation weighting coefficients satisfy the following relationship:

[0066] in, For the first The first candidate production batch Physical cause weighting coefficients for process parameter sequences. For the first The first candidate production batch Cumulative effective load value of a series of process parameters The standard deviation of the drift directionality coefficients for all types of process parameter sequences in a pre-obtained defect-free production batch is given. It is the standard deviation of the cumulative effective values ​​of the sequence of process parameters of all types in a pre-acquired defect-free production batch.

[0067] In this formula, the standard deviation of a defect-free baseline is used. and A physical activation threshold was set. The weighting coefficients rapidly saturate to 1 only when the process parameters exhibit significant anomalies in both total load and offset direction. Such parameters are identified as potential physical causative factors. (First term) In the middle, when much smaller When (i.e., within the normal fluctuation range), this term approaches 0. Much larger When a significant anomaly occurs, this term approaches 1; the second term In the middle, when much smaller When (i.e., within the normal fluctuation range), this term approaches 0. Much larger When there is a significant abnormality, this item approaches 1.

[0068] The multiplication operation implements the functionality of logical AND, meaning that it only applies when the cumulative load of the process parameters significantly exceeds the limit and the drift directionality is significant. Only then will it approach 1. The squaring operation is used to amplify the nonlinear response of the deviation feature and strengthen the significance of abnormal fluctuations. This coefficient suppresses false signals. Only parameter changes that simultaneously possess high energy and strong trend characteristics in terms of physical mechanism are judged to be high-risk, thus avoiding misjudgments caused by a single feature.

[0069] The two terms of this relation construct a nonlinear threshold activation mechanism similar to the Hill equation, enabling the rapid identification and extraction of process parameters that truly cause physical damage. The multiplication of the two terms constructs a logical AND gate, employing a square operation instead of a simple linear ratio to create a steep activation curve. This means that only when the process deviation simultaneously crosses [a threshold value] in both load and direction... The weights are activated instantly only when the threshold is reached; otherwise, they will be suppressed to the noise level, thus accurately eliminating spurious correlations.

[0070] S5: Based on the Granger causality test principle, a sequence of process parameters for each type is introduced to predict the data sequence of spring defects to be traced, and the prediction error is corrected by the physical cause weighting coefficient.

[0071] This step employs a binary regression model, using the defect data sequence of the springs to be traced as the dependent variable and the process parameter sequence of the candidate production batches and its lag terms as the independent variables to construct a prediction model. Unlike conventional statistical prediction, this step introduces a physical causal weighting coefficient as a correction factor after obtaining the prediction error at the mathematical level. The core logic is that a process parameter that truly causes a defect should not only significantly reduce the prediction residual in mathematical regression (i.e., high fit), but also must have sufficient energy accumulation and directional characteristics in physics. By using the reciprocal of the physical coefficient as a penalty term to apply to the prediction error, the model is forced to forget those pseudo-causal parameters that only have statistical correlation but lack physical destructive power.

[0072] In one embodiment, based on the Granger causality test principle, each type of process parameter sequence is introduced to predict the data sequence of spring defects to be traced, and the prediction error of that type of process parameter sequence is obtained, including: For each value in the traceable spring defect data sequence, according to the index of the value, all values ​​before the index are divided in both the traceable spring defect data sequence and the process parameter sequence of this type, forming two sequences. The two sequences are used to perform binary regression fitting on the value to obtain the first predicted value of the value. The difference between the value and the first predicted value is calculated to obtain the first prediction residual. The variance of the first prediction residuals of all values ​​is used as the prediction error of this type of process parameter sequence.

[0073] Specifically: Define the spring defect data sequence to be traced as follows: ,in, For indexing, For the first The first candidate production batch The first of the process parameter sequence A numerical value is used to set the lag order for the regression analysis. Regarding the first One index is used to extract historical items from the data sequence of spring defects to be traced. and historical items of process parameters Construct feature vectors.

[0074] Based on these two historical terms, a linear regression prediction equation is constructed according to the construction method of the classic linear regression prediction model. The least squares method is used to estimate the parameters of the above equation. After solving the linear regression prediction equation, the predicted value (first predicted value) of the t-th value of the spring defect data sequence to be traced is obtained.

[0075] Subtract the predicted value from the t-th value to obtain the prediction residual (first prediction residual) of the t-th value. Calculate the variance of the prediction residuals of all values ​​in the spring defect data sequence to be traced, and use this variance as the first prediction residual. The first candidate production batch The prediction error of a series of process parameters. The smaller this value, the higher the degree of fit of the process parameter to the defect fluctuation in pure statistics.

[0076] In one embodiment, the method of correcting the prediction error of a process parameter sequence using the physical cause weighting coefficient of each type of process parameter sequence is as follows: multiply the prediction error of each type of process parameter sequence by the reciprocal of the physical cause weighting coefficient of that type of process parameter sequence to obtain the corrected prediction error of that type of process parameter sequence.

[0077] For batches that are only correlated at the data level but have low physical causation weighting, the penalty factor amplifies their prediction error, causing them to be automatically downweighted when calculating Granger causality; while the causality scores of anomalous batches that truly conform to the physical mechanism are retained. This mechanism not only solves the problem of misjudgment caused by thermal inertia and physical hysteresis in the spring heat treatment process, but also ensures that the traceability results can accurately pinpoint the responsible batch and specific process parameters that caused the finished product performance defects.

[0078] If the physical cause weighting coefficient of a certain type of process parameter is extremely small, its reciprocal will act as a large penalty factor to amplify the original error; if the physical cause weighting coefficient of a certain type of process parameter is close to 1 (physical cause is significant), the error remains basically unchanged.

[0079] For example, if the source to be traced is a batch of springs with gradually decreasing stiffness, and the system performs binary regression fitting on the spring speed sequence, it finds that the high-frequency fluctuations in the spring speed sequence can barely match the defect fluctuations, and the original prediction error is calculated to be... However, due to the lack of effective energy accumulation in the coil spring velocity sequence, the weighting coefficient of the physical cause calculated in the preceding steps is... The prediction error is corrected by using a physical cause weighting coefficient based on the spring speed sequence. As can be seen, the prediction error was amplified by 10 times, indicating that under physical constraints, this parameter increased the uncertainty of the system.

[0080] For the tempering furnace temperature series, the system performed binary regression fitting and found that the gradual decrease in temperature could reflect stiffness degradation. The original prediction error was [missing value]. Meanwhile, due to its significant thermal inertia, the physical causative weighting coefficient is... The corrected prediction error is: As can be seen, the prediction error remained at an extremely low level, confirming its dual significance at both the statistical and physical levels.

[0081] S6: Calculate the Granger causality of each type of process parameter sequence and the spring defect data sequence to be traced based on the corrected prediction error, and determine the tracing result based on the Granger causality.

[0082] This step aims to establish a benchmark for causality determination, and to judge variables based on the Granger causality test principle. Is it a variable? The key reason lies in comparing "only using" "Predictive effect of its own historical information" and "Introduction" To assess the predictive effectiveness, this step first uses an autoregressive model to calculate the inherent fluctuation error of the defect sequence to be traced, using it as the system's baseline prediction error, i.e., the uncertainty under the null hypothesis. Subsequently, by calculating the logarithmic ratio of the baseline error to the corrected prediction error in step S5, Granger causality is obtained, which intuitively reflects how much additional information gain the introduction of this process parameter contributes to the interpretation of defect data, considering the physical mechanism.

[0083] In one embodiment, Granger causality is determined based on the following: For each value in the traceable spring defect data sequence, according to the index of that value, all values ​​before that index are divided into a sequence. This sequence is then used to perform an autoregressive fit on the value to obtain a second predicted value. The difference between the predicted value and the second predicted value is calculated to obtain a second prediction residual. The variance of the second prediction residuals for all values ​​is taken as the prediction error of the traceable spring defect data sequence. The ratio of the prediction error of the traceable spring defect data sequence to the corrected prediction error of the process parameter sequence is taken as the Granger causality between the process parameter sequence and the traceable spring defect data sequence.

[0084] Specifically, to establish a benchmark, the system excludes all external process parameters and performs autoregressive operations only on the spring defect data sequence to be traced. Based on this sequence, an autoregressive model is constructed using the classic autoregressive modeling method. This model simulates the fluctuation of spring mass data due to its own inertia without any external process parameter interference. The parameters of this model are estimated by fitting using the least squares method, thereby obtaining the predicted value (second predicted value) of each value in the sequence. Then, each value is subtracted from its predicted value to obtain the predicted residual (second predicted residual). The variance of all predicted residuals is calculated as the prediction error of the spring defect data sequence to be traced. This is the prediction error of the defect data sequence itself without introducing any process parameter sequence, which is the benchmark prediction error of the defect data sequence.

[0085] Based on information theory principles, the following formula for the ratio of the natural logarithm is used to calculate the first... The first batch Granger causality of process parameter sequences :

[0086] in, It is the prediction error of the spring defect data sequence to be traced (the baseline prediction error of the defect data sequence), reflecting the inherent fluctuations of the defect data itself. It is the first The first batch The prediction error after correction of the process parameter sequence reflects the residual fluctuation after introducing process parameters and undergoing physical correction. If This indicates that there is no obvious causal relationship between the process parameters and the spring defect; if A positive value and a larger value indicate that the introduction of this process parameter significantly reduces the prediction error. This process parameter and the spring defect have a clear causal relationship and belong to the core physical cause.

[0087] The construction of this relation is based on Shannon's information theory and the log-likelihood ratio test. The ratio of the predicted residual variance is taken as logarithm, which essentially measures "the amount of reduction in system uncertainty (information entropy) after introducing specific process parameters". This is an accurate criterion for proving the strength of causality.

[0088] In one embodiment, determining the tracing result of the spring defect data sequence to be traced based on the Granger causality of the process parameter sequences of all candidate production batches and the spring defect data sequence to be traced includes: uniformly sorting the Granger causality of all process parameter sequences of all candidate production batches, determining the process parameter category of the process parameter sequence corresponding to the Granger causality with the largest value as the causative parameter causing the spring defect, and determining the batch corresponding to the process parameter sequence corresponding to the Granger causality with the largest value as the responsible batch causing the spring defect.

[0089] like Figure 4 As shown, the comparison of Granger causality before and after the introduction of the physical causality weighting coefficient is demonstrated, verifying the technical effectiveness of this invention in solving false causality misjudgments. Based on the causality calculated by the traditional Granger causality test algorithm, it is shown that the causality of the coil spring speed sequence is as high as 0.78, while the causality of the tempering furnace temperature sequence is 0.42, which means that the probability of the coil spring speed sequence causing defects is higher than that of the tempering furnace temperature sequence. This invention introduces a physical causation weighting correction to calculate causality. The causality of the spring speed sequence drops significantly to 0.12, while the causality of the tempering furnace temperature sequence increases to 0.91. The root cause of this difference is that existing technologies are based solely on statistical principles, calculating causality by simply comparing the reduction in variance of the prediction residuals. Therefore, they are easily misled by the spurious correlations presented by the high-frequency and violent fluctuations of the spring speed sequence. This invention uses a physical causation weighting coefficient to perform nonlinear correction on the prediction error. Since the spring speed lacks effective physical load accumulation, its weighting coefficient is extremely low, causing it to be penalized in the corrected algorithm, thus significantly reducing its final Granger causality value. This means that the probability that the spring speed sequence is the cause of the defect is lower than the probability that the tempering furnace temperature is the cause of the defect.

[0090] This invention upgrades Granger causality from a purely statistical indicator to an indicator that couples physics and statistics, thereby suppressing interference from fluctuations in process parameters and amplifying the physical causes of real defects, eliminating the pseudo-causal interference that is common in industrial settings.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for tracing spring defect data related to a production batch, characterized in that, include: Obtain all candidate production batches corresponding to the spring defect data sequence to be traced, and collect multiple process parameter sequences for each candidate production batch; For each type of process parameter sequence, the cumulative effective load value of the process parameter sequence is calculated based on the numerical drift of the process parameter sequence, and the drift directionality coefficient of the process parameter sequence is determined based on the cumulative effective load value. By combining the cumulative effective load value and the drift directionality coefficient, the physical causal weighting coefficient of this type of process parameter sequence is determined; Based on the Granger causality test principle, each type of process parameter sequence is introduced to predict the data sequence of spring defects to be traced, and the prediction error of the process parameter sequence is obtained. The prediction error is corrected by the physical cause weighting coefficient of the process parameter sequence. The Granger causality of the process parameter sequence and the spring defect data sequence to be traced is calculated based on the corrected prediction error. The traceability result of the spring defect data sequence to be traced is determined based on the Granger causality of the process parameter sequence and the spring defect data sequence to be traced for all candidate production batches. The cumulative effective payload value satisfies the following relationship: ; For the first The first candidate production batch The first of the process parameter sequence A number of effective payload values, For the first The first candidate production batch The first of the process parameter sequence A number, , and These are the first and second batches of the pre-obtained defect-free production batches. The mean, standard deviation, and time decay constant of the process parameter series. For the first The first candidate production batch The first of the process parameter sequence The cumulative drift time at each value, It is the absolute value symbol. For the natural exponential function; let the first... The first candidate production batch The effective loads of all values ​​in the sequence of process parameters are summed to obtain the first value. The first candidate production batch Cumulative effective load value of a series of process parameters The preset anti-zero parameter; The drift directionality coefficient satisfies the following relationship: ;in, and The first The first candidate production batch The drift directionality coefficient and cumulative effective load of the process parameter sequence. and They are the first The first candidate production batch The index and length of the process parameter sequence. For symbolic functions, For the first The reciprocal of the average cumulative value of the effective load of all types of process parameter sequences for a candidate production batch. It is the hyperbolic tangent function.

2. The spring defect data tracing method according to claim 1, characterized in that, The time decay constant and cumulative drift duration are determined based on the following methods: Calculate the first defective production batch obtained in advance. The autocorrelation coefficients of process parameter sequences at different lag times are decayed from 1 to... The corresponding lag time length determines the first The time decay constant of the process parameter sequence, where, It is a natural constant; From the perspective of the first The first candidate production batch A series of process parameters, from Start tracing back, according to Each previous value and The sign of the difference indicates the sum of each value. The deviation polarity is statistically consistent for a length of time, and the cumulative drift duration is taken as the length of time the deviation polarity remains consistent.

3. The spring defect data tracing method according to claim 1, characterized in that, The physical cause weighting coefficients satisfy the following relationship: ;in, For the first The first candidate production batch Physical cause weighting coefficients for process parameter sequences. For the first The first candidate production batch Cumulative effective load value of a series of process parameters The standard deviation of the drift directionality coefficients for all types of process parameter sequences in a pre-obtained defect-free production batch is given. It is the standard deviation of the cumulative effective values ​​of the sequence of process parameters of all types in a pre-acquired defect-free production batch.

4. The spring defect data tracing method according to claim 1, characterized in that, Based on the Granger causality test principle, each type of process parameter sequence is used to predict the data sequence of spring defects to be traced, and the prediction error of this type of process parameter sequence is obtained, including: For each value in the traceable spring defect data sequence, according to the index of the value, all values ​​before the index are divided in both the traceable spring defect data sequence and the process parameter sequence of this type, forming two sequences. The two sequences are used to perform binary regression fitting on the value to obtain the first predicted value of the value. The difference between the value and the first predicted value is calculated to obtain the first prediction residual. The variance of the first prediction residuals of all values ​​is used as the prediction error of this type of process parameter sequence.

5. The spring defect data tracing method according to claim 1, characterized in that, The method for correcting prediction errors is as follows: The corrected prediction error for each type of process parameter sequence is obtained by multiplying the prediction error of each type of process parameter sequence by the reciprocal of the physical cause weighting coefficient of that type of process parameter sequence.

6. The spring defect data tracing method according to claim 1, characterized in that, Granger causality is determined based on the following method: For each value in the spring defect data sequence to be traced, according to the index of the value, all values ​​before the index are divided in the spring defect data sequence to be traced, forming a sequence. The sequence is used to perform autoregression fitting on the value to obtain the second predicted value of the value. The difference between the value and the second predicted value is calculated to obtain the second prediction residual. The variance of the second prediction residuals of all values ​​is used as the prediction error of the spring defect data sequence to be traced. The ratio of the prediction error of the spring defect data sequence to the corrected prediction error of the process parameter sequence is used as the logarithmic function value of the ratio to determine the Granger causality between the process parameter sequence and the spring defect data sequence.

7. The spring defect data tracing method according to claim 1, characterized in that, The tracing results of the spring defect data sequence to be traced were determined based on the following method: The Granger causality of all process parameter sequences in all candidate production batches is uniformly sorted. The process parameter category corresponding to the process parameter sequence with the largest Granger causality value is determined as the causative parameter causing the spring defect. The batch corresponding to the process parameter sequence with the largest Granger causality value is determined as the responsible batch causing the spring defect.

8. The spring defect data tracing method according to claim 1, characterized in that, The sequence of process parameters for multiple categories is determined based on the following methods: The time series of spring quality defect indicators detected in the terminal quality inspection process is used as the spring defect data sequence to be traced. The defect indicators include stiffness, free height, and fatigue life. Based on the time period corresponding to the spring defect data sequence to be traced, the production process is traced back to identify all production batches that overlap with the time period as candidate production batches. The various process parameters of each candidate production batch throughout the entire time period are arranged into multiple categories of process parameter sequences according to the time sequence. The process parameter sequences of each category are aligned with the spring defect data sequence to be traced through time sequence alignment and resampling operations. Among them, the various process parameters include tempering furnace temperature, spring winding speed, cutting force, and pressure.

Images