Method for predicting conducted interference test curve interval of mass production equipment

Abstract

The invention relates to the technical field of interval prediction, in particular to a conducted interference test curve interval prediction method for mass production equipment, which comprises the following steps: selecting part of equipment to be tested to perform conducted interference test in a frequency band to be tested to obtain a test curve, and dividing a prediction frequency band and a test frequency band based on the test curve, establishing a prediction data set according to the test data of the to-be-tested equipment subjected to the conducted interference test and the industrial big data, and establishing an interval prediction model for conducting conducted interference curve prediction on the prediction frequency bands of the remaining to-be-tested equipment; inputting the test data of the test frequency band into the interval prediction model, and predicting the conducted interference curve of the prediction frequency band; evaluating the interval prediction model based on the prediction result; the division standards of the prediction frequency band and the test frequency band are corrected based on the evaluation result, selective prediction of conducted interference of equipment produced in batches is achieved, and then the test cost is reduced while the accuracy of the prediction result is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of interval prediction, and in particular to a method for predicting an interval of a conducted interference test curve of mass-produced equipment. Background Art

[0002] For mass-produced equipment, it is often necessary to test its electromagnetic compatibility (such as sensitivity and conducted interference). Electromagnetic characteristics are usually complex and changeable. Factors such as the type, location, signal strength, and frequency of the interference source may cause the conducted interference test curve to present uncertain patterns and changes, which usually have nonlinear relationships and complex dependencies. Existing mathematical models cannot accurately capture the trends and changing laws of these interferences. In the process of capturing interference trends and laws, the quality and quantity of feature data extraction are crucial to obtaining accurate analysis results. Currently, statistical analysis methods are often used to obtain feature data for single-feature data analysis, such as mean, variance, standard deviation, kurtosis, skewness, maximum value, minimum value, etc. However, this type of feature data extracted based on basic statistical and mathematical methods is usually not suitable for complex sequence data. The extracted features are static and it is difficult to capture dynamic feature changes in the sequence. At the same time, when these features are applied to machine learning or deep learning models, it is often difficult to fully mine the important information in the data, and some key information will inevitably be lost in the process of manually selecting features. Summary of the invention

[0003] The purpose of the present invention is to provide a method for predicting the interval of conducted interference test curves of mass-produced equipment, so as to achieve selective prediction of conducted interference of mass-produced equipment, thereby reducing test costs while ensuring the accuracy of prediction results.

[0004] To this end, the present invention provides a method for predicting the interval of a conducted interference test curve of a mass-produced device, the method comprising:

[0005] Select some devices under test from a single batch to conduct conducted interference test in the frequency band under test to obtain the test curve;

[0006] Divide the frequency band to be tested into a predicted frequency band and a test frequency band based on the test curve of the device to be tested that has been tested for conducted interference;

[0007] Establish a prediction data set based on the test data of the equipment under test that has been tested for conducted interference and industrial big data;

[0008] Decomposing the predicted data set into sequence features by sliding average decomposition and SSA decomposition to obtain trend items, seasonal items, short-term items and noise items of each sequence in the predicted data set;

[0009] Perform stationarity tests on the seasonal term, short-cycle term and noise term of each series;

[0010] Based on the stationarity test result, the sequences decomposed by the sequence features are respectively input into the corresponding models for training, so as to obtain an interval prediction model for predicting the conducted interference curve of the predicted frequency band of the remaining equipment to be tested;

[0011] For each of the remaining devices to be tested, the test data of the test frequency band is input into the interval prediction model to predict the conducted interference curve of the predicted frequency band;

[0012] Evaluating the interval prediction model based on the prediction results, the interval coverage of the prediction results, and the average width of the intervals;

[0013] The division criteria of the predicted frequency band and the test frequency band are modified based on the evaluation result.

[0014] As a preferred technical solution of the conducted interference test curve interval prediction method for mass-produced equipment, the division of the predicted frequency band and the test frequency band specifically includes:

[0015] Extracting a portion of the frequency band to be tested whose test value is continuously higher than the first threshold as the test frequency band;

[0016] Dividing the remaining frequency bands to be tested except the test frequency band into a plurality of sub-frequency bands;

[0017] For the device under test that has been subjected to the conducted interference test, the number of times each sub-frequency band exceeds the limit is counted, and the sub-frequency band whose proportion of times exceeding the limit is greater than the second threshold is used as the test frequency band;

[0018] The frequency band to be tested except the test frequency band is used as the predicted frequency band;

[0019] The first threshold value changes dynamically with the test curve of the device under test that has been subjected to the conducted interference test, so that the proportion of the test frequency band in the frequency band to be tested is a preset ratio.

[0020] As a preferred technical solution for the conducted interference test curve interval prediction method of mass-produced equipment, the method of establishing a prediction data set based on the test data of the equipment under test that has undergone conducted interference testing and industrial big data includes:

[0021] Obtain historical conducted interference test data of devices that use the same power supply mode and signal conduction path as the device under test as prediction data;

[0022] Obtain historical conducted interference test data of a device having the same oscillation frequency and impedance characteristics as the device under test as prediction data;

[0023] Obtain historical conducted interference test data of the same model of equipment as the equipment to be tested as prediction data;

[0024] The above prediction data are integrated into the prediction data set.

[0025] As a preferred technical solution for the conducted interference test curve interval prediction method of mass-produced equipment, the sequence feature decomposition and extraction of the predicted data set by sliding average decomposition and SSA decomposition include:

[0026] Extracting trend items of each sequence in the forecast data set by sliding average decomposition;

[0027] For each sequence in the forecast data set, the remaining terms except the trend term are extracted based on SSA decomposition to obtain seasonal terms, short-term terms and noise terms.

[0028] As a preferred technical solution for the conducted interference test curve interval prediction method of mass-produced equipment, the stability test specifically includes:

[0029] Perform ADF test and KPSS test on the seasonal term, short cycle term and noise term of the sequence respectively;

[0030] If both the ADF test and the KPSS test determine that the sequence is stationary, the sequence is considered to be a stationary sequence, otherwise it is a non-stationary sequence.

[0031] As a preferred technical solution for the conducted interference test curve interval prediction method of mass-produced equipment, the sequence decomposed by the sequence features is input into the corresponding model for training based on the stationarity test result, which specifically includes:

[0032] For the sequences extracted as trend items, and the sequences whose stationarity tests of seasonal items, short-cycle items and noise items are all stationary sequences, they are input into the ARIMA model for prediction.

[0033] As a preferred technical solution for the conducted interference test curve interval prediction method of mass-produced equipment, the sequence decomposed by the sequence features based on the stationarity test results is input into the corresponding model for training, and further includes: for the sequence that is judged to be a stationary sequence with a frequency of 2 in the stationarity test of seasonal terms, short-cycle terms and noise terms, it is input into the LSTM-Bootstrap model for prediction.

[0034] As the preferred technical solution for the interval prediction method of the conducted interference test curve of mass-produced equipment, the sequences that are judged to be stationary sequences with a frequency less than or equal to 1 in the stationarity test of seasonal terms, short-cycle terms and noise terms are input into the CN-LSTM-Bootstrap model for prediction.

[0035] As a preferred technical solution for the interval prediction method of the conducted interference test curve of mass-produced equipment, the evaluation of the interval prediction model based on the prediction results, the interval coverage rate of the prediction results and the average width of the interval includes:

[0036] The evaluation index f1 for the interval coverage is obtained by the following formula:

[0037]

[0038] The evaluation index f2 for the average width of the interval is obtained by the following formula:

[0039] f2=u i -s i ,

[0040] The final evaluation index f is obtained by the following formula: i As a result of the evaluation,

[0041] f i =k1·f1+k2·f2,

[0042] Among them, β is the penalty weight factor, u i and i Represent the upper and lower bounds of the prediction range, y i is the predicted value, k1 and k2 are the consideration weights of interval coverage and interval average width respectively, and k1 decreases gradually with the running progress of the interval prediction model, and k2 increases gradually with the running progress of the interval prediction model.

[0043] As a preferred technical solution of the conducted interference test curve interval prediction method for mass-produced equipment, the modification of the division criteria of the predicted frequency band and the test frequency band based on the evaluation results specifically includes:

[0044] The evaluation index f i Compare with the standard index f0, if f i is greater than f0, the preset ratio is based on the evaluation index f i The difference with the standard index f0 increases, and the increase is positively correlated with the difference.

[0045] The beneficial effects of the present invention are:

[0046] The present invention uses a method to automatically extract meaningful features or attributes from the prediction data set, and these features can effectively describe the core dynamic characteristics of the data. By introducing machine learning and deep learning models, suitable methods are used to predict the characteristics of different components separately, and in the process of separate predictions, the potential correlation between components is used through joint learning and training to further improve the accuracy and efficiency of data analysis and model prediction. Through the frequency band division of prediction and test, the model division of prediction, the evaluation of the interval prediction model, and the correction of the prediction weight based on the evaluation results, the full process adaptive control of the conducted interference characteristics of mass-produced equipment is realized, and the steps in the scheme are closely related, realizing closed-loop analysis and control of prediction and test, and having good reliability in use.

[0047] Furthermore, the present invention can predict the complete detection curve and its changing trend in the conducted interference test project, thereby determining the specific range of the test points under different confidence levels and evaluating whether the product performance and test results are within the allowable threshold.

[0048] Furthermore, the present invention can significantly reduce the amount of testing required for electromagnetic compatibility test items in product research and development and production. By using partial test data to replace the complex electromagnetic model simulation process, the complete test curve of the product can be estimated in advance, helping engineers to quickly identify frequency bands or key parameters that need to be focused on, which is helpful to discover potential problems in advance, improve the efficiency of design verification, reduce trial and error costs, and shorten the development cycle.

[0049] Furthermore, the model is continuously optimized through iterative feedback of test data. As the method is improved and verified in static test scenarios, it can also be migrated to more complex application scenarios to achieve real-time monitoring of equipment electromagnetic interference and capture the electromagnetic laws during equipment operation, thereby providing stronger protection for equipment safety and performance stability.

[0050] Furthermore, the present invention divides the frequency band to be tested into predicted frequency bands and test frequency bands through partial tests. Taking into account that the accuracy of prediction is lower than that of actual testing, prediction is more applied to the frequency band to be tested with relatively smaller impact. The present invention designs an evaluation step for the model and adjusts the division standard of the predicted frequency band and the test frequency band according to the evaluation results, so that the prediction is more in line with the actual scenario, avoiding the impact of inaccurate prediction on equipment quality control.

[0051] Furthermore, in the evaluation of the model, the present invention adopts an evaluation method designed for the characteristics of electromagnetic compatibility, and designs phased evaluation indicators based on the concern about the upper limit value exceeding the conducted interference test threshold in the conducted interference test curve of electromagnetic compatibility. Considering that the standard in the electromagnetic compatibility test mainly defines the qualified range through a threshold curve, during the test, more attention is paid to whether the test value exceeds the upper limit threshold. The obtained evaluation indicator is more affected by the prediction result of the upper limit part, and the obtained evaluation result is more applicable. The modification of the division standard based on the evaluation result is correspondingly more applicable. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of a method for predicting a conducted interference test curve interval of mass-produced equipment in an embodiment of the present invention;

[0053] Figure 2 This is a workflow diagram of the model training phase in an embodiment of the present invention;

[0054] Figure 3 The figure is a flow chart for constructing the CN-LSTM-Bootstrap prediction model in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] In the description of the present invention, it should be noted that the terms “first” and “second” are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0057] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0058] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0059] See also Figure 1 and Figure 2 As shown, this embodiment provides a method for predicting the interval of a conducted interference test curve of a mass-produced device, comprising the following steps:

[0060] Step S1, selecting a portion of the devices to be tested in a single batch to perform a conducted interference test in a frequency band to be tested to obtain a test curve;

[0061] Step S2, dividing the predicted frequency band and the test frequency band in the test frequency band based on the test curve of the device under test that has been tested for conducted interference;

[0062] Step S3, establishing a prediction data set based on the test data of the device under test that has been tested for conducted interference and the industrial big data;

[0063] Step S4, performing sequence feature decomposition on the prediction data set by sliding average decomposition and SSA decomposition to obtain trend items, seasonal items, short-term items and noise items of each sequence in the prediction data set;

[0064] Step S5, performing a stationarity test on the seasonal term, short-cycle term and noise term of each series;

[0065] Step S6, based on the stationarity test result, the sequences decomposed by the sequence features are respectively input into the corresponding models for training, so as to obtain an interval prediction model for predicting the conducted interference curve of the predicted frequency band of the remaining equipment under test;

[0066] Step S7, for each of the remaining devices to be tested, input the test data of the test frequency band into the interval prediction model to predict the conducted interference curve of the predicted frequency band;

[0067] Step S8, evaluating the interval prediction model based on the prediction results, the interval coverage rate of the prediction results, and the average width of the intervals;

[0068] Step S9: modifying the division criteria of the prediction frequency band and the test frequency band based on the evaluation result.

[0069] In the above embodiments, by automatically extracting meaningful features or attributes from the prediction data set, these features can effectively describe the core dynamic characteristics of the data. By introducing machine learning and deep learning models, suitable methods are used to predict the characteristics of different components separately, and in the process of separate predictions, the potential correlation between components is used through joint learning and training to further improve the accuracy and efficiency of data analysis and model prediction. Through the frequency band division of prediction and test, the model division of prediction, the evaluation of the interval prediction model, and the repair of the prediction model based on the evaluation results, the full process adaptive control of the conducted interference characteristics of mass-produced equipment is realized, and the steps in the scheme are closely related, realizing closed-loop analysis and control of prediction and test, and having good reliability in use.

[0070] Furthermore, the prediction method of this embodiment can predict the complete detection curve and its changing trend in the conducted interference test item, thereby determining the specific range of the test points under different confidence levels and evaluating whether the product performance and test results are within the allowable threshold.

[0071] On the basis of the above effects, the prediction method of this embodiment can significantly reduce the amount of testing required for electromagnetic compatibility test items in product development and production. By using partial test data to replace the complex electromagnetic model simulation process, the complete test curve of the product can be estimated in advance, helping engineers to quickly identify the frequency bands or key parameters that need to be focused on, which helps to discover potential problems in advance, improve the efficiency of design verification, reduce trial and error costs, and shorten the development cycle. Then, through iterative feedback of test data, the model is continuously optimized. As the method is improved and verified in static test scenarios, it can also be migrated to more complex application scenarios, realizing real-time monitoring of electromagnetic interference of equipment, capturing electromagnetic laws during equipment operation, and thus providing stronger protection for the safety and performance stability of the equipment.

[0072] Moreover, this embodiment divides the frequency band to be tested into predicted frequency band and test frequency band through partial tests. Taking into account that the accuracy of prediction is lower than that of actual test, prediction is more applied to the frequency band to be tested with relatively small impact. The present invention designs an evaluation step for the model and adjusts the division standard of the predicted frequency band and the test frequency band according to the evaluation results, so that the prediction is more in line with the actual scenario, avoiding the impact of inaccurate prediction on equipment quality control.

[0073] Specifically, the division of the prediction frequency band and the test frequency band includes:

[0074] Step S21, extracting the part of the frequency band to be tested whose test value (in dBμA or dBμV) is continuously higher than the first threshold as the test frequency band;

[0075] Step S22, evenly divide the remaining frequency bands to be tested except the test frequency band into a plurality of sub-frequency bands;

[0076] Step S23, counting the number of times each sub-frequency band exceeds the limit value in the device under test that has been tested for conducted interference, and taking the sub-frequency band whose proportion of the number of times exceeding the limit value (i.e., the number of times the sub-frequency band exceeds the limit / the number of tests) is greater than the second threshold as the test frequency band;

[0077] Step S24, taking the frequency band to be tested except the test frequency band as the predicted frequency band;

[0078] Among them, the first threshold value changes dynamically with the test curve of the device under test that has been tested for conducted interference, so that the proportion of the test frequency band in the frequency band to be tested is a preset ratio. In detail, the dynamic change of the first threshold value makes the ratio of the test frequency band and the prediction frequency band controllable. Maintaining a fixed ratio between the two can make the model operation effect and the prediction accuracy controllable. The purpose of setting the second threshold value is to test the frequency band that is prone to problems and avoid ignoring it in the prediction. In implementation, there is a situation where the test frequency band determined by the second threshold value has exceeded the preset ratio. In this case, the ratio is not controlled by the first threshold value, and the test frequency band determined by the second threshold value is directly tested, and the remaining frequency bands are predicted. The optional value of the second threshold is 0.05, which can be adjusted in combination with the scenario during implementation.

[0079] Specifically, the prediction data set established based on the test data of the equipment under test that has been tested for conducted interference and industrial big data includes:

[0080] Step S31, obtaining historical conducted interference test data of devices that use the same power supply mode and signal conduction path as the device under test as predicted data;

[0081] Step S32, obtaining historical conducted interference test data of a device having the same oscillation frequency and impedance characteristics as the device to be tested as prediction data;

[0082] Step S33, obtaining historical conducted interference test data of a device of the same model as the device to be tested as prediction data;

[0083] Step S34, integrating the above prediction data into a prediction data set.

[0084] Furthermore, the overall trend of the test curve is mainly affected by factors such as the product's filter circuit, parasitic inductance and capacitance, and cables. Before predicting the harmonic terms and coupling terms, the most important thing is to accurately extract and grasp the overall trend changes. Other peaks and fluctuation components are more superimposed in the trend terms. Therefore, the prediction data set is decomposed and extracted through sliding average decomposition and SSA decomposition, including:

[0085] Step S41, extracting the trend item of each sequence in the prediction data set by sliding average decomposition;

[0086] Step S42, for each sequence in the forecast data set, the remaining items except the trend item are extracted based on SSA decomposition to obtain seasonal items, short-term items and noise items. In detail, sliding average decomposition, as a smoothing technique, can effectively extract the overall trend from time series data. By calculating the average value of the data within the sliding window, the impact of short-term fluctuations can be reduced, thereby highlighting the long-term trend. The stability of the trend item is obtained and improved through STL decomposition, which reflects the long-term change trend of the sequence and is the core influencing factor of the future trend of the sequence. It will be directly input into the ARIMA model for training and prediction.

[0087] Specifically, the stability test includes:

[0088] Step S51, performing ADF test and KPSS test on the seasonal term, short-cycle term and noise term of the sequence respectively;

[0089] Step S52: If both the ADF test and the KPSS test determine that the sequence is stationary, the sequence is determined to be a stationary sequence, otherwise it is a non-stationary sequence.

[0090] Specifically, based on the stationarity test results, the sequences that have undergone sequence feature decomposition are input into the corresponding models for training, which specifically includes:

[0091] For the sequences extracted as trend items, and the sequences whose stationarity tests of seasonal items, short-cycle items and noise items are all stationary sequences, they are input into the ARIMA model for prediction.

[0092] Specifically, based on the stationarity test results, the sequences that have undergone sequence feature decomposition are input into the corresponding models for training, which also includes: for the sequences that are judged to be stationary sequences with a frequency of 2 in the stationarity tests of seasonal terms, short-cycle terms and noise terms, they are input into the LSTM-Bootstrap model for prediction.

[0093] In detail, the LSTM (Long Short-Term Memory) model is a variant of a recursive neural network (RNN) specifically designed to handle dependency problems of long sequence data. The structure of LSTM manages information flow based on cell state and multiple gating mechanisms, thereby maintaining long-term dependencies in sequence modeling. In this embodiment, the LSTM model input_features used, that is, the number of input features per time step is 60, the number of model layers is 4, and finally a fully connected layer is added to convert the output of LSTM to target shape 1, that is, single-step prediction, using the complete conduction interference test curve in the historical test of the same type of product as training data, and they are input into the corresponding model after including secondary decomposition and stationarity test to complete model training.

[0094] Specifically, for the sequences that are determined to be stationary in the stationarity test of seasonal terms, short-term terms, and noise terms, the number of times is less than or equal to 1 is input into the CN-LSTM-Bootstrap model for prediction. In detail, the CN-LSTM-Bootstrap prediction model is similar to the LSTM-Bootstrap model in terms of interval prediction, but the similarity influence matrix of Cross-stitch Networks is introduced to jointly interact with multiple models. The construction flow chart is as follows Figure 3 As shown:

[0095] Taking two sequences as an example, two LSTM models with the same basic structure as in the previous section are used to process sequences X and Y respectively. Although their changing rules and characteristics are different, they jointly affect the final performance of the sequence. Especially in the electromagnetic compatibility test scenario, the same interference source will show various forms in the test results with different coupling modes and propagation paths. Ignoring the potential correlation between them will lead to distortion of the prediction of the non-stationary part of the sequence. To solve this problem, the CN-LSTM model can not only fully capture the independent characteristics of each sequence through a joint training mechanism, but also reveal the dependencies and interactions between sequences. Through this collaborative learning, the model is more coordinated when processing the prediction of seasonal terms, short-term terms and noise terms, thereby significantly improving the prediction accuracy of the overall time series. In specific implementation, CN-LSTM introduces an interaction mechanism in the weight matrix of the output gate of each LSTM model. The degree of interaction is expressed by the similarity measurement formula. The output gate weights of the two LSTM models are W1 and W2, respectively. According to the similarity influence matrix (SIM), they interact with each other, then there is

[0096]

[0097] W1′ and W2′ are the output gate weight matrices after interactive update. a in the SIM matrix 11 and a 22 The degree to which each model retains its own information, and a 12 and a 21 is the degree of mutual influence between the two models. If we replace it with the similarity index, we get:

[0098]

[0099] Then W1′ and W2′ are updated as follows:

[0100]

[0101] The calculation method of the indicator p(l) mentioned in the above formula is as follows:

[0102] First, the numerical sequence is symbolized, and each data point is simplified into three basic categories according to its change state: "increase", "decrease" or "unchanged". This process aims to simplify the complexity of the numerical sequence and focus on the formal change relationship between local sequences in order to more effectively capture the trend and fluctuation of the sequence. The specific operation is: by comparing the size relationship between each data point and its previous data point, each data point is classified into one of three states: "increase", "decrease" or "unchanged". Further, based on the magnitude of the change, each state is refined into two situations: "dramatic change" and "gentle change". Through the above processing, the values ​​in each sequence are converted into the following five states, which are converted into clear structured data, where the parameter δ is used to distinguish between drastic changes and gentle changes.

[0103]

[0104] Among them, the threshold parameter δ is used to distinguish the severity and gentleness of spectrum changes. The selection of this value is determined by the difference between the amplitude of each frequency point in the complete test curve and the standard threshold. The smaller the difference, the more sensitive the product is to the conducted interference test at this frequency point, and the greater the potential risk. Therefore, a smaller threshold should be set to more keenly capture changes and make more stringent judgments. Considering that frequency points with large amplitude differences usually represent safer areas, these frequency points are not involved in the determination of parameter δ. Finally, the calculation of parameter δ is based only on those frequency points whose amplitude difference from the standard threshold is less than one tenth of it. The amplitudes of these frequency points are summed and averaged to obtain the final threshold parameter. A s , A m , U, D s , D m They represent a sharp increase, a gentle increase, no change, a sharp decrease, and a gentle decrease. i -x i-1Represents the change between two adjacent data points in a numerical sequence. Specifically, it is the change in the current data point x i With the previous data point x i-1 The difference.

[0105] After discretizing the sequences, the edit distance method in biological genetics is used to measure the similarity between sequences. The similarity of the change patterns between two sequences is studied from the perspective of macroscopic change patterns. The edit distance can quantify their similarity by calculating the minimum number of operations required between sequences (such as insertion, deletion or substitution). After completing the macroscopic pattern analysis, the covariance coefficient Cov(X,Y) of the two sequences is further calculated to quantitatively judge the similarity of the numerical change degree between the sequences. The covariance coefficient is used to measure the synchronous changes of the two sequences, that is, their numerical correlation. Combining the two similarity measures of edit distance and covariance coefficient can comprehensively judge the overall similarity between sequences. On this basis, as the depth of the model increases, the sharing ratio is exponentially decayed, and multiplied by the decay coefficient to finally obtain the sharing ratio index p(l) of the two training models:

[0106]

[0107] Where d(X,Y) is the edit distance between sequences, n is the sequence length, l is the depth of the model, and α is the similarity weight ratio. When it is larger, it pays more attention to the changing trend of the sequence, and when it is smaller, it pays more attention to the quantitative change relationship of the sequence value. λ is the decay rate, and α and λ participate in the iterative training of the model. σ X and σ Y are the standard deviations of series X and Y, respectively.

[0108] In each layer of LSTM, the weight of the output gate is interactively adjusted through the similarity influence matrix, so that the two LSTM models can not only train their own sequences independently, but also learn the information of each other's sequences through the interaction of the output gate weights.

[0109] As the number of network layers increases, the mutual influence between models gradually weakens. By designing a reasonable attenuation mechanism, the interaction between high-level layers can be gradually reduced. This design can ensure that the model fully utilizes the correlation information between sequences at a lower level, while maintaining a certain degree of independence at a high level, so that different non-stationary sequences can be predicted more accurately. The introduction of this similarity influence matrix and the layer-by-layer attenuation mechanism balances the information sharing and independence in the multi-sequence prediction task, reduces the prediction error caused by the interference of irrelevant information between sequences, and improves the stability and overall performance of model training.

[0110] In the above embodiment, by further decomposing the sequence to extract its features, maximizing and utilizing the advantages of different models is the core goal of this embodiment. This method aims to improve the accuracy and reliability of prediction. The ARIMA model (autoregressive integrated moving average model) is suitable for processing stationary sequences. However, if it is directly applied to non-stationary sequences, it may cause the model to fit the data too complex, thereby causing overfitting and reducing the prediction performance. On the other hand, the LSTM model (long short-term memory network) is a powerful nonlinear model that is good at capturing complex patterns, long-term dependencies and dynamic changes in sequences. For the electromagnetic performance test of electronic products with high integration, due to the complex factors such as harmonics, oscillators, clock signals, differential mode and common mode interference, it is often manifested as a reciprocating fluctuation phenomenon with different densities in different frequency bands. Therefore, in order to improve the purity of the features and refine the remaining signal components, it is necessary to decompose the sequence twice.

[0111] In the above embodiment, the three sequences obtained by the previous step are tested for stationarity. This embodiment uses a consistency test, combined with the ADF test (Augmented Dickey-Fuller) and the KPSS test (Kwiatkowski-Phillips-Schmidt-Shin). When the test results of both indicate that the sequence is stationary, the sequence can be judged to be a stationary sequence. The stationary sequence is input into the ARIMA model for prediction, and the LSTM-Bootstrap model is selected for the non-stationary sequence to make full use of its advantages in capturing nonlinearity and long-term dependencies. If there are two or more non-stationary sequences, these sequences will be input into the CN-LSTM-Bootstrap model for joint training to enhance the model's ability to learn the correlation between sequences and further improve the accuracy and efficiency of the prediction.

[0112] For the evaluation of the model, the evaluation of model performance in interval prediction usually combines two key indicators: interval coverage (Prediction Interval Coverage Probability, PICP) and mean interval width (Mean Prediction Interval Width, MPIW). Interval coverage is used to ensure the reliability of the model and ensure that the prediction interval can contain the true value; while the mean interval width reflects the accuracy of the model. The narrower the interval, the higher the accuracy. A good interval prediction evaluation standard must take both indicators into account at the same time, and neither is indispensable. However, these two indicators generally have a competitive relationship. For this reason, this embodiment designs a phased evaluation indicator based on the specific requirements of electromagnetic compatibility testing, especially the concern about the upper limit value exceeding the conducted interference test threshold. The interval prediction model is evaluated based on the prediction results, the interval coverage of the prediction results, and the mean interval width, including:

[0113] Step S81, obtain the evaluation index f1 for the interval coverage rate by the following formula:

[0114]

[0115] Step S82, obtain the evaluation index f2 for the average width of the interval by the following formula:

[0116] f2=u i -s i ,

[0117] Step S83, the final evaluation index f is obtained by the following formula: i As a result of the evaluation,

[0118] f i =k1·f1+k2·f2,

[0119] Among them, β is the penalty weight factor, u i and i Represent the upper and lower bounds of the prediction range, y iis the predicted value, k1 and k2 are the weights of interval coverage and average width of interval respectively (the specific value and the increase or decrease are set in combination with the actual scenario), and k1 gradually decreases with the running process of the interval prediction model, and k2 gradually increases with the running process of the interval prediction model. In the early stage, the focus is on the guarantee of interval coverage to ensure the reliability of the model; in the later stage, the consideration of the average width of the interval is gradually increased to optimize the accuracy of the model, so as to achieve higher accuracy in the upper limit prediction in the electromagnetic compatibility test. In detail, in the formula, f1 and f2 consider the interval coverage and the average width of the interval respectively. f1 first calculates the difference between the predicted value and the midpoint of the interval. If the predicted point does not fall within the interval, the second part will be additionally punished. β is the penalty weight factor. Considering that the national and industry standards in the electromagnetic compatibility test mainly define the qualified range through a threshold curve, during the test, whether the test value exceeds the upper limit threshold is more concerned. In this embodiment, an exponential function is used to distinguish between the situations of exceeding the upper and lower limits of the interval. This phased evaluation mechanism can effectively balance the contradiction between model reliability and accuracy, and meet the dual requirements of electromagnetic compatibility testing for high reliability and accuracy.

[0120] Specifically, the division criteria of the prediction frequency band and the test frequency band are revised based on the evaluation results, including:

[0121] The evaluation index f i Compared with the standard index f0,

[0122] If f i is greater than f0, the preset ratio is based on the evaluation index f i The difference with the standard index f0 increases, and the increase is positively correlated with the difference. i If it is less than or equal to f0, the original preset ratio is maintained or restored to the initial value of the preset ratio.

[0123] Optionally, the correction process specifically includes:

[0124] If f i If f0 is greater than f0, the preset ratio R0 is increased to the corrected preset ratio R by the following formula, and the preset ratio R shall not exceed the upper limit R min and the lower limit R max ,

[0125]

[0126] R = min(max(R, R min ), R max ),

[0127] Specifically, the larger the evaluation index is, the less accurate the prediction result is. At this time, reducing the prediction frequency band can reduce the impact on the conducted interference analysis. As the training data gradually increases, the interval prediction model can also restore the accuracy of the prediction results. After recovery, the preset ratio can be restored, that is, restored to the initial value. Optionally, the initial value of the preset ratio is 0.4.

[0128] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for predicting the interval of a conducted interference test curve of a mass-produced device, characterized in that: include: Select some devices under test from a single batch to conduct conducted interference test in the frequency band under test to obtain the test curve; Divide the frequency band to be tested into a predicted frequency band and a test frequency band based on the test curve of the device to be tested that has been tested for conducted interference; Establish a prediction data set based on the test data of the equipment under test that has been tested for conducted interference and industrial big data; Decomposing the predicted data set into sequence features by sliding average decomposition and SSA decomposition to obtain trend items, seasonal items, short-term items and noise items of each sequence in the predicted data set; Perform stationarity tests on the seasonal term, short-cycle term and noise term of each series; Based on the stationarity test result, the sequences decomposed by the sequence features are respectively input into the corresponding models for training, so as to obtain an interval prediction model for predicting the conducted interference curve of the predicted frequency band of the remaining equipment to be tested; For each of the remaining devices to be tested, the test data of the test frequency band is input into the interval prediction model to predict the conducted interference curve of the predicted frequency band; Evaluating the interval prediction model based on the prediction results, the interval coverage of the prediction results, and the average width of the intervals; The division criteria of the predicted frequency band and the test frequency band are modified based on the evaluation result.

2. The method for predicting the interval of the conducted interference test curve of mass-produced equipment according to claim 1, characterized in that: The division of the predicted frequency band and the test frequency band specifically includes: Extracting a portion of the frequency band to be tested whose test value is continuously higher than the first threshold as the test frequency band; Dividing the remaining frequency bands to be tested except the test frequency band into a plurality of sub-frequency bands; For the device under test that has been subjected to the conducted interference test, the number of times each sub-frequency band exceeds the limit is counted, and the sub-frequency band whose proportion of times exceeding the limit is greater than the second threshold is used as the test frequency band; The frequency band to be tested except the test frequency band is used as the predicted frequency band; The first threshold value changes dynamically with the test curve of the device under test that has been subjected to the conducted interference test, so that the proportion of the test frequency band in the frequency band to be tested is a preset ratio.

3. The method for predicting the interval of the conducted interference test curve of mass production equipment according to claim 1, characterized in that: The establishment of a prediction data set based on the test data of the device under test that has been tested for conducted interference and the industrial big data includes: Obtain historical conducted interference test data of devices that use the same power supply mode and signal conduction path as the device under test as prediction data; Obtain historical conducted interference test data of a device having the same oscillation frequency and impedance characteristics as the device under test as prediction data; Obtain historical conducted interference test data of the same model of equipment as the equipment to be tested as prediction data; The above prediction data are integrated into the prediction data set.

4. The method for predicting the interval of the conducted interference test curve of mass production equipment according to claim 1, characterized in that: The extracting of sequence features from the prediction data set by sliding average decomposition and SSA decomposition includes: Extracting trend items of each sequence in the forecast data set by sliding average decomposition; For each sequence in the forecast data set, the remaining terms except the trend term are extracted based on SSA decomposition to obtain seasonal terms, short-term terms and noise terms.

5. The method for predicting the interval of the conducted interference test curve of mass production equipment according to claim 4, characterized in that: The stability test specifically includes: Perform ADF test and KPSS test on the seasonal term, short cycle term and noise term of the sequence respectively; If both the ADF test and the KPSS test determine that the sequence is stationary, the sequence is considered to be a stationary sequence, otherwise it is a non-stationary sequence.

6. The method for predicting the interval of the conducted interference test curve of mass production equipment according to claim 5, characterized in that: The step of inputting the sequences decomposed by the sequence features into corresponding models for training based on the stationarity test results specifically includes: For the sequences extracted as trend items, and the sequences whose stationarity tests of seasonal items, short-cycle items and noise items are all stationary sequences, they are input into the ARIMA model for prediction.

7. The method for predicting the interval of the conducted interference test curve of mass-produced equipment according to claim 6, characterized in that: The step of inputting the sequences decomposed by the sequence features into corresponding models for training based on the stationarity test results also includes: for the sequences that are determined to be stationary sequences with a frequency of 2 in the stationarity tests of seasonal terms, short-cycle terms and noise terms, inputting them into the LSTM-Bootstrap model for prediction.

8. The method for predicting the interval of the conducted interference test curve of mass-produced equipment according to claim 7, characterized in that: For the sequences that are judged to be stationary sequences with a frequency less than or equal to 1 in the stationarity test of seasonal terms, short-cycle terms, and noise terms, they are input into the CN-LSTM-Bootstrap model for prediction.

9. The method for predicting the interval of the conducted interference test curve of mass production equipment according to claim 2, characterized in that: The evaluating the interval prediction model based on the prediction results, the interval coverage rate of the prediction results and the average width of the intervals includes: The evaluation index f1 for the interval coverage is obtained by the following formula: The evaluation index f2 for the average width of the interval is obtained by the following formula: f2=u i -s i , The final evaluation index f is obtained by the following formula: i As a result of the evaluation, <h2 style=";text-align:left;direction:ltr">f<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> (k1 f1 + k2 f2) Among them, β is the penalty weight factor, u i and i Represent the upper and lower bounds of the prediction range, y i is the predicted value, k1 and k2 are the consideration weights of interval coverage and interval average width respectively, and k1 decreases gradually with the running progress of the interval prediction model, and k2 increases gradually with the running progress of the interval prediction model.

10. The method for predicting the interval of the conducted interference test curve of mass-produced equipment according to claim 9, characterized in that: The modifying of the division standard of the predicted frequency band and the test frequency band based on the evaluation result specifically includes: The evaluation index f i Compared with the standard index f0, If f i is greater than f0, the preset ratio is based on the evaluation index f i The difference with the standard index f0 increases, and the increase is positively correlated with the difference.

Images