A lightweight neural network assisted emi filtering system

Abstract

The application relates to the technical field of electromagnetic compatibility and intelligent filter control, and discloses a light-weight neural network assisted EMI filter system, which comprises a collection module, a conversion module, an extraction module, a judgment module, a correction module and a control module. First EMI observation data under a first measurement method is collected, and second EMI observation data is obtained through conversion of a measurement mapping relationship; features are extracted and input into a light-weight neural network model respectively, so that a first compliance margin result, a second compliance margin result and a measurement method difference result are obtained; boundary offset correction is performed on the smaller one of the compliance margin results, so that a target compliance margin result is obtained, and the target compliance margin result is used to control EMI filter mode switching, filter parameter adjustment or compensation signal output. The application can reduce the influence of differences between different measurement methods on boundary judgment and subsequent control, and improve filter control consistency.

Description

Technical Field

[0001] This application relates to the field of electromagnetic compatibility and intelligent filtering control technology, and more specifically, to a lightweight neural network-assisted EMI filtering system. Background Technology

[0002] Existing electromagnetic interference testing typically requires measuring the target equipment according to the measurement methods specified in the standards. In actual R&D, debugging, in-plant self-testing, and certification testing, different conductive emission measurement methods often coexist under the same standard system. Different measurement methods differ in terms of the observed object, observation location, detection path, and frequency response characteristics, which makes the observation results obtained for the same target equipment under the same operating conditions and the same frequency point not completely consistent. For lightweight neural network-assisted EMI filtering systems, if training, calibration, or online judgment mainly relies on the data corresponding to one of the measurement methods, the judgment results formed under that measurement method are difficult to stably characterize the boundary state under another measurement method.

[0003] Especially at the boundary frequency points close to the limit, even small differences between different measurement methods may cause the compliance margin to change from a positive value to a zero or negative value. This can lead to the system determining that there is still a safety margin under one measurement method, while it is already in an over-limit state under another measurement method. Existing solutions usually lack the ability to handle the differences between different measurement methods in a coordinated manner. It is difficult to effectively convert the observation results obtained under one measurement method to another measurement method for unified judgment, and it is also difficult to further quantify the differences between the two judgment results and correct the boundary judgment results after obtaining the judgment results of different methods.

[0004] For example, during the internal debugging phase of a target device, a single input observation aperture was used to collect data and train a lightweight neural network. The model output showed that several boundary frequency points still had compliance margins. However, during the subsequent certification phase, when a different input conductor observation aperture was used for testing, the test results of the aforementioned boundary frequency points exceeded the limits due to the different measurement links and observation locations. At this point, if the system still maintains the existing filtering mode, filtering parameters, or compensation strategy based on the judgment results under the aforementioned single measurement aperture, it is easy to cause inconsistencies between the judgment basis and the actual boundary state.

[0005] If the above problems are not effectively resolved, inconsistencies will arise between the product development stage, the in-house testing stage, and the third-party certification stage, increasing the number of repeated EMI rectifications, resulting in a lack of stable basis for subsequent filter control, and thus affecting the certification cycle and mass production schedule. Summary of the Invention

[0006] To address the technical problem in existing technologies where differences in measurement aperture between different measurement methods make it difficult for lightweight neural network-assisted EMI filtering systems to stably characterize boundary states under different measurement methods, thereby affecting boundary judgment results and the consistency of subsequent filtering control, this application provides the following technical solution to achieve the above objective: This application discloses a lightweight neural network-assisted EMI filtering system, comprising: The acquisition module is used to acquire the first EMI observation data collected by the target device in its current operating state according to the first measurement method; The conversion module is used to perform an equivalent conversion on the first EMI observation data based on the pre-established measurement mapping relationship between the first measurement method and the second measurement method, so as to obtain the second EMI observation data corresponding to the second measurement method. The extraction module is used to perform feature extraction on the first EMI observation data and the second EMI observation data respectively, to obtain the first EMI feature results and the second EMI feature results; The determination module is used to input the first EMI feature result and the second EMI feature result into the lightweight neural network model to obtain the first compliance margin result and the second compliance margin result, and to generate the measurement method difference result based on the difference between the first compliance margin result and the second compliance margin result. The correction module is used to perform boundary offset correction on the smaller of the first compliance margin result and the second compliance margin result based on the difference in measurement methods, to obtain the target compliance margin result. The control module is used to control the EMI filtering mode switching, filtering parameter adjustment, or compensation signal output of the target device based on the target compliance margin result.

[0007] Compared with related technologies, this application has the following advantages: Compared with related technologies, this application obtains first EMI observation data collected by the target device under the current operating state according to the first measurement method, and performs equivalent transformation on the first EMI observation data according to the pre-established measurement mapping relationship between the first and second measurement methods to obtain second EMI observation data corresponding to the second measurement method; then performs feature extraction on the first and second EMI observation data respectively, and inputs them into a lightweight neural network model to obtain the first compliance margin result and the second compliance margin result; further, generates a measurement method difference result based on the difference between the two compliance margin results, performs boundary offset correction on the smaller of the two compliance margin results to obtain the target compliance margin result, and controls the EMI filtering mode switching, filtering parameter adjustment or compensation signal output of the target device according to the target compliance margin result, thereby reducing the impact of differences between different measurement methods on boundary judgment and subsequent control.

[0008] This application establishes a measurement mapping relationship between the first measurement method and the second measurement method in advance, and performs an equivalent transformation on the first EMI observation data to obtain the second EMI observation data corresponding to the second measurement method. This integrates data that originally belonged to different measurement calibers into the same processing link, reduces data breakpoints between R&D debugging calibers and certification testing calibers, and improves the data correspondence between different measurement methods.

[0009] This application extracts features from the first EMI observation data and the second EMI observation data respectively, and obtains the first compliance margin result and the second compliance margin result respectively. Therefore, instead of relying solely on a single measurement aperture for boundary judgment, it can simultaneously characterize the boundary state under different measurement methods, thereby improving the completeness of the characterization of compliance margin changes at the boundary frequency point.

[0010] This application generates a measurement method difference result based on the difference between the first compliance margin result and the second compliance margin result, and performs boundary offset correction on the smaller compliance margin result according to the measurement method difference result to obtain the target compliance margin result. This allows the difference brought by different measurement methods to be quantified and introduced into the boundary judgment process, reducing the risk of judgment reversal caused by measurement method differences when approaching the limit, and improving the consistency and robustness of the boundary judgment result.

[0011] This application controls the EMI filtering mode switching, filtering parameter adjustment, or compensation signal output of the target equipment based on the target compliance margin result, and reviews the corresponding control results after the control is completed, so that the boundary judgment result can further affect the actual filtering control process, improve the consistency between the subsequent control process and the actual boundary state, and reduce the debugging cost and certification cycle fluctuation caused by repeated rectification. Attached Figure Description

[0012] Figure 1 This application provides a block diagram of the operation control structure of a lightweight neural network-assisted EMI filtering system. Figure 2 The flowchart for establishing measurement mapping relationships and equivalent transformation provided in this application; Figure 3 The flowchart for compliance margin difference correction and control provided for this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] Please see Figure 1 As shown, this embodiment provides a lightweight neural network-assisted EMI filtering system, including an acquisition module, a conversion module, an extraction module, a judgment module, a correction module, and a control module. Each module is connected via wired and / or wireless means to achieve data transmission between modules.

[0015] The conversion module, used to acquire the first EMI observation data collected by the first measurement method under the current operating state of the target device, aims to generate first EMI observation data corresponding one-to-one with the current operating state of the target device using a unified first measurement method. This provides the original observation basis for subsequent equivalent conversion based on the measurement mapping relationship. Simultaneously, it ensures that the first EMI observation data includes operating state information, frequency information, amplitude information, and time information, so that it can be directly called upon in the subsequent second EMI observation data generation process. Specific implementation methods include: Step 101: Determine the current operating status of the target device; in some embodiments, the target device is an electrical device equipped with an electromagnetic interference filtering structure, and the target device can be a power supply device, a drive device, a conversion device, or an electronic device with switching devices.

[0016] First, connect the power supply to the target device and put it into the test operation phase. Then, read the target device's operating mode, load level, power supply status, and switching parameters. The operating mode refers to the category of operating functions performed by the target device, represented by a function category number. The load level refers to the preset range of the target device's output load, represented by a preset load range number. The power supply status refers to the preset range of the target device's input voltage, represented by a preset input voltage range number. The switching parameters refer to the switching frequency range and duty cycle parameter range corresponding to the internal switching devices of the target device, represented by a combination of the switching frequency range number and the duty cycle parameter range number. Then, combine the operating mode, load level, power supply status, and switching parameters in a fixed order to obtain the current operating status identifier. The fixed order is: operating mode, load level, power supply status, switching frequency range number, and duty cycle parameter range number. This current operating status identifier is used as the status field when combining the first EMI observation data in subsequent steps, and also as a retrieval field when matching measurement mapping relationships in subsequent steps.

[0017] Step 102: Set the observation conditions corresponding to the first measurement method; the first measurement method refers to the measurement method that takes the observed voltage at the input terminal of the target device as the object of observation; the second measurement method mentioned below refers to the measurement method that takes the observed current in the input conductor of the target device as the object of observation.

[0018] A first measurement position is set at the input end of the target device, defined as the observation position between the conductor at the input end of the target device and the reference ground. A line impedance stabilizing device is set on the line input side. The line impedance stabilizing device is used to provide a predetermined observation impedance at the first measurement position, so that the acquisition results of different batches have a unified observation benchmark. A voltage observation device is connected to the first measurement position, and the output end of the voltage observation device is connected to an analog-to-digital converter. The output end of the analog-to-digital converter is connected to a data processor. If the target device has a multi-phase input structure, a corresponding first measurement position is set at each phase input end, and a phase identifier is set at each phase input end. The phase identifier is subsequently incorporated into the first measurement condition identifier. After the above settings, the observation conditions corresponding to the first measurement method are formed.

[0019] Step 103: Establish the first measurement frequency point sequence; based on the electromagnetic interference frequency range that the target device needs to observe, set the first measurement frequency range; then generate the first measurement frequency point sequence within the first measurement frequency range; the first measurement frequency point sequence is generated using any of the following two methods: Firstly, multiple frequency points are generated sequentially at fixed frequency intervals within the first measurement frequency range; Secondly, the first measurement frequency range is divided into multiple continuous frequency bands, and then multiple frequency points are generated at fixed multiple intervals within each continuous frequency band.

[0020] If the second method is adopted, a denser frequency point distribution is set for continuous frequency bands with rapid amplitude changes, and a sparser frequency point distribution is set for continuous frequency bands with slow amplitude changes. The method for setting the density of frequency points is determined based on the average absolute value of the amplitude difference between adjacent frequency points in the historical acquisition results. When the average absolute value of the amplitude difference between adjacent frequency points is large, a denser distribution is used. This yields the first measurement frequency point sequence. The first measurement frequency point sequence serves as the execution order for subsequent point-by-point acquisition steps and also serves as the frequency field of the first EMI observation data in the next step.

[0021] Step 104: Under the current operating state, acquire the original amplitude at each frequency point; control the target device to maintain the current operating state, and acquire the original amplitude corresponding to each frequency point in sequence according to the first measurement frequency point sequence; for any frequency point, firstly, continuously acquire multiple voltage sample values ​​within a preset sampling time, and then retain the multiple voltage sample values ​​together with the corresponding sampling time; the method for setting the preset sampling time is determined based on the time fluctuation degree of the target device under the same operating state, and a longer sampling time is used when the time fluctuation degree is large; the number of multiple voltage sample values ​​is determined based on the sampling frequency of the analog-to-digital converter and the preset sampling time; after repeating the above acquisition action for all frequency points in the first measurement frequency point sequence, the original amplitude sequence and the acquisition time sequence are obtained.

[0022] Step 105: Organize the original amplitudes to form a first amplitude sequence; group the obtained original amplitude sequence according to frequency points; for any group of original amplitudes corresponding to any frequency point, first calculate the median value of the group of original amplitudes, and then calculate the deviation between each original amplitude and the median value; if the deviation of a certain original amplitude exceeds the preset fluctuation tolerance range, the original amplitude is removed from the representative amplitude calculation object; the preset fluctuation tolerance range is determined based on the background fluctuation of the voltage observation device, the quantization error of the analog-to-digital converter, and the amplitude fluctuation of the target device under repeated acquisition; after the removal is completed, calculate the arithmetic mean of the remaining original amplitudes to obtain the representative amplitude of the corresponding frequency point; repeat the above processing for all frequency points to obtain the first amplitude sequence that corresponds one-to-one with the first measurement frequency point sequence; in this embodiment, the representative amplitude is defined as the statistical value used to characterize the observation result of a frequency point after outlier removal.

[0023] In some implementations, to illustrate the process of forming the first amplitude sequence in steps 103 to 105, it is assumed that the current operating state is identified as M2, L3, U2, F4, D2, and the corresponding first measurement frequency point sequence includes 150 kHz, 300 kHz, 450 kHz, and 600 kHz. Taking a set of original amplitudes corresponding to 300 kHz as an example, five voltage sampling values ​​are continuously acquired within a preset sampling time, which are 52.1 dB / µV, 51.8 dB / µV, 52.4 dB / µV, 59.9 dB / µV, and 52.0 dB / µV, respectively. First, the median value of this set of original amplitudes is calculated, which is 52.1 dB / µV. Then, the deviation between each original amplitude and the median value is calculated, which is 0, 0.3, 0.3, 7.8, and 0.1 dB / µV, respectively. For the 5 The deviation corresponding to 9.9 dB / µV exceeds the preset fluctuation tolerance range determined by the background fluctuation of the voltage observation device, the quantization error of the analog-to-digital converter, and the fluctuation of repeated acquisition by the target device. Therefore, 59.9 dB / µV is removed from the representative amplitude calculation. Subsequently, the arithmetic mean of the remaining 52.1 dB / µV, 51.8 dB / µV, 52.4 dB / µV, and 52.0 dB / µV is calculated to obtain the representative amplitude of 52.075 dB / µV corresponding to 300 kHz. Using the same method, the representative amplitudes corresponding to 150 kHz, 450 kHz, and 600 kHz can be obtained respectively, thus forming the first amplitude sequence that corresponds one-to-one with the first measurement frequency point sequence. The formed first amplitude sequence is used as the amplitude field when combining the first EMI observation data in step 106.

[0024] Step 106: Combine to form the first EMI observation data; combine the obtained current operating status identifier, first measurement frequency point sequence, first amplitude sequence, acquisition time sequence and first measurement condition identifier to obtain the first EMI observation data; the first measurement condition identifier includes at least the first measurement location identifier, phase identifier, power supply mode identifier and acquisition batch identifier; the first measurement location identifier is used to indicate the location of the voltage observation device; the power supply mode identifier is used to indicate whether the target device is powered by a single phase or a multi-phase power supply; the acquisition batch identifier is used to distinguish different rounds of repeated acquisition.

[0025] Step 107: Perform repeated acquisition verification on the first EMI observation data; under the current operating state, repeat the above steps at least once to obtain the reference first EMI observation data formed by repeated acquisition; then compare the first EMI observation data formed in step 106 with the reference first EMI observation data point by point according to frequency points, and calculate the amplitude deviation at each frequency point; if the amplitude deviation at all frequency points is within the preset repeated acquisition allowable range, then retain the first EMI observation data formed in step 106; if the amplitude deviation at any frequency point exceeds the preset repeated acquisition allowable range, then reconfirm whether the current operating state in step 101 has changed, and re-execute steps 101 to 106. The method for setting the preset repeated acquisition allowable range is determined based on the remaining amplitude fluctuation distribution after removing outliers in step 105.

[0026] The conversion module is used to perform equivalent conversion on the first EMI observation data according to the pre-established measurement mapping relationship between the first measurement method and the second measurement method, so as to obtain the second EMI observation data corresponding to the second measurement method. The purpose is to convert the first EMI observation data obtained under the first measurement method into the second EMI observation data with the same diameter as the second measurement method, so that the feature results extracted from the first EMI observation data and the second EMI observation data can jointly characterize the electromagnetic interference distribution under different measurement methods, and provide a unified basis for the subsequent lightweight neural network model to process the two measurement diameter data at the same time.

[0027] See Figure 2As shown, step 201 involves pre-establishing a measurement mapping relationship. To explain the subsequent equivalent conversion, a measurement mapping relationship between the first measurement method and the second measurement method is established before formal operation. The second measurement method is the current observation method under the second measurement aperture. The establishment process includes: selecting multiple reference operating states; synchronously acquiring data for the same target device or similar devices according to the first and second measurement methods under each reference operating state; obtaining first and second reference observation data; aligning the first and second reference observation data according to the reference operating state identifier, frequency point, and acquisition sequence; based on the alignment results, for the same reference operating state and the same frequency point, instead of determining the proportional coefficient and offset coefficient for a single first reference amplitude and a single second reference amplitude, multiple repeated observation results are acquired first to form a first reference amplitude sequence and a second reference amplitude sequence, and ensuring that each amplitude in the first reference amplitude sequence corresponds one-to-one with each amplitude at the corresponding position in the second reference amplitude sequence.

[0028] Subsequently, the candidate ranges for the scaling factor and the offset factor are determined. The candidate range for the scaling factor is determined based on the proportional distribution between corresponding amplitudes in the first and second reference amplitude sequences; the candidate range for the offset factor is determined based on the difference distribution between corresponding amplitudes in the first and second reference amplitude sequences. Then, multiple candidate scaling factors are selected sequentially within the candidate range for the scaling factor at a preset scaling step size, and multiple candidate offset factors are selected sequentially within the candidate range for the offset factor at a preset offset step size, thus forming multiple sets of candidate coefficient combinations. The preset scaling step size and preset offset step size are determined based on the measurement accuracy requirements and the computational capacity allowed by the processor. For any set of candidate coefficient combinations, each first reference amplitude in the first reference amplitude sequence is first converted by multiplying it by the candidate scaling factor and adding the candidate offset factor to obtain the corresponding converted amplitude. This converted amplitude is then compared with the corresponding second reference amplitude to obtain the deviation value for that set of corresponding amplitudes. After repeating the above process for all corresponding positions, all deviation values ​​are accumulated to obtain the cumulative deviation value corresponding to that set of candidate coefficient combinations.

[0029] Subsequently, the cumulative deviation value calculation is performed on all candidate coefficient combinations, and the group with the smallest cumulative deviation value is selected from all candidate coefficient combinations. The candidate proportional coefficient in this group is determined as the proportional coefficient corresponding to the reference operating state and the frequency point, and the candidate offset coefficient in this group is determined as the offset coefficient corresponding to the reference operating state and the frequency point. The proportional coefficient characterizes the amplitude change ratio between the first and second measurement methods, and the offset coefficient characterizes the overall offset relationship between the first and second measurement methods. When the range of change of the first reference amplitude sequence at the same frequency point is too small to stably distinguish different candidate proportional coefficient combinations, the proportional coefficient is set to a preset default proportional value. Under the condition of fixing the preset default proportional value, the offset coefficient is determined according to the above-mentioned candidate offset coefficient combination comparison method. The preset default proportional value is determined based on the statistical results of the proportional coefficients at historical frequency points. After repeating the above processing on all reference operating states and all frequency points, a measurement mapping relationship table is obtained. The measurement mapping relationship table includes at least the reference operating state identifier, frequency point, proportional coefficient, and offset coefficient. Subsequent steps 202 to 206 call the measurement mapping relationship table to complete the equivalent conversion.

[0030] Step 202: Read the first EMI observation data and match it with the target mapping entry set; read the first EMI observation data retained in step 107, and extract the current operating status identifier, the first measurement frequency point sequence, and the first amplitude sequence from the first EMI observation data; then, retrieve the mapping entry set corresponding to the current operating status identifier from the measurement mapping relationship table formed in step 201; when there is a mapping entry set in the measurement mapping relationship table that is completely consistent with the current operating status identifier, determine the mapping entry set as the target mapping entry set.

[0031] When there is no set of mapping entries in the measurement mapping table that completely matches the current operating status identifier, nearest neighbor matching is performed in the order of operating mode, load level, power supply status, and switch parameters. Specifically, first, it is compared whether the operating mode corresponding to the current operating status is consistent with the operating modes corresponding to each reference operating status, and the entries corresponding to the reference operating statuses with consistent operating modes are retained; then, among the entries corresponding to the reference operating statuses with consistent operating modes, it is compared whether the load level corresponding to the current operating status is in an adjacent interval with the load levels corresponding to each reference operating status.

[0032] The method for determining whether a load level is in an adjacent interval is as follows: First, read the load level interval number corresponding to the current operating state and the load level interval number corresponding to the reference operating state, and then compare the difference between the two interval numbers. When the two interval numbers are the same, or the difference between the two interval numbers is equal to the preset adjacent number value, the load level is determined to be in an adjacent interval. When the difference between the two interval numbers is greater than the preset adjacent number value, the load level is determined not to be in an adjacent interval. The preset adjacent number value is determined based on the granularity of the load level division. When the load levels are divided into consecutive levels, the preset adjacent number value is taken as the difference between the numbers corresponding to the adjacent level.

[0033] After completing the load level comparison, the system continues to compare the power supply status corresponding to the current operating state with the power supply status corresponding to each reference operating state in the retained entries to determine whether they are in adjacent intervals. The method for determining whether power supply statuses are in adjacent intervals is as follows: first, read the power supply status interval number corresponding to the current operating state and the power supply status interval number corresponding to the reference operating state; then compare the difference between the two interval numbers. When the two interval numbers are the same, or the difference between the two interval numbers is equal to a preset adjacent number value, the power supply status is determined to be in an adjacent interval; when the difference between the two interval numbers is greater than the preset adjacent number value, the power supply status is determined not to be in an adjacent interval. The preset adjacent number value is determined based on the granularity of the power supply status interval division.

[0034] After completing the power supply status comparison, the distance between the switch parameter interval corresponding to the current operating state and the switch parameter intervals corresponding to each reference operating state is compared in the last remaining entry. The distance is represented by a weighted sum of the differences between the center values ​​of the switch frequency intervals and the center values ​​of the duty cycle parameter intervals. Specifically, the switch frequency interval and duty cycle parameter interval corresponding to the current operating state, as well as the switch frequency interval and duty cycle parameter interval corresponding to each reference operating state, are first read. Then, the average value of the upper and lower boundaries of the switch frequency interval corresponding to the current operating state is taken as the center value of the current switch frequency interval, and the average value of the upper and lower boundaries of the switch frequency interval corresponding to the reference operating state is taken as the center value of the reference switch frequency interval, and the difference between the two is calculated. At the same time, the average value of the upper and lower boundaries of the duty cycle parameter interval corresponding to the current operating state is taken as the center value of the current duty cycle parameter interval, and the average value of the upper and lower boundaries of the duty cycle parameter interval corresponding to the reference operating state is taken as the center value of the reference duty cycle parameter interval, and the difference between the two is calculated.

[0035] After obtaining the center value difference of the switching frequency interval and the center value difference of the duty cycle parameter interval, the switching frequency weight and duty cycle parameter weight are determined respectively. The weights are formed as follows: First, during the training or pre-calibration phase, the duty cycle parameter interval is fixed while the switching frequency interval is changed. The amplitude changes corresponding to the observed results are statistically analyzed to obtain the switching frequency influence. Then, the switching frequency interval is fixed while the duty cycle parameter interval is changed, and the amplitude changes corresponding to the observed results are statistically analyzed to obtain the duty cycle parameter influence. Subsequently, based on the proportion of the switching frequency influence and the duty cycle parameter influence in their sum, the switching frequency weight and duty cycle parameter weight are determined respectively, so that a larger switching frequency influence corresponds to a larger switching frequency weight, and a larger duty cycle parameter influence corresponds to a larger duty cycle parameter weight. The switching frequency weight and duty cycle parameter weight are then normalized so that their sum equals the preset total weight value. The preset total weight value can be set to one.

[0036] Next, the difference in the center value of the switching frequency interval is multiplied by the switching frequency weight, and the difference in the center value of the duty cycle parameter interval is multiplied by the duty cycle parameter weight. The results of the two products are then summed to obtain the distance between the switching parameter intervals between the current operating state and the corresponding reference operating state. After completing the above comparison, the entries corresponding to the reference operating states with the same working mode, load level, and power supply status are first retained. Then, the entry corresponding to the reference operating state with the smallest switching parameter interval distance is selected from the retained entries. The selection result is determined as the target mapping entry set. The target mapping entry set is called by steps 203 and 204.

[0037] Step 203: Perform frequency aperture alignment on the first amplitude sequence. To ensure that the equivalent converted second EMI observation data has a clear frequency field, a second target frequency point sequence is first established. The second target frequency point sequence can use the same frequency point distribution as the first measurement frequency point sequence, or it can use a frequency point distribution pre-set under the second measurement method. If the second target frequency point sequence is consistent with the first measurement frequency point sequence, the first amplitude sequence is directly used as the first aligned amplitude sequence. If the second target frequency point sequence is inconsistent with the first measurement frequency point sequence, frequency interpolation processing is performed on the first amplitude sequence. The frequency interpolation processing is completed using adjacent point linear interpolation. For any second target frequency point, first find the first left frequency point and the first right frequency point located on the left and right sides of the second target frequency point in the first measurement frequency point sequence, then read the left amplitude corresponding to the first left frequency point and the right amplitude corresponding to the first right frequency point, and then calculate the corresponding first aligned amplitude based on the position ratio of the second target frequency point between the first left frequency point and the first right frequency point. After repeating the above processing for all second target frequency points, the first aligned amplitude sequence corresponding to the second target frequency point sequence is obtained.

[0038] Step 204: Perform equivalent conversion on a frequency-by-frequency basis according to the target mapping entry set; for any target frequency point in the second target frequency point sequence, retrieve the mapping entry corresponding to the target frequency point from the target mapping entry set; when there is a mapping frequency point in the target mapping entry set that is completely consistent with the target frequency point, read the scaling factor and offset factor corresponding to the mapping frequency point from the step measurement mapping relationship table, then read the first aligned amplitude corresponding to the target frequency point from the first aligned amplitude sequence, and calculate the second equivalent amplitude corresponding to the target frequency point based on the scaling factor and offset factor, wherein the scaling factor and offset factor are the mapping parameters determined by the candidate coefficient combination comparison in step 201 and recorded in the measurement mapping relationship table.

[0039] When there is no mapping frequency point in the target mapping entry set that is completely consistent with the target frequency point, first determine whether the target frequency point is within the mapping frequency range corresponding to the target mapping entry set; the mapping frequency range is defined by the minimum and maximum mapping frequency points in the target mapping entry set; when the target frequency point is within the mapping frequency range, select the mapping frequency point in the target mapping entry set that is smaller than the target frequency point and closest to the target frequency point as the left mapping frequency point, and select the mapping frequency point in the target mapping entry set that is larger than the target frequency point and closest to the target frequency point as the right mapping frequency point; then read the left proportional coefficient and left offset coefficient corresponding to the left mapping frequency point, and the right proportional coefficient and right offset coefficient corresponding to the right mapping frequency point from the measurement mapping relationship table formed in step 201; then, based on the positional relationship between the target frequency point and the left and right mapping frequency points, interpolate the left proportional coefficient and the right proportional coefficient to obtain the interpolated proportional coefficient, and then perform interpolation processing on the left and right proportional coefficients. The side offset coefficient and the right offset coefficient are interpolated to obtain the interpolation offset coefficient. Then, the first aligned amplitude corresponding to the target frequency point is read, and the second equivalent amplitude corresponding to the target frequency point is calculated based on the interpolation scaling factor and the interpolation offset coefficient. When the target frequency point is less than the minimum mapped frequency point in the target mapping entry set, the scaling factor and offset coefficient corresponding to the minimum mapped frequency point are read from the measurement mapping relationship table formed in step 201, and the second equivalent amplitude is calculated in combination with the first aligned amplitude corresponding to the target frequency point. When the target frequency point is greater than the maximum mapped frequency point in the target mapping entry set, the scaling factor and offset coefficient corresponding to the maximum mapped frequency point are read from the measurement mapping relationship table formed in step 201, and the second equivalent amplitude is calculated in combination with the first aligned amplitude corresponding to the target frequency point. After repeating the above processing for all second target frequency points, the second equivalent amplitude sequence is obtained. The second equivalent amplitude sequence output in step 204 serves as the amplitude basis for forming the second EMI observation data in step 205.

[0040] For example, to illustrate the frequency aperture alignment and equivalent conversion process in steps 203 and 204, assume that the second target frequency point sequence formed in step 203 includes 320 kHz, and that there is no completely identical mapping frequency point for 320 kHz in the target mapping entry set; in this case, select 300 kHz as the left mapping frequency point and 340 kHz as the right mapping frequency point from the target mapping entry set; the left scaling factor corresponding to 300 kHz is 0.42, and the left offset factor is 1.6; the right scaling factor corresponding to 340 kHz is 0.48, and the right offset factor is 1.2; then subtract 300 kHz from 320 kHz to obtain a left frequency difference of 20 kHz; subtract 3 from 340 kHz... 00 kHz, the total frequency difference is 40 kHz; divide 20 by 40 to get a position ratio of 0.5; then calculate the interpolation ratio coefficient, which is 0.42 plus the difference between 0.5 multiplied by 0.48 and 0.42, i.e., 0.45; then calculate the interpolation offset coefficient, which is 1.6 plus the difference between 0.5 multiplied by 1.2 and 1.6, i.e., 1.4; if the first alignment amplitude formed at 320 kHz in step 203 is 53.0 dB / µV, then multiply 0.45 by 53.0 and add 1.4 to get the second equivalent amplitude corresponding to 320 kHz as 25.25; repeat the same process for the remaining target frequency points in the second target frequency point sequence to obtain the second equivalent amplitude sequence.

[0041] Step 205: Combine to form the second EMI observation data; combine the current operating status identifier extracted in step 202, the second target frequency point sequence formed in step 203, the second equivalent amplitude sequence formed in step 204, the conversion time sequence corresponding to the equivalent conversion time, and the second measurement condition identifier to obtain the second EMI observation data; wherein, the conversion time sequence is formed as follows: based on the arrangement order of the second target frequency point sequence in step 203, record the time when the corresponding second equivalent amplitude is calculated for each second target frequency point to form a conversion time sequence that corresponds one-to-one with the second target frequency point sequence; the second measurement condition identifier includes at least the second measurement position identifier, the observation conductor identifier, the phase identifier, and the conversion batch identifier; the second measurement position identifier is used to indicate the observation position corresponding to the second measurement method; the observation conductor identifier is used to distinguish the observed input conductor; the phase identifier is used to distinguish the input phase; the conversion batch identifier is used to distinguish the equivalent conversion processing of different rounds.

[0042] During the combination process, firstly, the current operating status identifier is mapped to each of the second target frequency points according to the order of the second target frequency point sequence; then, the second equivalent amplitude sequence and the conversion time sequence are mapped item by item in the same order; finally, the second measurement condition identifier is added to form complete second EMI observation data; the second EMI observation data includes at least the current operating status identifier field, the second target frequency point field, the second equivalent amplitude field, the conversion time field, and the second measurement condition field. The second EMI observation data serves as the consistency verification object in step 206 and as one of the input objects for the feature extraction module.

[0043] Step 206: Perform a consistency check on the second EMI observation data. To confirm the usability of the second EMI observation data formed in step 205, first select a set of check frequency points from the second target frequency point sequence. The selection method for the check frequency point set is as follows: according to the arrangement order of the second target frequency point sequence, select at least one frequency point in each of the first, middle, and last frequency regions. When the number of second target frequency points is large, continue to select multiple frequency points in each frequency region at fixed intervals. The fixed interval is determined based on the total number of second target frequency points and the frequency coverage area. Then read... Take the second reference observation data retained in step 201 during the determination of the proportional coefficient and offset coefficient, and extract the second reference amplitude corresponding to the set of verification frequency points under the reference operating state consistent with the current operating state identifier; then, calculate the deviation between the second equivalent amplitude corresponding to the set of verification frequency points in the second EMI observation data obtained in step 205 and the second reference amplitude point by point; the deviation calculation method is as follows: for any verification frequency point, first read the second equivalent amplitude and the second reference amplitude corresponding to the verification frequency point, then subtract the second reference amplitude from the second equivalent amplitude and take the absolute value to obtain the amplitude deviation corresponding to the verification frequency point.

[0044] After repeating the above deviation calculation for all frequency points in the verification frequency point set, a verification deviation sequence is obtained. Then, each value in the verification deviation sequence is compared with the preset conversion allowable range. The preset conversion allowable range is determined based on the second reference amplitude fluctuation range retained after the candidate coefficient combination comparison in step 201. When all values ​​in the verification deviation sequence are within the preset conversion allowable range, the second EMI observation data formed in step 205 is retained. When at least one value in the verification deviation sequence exceeds the preset conversion allowable range, the target mapping item set selection result in step 202, the frequency alignment result in step 203, and the second equivalent amplitude calculation result in step 204 are sequentially reviewed, and steps 202 to 205 are re-executed until the verification deviation sequence corresponding to the newly formed second EMI observation data is within the preset conversion allowable range. Step 206 outputs the second EMI observation data after consistency verification.

[0045] The extraction module is used to perform feature extraction on the first EMI observation data and the second EMI observation data respectively, to obtain the first EMI feature results and the second EMI feature results. The purpose is to extract feature information that can characterize frequency distribution, amplitude distribution, and time fluctuation from the first EMI observation data output in step 107 and the second EMI observation data output in step 206, respectively, and to organize the extraction results into a unified field structure so that the subsequent lightweight neural network model can jointly process the first EMI feature results and the second EMI feature results. Specific implementation methods include: Step 301: Read the first EMI observation data and the second EMI observation data, and perform basic alignment; first read the first EMI observation data output in step 107, and then read the second EMI observation data output in step 206; then extract the current operating status identifier, the first measurement frequency point sequence, the first amplitude sequence and the acquisition time sequence from the first EMI observation data, and the current operating status identifier, the second target frequency point sequence, the second equivalent amplitude sequence and the conversion time sequence from the second EMI observation data.

[0046] Then compare the current operating status identifiers in the first EMI observation data and the second EMI observation data; if the current operating status identifiers are the same, continue to perform frequency alignment; if the current operating status identifiers are not the same, return to steps 101 to 107 or steps 202 to 206 to reconfirm the source of the input data.

[0047] Under the premise that the current operating status identifiers are consistent, the first measurement frequency point sequence and the second target frequency point sequence are compared. If the first measurement frequency point sequence and the second target frequency point sequence are consistent item by item, the first amplitude sequence is determined as the first aligned amplitude sequence, and the second equivalent amplitude sequence is determined as the second aligned amplitude sequence. If the first measurement frequency point sequence and the second target frequency point sequence are inconsistent, the set of frequency point sequences with more points is taken as the unified target frequency point sequence, and linear interpolation is performed on the other set of amplitude sequences to form an aligned amplitude sequence that corresponds one-to-one with the unified target frequency point sequence.

[0048] After processing in step 301, a unified target frequency point sequence, a first aligned amplitude sequence, and a second aligned amplitude sequence are obtained; the unified target frequency point sequence, the first aligned amplitude sequence, and the second aligned amplitude sequence serve as the common input basis for steps 302 to 307.

[0049] Step 302: Perform smoothing processing on the first aligned amplitude sequence and the second aligned amplitude sequence respectively; firstly, perform moving average processing on the first aligned amplitude sequence according to the arrangement order of the unified target frequency point sequence; for any unified target frequency point, read the first aligned amplitude corresponding to the unified target frequency point, as well as the first aligned amplitude corresponding to the frequency points to the left and right of the unified target frequency point; then calculate the arithmetic mean of all read amplitudes to obtain the first basic amplitude corresponding to the unified target frequency point; if the unified target frequency point is located at the beginning or end of the sequence, only calculate the arithmetic mean of the actual amplitudes.

[0050] Then, using the same processing method, a moving average process is performed on the second aligned amplitude sequence to obtain the second basic amplitude sequence.

[0051] The window length used in the moving average processing is set based on the inter-point interval and amplitude change rate of the unified target frequency point sequence; step 302 outputs the first basic amplitude sequence and the second basic amplitude sequence, and the two results are used as input objects for steps 303 to 307 respectively.

[0052] Step 303: Extract the first characteristic frequency point sequence and the first characteristic amplitude sequence from the first EMI observation data; scan the first basic amplitude sequence point by point according to the arrangement order of the unified target frequency point sequence; for any unified target frequency point that is neither the beginning nor the end, read the first basic amplitude corresponding to the previous frequency point, the current frequency point, and the next frequency point respectively; when the first basic amplitude corresponding to the current frequency point is greater than the first basic amplitude corresponding to the previous frequency point, and the first basic amplitude corresponding to the current frequency point is not less than the first basic amplitude corresponding to the next frequency point, record the current frequency point as the first candidate characteristic frequency point, and record the first basic amplitude corresponding to the current frequency point as the first candidate characteristic amplitude.

[0053] After completing the scanning of all unified target frequency points, all first candidate feature amplitudes are sorted according to their numerical values; then, the first candidate feature frequency points and first candidate feature amplitudes are selected according to preset retention rules. The preset retention rules can be any of the following methods: One approach is to retain the first N items according to their numerical values; another approach is to retain a preset number of candidate items in each consecutive frequency band. The value of N or the number of items to be retained in each consecutive frequency band is determined based on the input length that the subsequent lightweight neural network model is allowed to receive.

[0054] After processing in step 303, a first characteristic frequency point sequence and a first characteristic amplitude sequence corresponding one-to-one with the first characteristic frequency point sequence are obtained; these are the components of the first EMI characteristic result formed in step 308.

[0055] Step 304: Extract the second characteristic frequency point sequence and the second characteristic amplitude sequence from the second EMI observation data; apply the same processing rules as in step 303 to the second fundamental amplitude sequence; for any unified target frequency point that is neither the beginning nor the end, read the second fundamental amplitude corresponding to the previous frequency point, the current frequency point, and the next frequency point respectively; when the second fundamental amplitude corresponding to the current frequency point is greater than the second fundamental amplitude corresponding to the previous frequency point, and the second fundamental amplitude corresponding to the current frequency point is not less than the second fundamental amplitude corresponding to the next frequency point, record the current frequency point as the second candidate characteristic frequency point, and record the second fundamental amplitude corresponding to the current frequency point as the second candidate characteristic amplitude.

[0056] After completing the scanning of all unified target frequency points, all second candidate feature amplitudes are sorted according to their numerical values, and then filtered according to the preset retention rules consistent with step 303 to obtain the second feature frequency point sequence and the second feature amplitude sequence corresponding to the second feature frequency point sequence.

[0057] The second characteristic frequency point sequence and the second characteristic amplitude sequence output in step 304 are used as components of the second EMI characteristic result formed in step 309.

[0058] Step 305: Extract frequency band energy sequences from the first EMI observation data and the second EMI observation data respectively; firstly, divide multiple continuous frequency bands based on a unified target frequency point sequence; each continuous frequency band is defined by a start frequency point and an end frequency point; then, perform energy accumulation processing on the first basic amplitude within each continuous frequency band to obtain the first frequency band energy of the corresponding continuous frequency band; the energy accumulation processing method is as follows: if the amplitude is represented by linear amplitude, first square the first basic amplitude corresponding to each frequency point within the continuous frequency band, and then accumulate the squared results item by item to obtain the first frequency band energy; if the amplitude is represented by logarithmic amplitude, first convert the logarithmic amplitude to linear energy value, and then accumulate the linear energy values ​​item by item to obtain the first frequency band energy; repeat the above processing for all continuous frequency bands to form the first frequency band energy sequence; then, use the same processing method to perform continuous frequency band energy accumulation processing on the second basic amplitude sequence to form the second frequency band energy sequence; Step 305 outputs the first frequency band energy sequence and the second frequency band energy sequence as components of steps 308 and 309.

[0059] Step 306: Extract adjacent variation sequences from the first EMI observation data and the second EMI observation data respectively; perform adjacent difference processing on the first basic amplitude sequence according to the arrangement order of the unified target frequency point sequence; for any non-first unified target frequency point, read the first basic amplitude corresponding to the current frequency point and the first basic amplitude corresponding to the previous frequency point, and then subtract the first basic amplitude corresponding to the previous frequency point from the first basic amplitude corresponding to the current frequency point to obtain the first adjacent variation corresponding to the current frequency point. For the first unified target frequency point, set the corresponding first adjacent variation to zero; repeat the above processing for all unified target frequency points to obtain the first adjacent variation sequence.

[0060] Then, using the same processing method, the second basic amplitude sequence is subjected to adjacent difference processing to obtain the second adjacent change sequence; step 306 outputs the first adjacent change sequence and the second adjacent change sequence, which are respectively used as components of steps 308 and 309.

[0061] Step 307: Extract time-series fluctuation sequences from the first EMI observation data and the second EMI observation data respectively; for any unified target frequency point in the first EMI observation data, if multiple sampled values ​​are retained at the unified target frequency point or the corresponding interpolation source frequency point in step 104, first calculate the arithmetic mean of the multiple sampled values, then calculate the absolute difference between each sampled value and the arithmetic mean, and finally calculate the arithmetic mean of all absolute differences to obtain the first time-series fluctuation amount corresponding to the unified target frequency point; if there is only one sampled value, then set the corresponding first time-series fluctuation amount to zero; repeat the above processing for all unified target frequency points to form the first time-series fluctuation sequence.

[0062] For any uniform target frequency point in the second EMI observation data, the first aligned amplitude source sample value called in step 204 when calculating the second equivalent amplitude corresponding to the uniform target frequency point is used as the time series basis, and the second time series fluctuation quantity is formed according to the same calculation method as the first time series fluctuation sequence. After repeating the above processing for all uniform target frequency points, the second time series fluctuation sequence is formed.

[0063] Step 307 outputs the first time series fluctuation sequence and the second time series fluctuation sequence, which are used as components of steps 308 and 309, respectively.

[0064] As another example, to illustrate the feature extraction process in steps 302 to 307, assume that the unified target frequency point sequence includes 280 kHz, 300 kHz, 320 kHz, 340 kHz, and 360 kHz, with corresponding first basic amplitude sequences of 49, 52, 55, 51, and 50, and corresponding second basic amplitude sequences of 23, 24, 25, 24, and 23, respectively. For the first basic amplitude sequence, at 320 kHz, the first basic amplitude of 55 corresponding to the current frequency point is greater than that of the previous frequency point. The first fundamental amplitude is 52, and it is not less than the first fundamental amplitude 51 corresponding to the next frequency point. Therefore, 320 kHz is determined as the first candidate characteristic frequency point, and 55 is determined as the corresponding first candidate characteristic amplitude. For the second fundamental amplitude sequence, at 320 kHz, 25 is greater than 24 corresponding to the previous frequency point and not less than 24 corresponding to the next frequency point. Therefore, 320 kHz is determined as the second candidate characteristic frequency point, and 25 is determined as the corresponding second candidate characteristic amplitude. Furthermore, if 300 kHz to 340 kHz is divided into... Dividing the frequencies into a continuous band, the energy of the first band can be obtained by successively adding the squares of 52, 55, and 51 to get 8330, and the energy of the second band can be obtained by successively adding the squares of 24, 25, and 24 to get 1777. For adjacent changes, taking 340 kHz as an example, the first adjacent change is 51 minus 55, resulting in -4, and the second adjacent change is 24 minus 25, resulting in -1. For time-series fluctuations, if the three sampled values ​​retained at 320 kHz are 54.8, 55.2, and 5... If the value is 5.0, the arithmetic mean is first calculated to obtain 55.0. Then, the absolute difference between each sampled value and 55.0 is calculated to obtain 0.2, 0.2 and 0. Finally, the arithmetic mean of all absolute differences is calculated to obtain the first timing fluctuation corresponding to 320 kHz as 0.133. After the above processing, the obtained characteristic frequency points, characteristic amplitudes, frequency band energy, adjacent changes and timing fluctuations can be combined according to the fixed field order of steps 308 and 309 to form the first EMI characteristic result and the second EMI characteristic result, respectively.

[0065] Step 308: Form the first EMI characteristic result; combine the current operating status identifier confirmed in step 301, the first characteristic frequency point sequence obtained in step 303, the first characteristic amplitude sequence obtained in step 303, the first frequency band energy sequence obtained in step 305, the first adjacent change quantity sequence obtained in step 306, and the first time series fluctuation sequence obtained in step 307 in a fixed field order to obtain the first EMI characteristic result.

[0066] The fixed field order is as follows: current running status identifier, first characteristic frequency point sequence, first characteristic amplitude sequence, first frequency band energy sequence, first adjacent change quantity sequence, and first time series fluctuation sequence.

[0067] Step 308 outputs the first EMI feature result, which serves as one of the input objects for the subsequent lightweight neural network model.

[0068] Step 309: Form the second EMI characteristic result; combine the current operating status identifier confirmed in step 301, the second characteristic frequency point sequence obtained in step 304, the second characteristic amplitude sequence obtained in step 304, the second frequency band energy sequence obtained in step 305, the second adjacent change quantity sequence obtained in step 306, and the second time series fluctuation sequence obtained in step 307 in the same fixed field order as in step 308 to obtain the second EMI characteristic result.

[0069] Step 309 outputs the second EMI feature result, which serves as one of the input objects for the subsequent lightweight neural network model.

[0070] Step 310: Perform a structural consistency check on the first EMI feature result and the second EMI feature result; first compare the number of fields, field order and field length in the first EMI feature result and the second EMI feature result, and then compare whether the current running status identifiers corresponding to the two results are consistent.

[0071] When the number of fields, the order of fields, the length of fields, and the current running status identifier are consistent, the first EMI feature result formed in step 308 and the second EMI feature result formed in step 309 are retained; when there are inconsistencies in the number of fields, the order of fields, the length of fields, or the current running status identifier, the basic alignment is re-executed in step 301, or the feature extraction and combination are re-executed in steps 303 to 309.

[0072] Step 310 outputs the first EMI feature result and the second EMI feature result with unified structure, which together serve as the input basis for the subsequent lightweight neural network model.

[0073] The determination module is used to input the first EMI feature result and the second EMI feature result into the lightweight neural network model to obtain the first compliance margin result and the second compliance margin result, and to generate the measurement method difference result based on the difference between the first compliance margin result and the second compliance margin result.

[0074] The first and second EMI feature results are input into a lightweight neural network model to obtain the first and second compliance margin results. The purpose is to obtain, based on the first and second EMI feature results, the first compliance margin result under the first measurement method caliber and the second compliance margin result under the second measurement method caliber, respectively, providing direct input for subsequent generation of measurement method difference results and execution of boundary offset correction. Specific implementation methods include: Step 4011: Construct the first model input vector and the second model input vector; first, read the first EMI feature result and the second EMI feature result; then, according to a fixed field order, sequentially expand the current operating status identifier, the first characteristic frequency point sequence, the first characteristic amplitude sequence, the first frequency band energy sequence, the first adjacent change quantity sequence, and the first time series fluctuation sequence in the first EMI feature result to form the first model input vector; then, according to the same field order, sequentially expand the current operating status identifier, the second characteristic frequency point sequence, the second characteristic amplitude sequence, the second frequency band energy sequence, the second adjacent change quantity sequence, and the second time series fluctuation sequence in the second EMI feature result to form the second model input vector.

[0075] The expansion method is as follows: first, retain the status code value corresponding to the current running status identifier, and then append each value item sequentially according to the internal arrangement order of the field; if the length of a certain field is less than the preset input length, then a preset padding value is appended to the end of the field; if the length of a certain field is greater than the preset input length, then a fixed number of values ​​at the beginning are retained and the remaining values ​​are discarded. The preset input length is determined based on the maximum vector length that the input end of the lightweight neural network model can receive; the preset padding value is the median value or zero value of the same field value distribution during the training phase; the first model input vector and the second model input vector are used as the input objects for step 402.

[0076] Step 4012: Perform normalization processing on the first model input vector and the second model input vector; read the field normalization parameter table retained during the training phase; the field normalization parameter table includes at least the mean and standard deviation, or minimum and maximum values, corresponding to each input field; then perform normalization processing on each numerical field in the first model input vector; if the field normalization parameter table uses the mean and standard deviation, subtract the corresponding mean from the current field value, and then divide the difference by the corresponding standard deviation to obtain the first normalized input vector; if the field normalization parameter table uses the minimum and maximum values, subtract the corresponding minimum value from the current field value, and then divide the difference by the difference between the corresponding maximum and minimum values ​​to obtain the first normalized input vector; perform normalization processing on each numerical field in the second model input vector in the same way to obtain the second normalized input vector.

[0077] When the corresponding standard deviation is zero, or the difference between the corresponding maximum and minimum values ​​is zero, the original field value is directly retained to avoid division by zero; the first normalized input vector and the second normalized input vector are used as the input objects for step 4013.

[0078] Step 4013: Append measurement method identifiers to the first normalized input vector and the second normalized input vector; In order to enable the lightweight neural network model to distinguish different measurement method calibers under the same network structure, append a first measurement method identifier value to the end of the first normalized input vector and append a second measurement method identifier value to the end of the second normalized input vector; The first measurement method identifier value adopts a first integer encoding value, and the second measurement method identifier value adopts a second integer encoding value, and the first integer encoding value and the second integer encoding value are different.

[0079] After the addition is completed, a first inference input vector and a second inference input vector are formed, which serve as the input objects for steps 4014 and 4015, respectively.

[0080] Step 4014: Input the first inference input vector into the lightweight neural network model to obtain the first compliance margin value; input the first inference input vector into the lightweight neural network model; the lightweight neural network model first performs a first-layer linear transformation on the first inference input vector to obtain the first hidden vector; then performs non-linear activation processing on the first hidden vector term by term to obtain the first activation vector; subsequently performs a second-layer linear transformation on the first activation vector to obtain the second hidden vector; then performs non-linear activation processing on the second hidden vector term by term to obtain the second activation vector; finally, performs an output-layer linear transformation on the second activation vector to obtain the first compliance margin value.

[0081] The first-level linear transformation refers to multiplying the first-level weight parameter matrix by the first inference input vector and adding the first-level bias parameter vector; the second-level linear transformation and the output-level linear transformation use the same calculation method; the nonlinear activation processing uses the modified linear unit function, and the processing method is: when the input value is greater than zero, the input value is retained, and when the input value is less than or equal to zero, the zero value is output; the first compliance margin value is the core value for forming the first compliance margin result in step 4016.

[0082] Step 4015: Input the second inference input vector into the lightweight neural network model to obtain the second compliance margin value; input the second inference input vector into the lightweight neural network model with the same parameters as in step 4014, and use the same inference process as in step 4014 to sequentially perform the first layer linear transformation, the first nonlinear activation process, the second layer linear transformation, the second nonlinear activation process, and the output layer linear transformation to obtain the second compliance margin value, which serves as the core value for forming the second compliance margin result in step 4016.

[0083] Step 4016: Form a first compliance margin result and a second compliance margin result; combine the first compliance margin value obtained in step 4014, the first inference time, and the first result source identifier to form the first compliance margin result. The first inference time is defined as the time when the output processing of step 4014 is completed; the first result source identifier is used to indicate that the first compliance margin result originates from the first measurement method caliber.

[0084] The second compliance margin value, the second inference time, and the second result source identifier obtained in step 4015 are then combined to form the second compliance margin result; the second inference time is defined as the time when the output processing of step 4015 is completed; the second result source identifier is used to indicate that the second compliance margin result comes from the second measurement method caliber.

[0085] The purpose of generating measurement method difference results based on the difference between the first and second compliance margin results is to form measurement method difference results that characterize the degree and direction of the difference in caliber between the two measurement methods, based on the numerical difference between the first and second compliance margin results, thus providing a direct basis for subsequent boundary offset correction. Specific implementation methods include: See Figure 3 As shown, in step 4021, read the first compliance margin result and the second compliance margin result; first read the first compliance margin result and the second compliance margin result output in step 4016, and extract the first compliance margin value, the first inference time, the first result source identifier, the second compliance margin value, the second inference time, and the second result source identifier, respectively, as the input objects for step 4022.

[0086] Step 4022: Calculate the difference magnitude and difference direction. First, subtract the second compliance margin value from the first compliance margin value to obtain the original difference value. Then, take the absolute value of the original difference value to obtain the difference magnitude. Subsequently, determine the difference direction based on the sign of the original difference value. When the original difference value is greater than zero, it indicates that the first compliance margin value is greater than the second compliance margin value, and the difference direction is marked as the first direction. When the original difference value is less than zero, it indicates that the first compliance margin value is less than the second compliance margin value, and the difference direction is marked as the second direction. When the original difference value is equal to zero, the difference direction is marked as the zero difference direction. The difference magnitude and difference direction serve as the input objects for steps 503 and 4024.

[0087] Step 4023: Determine the difference interval identifier based on the difference range; first, establish a difference interval table. The process of establishing the difference interval table includes: reading the first compliance margin value and the second compliance margin value corresponding to each sample in the training phase, and calculating the difference amplitude between the two for each sample; then sorting all the difference amplitudes according to their numerical size to form a difference amplitude sequence; then determining the number of intervals based on the number of training samples and the dispersion of the difference amplitude distribution. The number of intervals is determined by a preset interval selection rule, where fewer consecutive difference intervals are selected when the number of training samples is small and the difference amplitudes are concentrated, and more consecutive difference intervals are selected when the number of training samples is large and the difference amplitudes are dispersed; after determining the number of intervals, the difference amplitude sequence is divided into multiple consecutive groups according to the principle of balanced distribution of sample numbers, so that the number of samples contained in each consecutive group is as close as possible; then reading the minimum and maximum difference amplitudes in each consecutive group, and determining the position between the maximum difference amplitude of the previous consecutive group and the minimum difference amplitude of the next consecutive group as the boundary position, thus forming multiple consecutive difference intervals; finally, assigning interval numbers to each consecutive difference interval according to the difference amplitude size to obtain the difference interval table.

[0088] After establishing the difference interval table, the difference magnitude corresponding to the current sample is compared with each consecutive difference interval in the difference interval table, and the interval number of the consecutive difference interval into which the difference magnitude corresponding to the current sample falls is determined as the difference interval identifier. When the difference magnitude coincides with a certain boundary position, the difference magnitude is assigned to the adjacent consecutive difference interval with a larger value, so as to keep the difference interval identifier monotonically changing along the direction of increasing difference magnitude. The difference interval identifier serves as the input object for subsequent boundary offset coefficient retrieval.

[0089] Step 4024: Generate measurement method difference results; combine the difference amplitude obtained in step 4022, the difference direction obtained in step 4022, the difference interval identifier obtained in step 4023, and the difference generation time to obtain the measurement method difference results; the difference generation time is defined as the time when the combination process is completed in step 4024; the measurement method difference results serve as the input object of the correction module.

[0090] In some implementations, the training process of the lightweight neural network model includes: collecting first EMI observation data and second EMI observation data from multiple target devices or the same target device in multiple current operating states, and extracting the corresponding first EMI feature results and second EMI feature results; calculating the true compliance margin value corresponding to each observation sample according to the limit table; forming training input samples by expanding, completing, normalizing, and adding measurement method identifiers to the first EMI feature results and second EMI feature results; forming training output labels with the corresponding true compliance margin values; dividing the training input samples and training output labels into training datasets, validation datasets, and test datasets; constructing a lightweight neural network model including an input layer, a first fully connected layer, a second fully connected layer, and an output layer; and using quantization-aware processing for weight parameters and activation values ​​during the training phase; using the sum of the weighted mean square error term and parameter constraint term as the training error function, updating the model parameters through mini-batch iteration, and judging training convergence based on the change in the validation error function value or the number of training rounds; after training convergence, inputting the test dataset into the trained lightweight neural network model for testing and verification; and obtaining a lightweight neural network model for deployment when the test error is within the preset test allowable range.

[0091] In some implementations, the method for calculating the true compliance margin value corresponding to each observation sample based on the limit table includes: first, reading the measurement method identifier, target equipment category identifier, and frequency point sequence corresponding to the observation sample; then, using the measurement method identifier and target equipment category identifier as search conditions, extracting the limit sequence corresponding one-to-one with the frequency point sequence from the limit table, wherein the limit table is a pre-established set of limit data according to the measurement method, frequency point, and target equipment category; subsequently, for each frequency point in the observation sample, reading the observed amplitude corresponding to that frequency point and the limit sequence corresponding to the frequency point. The limit value for each frequency point is determined, and the corresponding observed amplitude is subtracted from the corresponding limit value to obtain the single-point compliance margin for that frequency point. After repeating the above process for all frequency points, a single-point compliance margin sequence is formed. Then, the item with the smallest value in the single-point compliance margin sequence is selected as the true compliance margin value corresponding to the observed sample. A positive single-point compliance margin indicates that the observed amplitude of the frequency point is within the limit value, a zero single-point compliance margin indicates that the observed amplitude of the frequency point is at the limit value boundary, and a negative single-point compliance margin indicates that the observed amplitude of the frequency point exceeds the limit value.

[0092] The correction module is used to perform boundary offset correction on the smaller of the first compliance margin result and the second compliance margin result based on the measurement method difference results, to obtain the target compliance margin result. The purpose is to use the smaller of the first and second compliance margin results as the basis for boundary correction, and then combine it with the measurement method difference results to perform conservative correction, forming a target compliance margin result for actual control. This avoids outputting control commands based solely on a single measurement method. Specific implementation methods include: Step 501: Read the first compliance margin result, the second compliance margin result, and the measurement method difference result; first read the first compliance margin result and the second compliance margin result output in step 4016, and then read the measurement method difference result output in step 4024; then extract the first compliance margin value, the second compliance margin value, the difference magnitude, the difference direction, and the difference interval identifier respectively.

[0093] Step 501 outputs the first compliance margin value, the second compliance margin value, the difference magnitude, the difference direction, and the difference interval identifier, which serve as the input objects for steps 502 to 505.

[0094] Step 502: Determine the boundary correction baseline value; compare the first compliance margin value with the second compliance margin value; when the first compliance margin value is less than or equal to the second compliance margin value, determine the first compliance margin value as the boundary correction baseline value, and determine the first result source identifier as the correction source identifier; When the second compliance margin value is less than the first compliance margin value, the second compliance margin value is determined as the boundary correction base value, and the second result source identifier is determined as the correction source identifier; the boundary correction base value and the correction source identifier are used as input objects for steps 505 and 506.

[0095] Step 503: Determine the boundary offset coefficient based on the difference interval identifier; first, establish a boundary offset coefficient table; the process of establishing the boundary offset coefficient table includes: reading the difference interval identifier, first compliance margin value, second compliance margin value, first true compliance margin value, and second true compliance margin value corresponding to each sample in the training phase; then, for each training sample, compare the first compliance margin value with the first true compliance margin value, and the second compliance margin value with the second true compliance margin value, and select the deviation amount where the predicted value is greater than the true value as the prediction deviation value of the current sample; if neither the first compliance margin value nor the second compliance margin value is greater than the corresponding true compliance margin value, then the current sample's... The prediction bias is recorded as zero. Then, the prediction bias of the current sample is correlated with the corresponding difference magnitude of the current sample to obtain the deviation magnitude correspondence of the current sample. Then, all training samples are grouped according to the difference interval label, so that training samples with the same difference interval label are grouped into the same interval sample set. For any interval sample set, the prediction bias and difference magnitude of all samples in it are read, and the deviation ratio of the prediction bias value to the difference magnitude is calculated one by one. Then, all deviation ratios are sorted according to the numerical size, and according to the preset conservative selection rule, the ratio value closer to the larger end of the sorted deviation ratios is selected as the boundary offset coefficient corresponding to the difference interval label.

[0096] After determining the boundary offset coefficients corresponding to all difference interval identifiers, monotonicity correction is performed on the boundary offset coefficients according to the size pattern of the difference interval identifiers. When the boundary offset coefficient corresponding to the next difference interval identifier is less than the boundary offset coefficient corresponding to the previous difference interval identifier, the boundary offset coefficient corresponding to the next difference interval identifier is adjusted to be no less than the boundary offset coefficient corresponding to the previous difference interval identifier, so that the boundary offset coefficient does not decrease as the difference increases. This forms the boundary offset coefficient table. After the boundary offset coefficient table is established, the boundary offset coefficient corresponding to the same difference interval identifier is retrieved from the boundary offset coefficient table according to the difference interval identifiers obtained in step 501. The boundary offset coefficient is used as the input object for subsequent boundary offset calculation.

[0097] Step 504: Calculate the boundary offset; read the difference magnitude obtained in step 501 and the boundary offset coefficient obtained in step 503, and then multiply the difference magnitude by the boundary offset coefficient to obtain the boundary offset; the boundary offset is used to characterize the conservative correction amount introduced by the difference in measurement methods; the boundary offset is used as the input object for steps 505 and 506.

[0098] Step 505: Generate target compliance margin value; read the boundary correction base value obtained in step 502, and then read the boundary offset obtained in step 504; then subtract the boundary offset from the boundary correction base value to obtain the target compliance margin value.

[0099] When the target compliance margin value is less than the preset output lower limit, the target compliance margin value is limited to the preset output lower limit; the preset output lower limit is determined based on the minimum margin range that the control execution terminal is allowed to receive; the target compliance margin value is the core value for forming the target compliance margin result in step 506.

[0100] Further, to explain the measurement method difference quantification and boundary offset correction process in steps 4022 to 505, assume that the first compliance margin value obtained in step 4016 is 3.2 dB and the second compliance margin value is 1.8 dB; firstly, step 4022 subtracts the second compliance margin value from the first compliance margin value to obtain an original difference value of 1.4 dB; then, the absolute value of the original difference value is taken to obtain a difference amplitude of 1.4 dB; since the original difference value is greater than zero, the difference direction is marked as the first direction; subsequently, step 4023 compares 1.4 dB with each continuous difference interval in the difference interval table to determine the continuous difference interval to which 1.4 dB belongs, and determines the number corresponding to this continuous difference interval as the difference interval identifier; then, step 502 compares... Compared to the first compliance margin value and the second compliance margin value, since the second compliance margin value is smaller, 1.8 dB is determined as the baseline value for boundary correction. Then, in step 503, the corresponding boundary offset coefficient is retrieved from the boundary offset coefficient table based on the difference interval identifier. Assuming that the retrieved boundary offset coefficient is 0.35, in step 504, the difference amplitude of 1.4 dB is multiplied by the boundary offset coefficient of 0.35 to obtain a boundary offset of 0.49 dB. Finally, in step 505, the boundary offset of 0.49 dB is subtracted from the baseline value of boundary correction of 1.8 dB to obtain the target compliance margin value of 1.31 dB. In step 506, the target compliance margin value, the correction source identifier, the boundary offset, and the target result generation time can be further combined to form the target compliance margin result.

[0101] Step 506: Form the target compliance margin result; combine the target compliance margin value obtained in step 505, the correction source identifier obtained in step 502, the boundary offset obtained in step 504, and the target result generation time to obtain the target compliance margin result; the target result generation time is defined as the time when the combination process is completed in step 506; the target compliance margin result serves as the input object of the control module.

[0102] The control module is used to control the EMI filtering mode switching, filtering parameter adjustment, or compensation signal output of the target equipment based on the target compliance margin result. Its purpose is to convert the target compliance margin value into actual control actions for the target equipment, enabling the target equipment to perform EMI filtering mode switching, filtering parameter adjustment, or compensation signal output under different compliance margin states. Specific implementation methods include: See Figure 3As shown, in step 601, read the target compliance margin result and the current operating status identifier; first read the target compliance margin result and extract the target compliance margin value, correction source identifier and boundary offset; then read the current operating status identifier confirmed in step 301.

[0103] Step 602: Determine the target control interval. First, establish a margin control interval table. The process of establishing the margin control interval table includes: reading historical operating data, which includes at least the compliance margin value corresponding to the historical sample, the control action category executed by the historical sample, and the result change data after the control action is executed; among which, the control action category includes at least one of EMI filter mode switching, filter parameter adjustment, and compensation signal output, and the result change data includes at least one of the compliance margin change after the control action is executed, the stable holding time, and the number of repeated adjustments; then, sort the compliance margin values ​​corresponding to all historical samples according to the numerical order to form a historical margin sequence; then, determine the number of initial intervals based on the distribution range of the historical margin sequence and the number of historical samples, and divide the historical margin sequence into multiple continuous initial intervals; then, for each continuous initial interval, statistically analyze the result change data corresponding to different control action categories to determine the compliance margin improvement within that continuous initial interval. For control action categories with larger increases and fewer repeated adjustments, this category is determined as the dominant control action category for the continuous initial interval. Next, the dominant control action categories of adjacent continuous initial intervals are compared. When the dominant control action categories of adjacent continuous initial intervals are the same, they are merged into the same continuous margin interval. When the dominant control action categories of adjacent continuous initial intervals are different, the position between them is determined as the interval boundary, thus forming multiple continuous margin intervals. Finally, the continuous margin intervals are arranged sequentially according to the size of the target compliance margin value, and each continuous margin interval is assigned an interval number, forming a margin control interval table. After the margin control interval table is established, the target compliance margin value obtained in step 601 is compared item by item with each continuous margin interval in the margin control interval table, and the interval number corresponding to the continuous margin interval into which the target compliance margin value falls is determined as the target control interval. The target control interval serves as the input object for steps 603 and 607.

[0104] Step 603: Determine the control command based on the target control interval and the current operating status identifier; first, establish a control parameter table; the process of establishing the control parameter table includes: reading historical operating data and debugging sample data, which at least include the target control interval, the current operating status identifier, the control action type, the target filtering mode, the filtering parameters, the compensation signal parameters, and the result change data after the control action is executed; then, group all samples according to the target control interval and the current operating status identifier, so that samples with the same target control interval and the same current operating status identifier are grouped into the same parameter candidate set; for any parameter candidate set, statistically analyze the result change data corresponding to different target filtering modes, different combinations of filtering parameters, and different combinations of compensation signal parameters, and determine those that result in a larger increase in compliance margin, fewer repeated adjustments, and a longer stable maintenance time. A long set of parameter combinations serves as the target control command corresponding to the parameter candidate set; wherein, the target filtering mode is used to indicate that the target device is in the first filtering mode, the second filtering mode, or the third filtering mode, and the filtering parameters include at least one of the following: capacitor branch access information, inductor tap selection information, and damping element access information; the compensation signal parameters include at least the compensation enable flag, the compensation frequency band parameters, and the compensation amplitude parameters and compensation phase parameters corresponding to each discrete compensation frequency point within the compensation frequency band; then, the target control interval, the current operating status flag, and the corresponding target control command are associated to form a control parameter table; after the control parameter table is established, the corresponding control command is retrieved from the control parameter table based on the target control interval obtained in step 602 and the current operating status flag obtained in step 601; the control command serves as the input object for steps 604 to 606.

[0105] Step 604: Perform EMI filtering mode switching. Based on the target filtering mode output in step 603, perform EMI filtering mode switching processing on the target device. When the target filtering mode is inconsistent with the current filtering mode used by the target device, first control the corresponding mode switching device to operate, and then switch the target device to the target filtering mode. The mode switching device is at least one of a relay, an electronic switch, or a programmable logic device control port. When the target filtering mode is the first filtering mode, keep the basic filtering branch in the connected state. When the target filtering mode is the second filtering mode, connect the additional filtering branch or switch to the enhanced damping state. When the target filtering mode is the third filtering mode, keep the basic filtering branch in the connected state and provide enable conditions for subsequent compensation signal output. Step 604 completes the EMI filtering mode switching of the target device and outputs the target filtering mode switching result, which serves as one of the comparison bases for review in step 607.

[0106] Step 605: Perform filter parameter adjustment; based on the filter parameters output in step 603, perform parameter adjustment processing on the adjustable filter components in the target device; when the filter parameters include capacitor branch access information, control the corresponding electronic switch to turn on or off, so that the corresponding capacitor branch is connected or disconnected; when the filter parameters include inductor tap selection information, control the corresponding switching device to select the corresponding inductor tap position; when the filter parameters include damping element access information, control the corresponding electronic switch to change the damping element access status.

[0107] After completing all parameter adjustments, the actual connection results of each adjustable filter device are read, and the read results are matched with the filter parameters output in step 603 to obtain the filter parameter status. The filter parameter status includes at least one of the following: capacitor branch connection status, inductor tap position status, and damping element connection status.

[0108] Step 606: Execute compensation signal output; based on the compensation signal parameters, perform compensation signal output processing on the target device; when the compensation enable flag in the compensation signal parameters indicates that a compensation signal needs to be output, first read the compensation frequency band parameters, and extract the compensation start frequency, compensation end frequency, and compensation frequency interval rule from the compensation frequency band parameters; then, starting from the compensation start frequency, select multiple frequency points sequentially between the compensation start frequency and the compensation end frequency according to the compensation frequency interval rule, and determine the selected multiple frequency points as discrete compensation frequency points; when the next frequency point selected according to the compensation frequency interval rule exceeds the compensation end frequency, stop frequency point selection; then read the compensation amplitude parameters and compensation phase parameters corresponding to each discrete compensation frequency point respectively; then establish a preset output clock, which is formed by: first reading the maximum frequency point among all discrete compensation frequency points, and then determining the output clock frequency according to the preset sampling point number rule per cycle, so that a complete cycle corresponding to the maximum frequency point contains at least a preset number of sampling moments; the preset sampling point number rule per cycle is determined based on the compensation waveform reconstruction accuracy requirements.

[0109] After determining the preset output clock, a corresponding sinusoidal digital sequence is generated for any discrete compensation frequency point, using the preset output clock as the sampling reference. Specifically, the number of sampling points corresponding to the discrete compensation frequency point within a complete cycle is first determined based on the preset output clock. Then, based on the number of sampling points, the compensation amplitude parameter, and the compensation phase parameter, sinusoidal sample values ​​corresponding to each sampling time within a complete cycle are generated sequentially, thus forming a single-frequency digital sequence corresponding to the discrete compensation frequency point. The amplitude of the sinusoidal digital sequence is determined by the compensation amplitude parameter, and the phase is determined by the compensation phase parameter. After performing the above processing on all discrete compensation frequency points, the single-frequency digital sequences corresponding to all discrete compensation frequency points are then added point by point according to the sampling time, forming... A digital compensation sequence is generated. The sequence length of the digital compensation sequence is determined based on the common repetition period of all single-frequency digital sequences. When all single-frequency digital sequences have a common repetition period within a preset maximum sequence length range, the number of sampling points corresponding to that common repetition period is determined as the sequence length of the digital compensation sequence. When all single-frequency digital sequences do not have a common repetition period within a preset maximum sequence length range, the number of sampling points corresponding to the preset maximum sequence length is determined as the sequence length of the digital compensation sequence. The update period of the digital compensation sequence is determined by the sequence length and a preset output clock, i.e., after each complete digital compensation sequence is output, the next update period begins, and the digital compensation sequence is regenerated in the next update period. Subsequently, the digital compensation sequence is input to a digital-to-analog converter to obtain an analog compensation signal. Finally, the analog compensation signal is applied to the compensation injection position. The compensation injection position is set at least one of the following: the input side of the target device, the reference ground side, or the front end of the filter device.

[0110] In some implementations, when the compensation enable flag in the compensation signal parameters indicates that a compensation signal needs to be output, it can be represented by a numerical flag or a logical flag. Specifically, the compensation enable flag is first read from the compensation signal parameters obtained in step 603. If the compensation enable flag takes a first preset value, it is determined that a compensation signal needs to be output. The first preset value can be set to a numerical value "1", a logic high level, or an "enable" state code. To avoid malfunctions caused by a single flag, the compensation frequency band parameters, compensation amplitude parameters, and compensation phase parameters can be combined for joint confirmation. That is, when the compensation enable flag takes a first preset value, the compensation frequency band range corresponding to the compensation frequency band parameter is not empty, and there is at least one non-zero amplitude term in the compensation amplitude parameter, the compensation enable flag is determined to indicate that a compensation signal needs to be output, and the process proceeds to digital compensation sequence generation and analog compensation signal output processing.

[0111] When the compensation enable flag in the compensation signal parameters indicates that no compensation signal is output, the compensation injection device is kept in the off state; step 606 completes the compensation signal output to the target device and outputs the compensation execution result, which includes at least one of the following: compensation enable state, compensation frequency band range, and compensation output time.

[0112] In some implementations, when the compensation enable flag in the compensation signal parameters indicates that no compensation signal is output, it can also be represented by a numerical flag or a logical flag. Specifically, the compensation enable flag in the compensation signal parameters retrieved in step 603 is first read; if the compensation enable flag takes a second preset value, it is determined that no compensation signal is output, wherein the second preset value can be set to a numerical value "0", a logic low level, or a "closed" state code. To improve the stability of the determination, it can also be further checked in conjunction with the compensation frequency band parameters, compensation amplitude parameters, and compensation phase parameters. That is, when the compensation enable flag takes a second preset value, or the compensation frequency band range corresponding to the compensation frequency band parameter is empty, or all compensation amplitude parameters are zero, the compensation enable flag is determined to indicate that no compensation signal is output, and the compensation injection device is kept in the closed state, and the digital compensation sequence generation and analog compensation signal application processing are not performed.

[0113] Step 607: Perform control result verification; after steps 604 to 606 are completed, re-perform observation, feature extraction, compliance margin inference, difference generation, and boundary offset correction processing on the target equipment to obtain updated target compliance margin results; then extract the updated target compliance margin value and determine the updated target control interval according to the margin control interval determination method in step 602; then compare the updated target control interval with the target control interval obtained in step 602; when the updated target control interval is consistent with the target control interval obtained in step 602, continue with step 604. The steps include: EMI filter mode switching results, filter parameter adjustment results executed in step 605, and compensation signal output results executed in step 606; when the updated target control range is inconsistent with the target control range obtained in step 602, the process returns to step 603, and the control command is re-determined based on the updated target control range and the current operating status identifier, and steps 604 to 606 are executed sequentially until the updated target control range is consistent with the re-determined target control range; step 607 completes the verification process for EMI filter mode switching, filter parameter adjustment, or compensation signal output of the target device.

[0114] This example illustrates the control execution process in steps 602 to 607. Assuming the target compliance margin value formed in step 506 is 1.31 dB, after comparing 1.31 dB with each consecutive margin interval in the margin control interval table, the target control interval to which it belongs is determined to be the second interval. Subsequently, based on the second interval and the current operating status identifier, the corresponding control command is retrieved from the control parameter table. The control command includes the second filtering mode, target capacitor branch access information, target inductor tap selection information, target damping element access information, and compensation signal parameters. After executing EMI filtering mode switching, filtering parameter adjustment, and compensation signal output according to the control command, the target device is then re-observed, feature extraction, compliance margin inference, difference generation, and boundary offset correction processing are performed. Assuming the updated target compliance margin value obtained after verification is 2.6 dB, and 2.6 dB still falls within the second range, then the aforementioned EMI filter mode switching results, filter parameter adjustment results, and compensation signal output results are maintained. If the updated target compliance margin value obtained after verification falls into other control ranges, then return to step 603 to re-search for the corresponding control command, and execute steps 604 to 606 again until the updated target control range is consistent with the re-determined target control range.

[0115] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A lightweight neural network-assisted EMI filtering system, characterized in that, include: The acquisition module is used to acquire the first EMI observation data collected by the target device in its current operating state according to the first measurement method; The conversion module is used to perform an equivalent conversion on the first EMI observation data based on the pre-established measurement mapping relationship between the first measurement method and the second measurement method, so as to obtain the second EMI observation data corresponding to the second measurement method. The extraction module is used to perform feature extraction on the first EMI observation data and the second EMI observation data respectively, to obtain the first EMI feature results and the second EMI feature results; The determination module is used to input the first EMI feature result and the second EMI feature result into the lightweight neural network model to obtain the first compliance margin result and the second compliance margin result, and to generate the measurement method difference result based on the difference between the first compliance margin result and the second compliance margin result. The correction module is used to perform boundary offset correction on the smaller of the first compliance margin result and the second compliance margin result based on the difference in measurement methods, to obtain the target compliance margin result. The control module is used to control the EMI filtering mode switching, filtering parameter adjustment, or compensation signal output of the target device based on the target compliance margin result.

2. The lightweight neural network-assisted EMI filtering system according to claim 1, characterized in that, The acquisition module is used to obtain the first EMI observation data through the following method: Read the target device's operating mode, load level, power supply status, and switching parameters, and combine them in a fixed order to form the current operating status identifier; Based on the first measurement location and the first measurement frequency point sequence corresponding to the first measurement method, multiple sample values ​​are collected point by point at each frequency. Outlier removal and statistical processing are performed on the multiple sample values ​​corresponding to each frequency point to obtain the first amplitude sequence. Under the same current operating state, the current operating state identifier, the first measurement frequency point sequence, and the first amplitude sequence are correlated accordingly, and the results after correlation are repeatedly collected and verified to obtain the first EMI observation data.

3. The lightweight neural network-assisted EMI filtering system according to claim 1, characterized in that, The conversion module is used to establish a measurement mapping relationship between the first measurement method and the second measurement method through the following method: Multiple reference operating states are selected, and data are collected under each reference operating state according to the first measurement method and the second measurement method to obtain the first reference observation data and the second reference observation data. Align the first and second reference observation data according to the reference operating status identifier, frequency point, and acquisition time, and calculate the scaling factor and offset factor based on the first and second reference amplitudes corresponding to any reference operating status and any frequency point. The reference operating status identifier, frequency point, proportional coefficient, and offset coefficient are combined to form a measurement mapping relationship table to establish the measurement mapping relationship between the first measurement method and the second measurement method.

4. The lightweight neural network-assisted EMI filtering system according to claim 2, characterized in that, The conversion module is used to obtain the second EMI observation data through the following method: Extract the current operating status identifier, the first measurement frequency point sequence, and the first amplitude sequence from the first EMI observation data, and match the target mapping entry set based on the current operating status identifier; When there is no mapping frequency point in the target mapping entry set that is completely consistent with the target frequency point, select the left and right mapping frequency points located on the left and right sides of the target frequency point, calculate the interpolation scaling factor and interpolation offset factor, and perform frequency alignment processing on the first amplitude sequence. Then, calculate the second equivalent amplitude based on the amplitude after frequency alignment processing, the interpolation scaling factor and the interpolation offset factor. Second EMI observation data are generated based on the second equivalent amplitude.

5. The lightweight neural network-assisted EMI filtering system according to claim 1, characterized in that, The extraction module is used to obtain the first EMI feature result and the second EMI feature result through the following methods: The first EMI observation data and the second EMI observation data are aligned to obtain the first aligned amplitude sequence and the second aligned amplitude sequence; Smoothing processes are performed on the first aligned amplitude sequence and the second aligned amplitude sequence respectively to obtain the first basic amplitude sequence and the second basic amplitude sequence; Based on the first basic amplitude sequence, the first characteristic frequency point sequence, the first characteristic amplitude sequence, the first frequency band energy sequence, the first adjacent change sequence, and the first time series fluctuation sequence are extracted respectively. Based on the second basic amplitude sequence, the second characteristic frequency point sequence, the second characteristic amplitude sequence, the second frequency band energy sequence, the second adjacent change sequence, and the second time series fluctuation sequence are extracted respectively. The first EMI characteristic result and the second EMI characteristic result are generated according to a unified structure.

6. The lightweight neural network-assisted EMI filtering system according to claim 1, characterized in that, The determination module is used to obtain the first compliance margin result and the second compliance margin result through the following methods: The first EMI feature result is expanded, completed, and normalized according to a fixed arrangement order to form a first normalized input vector. The first measurement method identifier value is appended to the end of the first normalized input vector to obtain the first inference input vector. The second EMI feature result is expanded, completed, and normalized using the same method to form a second normalized input vector. The second measurement method identifier value is then appended to the end of the second normalized input vector to obtain the second inference input vector. The first inference input vector and the second inference input vector are then input into the lightweight neural network model, which outputs the first compliance margin value and the second compliance margin value, and forms the first compliance margin result and the second compliance margin result, respectively.

7. The lightweight neural network-assisted EMI filtering system according to claim 1, characterized in that, The lightweight neural network model is obtained through the following training method: Collect first and second EMI observation data during the training phase, extract the corresponding first and second EMI feature results, and calculate the true compliance margin value corresponding to each observation sample according to the limit table to form training input samples and training output labels. A regression structure consisting of an input layer, a first fully connected layer, a second fully connected layer, and an output layer is constructed, and quantization-aware processing is used for weight parameters and activation values ​​during the training phase. Then, the sum of the weighted mean square error term and the parameter constraint term is used as the training error function to perform model training and training convergence judgment, so as to obtain a lightweight neural network model.

8. The lightweight neural network-assisted EMI filtering system according to claim 1, characterized in that, The determination module is used to generate measurement method difference results using the following methods: Extract the first compliance margin value from the first compliance margin result and the second compliance margin value from the second compliance margin result respectively. Subtract the second compliance margin value from the first compliance margin value to obtain the original difference. Take the absolute value of the original difference to obtain the difference amplitude, and determine the difference direction based on the sign of the original difference. Compare the difference magnitude with each consecutive difference interval in the difference interval table to determine the consecutive difference interval to which the difference magnitude belongs, and determine the number corresponding to the consecutive difference interval as the difference interval identifier; The difference magnitude, difference direction, and difference interval identifiers are combined to generate the measurement method difference results.

9. The lightweight neural network-assisted EMI filtering system according to claim 8, characterized in that, The calibration module is used to obtain the target compliance margin result through the following methods: Compare the first compliance margin value and the second compliance margin value, and determine the smaller one as the boundary correction baseline value; based on the difference interval identifier in the measurement method difference results, retrieve the corresponding boundary offset coefficient from the preset boundary offset coefficient table; Multiply the difference magnitude by the boundary offset coefficient to obtain the boundary offset, and subtract the boundary offset from the boundary correction base value to obtain the target compliance margin value; By associating the target compliance margin value with the boundary offset, the target compliance margin result is obtained.

10. The lightweight neural network-assisted EMI filtering system according to claim 9, characterized in that, The control module is used to control the EMI filtering mode switching, filtering parameter adjustment, or compensation signal output of the target device through the following methods: The target control interval is determined based on the comparison results between the target compliance margin value and each continuous margin interval in the preset margin control interval table. The corresponding control command is retrieved from the control parameter table based on the target control interval and the current operating status identifier. The control commands include the target filtering mode, filtering parameters, and compensation signal parameters; The system performs EMI filtering mode switching based on the target filtering mode, filters parameter adjustment based on the filtering parameters, and compensation signal output based on the compensation signal parameters. After completing EMI filtering mode switching, filters parameter adjustment, or compensation signal output, the corresponding control results are reviewed.

Images