Large language model assisted test measurement method and system

Abstract

The present invention provides a large language model-assisted test and measurement method and system, relating to the field of test and measurement technology. The method comprises collecting raw measurement data, drawing images, and calculated parameters of a circuit under test through a measuring test device; determining the data content to be input into the large language model according to the type of question selected by the user; when the question type is a waveform quality-related question, inputting the drawing images and calculated parameters into the large language model; and when the question type is a data calculation and analysis-related question, inputting the raw measurement data and calculated parameters into the large language model; obtaining waveform change information of the circuit under test during the entire test and measurement process through a dynamic time warping algorithm based on the waveform similarity of the same channel at different times recorded over a full cycle; and generating a test and measurement analysis report, the analysis report including waveform quality analysis, circuit noise source analysis, existing problems and solutions, parameter information, and data calculation synthesis results.

Description

Technical Field

[0001] The present invention relates to test and measurement technology, and in particular to a large language model-assisted test and measurement method and system. Background Art

[0002] Electronic measurement and testing technology has become an indispensable component of modern scientific research, industrial production, and education. As an innovative tool in the field of electronic measurement, desktop test and measurement software has gained widespread popularity due to its convenience, efficiency, and accuracy.

[0003] Desktop test and measurement software displays complex electronic signals in intuitive, easy-to-understand waveforms, helping users quickly analyze and identify common signal characteristics in circuits. It not only displays key parameters such as amplitude, frequency, and phase in real time, but also supports a variety of signal processing and filtering functions.

[0004] During measurement and testing, understanding unusual waveforms and the reasons behind them often confuses novice or non-professional users. Limited analysis methods are not targeted enough in different scenarios. Summary of the Invention

[0005] The embodiments of the present invention provide a large language model-assisted test measurement method and system, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides large language model-assisted testing and measurement methods, including:

[0008] Collecting raw measurement data, a drawing image, and calculated parameters of the circuit under test using a measurement test device, wherein the raw measurement data is multi-dimensional numerical data, the drawing image is generated by drawing based on the raw measurement data, and the calculated parameters include voltage parameters and time parameters calculated based on the raw measurement data, and inputting at least two of the raw measurement data, the drawing image, and the calculated parameters into a large language model;

[0009] Determine the data content to be input into the large language model according to the type of question selected by the user; when the question type is a waveform quality-related question, input the drawing image and the measurement parameters into the large language model; when the question type is a data calculation and analysis-related question, input the measurement raw data and the measurement parameters into the large language model;

[0010] Through a dynamic time warping algorithm, based on the waveform similarity of the same channel at different times recorded throughout the entire cycle, information on changes in the waveform of the circuit under test during the entire test and measurement process is obtained; through the multimodal image recognition capability of the large language model, the input drawing image is understood and analyzed to obtain waveform quality analysis results and circuit noise source analysis results; and / or, through the generalized reasoning capability of the large language model, based on the input measurement raw data and measurement parameters, multi-dimensional parameter information is obtained, and mathematical operations are performed on the measurement raw data of multiple channels to synthesize new waveforms; a test and measurement analysis report is generated, which includes waveform quality analysis, circuit noise source analysis, existing problems and solutions, parameter information, and data operation synthesis results.

[0011] In an optional embodiment,

[0012] The dynamic time warping algorithm is used to obtain waveform change information of the circuit under test during the entire test and measurement process based on the waveform similarity of the same channel at different times during the full cycle recording. This includes:

[0013] Acquire multiple timing waveform data collected at different times by the same measurement channel of the circuit under test;

[0014] Based on a dynamic time warping algorithm, finding an optimal matching path between each two time series waveform data in the plurality of time series waveform data through dynamic programming, wherein the optimal matching path maximizes the similarity between the two time series waveform data;

[0015] Calculating the similarity between each two time series waveform data according to the optimal matching path to obtain similarity measurement values ​​of the time series waveform data of the same measurement channel at different times;

[0016] Based on the similarity metric value, obtaining a change trend of the waveform sequence of the same measurement channel during the entire test measurement process;

[0017] Generate test measurement full cycle record data, the record data contains the characteristic information and similarity change information of the same measurement channel waveform at different times, so as to record the state change of the test measurement object during the entire test process.

[0018] In an optional embodiment,

[0019] Calculating the similarity between each pair of time series waveform data according to the optimal matching path to obtain similarity measurement values ​​of the time series waveform data of the same measurement channel at different times includes:

[0020] Initialize the distance matrix and create a matrix of size (N+1)x(M+1), where N and M are the lengths of the two timing waveforms respectively. Set the initial value of the matrix to infinity and the origin of the matrix to 0.

[0021] Iteratively calculate the distance between the corresponding elements of two time series waveforms and update the distance matrix;

[0022] The distance matrix is ​​traced back to find a path with the shortest cumulative distance, which is used as the optimal matching path of the two time series waveforms. The shortest distance value is the similarity measurement value of the two time series waveform data.

[0023] In an optional embodiment,

[0024] Through the generalized reasoning capability of the large language model, based on the input measurement raw data and measurement parameters, more dimensional parameter information is obtained, and mathematical operations are performed on the measurement raw data of multiple channels to synthesize new waveforms, including:

[0025] Obtaining measurement raw data collected by multiple measurement channels of the measured object and measurement parameters calculated based on the measurement raw data, wherein the measurement raw data is waveform data in the time domain or frequency domain, and the measurement parameters include voltage parameters and time parameters;

[0026] Inputting the raw measurement data and corresponding measurement parameters into a pre-trained large language model, wherein the large language model is trained based on massive test measurement data and corresponding analysis results, and has the ability to understand, generalize, and reason about the test measurement data;

[0027] Utilizing the generalized reasoning capability of the large language model, high-dimensional feature information contained in the raw measurement data is extracted, and combined with the measurement parameters, deep feature parameters of multiple measurement channels of the measured object under different working conditions are obtained;

[0028] According to the preset mathematical operation rules, the raw data of multiple measurement channels are fused and processed to generate a new synthetic waveform corresponding to the actual working state of the measured object;

[0029] The novel synthetic waveform is configured with different mathematical operation rules according to actual application requirements, including linear combination, nonlinear transformation, and time-frequency domain conversion of waveform data from different measurement channels.

[0030] In an optional embodiment,

[0031] Determining the data content to be input into the large language model according to the type of question selected by the user, when the question type is a waveform quality-related question, inputting a drawing image and measurement parameters into the large language model, and when the question type is a data calculation and analysis-related question, inputting raw measurement data and measurement parameters into the large language model includes:

[0032] Obtaining the question type selected by the user, wherein the question type includes waveform quality related questions and data calculation and analysis related questions;

[0033] Adaptively determining the test measurement data type and content input to the large language model based on the question type selected by the user;

[0034] When the problem type is a waveform quality-related problem, the waveform drawing image of the measured object and related measurement parameters are input into the large language model, where the measurement parameters include numerical values ​​representing waveform characteristics calculated based on the measurement data;

[0035] When the problem type is a data calculation and analysis related problem, the original data obtained by measurement and collection and the relevant measurement parameters are input into the large language model, and the original data is a time series numerical vector.

[0036] In an optional embodiment,

[0037] The method further comprises:

[0038] The large language model adaptively calls the corresponding data processing flow and algorithm model according to the input test measurement data type and content;

[0039] When the input data type is a waveform drawing image, the image recognition algorithm is called to extract features and understand semantics of the input image, and combined with the measurement parameters, the analysis results related to the waveform quality are output;

[0040] When the input data type is raw measurement data, numerical calculation and statistical analysis algorithms are called to process the raw data, and combined with the measurement parameters, the calculation and analysis results related to the performance indicators of the measured object are output;

[0041] Among them, the large language model adopts a learning paradigm that combines pre-training and fine-tuning. Through pre-training on massive test measurement-related corpus, it has basic knowledge and common sense in the field of test measurement. For different types of test measurement problems, it uses customized data for fine-tuning to enable it to quickly adapt to specific problems.

[0042] According to a second aspect of the embodiments of the present invention,

[0043] Provides a large language model-assisted test and measurement system, including:

[0044] A first unit is configured to collect raw measurement data, a drawing image, and calculated parameters of a circuit under test using a measurement and testing device, wherein the raw measurement data is multi-dimensional numerical data, the drawing image is generated by drawing based on the raw measurement data, and the calculated parameters include voltage parameters and time parameters calculated based on the raw measurement data, and input at least two of the raw measurement data, the drawing image, and the calculated parameters into a large language model;

[0045] The second unit is used to determine the data content to be input into the large language model according to the question type selected by the user. When the question type is a waveform quality-related question, the drawing image and measurement parameters are input into the large language model. When the question type is a data calculation and analysis-related question, the measurement raw data and measurement parameters are input into the large language model.

[0046] The third unit is used to obtain waveform change information of the tested circuit during the entire test and measurement process through a dynamic time warping algorithm based on the waveform similarity of the same channel at different times recorded throughout the entire cycle; understand and analyze the input drawing image through the multimodal image recognition capability of the large language model to obtain waveform quality analysis results and circuit noise source analysis results; and / or, through the generalized reasoning capability of the large language model, obtain parameter information of more dimensions based on the input measurement raw data and measurement parameters, perform mathematical operations on the measurement raw data of multiple channels to synthesize new waveforms; and generate a test measurement analysis report, which includes waveform quality analysis, circuit noise source analysis, existing problems and solutions, parameter information, and data operation synthesis results.

[0047] According to a third aspect of the embodiments of the present invention,

[0048] An electronic device is provided, comprising:

[0049] processor;

[0050] a memory for storing processor-executable instructions;

[0051] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0052] According to a fourth aspect of the embodiments of the present invention,

[0053] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0054] This application can improve the user's ability to interpret data during signal measurement; improve the efficiency of non-professional users in solving problems caused by unusual waveforms; and record debugging test and measurement behaviors throughout the entire cycle to quickly focus on problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Schematic diagram of the flow of a large language model-assisted test measurement method according to an embodiment of the present invention;

[0056] Figure 2 Schematic diagram of the structure of a large language model assisted test and measurement system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0058] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0059] Figure 1 FIG. 1 is a flow chart of a large language model-assisted test measurement method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] S101. Collecting raw measurement data, a drawing image, and calculated parameters of the circuit under test using a measurement test device, wherein the raw measurement data is multi-dimensional numerical data, the drawing image is generated based on the raw measurement data, and the calculated parameters include voltage parameters and time parameters calculated based on the raw measurement data. Inputting at least two of the raw measurement data, the drawing image, and the calculated parameters into a large language model;

[0061] S102. Determine the data content to be input into the large language model based on the question type selected by the user. When the question type is a waveform quality-related question, input the drawing image and measurement parameters into the large language model. When the question type is a data calculation and analysis-related question, input the measured raw data and measurement parameters into the large language model.

[0062] S103. Through the dynamic time warping algorithm, based on the waveform similarity of the same channel at different times recorded throughout the entire cycle, the waveform change information of the tested circuit during the entire test and measurement process is obtained; through the multimodal image recognition capability of the large language model, the input drawing image is understood and analyzed to obtain waveform quality analysis results and circuit noise source analysis results; and / or, through the generalized reasoning capability of the large language model, based on the input measurement raw data and measurement parameters, more dimensional parameter information is obtained, and mathematical operations are performed on the measurement raw data of multiple channels to synthesize new waveforms; a test measurement analysis report is generated, and the analysis report includes waveform quality analysis, circuit noise source analysis, existing problems and solutions, parameter information, and data operation synthesis results.

[0063] In an optional embodiment,

[0064] The dynamic time warping algorithm is used to obtain waveform change information of the circuit under test during the entire test and measurement process based on the waveform similarity of the same channel at different times during the full cycle recording. This includes:

[0065] Acquire multiple timing waveform data collected at different times by the same measurement channel of the circuit under test;

[0066] Based on a dynamic time warping algorithm, finding an optimal matching path between each two time series waveform data in the plurality of time series waveform data through dynamic programming, wherein the optimal matching path maximizes the similarity between the two time series waveform data;

[0067] Calculating the similarity between each two time series waveform data according to the optimal matching path to obtain similarity measurement values ​​of the time series waveform data of the same measurement channel at different times;

[0068] Based on the similarity metric value, obtaining a change trend of the waveform sequence of the same measurement channel during the entire test measurement process;

[0069] Generate test measurement full cycle record data, the record data contains the characteristic information and similarity change information of the same measurement channel waveform at different times, so as to record the state change of the test measurement object during the entire test process.

[0070] For example, this application proposes a full-cycle waveform similarity analysis method based on a dynamic time warping algorithm to obtain waveform variation information for a circuit under test throughout its entire measurement process. This method comprehensively monitors state changes of the object under test by analyzing waveform similarities across the same channel at different times. The following details the specific steps and technical content of this method:

[0071] First, it is necessary to obtain multiple time series waveform data collected at different times by the same measurement channel of the circuit under test. These data represent the different states of the object under test during the test and measurement process.

[0072] The specific steps are as follows:

[0073] 1) Set the sampling interval: Set an appropriate sampling interval based on the characteristics of the object being measured and the test requirements. For example, for a circuit that changes quickly, you can set it to 1ms; for a circuit that changes slowly, you can set it to 100ms.

[0074] 2) Determine the number of sampling points: Based on the test period and sampling frequency, determine the number of sampling points for each waveform. For example, if the test period is 1s and the sampling frequency is 1kHz, the number of sampling points for each waveform is 1000.

[0075] 3) Perform multiple sampling: During the entire test cycle, waveform data is collected multiple times at set time intervals. For example, if the test cycle is 1 hour and the sampling interval is 1 minute, a total of 60 waveform data points need to be collected.

[0076] 4) Data preprocessing: Preprocess the collected raw data, including denoising, normalization and other operations, for subsequent analysis.

[0077] For example, suppose a circuit has a 10-minute test cycle, waveform data is collected every 30 seconds, and each waveform contains 1000 sampling points. This results in 20 time-series waveforms, each of which is a one-dimensional array containing 1000 values.

[0078] Next, the dynamic time warping (DTW) algorithm is used to find the optimal matching path between each pair of time series waveform data. The DTW algorithm can handle time series of different lengths and is insensitive to local nonlinear scaling of the time axis, making it particularly suitable for waveform similarity analysis.

[0079] The implementation steps are as follows:

[0080] 1) Construct a distance matrix: For the two waveforms A and B to be compared, construct an m × n distance matrix D, where m and n are the lengths of the two waveforms. Each element D[i][j] in the matrix represents the distance between the i-th point in waveform A and the j-th point in waveform B.

[0081] 2) Initialize the cumulative distance matrix: Create a cumulative distance matrix C of the same size as the distance matrix D, and initialize the first element C[0][0] = D[0][0].

[0082] 3) Fill the cumulative distance matrix: Starting from the upper left corner, fill the cumulative distance matrix C by row or column. For each element C[i][j], its value is equal to D[i][j] plus the minimum value of its three adjacent elements to the left, above, and above the left.

[0083] 4) Backtrack to find the optimal path: Starting from the lower right corner of matrix C, by comparing the values ​​of adjacent elements, backtrack to the upper left until reaching the upper left corner, you can get the optimal matching path.

[0084] 5) Output the optimal matching path: record the path obtained by backtracking as the optimal matching result between the two waveforms.

[0085] Example: Suppose there are two simplified waveforms of length 5: A = [1, 2, 3, 4, 5] and B = [2, 3, 4, 5, 6]. Using the above steps, we can obtain an optimal matching path [(0, 0), (1, 1), (2, 2), (3, 3), (4, 4)]. This means that the first point of A matches the first point of B, the second point of A matches the second point of B, and so on.

[0086] Based on the obtained optimal matching path, the similarity between each two time series waveform data is calculated to obtain the similarity measurement values ​​of the time series waveform data of the same measurement channel at different times.

[0087] The calculation steps are as follows:

[0088] 1) Path distance calculation: Along the optimal matching path, accumulate the distance between each pair of matching points to obtain the total path distance.

[0089] 2) Normalization: Considering the influence of waveforms of different lengths, the total path distance is divided by the length of the matching path to obtain the normalized distance.

[0090] 3) Similarity conversion: Convert the normalized distance to a similarity value. You can use methods such as exponential functions to map the distance to a similarity value between 0 and 1, where larger values ​​indicate greater similarity.

[0091] 4) Construct a similarity matrix: Calculate the similarity of all collected waveforms pairwise and construct a similarity matrix.

[0092] Example: Assuming that a total of 5 waveforms are collected, the calculated similarity matrix is ​​as follows:

[0093] [1.00,0.95,0.88,0.82,0.75];

[0094] [0.95,1.00,0.93,0.87,0.80];

[0095] [0.88,0.93,1.00,0.94,0.86];

[0096] [0.82,0.87,0.94,1.00,0.91];

[0097] [0.75,0.80,0.86,0.91,1.00];

[0098] The value on the diagonal is 1, indicating that the similarity between the waveform and itself is 1. The matrix is ​​symmetrical. For example, 0.95 in the first row and second column indicates that the similarity between the first waveform and the second waveform is 0.95.

[0099] Based on the calculated similarity measurement value, the changing trend of the waveform sequence of the same measurement channel during the entire test measurement process is analyzed.

[0100] The analysis steps are as follows:

[0101] 1) Time series construction: The similarity values ​​of each waveform with all other waveforms form a time series. For example, for the first waveform, its time series is [1.00, 0.95, 0.88, 0.82, 0.75].

[0102] 2) Trend identification: Analyze the changing trend of each time series. Simple slope calculation or more complex time series analysis methods can be used.

[0103] 3) Threshold setting: Set the threshold for similarity change to determine whether the waveform change is significant. For example, a similarity drop of more than 0.1 can be considered a significant change.

[0104] 4) Change point detection: Detect significant change points in the waveform sequence based on the set threshold.

[0105] 5) Trend visualization: Use line graphs and other methods to visualize waveform change trends for intuitive understanding.

[0106] Example: Analyzing the first row of the similarity matrix [1.00, 0.95, 0.88, 0.82, 0.75], we can find that the waveform similarity shows a downward trend, and the similarity of the fourth and fifth waveforms relative to the first waveform decreases by more than 0.1, which can be determined as significant change points.

[0107] Finally, the full-cycle test measurement record data is generated, which contains the characteristic information and similarity change information of the waveform of the same measurement channel at different times, so as to record the state changes of the test measurement object during the entire test process.

[0108] The generation steps are as follows:

[0109] 1) Waveform feature extraction: Extract key features of each collected waveform, such as statistics such as maximum value, minimum value, mean value, standard deviation, as well as time domain features such as peak value, valley value, rise time, and fall time.

[0110] 2) Summary of similarity changes: Summarize the similarity changes between each waveform and other waveforms, including average similarity, maximum similarity drop, number of significant changes, etc.

[0111] 3) Status marking: Based on the similarity change and the preset threshold, the status at each moment is marked, such as "stable", "slight change", "significant change", etc.

[0112] 4) Data organization: Organize waveform features, similarity information and status marks in chronological order to form structured recorded data.

[0113] 5) Data storage: Save the generated full-cycle record data in an appropriate format (such as CSV, JSON, etc.) to facilitate subsequent analysis and processing.

[0114] Example: Assuming that 10 waveforms are collected during a test, the generated full-cycle recorded data may be as follows:

[0115] Time, maximum value, minimum value, mean value, standard deviation, average similarity, maximum similarity drop, number of significant changes, status;

[0116] 1,5.2,-3.1,1.5,2.3,1.00,0.00,0,stable;

[0117] 2,5.3,-3.0,1.6,2.2,0.95,0.05,0,stable;

[0118] 3,5.5,-2.8,1.8,2.1,0.91,0.09,0,slight change;

[0119] 4,5.8,-2.5,2.1,2.0,0.86,0.14,1,significant change;

[0120] 5,6.2,-2.1,2.5,1.9,0.81,0.19,2,significant change;

[0121] In this way, the state changes of the object under test during the entire test process can be fully recorded, providing an important basis for subsequent analysis and diagnosis.

[0122] In summary, the full-cycle waveform similarity analysis method based on dynamic time warping proposed in this application, through the steps of multi-moment waveform data acquisition, dynamic time warping algorithm implementation, similarity metric calculation, waveform change trend analysis, and full-cycle recorded data generation, achieves comprehensive analysis and recording of waveform changes throughout the entire test and measurement process of the circuit under test. This method can effectively capture the dynamic characteristics of the object under test, providing new ideas and methods for circuit performance evaluation and fault diagnosis.

[0123] In an optional embodiment,

[0124] Calculating the similarity between each pair of time series waveform data according to the optimal matching path to obtain similarity measurement values ​​of the time series waveform data of the same measurement channel at different times includes:

[0125] Initialize the distance matrix and create a matrix of size (N+1)x(M+1), where N and M are the lengths of the two timing waveforms respectively. Set the initial value of the matrix to infinity and the origin of the matrix to 0.

[0126] Iteratively calculate the distance between the corresponding elements of two time series waveforms and update the distance matrix;

[0127] The distance matrix is ​​traced back to find a path with the shortest cumulative distance, which is used as the optimal matching path of the two time series waveforms. The shortest distance value is the similarity measurement value of the two time series waveform data.

[0128] For example, this application details a method for calculating time series waveform similarity based on dynamic programming. This method accurately quantifies the similarity between any two time series waveform data sets by constructing a distance matrix, iteratively calculating element distances, and backtracking to find the optimal matching path. The following describes the specific implementation process and technical details of this method.

[0129] First, a distance matrix needs to be initialized to store the cumulative distance between each element of two time series waveforms.

[0130] The specific steps are as follows:

[0131] 1) Determine the matrix size: Assume that the two timing waveforms to be compared are A and B, with lengths N and M, respectively. Create a matrix D of size (N+1) × (M+1). The matrix size is one larger than the waveform length to handle edge cases.

[0132] 2) Initialize the matrix elements: Initialize all elements of the matrix D to a very large value, usually the largest floating point number in the programming language. This is to ensure that these initial values ​​do not interfere with the actual distance calculation in subsequent calculations.

[0133] 3) Set the starting point: Set the upper left corner element D[0][0] of matrix D to 0. This point represents the starting position of the two waveforms, and the cumulative distance is naturally 0.

[0134] 4) Processing boundary conditions: The first row and the first column can be selectively processed to reflect the matching of the starting part of the waveform.

[0135] Example:

[0136] Assume there are two simplified time series waveforms A = [1, 2, 3] and B = [2, 3, 4, 5]. The initialized distance matrix D (using inf to represent a very large number) is as follows:

[0137]

[0138] Next, the distance matrix is ​​gradually filled and updated by iteratively calculating the distance between the corresponding elements of the two time series waveforms.

[0139] The calculation steps are as follows:

[0140] 1) Define the distance metric: Choose an appropriate distance metric to calculate the distance between two waveform elements. Common methods include Euclidean distance and Manhattan distance. In this example, we use a simple absolute difference as the distance metric.

[0141] 2) Iteratively fill the matrix: Starting from the upper left corner, traverse the matrix D by row or column. For each element D[i][j], its value is equal to the distance between A[i-1] and B[j-1], plus the minimum value of the three adjacent elements D[i][j-1] on its left, D[i-1][j] above, and D[i-1][j-1] to the upper left.

[0142] 3) Boundary processing: Elements in the first row and column require special processing because they have no elements to the left or above them. Typically, their cumulative distance can be set to infinity, or appropriately adjusted based on the specific application.

[0143] 4) Complete filling: Repeat step 2 until the lower right corner element of matrix D is filled, at which point the entire matrix calculation is completed.

[0144] Example:

[0145] Continuing with the previous waveforms A = [1, 2, 3] and B = [2, 3, 4, 5], the filled distance matrix D is as follows:

[0146]

[0147] Finally, by backtracking the distance matrix, a path with the shortest cumulative distance is found as the optimal matching path of the two timing waveforms.

[0148] The backtracking steps are as follows:

[0149] 1) Determine the starting point: Start backtracking from the lower right corner element D[N][M] of the distance matrix D. The value of this element is the cumulative distance between the two complete waveforms.

[0150] 2) Select direction: For the current position (i, j), compare the values ​​of the three adjacent elements D[i-1][j] (above), D[i][j-1] (left), and D[i-1][j-1] (top left), and select the element with the smallest value as the next backtracking point.

[0151] 3) Record the path: Record the coordinates (i, j) of each step. These coordinates constitute the optimal matching path.

[0152] 4) Termination condition: When backtracking reaches the upper left corner of the matrix D[0][0], the backtracking process ends.

[0153] 5) Path reversal: Since backtracking is performed from the end point to the starting point, the recorded path needs to be reversed to obtain the optimal matching path from the starting point to the end point.

[0154] Example:

[0155] For the previously filled distance matrix, the optimal matching path obtained by backtracking is:

[0156] [(0,0),(1,1),(2,2),(3,3),(3,4)];

[0157] This means that the first point of waveform A matches the first point of B, the second point of A matches the second point of B, and the third point of A matches the third and fourth points of B.

[0158] Finally, the similarity measure between the two time series waveform data is determined based on the optimal matching path obtained by backtracking.

[0159] The steps to determine are as follows:

[0160] 1) Extract the shortest distance: The element D[N][M] in the lower right corner of the distance matrix D is the shortest cumulative distance.

[0161] 2) Normalization: Considering the influence of waveforms of different lengths, the accumulated distance can be divided by the path length for normalization.

[0162] 3) Similarity Conversion: Convert the normalized distance values ​​to similarity. A common method is to use an exponential function to map the distance to a similarity value between 0 and 1, where a larger value indicates greater similarity.

[0163] 4) Result output: Output the calculated similarity measurement value as a quantitative indicator of the similarity between the two time series waveform data.

[0164] Example:

[0165] For the previous example, the shortest cumulative distance is 2 (the value in the bottom right corner of the matrix).

[0166] Assume we use simple linear normalization and divide the distance by the path length of 5, resulting in a normalized distance of 0.4.

[0167] It is then converted into a similarity value using the exponential function e^(-0.4)≈0.67.

[0168] Therefore, the similarity measure value of waveforms A and B is 0.67, indicating that they have a moderate degree of similarity.

[0169] By repeating the above steps, the similarity between multiple time series waveform data collected by the same measurement channel at different times can be calculated, so as to obtain the similarity change of the time series waveform data of the measurement channel during the entire test and measurement process.

[0170] The specific analysis steps are as follows:

[0171] 1) Time window division: According to the test requirements, the entire test process is divided into multiple time windows.

[0172] 2) Waveform pairing: Pair the waveforms within each time window and the waveforms between adjacent time windows.

[0173] 3) Batch calculation: Use the above method to batch calculate the similarity measurement value between each pair of waveforms.

[0174] 4) Similarity matrix construction: All the calculated similarity metrics are organized into a matrix, where each element of the matrix represents the similarity between the waveforms at two specific moments.

[0175] 5) Trend analysis: Based on the similarity matrix, analyze the trend of waveform similarity over time and identify possible abnormal points or change patterns.

[0176] Example:

[0177] Assume that during a 10-minute test, waveform data is collected once a minute, resulting in a total of 10 waveforms. The calculated similarity matrix may be as follows:

[0178]

[0179] From this matrix, we can observe that the waveforms at adjacent moments have high similarity, but the similarity gradually decreases as the time interval increases. In particular, we can notice that starting from the 6th minute, the waveform similarity pattern changes significantly, which may indicate that some state transition has occurred in the circuit under test at this time.

[0180] This method not only quantifies the similarity between waveforms at any two moments, but also comprehensively analyzes the dynamic characteristics of the circuit under test throughout the entire test and measurement process. This provides powerful data support for circuit performance evaluation, fault diagnosis, and anomaly detection.

[0181] In summary, the dynamic programming-based timing waveform similarity calculation method proposed in this application achieves precise quantification of timing waveform similarity through steps such as distance matrix initialization, iterative calculation, and optimal path backtracking. Combined with multi-time waveform analysis, this method can effectively capture the state changes of the object under test during the test process, laying the foundation for further circuit analysis and diagnosis.

[0182] In an optional embodiment,

[0183] Through the generalized reasoning capability of the large language model, based on the input measurement raw data and measurement parameters, more dimensional parameter information is obtained, and mathematical operations are performed on the measurement raw data of multiple channels to synthesize new waveforms, including:

[0184] Obtaining measurement raw data collected by multiple measurement channels of the measured object and measurement parameters calculated based on the measurement raw data, wherein the measurement raw data is waveform data in the time domain or frequency domain, and the measurement parameters include voltage parameters and time parameters;

[0185] Inputting the raw measurement data and corresponding measurement parameters into a pre-trained large language model, wherein the large language model is trained based on massive test measurement data and corresponding analysis results, and has the ability to understand, generalize, and reason about the test measurement data;

[0186] Utilizing the generalized reasoning capability of the large language model, high-dimensional feature information contained in the raw measurement data is extracted, and combined with the measurement parameters, deep feature parameters of multiple measurement channels of the measured object under different working conditions are obtained;

[0187] According to the preset mathematical operation rules, the raw data of multiple measurement channels are fused and processed to generate a new synthetic waveform corresponding to the actual working state of the measured object;

[0188] The novel synthetic waveform is configured with different mathematical operation rules according to actual application requirements, including linear combination, nonlinear transformation, and time-frequency domain conversion of waveform data from different measurement channels.

[0189] In an optional embodiment,

[0190] Determining the data content to be input into the large language model according to the type of question selected by the user, when the question type is a waveform quality-related question, inputting a drawing image and measurement parameters into the large language model, and when the question type is a data calculation and analysis-related question, inputting raw measurement data and measurement parameters into the large language model includes:

[0191] Obtaining the question type selected by the user, wherein the question type includes waveform quality related questions and data calculation and analysis related questions;

[0192] Adaptively determining the test measurement data type and content input to the large language model based on the question type selected by the user;

[0193] When the problem type is a waveform quality-related problem, the waveform drawing image of the measured object and related measurement parameters are input into the large language model, where the measurement parameters include numerical values ​​representing waveform characteristics calculated based on the measurement data;

[0194] When the problem type is a data calculation and analysis related problem, the original data obtained by measurement and collection and the relevant measurement parameters are input into the large language model, and the original data is a time series numerical vector.

[0195] For example, this application details a method for adaptive data input based on question type. This method intelligently determines the data content for input to a large language model based on the question type selected by the user. This method can effectively improve the efficiency and accuracy of the large language model when handling different types of test and measurement questions. The following will gradually introduce the specific implementation process and technical details of this method.

[0196] First, a question type identification mechanism needs to be established to determine which category the question the user is concerned about belongs to.

[0197] The specific steps are as follows:

[0198] 1) Question type definition: Two major types of questions are defined in advance:

[0199] - Waveform quality related issues: mainly focus on visually discernible attributes such as signal shape and characteristics.

[0200] -Problems related to data calculation and analysis: mainly focus on the numerical characteristics, statistical properties and other indicators that need to be calculated.

[0201] 2) User interface design: Design an intuitive user interface that allows users to clearly select the question type.

[0202] 3) Keyword analysis: Perform keyword analysis on the question description entered by the user to assist in determining the question type.

[0203] 4) Historical data learning: Based on the user’s historical query records, a question type prediction model is established.

[0204] 5) Type confirmation: The question type is finally determined based on user selection, keyword analysis, and historical data prediction results.

[0205] Example:

[0206] Assume that the user selects "Waveform quality related issues" through the interface and enters the following problem description:

[0207] "Is there any overshoot on the rising edge of this signal? If so, approximately how much?"

[0208] The system performs the following steps:

[0209] 1. Record the user's direct selection: waveform quality related issues;

[0210] 2. Analyze the key words in the problem description: rising edge, overshoot;

[0211] 3. Query history: 70% of the user's previous queries were related to waveform quality issues;

[0212] 4. Comprehensive judgment: Confirm that this is a waveform quality-related issue;

[0213] Based on the determined question type, the system needs to adaptively select the appropriate data type to input into the large language model.

[0214] The selection steps are as follows:

[0215] 1) Data type mapping: Establish a mapping relationship between problem type and data type.

[0216] - Waveform quality related issues → Waveform drawing image + measurement parameters

[0217] -Data calculation and analysis related issues → original measurement data + measurement parameters

[0218] 2) Data availability check: Check whether the required data type is available. If not, consider alternatives or prompt the user.

[0219] 3) Data format conversion: Convert the selected data into an input format acceptable to the large language model.

[0220] 4) Data combination: Combine multiple types of data as needed, such as image data and text data.

[0221] 5) Input optimization: Perform necessary optimization on input data, such as compression and standardization, to improve model processing efficiency.

[0222] After the large language model processes the input data, the output results need to be processed and applied.

[0223] The processing steps are as follows:

[0224] 1) Result analysis: Analyze the output of the model and extract key information.

[0225] 2) Result verification: Compare and verify the analysis results of the model with those of traditional methods.

[0226] 3) Visualization processing: Convert the analysis results into intuitive charts or reports.

[0227] 4) Result storage: The analysis results are stored in the database for subsequent trend analysis and model optimization.

[0228] 5) Feedback mechanism: allows users to evaluate and provide feedback on analysis results to continuously improve system performance.

[0229] 1. Overshoot: The rising edge of the signal does have overshoot. The measured parameters show an overshoot amplitude of 4%, which is considered acceptable in most applications (usually less than 5% is considered acceptable).

[0230] 2. Cause of overshoot: The overshoot phenomenon is likely caused by the resonance formed by parasitic inductance and load capacitance in the signal path. The high slew rate (0.5V / ns) may exacerbate this phenomenon.

[0231] 3. Impact Assessment: Although 4% overshoot is acceptable in many cases, it may cause problems in some high-precision applications, such as the front-end signal processing of analog-to-digital converters.

[0232] 4. Suggestions for improvement:

[0233] -Optimize PCB layout to reduce parasitic inductance of signal paths.

[0234] -Consider adding appropriate damping networks in the signal path.

[0235] - If possible, try reducing the drive strength slightly to reduce the slew rate.

[0236] -Use a driver with pre-emphasis to provide extra drive strength at the beginning of the rising edge and then reduce the drive strength as it approaches the target voltage.

[0237] 5. Further analysis: It is recommended to perform frequency domain analysis to observe the high-frequency components caused by overshoot to more comprehensively evaluate its impact.

[0238] Overall, although overshoot is present, it is relatively mild and will not cause serious problems in most applications, but there is still room for optimization.

[0239] The analysis results include direct answers to questions, in-depth technical insights, and improvement suggestions. The system can further process the results, such as generating a PDF report containing the original waveform, a table of key parameters, and analysis conclusions.

[0240] In summary, the adaptive data input method based on question type proposed in this application significantly improves the efficiency and accuracy of large language models in handling test and measurement problems by intelligently identifying the problem type and selecting and processing the input data accordingly. This method can not only handle traditional numerical analysis problems, but also effectively analyze problems that require visual insight, such as waveform quality. By optimizing the quality and structure of the input data, this method fully realizes the potential of large language models and provides a powerful intelligent analysis tool for the test and measurement field. This method is expected to play an important role in many fields such as electronic testing, signal processing, and fault diagnosis, and promote the development of test and measurement technology in a smarter and more efficient direction.

[0241] In an optional embodiment,

[0242] The method further comprises:

[0243] The large language model adaptively calls the corresponding data processing flow and algorithm model according to the input test measurement data type and content;

[0244] When the input data type is a waveform drawing image, the image recognition algorithm is called to extract features and understand semantics of the input image, and combined with the measurement parameters, the analysis results related to the waveform quality are output;

[0245] When the input data type is raw measurement data, numerical calculation and statistical analysis algorithms are called to process the raw data, and combined with the measurement parameters, the calculation and analysis results related to the performance indicators of the measured object are output;

[0246] Among them, the large language model adopts a learning paradigm that combines pre-training and fine-tuning. Through pre-training on massive test measurement-related corpus, it has basic knowledge and common sense in the field of test measurement. For different types of test measurement problems, it uses customized data for fine-tuning to enable it to quickly adapt to specific problems.

[0247] Figure 2 FIG. 1 is a structural diagram of a large language model-assisted test and measurement system according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0248] A first unit is configured to collect raw measurement data, a drawing image, and calculated parameters of a circuit under test using a measurement and testing device, wherein the raw measurement data is multi-dimensional numerical data, the drawing image is generated by drawing based on the raw measurement data, and the calculated parameters include voltage parameters and time parameters calculated based on the raw measurement data, and input at least two of the raw measurement data, the drawing image, and the calculated parameters into a large language model;

[0249] The second unit is used to determine the data content to be input into the large language model according to the question type selected by the user. When the question type is a waveform quality-related question, the drawing image and measurement parameters are input into the large language model. When the question type is a data calculation and analysis-related question, the measurement raw data and measurement parameters are input into the large language model.

[0250] The third unit is used to obtain waveform change information of the tested circuit during the entire test and measurement process through a dynamic time warping algorithm based on the waveform similarity of the same channel at different times recorded throughout the entire cycle; understand and analyze the input drawing image through the multimodal image recognition capability of the large language model to obtain waveform quality analysis results and circuit noise source analysis results; and / or, through the generalized reasoning capability of the large language model, obtain parameter information of more dimensions based on the input measurement raw data and measurement parameters, perform mathematical operations on the measurement raw data of multiple channels to synthesize new waveforms; and generate a test measurement analysis report, which includes waveform quality analysis, circuit noise source analysis, existing problems and solutions, parameter information, and data operation synthesis results.

[0251] According to a third aspect of the embodiments of the present invention,

[0252] An electronic device is provided, comprising:

[0253] processor;

[0254] a memory for storing processor-executable instructions;

[0255] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0256] According to a fourth aspect of the embodiments of the present invention,

[0257] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0258] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0259] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A large language model-assisted test measurement method, characterized in that: include: Collecting raw measurement data, a drawing image, and calculated parameters of the circuit under test using a measurement test device, wherein the raw measurement data is multi-dimensional numerical data, the drawing image is generated by drawing based on the raw measurement data, and the calculated parameters include voltage parameters and time parameters calculated based on the raw measurement data, and inputting at least two of the raw measurement data, the drawing image, and the calculated parameters into a large language model; Determine the data content to be input into the large language model according to the type of question selected by the user; when the question type is a waveform quality-related question, input the drawing image and the measurement parameters into the large language model; when the question type is a data calculation and analysis-related question, input the measurement raw data and the measurement parameters into the large language model; Using a dynamic time warping algorithm, based on the waveform similarity of the same channel at different times during full-cycle recording, information about waveform changes of the circuit under test during the entire test and measurement process is obtained; using the multimodal image recognition capability of the large language model, the input drawing image is understood and analyzed to obtain waveform quality analysis results and circuit noise source analysis results; and / or, using the generalized reasoning capability of the large language model, based on the input raw measurement data and measured parameters, multi-dimensional parameter information is obtained, and mathematical operations are performed on the raw measurement data of multiple channels to synthesize new waveforms; Generate a test measurement analysis report, which includes waveform quality analysis, circuit noise source analysis, existing problems and solutions, parameter information, and data calculation synthesis results.

2. The method according to claim 1, characterized in that The dynamic time warping algorithm is used to obtain waveform change information of the circuit under test during the entire test and measurement process based on the waveform similarity of the same channel at different times during the full cycle recording. This includes: Acquire multiple timing waveform data collected at different times by the same measurement channel of the circuit under test; Based on a dynamic time warping algorithm, finding an optimal matching path between each two time series waveform data in the plurality of time series waveform data through dynamic programming, wherein the optimal matching path maximizes the similarity between the two time series waveform data; Calculating the similarity between each two time series waveform data according to the optimal matching path to obtain similarity measurement values ​​of the time series waveform data of the same measurement channel at different times; Based on the similarity metric value, obtaining a change trend of the waveform sequence of the same measurement channel during the entire test measurement process; Generate test measurement full cycle record data, the record data contains the characteristic information and similarity change information of the same measurement channel waveform at different times, so as to record the state change of the test measurement object during the entire test process.

3. The method according to claim 2, characterized in that Calculating the similarity between each pair of time series waveform data according to the optimal matching path to obtain similarity measurement values ​​of the time series waveform data of the same measurement channel at different times includes: Initialize the distance matrix and create a matrix of size (N+1)x(M+1), where N and M are the lengths of the two timing waveforms respectively. Set the initial value of the matrix to infinity and the origin of the matrix to 0. Iteratively calculate the distance between the corresponding elements of two time series waveforms and update the distance matrix; The distance matrix is ​​traced back to find a path with the shortest cumulative distance, which is used as the optimal matching path of the two time series waveforms. The shortest distance value is the similarity measurement value of the two time series waveform data.

4. The method according to claim 1, wherein Through the generalized reasoning capability of the large language model, based on the input measurement raw data and measurement parameters, more dimensional parameter information is obtained, and mathematical operations are performed on the measurement raw data of multiple channels to synthesize new waveforms, including: Obtaining measurement raw data collected by multiple measurement channels of the measured object and measurement parameters calculated based on the measurement raw data, wherein the measurement raw data is waveform data in the time domain or frequency domain, and the measurement parameters include voltage parameters and time parameters; Inputting the raw measurement data and corresponding measurement parameters into a pre-trained large language model, wherein the large language model is trained based on massive test measurement data and corresponding analysis results, and has the ability to understand, generalize, and reason about the test measurement data; Utilizing the generalized reasoning capability of the large language model, high-dimensional feature information contained in the raw measurement data is extracted, and combined with the measurement parameters, deep feature parameters of multiple measurement channels of the measured object under different working conditions are obtained; According to the preset mathematical operation rules, the raw data of multiple measurement channels are fused and processed to generate a new synthetic waveform corresponding to the actual working state of the measured object; The novel synthetic waveform is configured with different mathematical operation rules according to actual application requirements, including linear combination, nonlinear transformation, and time-frequency domain conversion of waveform data from different measurement channels.

5. The method according to claim 1, wherein Determining the data content to be input into the large language model according to the type of question selected by the user, when the question type is a waveform quality-related question, inputting a drawing image and measurement parameters into the large language model, and when the question type is a data calculation and analysis-related question, inputting raw measurement data and measurement parameters into the large language model includes: Obtaining the question type selected by the user, wherein the question type includes waveform quality related questions and data calculation and analysis related questions; Adaptively determining the test measurement data type and content input to the large language model based on the question type selected by the user; When the problem type is a waveform quality-related problem, the waveform drawing image of the measured object and related measurement parameters are input into the large language model, where the measurement parameters include numerical values ​​representing waveform characteristics calculated based on the measurement data; When the problem type is a data calculation and analysis related problem, the original data obtained by measurement and collection and the relevant measurement parameters are input into the large language model, and the original data is a time series numerical vector.

6. The method according to claim 5, characterized in that The method further comprises: The large language model adaptively calls the corresponding data processing flow and algorithm model according to the input test measurement data type and content; When the input data type is a waveform drawing image, the image recognition algorithm is called to extract features and understand semantics of the input image, and combined with the measurement parameters, the analysis results related to the waveform quality are output; When the input data type is raw measurement data, numerical calculation and statistical analysis algorithms are called to process the raw data, and combined with the measurement parameters, the calculation and analysis results related to the performance indicators of the measured object are output; Among them, the large language model adopts a learning paradigm that combines pre-training and fine-tuning. Through pre-training on massive test measurement-related corpus, it has basic knowledge and common sense in the field of test measurement. For different types of test measurement problems, it uses customized data for fine-tuning to enable it to quickly adapt to specific problems.

7. A large language model-assisted test and measurement system, configured to implement the method according to any one of claims 1 to 6, characterized in that: include: A first unit is configured to collect raw measurement data, a drawing image, and calculated parameters of a circuit under test using a measurement and testing device, wherein the raw measurement data is multi-dimensional numerical data, the drawing image is generated by drawing based on the raw measurement data, and the calculated parameters include voltage parameters and time parameters calculated based on the raw measurement data, and input at least two of the raw measurement data, the drawing image, and the calculated parameters into a large language model; The second unit is used to determine the data content to be input into the large language model according to the question type selected by the user. When the question type is a waveform quality-related question, the drawing image and measurement parameters are input into the large language model. When the question type is a data calculation and analysis-related question, the measurement raw data and measurement parameters are input into the large language model. The third unit is configured to obtain waveform change information of the circuit under test throughout the entire test and measurement process based on the waveform similarity of the same channel at different times recorded over a full cycle using a dynamic time warping algorithm; understand and analyze the input drawing image using the multimodal image recognition capability of the large language model to obtain waveform quality analysis results and circuit noise source analysis results; and / or, using the generalized reasoning capability of the large language model, derive parameter information in multiple dimensions based on the input raw measurement data and measured parameters, and perform mathematical operations on the raw measurement data of multiple channels to synthesize new waveforms; Generate a test measurement analysis report, which includes waveform quality analysis, circuit noise source analysis, existing problems and solutions, parameter information, and data calculation synthesis results.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images