Calibration data analysis method and system for mechanical sensors

Abstract

The present invention discloses a calibration data analysis method and system for mechanical sensors, belonging to the field of data analysis technology. By determining the association weight between the target mechanical sensor calibration device and each calibration deviation correction record to be processed, the present invention can accurately measure the degree of connection between the two, so that subsequent processing is more targeted; the attention semantics of the expected characteristics of sensor performance are optimized to obtain performance optimization output characteristics, which helps to focus on the more critical factors affecting sensor performance; the screening priority indication is obtained through feature integration, which comprehensively considers the association weight and performance optimization output characteristics, and comprehensively and reasonably reflects the importance of each calibration deviation correction record to be processed. According to the screening priority indication, the priority is sorted to obtain a calibration deviation correction record queue, and finally the target calibration deviation correction record is determined to assist in generating the current calibration correction plan, which can improve the efficiency and accuracy of calibration correction.

Description

Technical Field

[0001] This invention belongs to the field of data analysis technology, specifically relating to a calibration data analysis method and system for mechanical sensors. Background Technology

[0002] In the use of mechanical sensors, calibration is crucial to ensuring sensor accuracy. Traditional calibration methods often face numerous technical challenges. For example, when processing numerous calibration deviation correction records, there is a lack of effective means to accurately determine which records are most relevant to the target mechanical sensor calibration equipment. Because this correlation cannot be precisely measured, inappropriate calibration records may be used when generating calibration schemes, thus affecting the accuracy and efficiency of the calibration.

[0003] Furthermore, traditional methods are not precise enough in evaluating sensor performance. They fail to fully consider the complex quantitative relationships of sensor performance parameters and cannot focus on the most critical factors affecting performance, resulting in deviations between performance evaluation results and actual conditions. This deviation further affects the formulation of calibration and correction schemes, causing the measurement accuracy and reliability of mechanical sensors to fall short of ideal levels, making it difficult to meet the growing demand for high-precision mechanical sensors in industrial and scientific research fields. Summary of the Invention

[0004] This invention provides a calibration data analysis method and system for mechanical sensors, which can solve or partially solve the technical problems involved in the background art mentioned above.

[0005] This invention provides a calibration data analysis method for mechanical sensors, applied to a calibration data analysis system. The method includes: acquiring a calibration data knowledge vector corresponding to a target mechanical sensor calibration device; determining the association weight between the target mechanical sensor calibration device and each calibration deviation correction record based on the calibration data knowledge vector and the deviation correction state vector corresponding to each calibration deviation correction record to be processed; optimizing the attention semantics of the sensor performance expectation features corresponding to each calibration deviation correction record to be processed, obtaining performance optimization output features of each sensor performance expectation feature; the sensor performance expectation features reflect the quantization relationship vector corresponding to the sensor performance parameters; integrating the association weight between each calibration deviation correction record to be processed and the target mechanical sensor calibration device with the corresponding performance optimization output features of the sensor performance expectation features, obtaining a screening priority indicator corresponding to each calibration deviation correction record to be processed; prioritizing each calibration deviation correction record to be processed according to the screening priority indicator, obtaining a calibration deviation correction record queue; the calibration deviation correction record queue is used to determine target calibration deviation correction records, and the target calibration deviation correction records are used to assist in generating the current calibration correction scheme for the target mechanical sensor calibration device.

[0006] The present invention provides a calibration data analysis system, including at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, causing the at least one processor to perform the above-described method.

[0007] The present invention provides a readable storage medium on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0008] In this invention, by determining the correlation weight between the target mechanical sensing calibration device and each calibration deviation correction record to be processed, the closeness of the connection between the two can be accurately measured, making subsequent processing more targeted. This process, based on calibration data knowledge vectors and deviation correction state vectors, makes full use of the rich information in the device calibration data and correction records, avoiding blind processing of correction records.

[0009] Optimizing the attention semantics of the desired sensor performance features yields performance-optimized output features, which helps focus on factors that are more critical to sensor performance. This optimization approach takes into account the quantified relationship vectors of sensor performance parameters, making the performance evaluation more consistent with actual working conditions and improving the accuracy of sensor performance assessment.

[0010] Feature integration yields a screening priority indicator, comprehensively considering correlation weights and performance optimization output features to fully and reasonably reflect the importance of each calibration deviation correction record to be processed. Based on the screening priority indicator, a priority sorting process is performed to obtain a calibration deviation correction record queue. Finally, the target calibration deviation correction record is determined to assist in generating the current calibration correction scheme. This entire process ensures that the generation of the correction scheme is based on the most relevant and effective correction records, thereby improving the efficiency and accuracy of calibration correction. It can significantly enhance the performance of mechanical sensor calibration equipment, giving it higher accuracy and reliability in measurements, and can be widely applied in various industrial and scientific research fields with high accuracy requirements for mechanical sensors. Attached Figure Description

[0011] Figure 1 This is a flowchart of a calibration data analysis method for mechanical sensors provided by the present invention.

[0012] Figure 2 This is a schematic diagram of the structure of a calibration data analysis system provided by the present invention. Detailed Implementation

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

[0014] In this invention, the terms "first," "second," etc., are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, in this invention, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects have an "or" relationship.

[0015] Figure 1 A calibration data analysis method for mechanical sensors is shown, which is applied to a calibration data analysis system. The method includes steps S110-S140.

[0016] S110. Obtain the calibration data knowledge vector corresponding to the target mechanical sensing calibration device, and determine the association weight between the target mechanical sensing calibration device and each calibration deviation correction record to be processed based on the calibration data knowledge vector and the deviation correction state vector corresponding to each calibration deviation correction record to be processed.

[0017] In S110, the core objective of the calibration data analysis system is to determine the correlation weight between the target mechanical sensor calibration equipment and each calibration deviation correction record to be processed. Based on this, the calibration data knowledge vector is first obtained. The calibration data knowledge vector is a comprehensive representation of the calibration data of the target mechanical sensor calibration equipment. It contains various information about the calibration equipment, such as the input physical quantities during the calibration process (e.g., applied force, pressure, torque, etc.), output signals (e.g., voltage, current, etc.), calibration environmental conditions (temperature, humidity, etc.), and some inherent attributes of the calibration equipment itself (e.g., sensor type, range, etc.). This information is organized in the form of a vector, with each element representing a specific attribute or data point. For example, the calibration data knowledge vector of a strain gauge force sensor may contain the following: [applied force value 1, applied force value 2, ..., output voltage 1, output voltage 2, ..., ambient temperature, ambient humidity, sensor range, sensor sensitivity nominal value].

[0018] The specific values ​​of the calibration data knowledge vector originate from the measurements and records taken during the actual calibration process. For example, during a force sensor calibration experiment, the corresponding output voltage values ​​are measured at 10 different force values ​​(from 1N to 10N), while simultaneously recording the ambient temperature as 25°C, the ambient humidity as 50%, the sensor range as 0-100N, and the nominal sensitivity as 2mV / N. This calibration data knowledge vector can then be represented as [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 0.002, 0.004, 0.006, 0.008, 0.01, 0.012, 0.014, 0.016, 0.018, 0.02, 25, 50, 100, 2] (here, for simplicity, the voltage values ​​are represented in a scaled-down manner).

[0019] Building upon the above, the calibration data analysis system further determines the deviation correction state vector and associated weights. Each calibration deviation correction record to be processed corresponds to a deviation correction state vector. This vector describes some state information from the previous calibration deviation correction process, such as the type of correction (zero-point correction, gain correction, or nonlinear correction), the degree of correction (e.g., the value of the correction coefficient), and the sensor performance indicators after correction (e.g., linearity, sensitivity).

[0020] To determine the association weight between the target mechanical sensing calibration device and each calibration deviation correction record to be processed, a distance-based algorithm, such as the Euclidean distance algorithm, can be used. Let the calibration data knowledge vector be (X=[x_1, x_2, cdots, x_n]), and the deviation correction state vector corresponding to the calibration deviation correction record to be processed be (Y=[y_1, y_2, cdots, y_m]) (where (n) and (m) are the dimensions of vectors (X) and (Y), respectively, and (n) and (m) may be different, requiring appropriate processing, such as padding or truncation). The association weight (W) can be calculated using the following formula:

[0021] [W=frac{1}{1+sqrt{sum_{i=1}^{min(n,m)}(x_i-y_i)^2}}]

[0022] For example, consider a calibration deviation correction record to be processed with a deviation correction state vector of [zero-point correction, correction coefficient 0.9, linearity after correction 0.99, sensitivity after correction 1.01] (elements are simplified here). The association weight is calculated between this vector and the calibration data knowledge vector of the aforementioned force sensor. After numerical calculation, an association weight value between 0 and 1 is obtained. This value represents the degree of association between the target force sensor calibration device and the calibration deviation correction record to be processed; the closer the value is to 1, the higher the degree of association.

[0023] In detail, the calibration data knowledge vector is a comprehensive representation vector of the calibration data of the target mechanical sensing calibration equipment. It covers many aspects of information in the calibration process, including but not limited to input physical quantities, output signals, calibration environmental conditions, and the attributes of the calibration equipment itself.

[0024] For example, the knowledge vector of calibration data for a force sensor might include elements such as the sequence of force values ​​applied at different calibration times (e.g., [10N, 20N, 30N, ...]), the corresponding output electrical signals (e.g., [0.1V, 0.2V, 0.3V, ...]), the temperature (e.g., 25°C), humidity (e.g., 50%) of the calibration environment, the sensor's range (e.g., 0-100N), and the nominal sensitivity value (e.g., 2mV / N). These elements are combined to form a vector that comprehensively describes the knowledge related to the calibration data.

[0025] The calibration data knowledge vector is the foundation for subsequent calculations of association weights. It provides basic information about the target mechanical sensing calibration device, and by comparing it with the deviation correction state vector, the degree of association between the two can be determined. In the algorithm, it acts as a reference template, used to measure the closeness of the relationship between other relevant vectors and the target device. For example, when calculating the association weight between the target device and a certain calibration deviation correction record to be processed, elements in the calibration data knowledge vector are operated on with corresponding elements in the deviation correction state vector to obtain a quantified value of the degree of association.

[0026] In addition, the deviation correction state vector describes the state information during the previous calibration deviation correction process. Its elements include the type of correction (such as zero-point correction, gain correction, etc., which can be represented by specific codes, such as 0 for zero-point correction, 1 for gain correction, etc.), the degree of correction (such as the offset during zero-point correction, the correction coefficient for gain correction, etc.), and the sensor performance indicators after correction (such as the linearity value, sensitivity value, etc. after correction).

[0027] For example, a bias correction state vector might be [0, 0.1, 0.99, 1.01], where 0 indicates that zero-point correction has been performed, 0.1 is the offset of the zero-point correction (e.g., a numerical value), 0.99 is the corrected linearity, and 1.01 is the corrected sensitivity.

[0028] The deviation correction state vector and the calibration data knowledge vector are used together to determine the association weights. It reflects key information from previous correction operations, and through association analysis with the calibration data knowledge vector, the relevance of past correction records to the current target device can be determined. Furthermore, in subsequent priority indicator calculations, the correction type and degree information it contains will indirectly affect the final priority ranking.

[0029] S120. Optimize the attention semantics of the sensor performance expectation features corresponding to each calibration deviation correction record to obtain the performance optimization output features of each sensor performance expectation feature; the sensor performance expectation features reflect the quantization relationship vector corresponding to the sensor performance parameters.

[0030] In this invention, S120 focuses on optimizing the attention semantics of the sensor performance expectation features to obtain performance-optimized output features.

[0031] Among them, the expected characteristics of sensor performance reflect the quantization relationship vector corresponding to the sensor performance parameters. For example, for a mechanical sensor, its performance parameters may include linearity, sensitivity, repeatability, hysteresis, etc. Taking linearity as an example, its quantization relationship vector can be represented as a vector composed of the deviation values ​​between the actual output and the ideal linear output under different input values. For example, under 5 different input force values ​​[1N, 2N, 3N, 4N, 5N], the deviations between the actual output and the ideal linear output are [0.01V, 0.02V, 0.015V, 0.025V, 0.03V], then the quantization relationship vector of this linearity is [0.01, 0.02, 0.015, 0.025, 0.03].

[0032] For sensitivity, the quantization vector may be a vector composed of the output voltage change rate under different input force values. For example, when the input force increases from 1N to 2N, the output voltage changes from 0.02V to 0.04V, and the sensitivity is ((0.04-0.02) / (2-1)=0.02V / N). The quantization vector of sensitivity can be obtained under multiple input force value changes.

[0033] When attention semantic optimization is involved, attention mechanism algorithms can be used to optimize the attention semantics of the desired sensor performance features. Attention mechanisms can help focus on factors that have a significant impact on sensor performance. For example, when calculating the performance-optimized output features, for the desired performance feature of linearity, analysis of previous calibration data revealed that certain input force ranges have a more critical impact on linearity. For instance, within the input force range of 2N-4N, the weight of linearity deviation on overall performance is 0.6 (this weight was obtained through analysis of a large amount of calibration data and the sensor's operating principle), while the weight of other force ranges is 0.4.

[0034] Therefore, when optimizing linearity using attention semantics, more attention is paid to deviation values ​​within the range of 2N-4N. For example, the original linearity quantification vector is [0.01, 0.02, 0.015, 0.025, 0.03]. After attention semantic optimization, the new vector might become [0.01×0.4, 0.02×0.6, 0.015×0.6, 0.025×0.6, 0.03×0.4] = [0.004, 0.012, 0.009, 0.015, 0.012]. This is the optimized performance output feature. The same method can be applied to other desired sensor performance features, such as sensitivity and repeatability.

[0035] In other examples, the expected characteristics of sensor performance reflect the quantization vectors corresponding to the sensor performance parameters. Sensor performance parameters typically include linearity, sensitivity, repeatability, and hysteresis. For each performance parameter, there is a corresponding quantization vector.

[0036] Taking linearity as an example, the quantization vector of linearity might be a vector composed of the deviations between the actual output and the ideal linear output under different input values. For instance, when the input force values ​​are [1N, 2N, 3N, 4N, 5N], the deviations between the actual output and the ideal linear output are [0.01V, 0.02V, 0.015V, 0.025V, 0.03V], then the quantization vector of linearity would be [0.01, 0.02, 0.015, 0.025, 0.03]. For sensitivity, the quantization vector might be a vector composed of the rate of change of output voltage under different input force values.

[0037] Sensor performance expectation features are a key basis for evaluating sensor performance. These feature vectors directly reflect the sensor's performance in different aspects and serve as the starting point for optimizing sensor performance. In the subsequent attention semantic optimization process, operations will be performed on these feature vectors to improve the evaluation and optimization effects of sensor performance.

[0038] Furthermore, the performance-optimized output feature is the result of attention-based semantic optimization of the sensor performance expectation feature. It is obtained by adjusting the elements in the sensor performance expectation feature vector according to their weights in relation to the overall performance.

[0039] For example, for the linearity performance expectation feature vector [0.01, 0.02, 0.015, 0.025, 0.03], if analysis shows that the deviation of the input force value in the range of 2N-4N has a weight of 0.6 on the overall linearity, and 0.4 in other ranges, then the optimized performance output feature may be [0.01 × 0.4, 0.02 × 0.6, 0.015 × 0.6, 0.025 × 0.6, 0.03 × 0.4] = [0.004, 0.012, 0.009, 0.015, 0.012].

[0040] The performance optimization output features more accurately reflect the actual impact of sensor performance parameters on overall performance. It is obtained after considering the differences in importance of different performance parameters across different input ranges, providing a more reasonable data foundation for subsequent feature integration and priority indicator calculations.

[0041] Next, the quantization relation vector is a representation of the expected characteristics of sensor performance. It quantifies the relationship between sensor performance parameters and factors such as input and output through specific numerical values.

[0042] For example, in terms of repeatability performance, the quantization relationship vector might be a vector composed of the dispersion of output values ​​obtained from multiple measurements of the same input value. For instance, if a 10N input force is measured five times, the resulting output voltages are [0.1V, 0.11V, 0.09V, 0.105V, 0.095V]. Then, the repeatability quantization relationship vector can be obtained by calculating statistical measures such as the standard deviation. This vector reflects the quantization relationship between the sensor's repeatability performance and the input force value.

[0043] Quantization relation vectors represent sensor performance parameters in a computable and comparable form. This allows for quantitative comparisons of the performance of different sensors or the same sensor under different conditions, providing concrete data support for sensor performance evaluation, optimization, and calculations with other relevant vectors (such as feature ensemble).

[0044] S130. The correlation weights between each calibration deviation correction record to be processed and the target mechanical sensing calibration device are respectively integrated with the performance optimization output features of the corresponding sensor performance expectation features to obtain the screening priority indicators corresponding to each calibration deviation correction record to be processed.

[0045] In S130, obtaining the screening priority indication corresponding to each calibration deviation correction record to be processed can be achieved through the following two sub-steps.

[0046] Sub-step 1 involves feature integration. In this step, the correlation weights between each calibration deviation correction record to be processed and the target mechanical sensing calibration device are integrated with the performance optimization output features of the corresponding sensor performance expectation features. For example, for a certain calibration deviation correction record to be processed, its correlation weight (W=0.8) (calculated through step S110) and its corresponding linearity performance optimization output features are [0.004, 0.012, 0.009, 0.015, 0.012] (obtained through step S120).

[0047] Feature integration is performed using a weighted summation method. For example, the weight of linearity in the overall performance evaluation is (α=0.3) (this weight is determined according to the specific application requirements and performance requirements of the sensor). Then, the integration result for the performance feature of linearity is (0.8×0.3×[0.004, 0.012, 0.009, 0.015, 0.012]=[0.00096, 0.00288, 0.00216, 0.0036, 0.00288]).

[0048] If there are other desired performance characteristics, such as sensitivity and repeatability, feature integration is performed using the same method. Then, the integration results of each performance characteristic are combined to obtain a comprehensive feature integration result. This result is part of the screening priority indication corresponding to the calibration deviation correction record to be processed.

[0049] Sub-step 2 involves determining the complete screening priority indicator: After performing the above feature integration operation on all the expected sensor performance features, the results are combined to form a complete vector. This vector is the screening priority indicator corresponding to the calibration deviation correction record to be processed. For example, after feature integration of the three expected performance features of linearity, sensitivity, and repeatability, the resulting screening priority indicator vector may be [0.00096, 0.00288, 0.00216, 0.0036, 0.00288, 0.001, 0.002, 0.0015, 0.0025, 0.002, 0.0005, 0.001, 0.0008, 0.0012, 0.001] (this is just an example vector; the actual length and value depend on the specific number of expected performance features and the integration result).

[0050] In detail, the screening priority indicator is a comprehensive vector, which is obtained by integrating the correlation weights between each calibration deviation correction record to be processed and the target mechanical sensing calibration device with the performance optimization output features of the corresponding sensor performance expectation features.

[0051] For example, for a calibration deviation correction record to be processed, its association weight is and the linearity performance optimization output feature is [0.004, 0.012, 0.009, 0.015, 0.012]. For example, the weight of linearity in the comprehensive performance evaluation is , then the integrated result for the performance feature of linearity is . The complete vector obtained by combining all performance features after performing similar operations is the screening priority indicator vector.

[0052] As can be seen, the screening priority indicator vector is used to determine the priority order of each calibration deviation correction record to be processed. It comprehensively considers multiple factors such as correlation weight and performance optimization output characteristics. By integrating these vectors, a vector that can comprehensively reflect the importance of the calibration deviation correction record to be processed to the target mechanical sensing calibration equipment is obtained, thus providing a basis for subsequent priority sorting and determination of the target calibration deviation correction record.

[0053] S140. Based on the filtering priority indication, the calibration deviation correction records to be processed are sorted by priority to obtain a calibration deviation correction record queue. The calibration deviation correction record queue is used to determine the target calibration deviation correction record. The target calibration deviation correction record is used to assist in generating the current calibration correction scheme of the target mechanical sensing calibration device.

[0054] In S140, the calibration data analysis system can obtain a calibration deviation correction record queue by prioritizing and determine the target calibration deviation correction record.

[0055] ① Prioritization

[0056] The priority of each calibration deviation correction record to be processed is determined based on the screening priority indicators obtained from S130. A sorting algorithm, such as bubble sort, can be used. For example, if there are 5 calibration deviation correction records to be processed, their screening priority indicator vectors are as follows:

[0057] Record 1: [0.00096, 0.00288, 0.00216, 0.0036, 0.00288, 0.001, 0.002, 0.0015, 0.0025, 0.002, 0.0005, 0.001, 0.0008, 0.0012, 0.001];

[0058] Record 2: [0.001, 0.003, 0.0025, 0.004, 0.0035, 0.0012, 0.0022, 0.0018, 0.003, 0.0022, 0.0008, 0.0012, 0.001, 0.0015, 0.0012];

[0059] Record 3: [0.0008, 0.0025, 0.002, 0.003, 0.0025, 0.0008, 0.0018, 0.0012, 0.002, 0.0018, 0.0004, 0.0008, 0.0006, 0.001, 0.0008];

[0060] Record 4: [0.0012, 0.0032, 0.0028, 0.0042, 0.0038, 0.0015, 0.0025, 0.002, 0.0032, 0.0025, 0.001, 0.0015, 0.0012, 0.0018, 0.0015];

[0061] Record 5: [0.0005, 0.002, 0.0015, 0.0025, 0.002, 0.0003, 0.0012, 0.0008, 0.0015, 0.0012, 0.0002, 0.0005, 0.0004, 0.0006, 0.0005].

[0062] The bubble sort algorithm is used to sort these vectors according to optional rules (such as the sum of the vector elements, or the size of a specific element). After sorting, a calibration deviation correction record queue is obtained. For example, if the queue order is obtained by sorting the vector elements in ascending order of sum of elements, the queue order is: record 5, record 3, record 1, record 2, record 4.

[0063] ② Determine the target calibration deviation correction record

[0064] The purpose of the calibration deviation correction record queue is to determine the target calibration deviation correction record. In the example above, record 5 at the very beginning is the target calibration deviation correction record. This target calibration deviation correction record will be used to assist in generating the current calibration correction scheme for the target force sensor calibration equipment. For example, record 5 may contain valid information about zero-point correction, such as the zero-point correction coefficient and environmental conditions during correction. This information can be extracted and combined with the current state of the target force sensor calibration equipment to formulate a suitable current calibration correction scheme, such as determining whether zero-point correction needs to be performed again and the magnitude of the correction, thereby improving the calibration accuracy and performance of the target force sensor calibration equipment.

[0065] In summary, this invention provides an innovative method for calibrating mechanical sensing devices, offering several beneficial effects. By determining the correlation weight between the target mechanical sensing calibration device and each calibration deviation correction record to be processed, the closeness of the connection between the two can be accurately measured, making subsequent processing more targeted. This process, based on calibration data knowledge vectors and deviation correction state vectors, fully utilizes the rich information in the device calibration data and correction records, avoiding blind processing of correction records.

[0066] Optimizing the attention semantics of the desired sensor performance features yields performance-optimized output features, which helps focus on factors that are more critical to sensor performance. This optimization approach takes into account the quantified relationship vectors of sensor performance parameters, making the performance evaluation more consistent with actual working conditions and improving the accuracy of sensor performance assessment.

[0067] Feature integration yields a screening priority indicator, comprehensively considering correlation weights and performance optimization output features to fully and reasonably reflect the importance of each calibration deviation correction record to be processed. Based on the screening priority indicator, a priority sorting process is performed to obtain a calibration deviation correction record queue. Finally, the target calibration deviation correction record is determined to assist in generating the current calibration correction scheme. This entire process ensures that the generation of the correction scheme is based on the most relevant and effective correction records, thereby improving the efficiency and accuracy of calibration correction. It can significantly enhance the performance of mechanical sensor calibration equipment, giving it higher accuracy and reliability in measurements, and can be widely applied in various industrial and scientific research fields with high accuracy requirements for mechanical sensors.

[0068] In some optional embodiments, obtaining the calibration data knowledge vector corresponding to the target mechanical sensing calibration device includes: obtaining the mechanical sensing calibration data corresponding to the target mechanical sensing calibration device; and performing knowledge vector mining on the mechanical sensing calibration data through the calibration data knowledge vector mining branch in the calibrated StackedLSTM network to obtain the calibration data knowledge vector corresponding to the target mechanical sensing calibration device.

[0069] In practical applications, the following is a detailed description of the technical solution of this embodiment.

[0070] I. Acquisition of mechanical sensing calibration data corresponding to the target mechanical sensing calibration equipment

[0071] ① Content and source of mechanical sensor calibration data

[0072] Mechanical sensor calibration data is a collection of various data related to the calibration process of a target mechanical sensor calibration device. This data covers multiple aspects, such as the precise values ​​of the mechanical inputs (force, pressure, torque, etc.) applied to the sensor during calibration, and the corresponding output signals of the sensor (such as voltage, current, or frequency). It also includes environmental parameters during the calibration process, such as temperature and humidity, as these environmental factors can affect the sensor's performance.

[0073] Taking the calibration of a force sensor as an example, during the calibration experiment, the sensor's output might be measured under a series of precisely controlled force values. For instance, the applied force could range from 1N to 100N, increasing in 1N increments, resulting in 100 different force values ​​for the sensor's output data. At each force value applied, the ambient temperature is recorded, for example, ranging from 20°C to 30°C, and the humidity fluctuates between 40% and 60%. These detailed measurement data collectively constitute the mechanical sensor calibration data.

[0074] ②Acquisition method and accuracy requirements

[0075] Acquiring this data requires specialized measuring instruments and equipment. For measuring mechanical inputs, high-precision force sensors (e.g., with an accuracy of ±0.01N), pressure sensors, or torque sensors are needed to ensure the accuracy of the input measurements. For measuring sensor output signals, high-precision data acquisition equipment is used, with voltage measurement accuracy potentially reaching ±0.001V. Environmental parameter measurements are performed using devices such as temperature and humidity sensors, with temperature measurement accuracy of ±0.1°C and humidity measurement accuracy of ±1%. These high-precision measuring instruments ensure that the acquired mechanical sensor calibration data has sufficient accuracy, providing a reliable data foundation for subsequent processing.

[0076] II. Knowledge vector mining using the calibration data knowledge vector mining branch in the calibrated StackedLSTM.

[0077] ①StackedLSTM (Stacked Long Short-Term Memory Network)

[0078] StackedLSTM is a deep learning network architecture consisting of multiple Long Short-Term Memory (LSTM) layers stacked together. LSTM is a special type of recurrent neural network (RNN) designed to address the vanishing and exploding gradient problems that traditional RNNs encounter when processing long sequences of data. By introducing gating mechanisms (input gate, forget gate, and output gate), LSTM can effectively handle long-term dependencies in long sequences of data.

[0079] In StackedLSTM, multiple LSTM layers are stacked together, with each layer processing the output of the layer above, thus enabling the extraction of more complex and advanced features. For example, the first LSTM layer might learn basic time-series patterns in calibration data, while the second LSTM layer can learn more abstract patterns based on the output of the first layer, such as the relationship between the changing trends of sensor output under different force values ​​and environmental factors.

[0080] ②Calibrating the knowledge vector mining branch

[0081] This branch is specifically designed for mechanical sensing calibration data and is located within a StackedLSTM network structure. Its structure and parameters have been tuned to adapt to the characteristics of the calibration data and the needs of data mining.

[0082] When training this branch, the input is mechanical sensing calibration data, such as a sequence of data consisting of force values, sensor outputs, ambient temperature, and humidity, as mentioned earlier. The output is a calibration data knowledge vector.

[0083] For example, the dimension of the calibrated data knowledge vector is (n), such as (n=20). The elements in this vector may include some statistical features, such as the mean and standard deviation of the sensor output within different force ranges, the average values ​​of ambient temperature and humidity, and the correlation coefficient between force and sensor output. Specifically, the first 5 elements may be the mean of the sensor output voltage in different force ranges (such as 0-20N, 20-40N, etc.), the next 3 elements may be the standard deviation of the corresponding range, the next 2 elements may be the average values ​​of ambient temperature and humidity, and the last 10 elements may be the values ​​of the correlation coefficient between different force values ​​and sensor output after optional transformation.

[0084] ③ Specific process and algorithm examples of knowledge vector mining

[0085] To illustrate with a simplified example, consider a mechanical sensor calibration data sequence of length (m) (e.g., (m=1000)). Each time step contains data points including force (F_t), sensor output voltage (V_t), temperature (T_t), and humidity (H_t) ((t=1, 2, cdots, m)).

[0086] In the knowledge vector mining branch of the calibration data in StackedLSTM, the data is first preprocessed to convert it into a format suitable for network input. For example, force values, voltage, temperature, and humidity are normalized to the interval ([-1, 1]).

[0087] For computation in the LSTM layer, taking a single LSTM unit as an example, the calculation formula for the forget gate (f_t) is: (f_t=sigma(W_fcdot[h_{t-1},x_t]+b_f)), where (W_f) is the weight matrix of the forget gate, (h_{t-1}) is the hidden state of the previous time step, (x_t) is the input of the current time step (containing normalized force, voltage, temperature and humidity information, etc.), (b_f) is the bias vector of the forget gate, and (sigma) is the sigmoid function.

[0088] The input gate (i_t) and output gate (o_t) also have similar calculation formulas, which are (i_t=sigma(W_icdot[h_{t-1},x_t]+b_i)) and (o_t=sigma(W_ocdot[h_{t-1},x_t]+b_o)).

[0089] The formula for updating the cell state (C_t) is: (C_t=f_t×C_{t-1}+i_t×tanh(W_Ccdot[h_{t-1},x_t]+b_C)).

[0090] The formula for calculating the hidden state (h_t) is: (h_t = o_t × tanh (C_t)).

[0091] After processing by multiple LSTM layers, the output of the last LSTM layer is converted into a labeled data knowledge vector through a fully connected layer. For example, if the weight matrix of the fully connected layer is (W_{out}) and the bias vector is (b_{out}), then the formula for calculating the labeled data knowledge vector (V) is: (V=W_{out}cdoth_m+b_{out}), where (h_m) is the hidden state at the last time step.

[0092] Supervised learning is performed using labeled, calibrated data samples when training this network. For example, mean squared error (MSE) is used as the loss function:

[0093] (MSE = frac{1}{n}sum_{i=1}^{n}(y_i-hat{y}_i)^2), where (y_i) is the true value in the calibration data knowledge vector (obtained manually from the original calibration data or using other accurate methods), and (hat{y}_i) is the value in the calibration data knowledge vector predicted by the network. By continuously adjusting the weights and biases of the network through the backpropagation algorithm, the loss function is minimized, thus obtaining the calibration data knowledge vector mining branch in the calibrated StackedLSTM, which can accurately mine the calibration data knowledge vector from the mechanical sensing calibration data.

[0094] This design, firstly, efficiently extracts calibration data knowledge vectors by acquiring comprehensive and accurate mechanical sensor calibration data and processing it using the calibration data knowledge vector mining branch of a calibrated StackedLSTM. This approach fully leverages the capabilities of StackedLSTM in processing long-sequence data and learning complex patterns, enabling deeper extraction of hidden information within the calibration data compared to traditional data analysis methods. For example, it accurately captures the subtle relationship between sensor output and environmental factors at different force values, a relationship that might be overlooked in traditional methods. Secondly, the extracted calibration data knowledge vectors provide a more accurate and richer foundation for subsequent operations such as determining the association weights between the target mechanical sensor calibration device and each calibration deviation correction record to be processed. This helps improve the accuracy and efficiency of the entire calibration deviation correction process, thereby enhancing the performance of the mechanical sensor calibration device and ensuring it provides more reliable measurement results in various application scenarios.

[0095] In some other embodiments, the method further includes: performing knowledge vector mining on each calibration deviation correction record in the deviation correction record set through the first deviation correction state vector mining branch or the second deviation correction state vector mining branch in the calibrated StackedLSTM network to obtain the initial deviation correction state vector corresponding to each calibration deviation correction record; and determining the deviation correction state vector corresponding to each calibration deviation correction record to be processed from each initial deviation correction state vector.

[0096] Furthermore, the specific implementation methods involved in this embodiment are as follows.

[0097] I. Deviation Correction Record Set and Initial Deviation Correction State Vector Mining

[0098] ① Deviation Correction Record Set

[0099] A calibration deviation correction record set is a collection of calibration deviation correction records performed on a mechanical sensing calibration device. Each calibration deviation correction record covers detailed information about the calibration process. This information includes the timestamp of the calibration, the specific sensor performance parameters targeted for calibration (such as linearity, sensitivity, repeatability, etc.), the calibration method used (such as zero-point calibration, gain calibration, etc.), the input parameters during the calibration process (such as the magnitude and frequency of the applied calibration signal, if applicable), and the evaluation of the calibration results (such as the actual measured values ​​of the sensor performance parameters after calibration).

[0100] For example, a deviation correction record might contain the following information: the correction time was 10:00 AM on May 10, 2023; the correction targeted the sensor's sensitivity parameters; the correction method used was gain correction; the applied correction gain coefficient was 1.05; and the actual measured sensitivity value obtained after correction, obtained by measuring a series of standard force values, was 2.1 mV / N (the original nominal value was 2 mV / N). This deviation correction record set provides a rich source of raw data for subsequent excavation operations.

[0101] ②The first bias correction state vector mining branch and the second bias correction state vector mining branch in StackedLSTM

[0102] The two mining branches in StackedLSTM are specifically designed for mining deviation correction state vectors from deviation correction record sets. They differ in structure and function, but both aim to extract useful information from the deviation correction records and convert it into deviation correction state vectors.

[0103] The first deviation correction state vector mining branch may focus more on mining the parameters and results of the correction operation itself. For example, it will focus on information such as the correction method, correction coefficients, and the direct numerical relationships between the corrected performance parameters.

[0104] The second deviation correction state vector mining branch may consider more complex information such as environmental factors of correction (if recorded in the deviation correction log) and the relationship between correction and other related operations (such as previous calibration operations).

[0105] ③ Knowledge Vector Mining Process

[0106] Taking the first deviation correction state vector mining branch as an example, for each calibration deviation correction record in the deviation correction record set, its relevant information is preprocessed for input into StackedLSTM. For example, the information of a calibration deviation correction record is organized into a vector form, such as ([t, p, m, i, r]), where (t) represents the correction time (in optional time encoding form, such as the number of seconds from a certain start time), (p) represents the performance parameter targeted by the correction (which can be encoded, such as 0 for linearity, 1 for sensitivity, etc.), (m) represents the correction method (which can also be encoded, such as 0 for zero-point correction, 1 for gain correction, etc.), (i) represents the correction input parameter (such as the gain coefficient in gain correction), and (r) represents the evaluation of the result after correction (such as the actual value of the corrected performance parameter).

[0107] These vectors are input into the StackedLSTM of the first bias correction state vector mining branch. Each LSTM unit of the StackedLSTM processes the data according to its internal computation mechanism. Taking the first LSTM unit as an example, for the input vector (x=[t, p, m, i, r]), its forget gate (f) is calculated as follows:

[0108] Let the weight matrix (W_f) of the forget gate be a (5×5) matrix (e.g., the input vector dimension is 5), and the bias vector (b_f) be a 5-dimensional vector.

[0109] First, calculate (z = W_fcdotx + b_f), for example (z = [z_1, z_2, z_3, z_4, z_5]), then use the forget gate (f = sigma(z)), where (sigma) is the sigmoid function, i.e. (f = [sigma(z_1), sigma(z_2), sigma(z_3), sigma(z_4), sigma(z_5)]).

[0110] The updates of the input gate (i), output gate (o), and cell state (C), as well as the calculation of the hidden state (h), are performed according to similar LSTM calculation rules. After processing through multiple LSTM layers, the output of the last LSTM layer is converted into an initial bias correction state vector through a fully connected layer. For example, if the weight matrix of the fully connected layer is (W_{out1}) and the bias vector is (b_{out1}), then the initial bias correction state vector (V_1) is calculated as: (V_1=W_{out1}cdoth_{last}+b_{out1}), where (h_{last}) is the hidden state at the last time step.

[0111] The processing for the second bias correction state vector mining branch is similar, but the preprocessing of the input information and the parameters such as the weight matrix and bias vector within the network may differ to adapt to its mining requirements for different types of information. For example, if the bias correction record includes ambient temperature (T) and humidity (H) information during correction, the input vector may become ([t, p, m, i, r, T, H]), and the corresponding network structure and parameters will be adjusted to correctly process this additional information and mine a suitable initial bias correction state vector.

[0112] II. Determine the deviation correction state vector corresponding to the calibration deviation correction record to be processed from the initial deviation correction state vector.

[0113] ① Determination of screening criteria

[0114] Determining the deviation correction state vector corresponding to each calibration deviation correction record to be processed from each initial deviation correction state vector requires establishing certain screening criteria. These screening criteria can be based on various factors, such as relevance to the target mechanical sensing calibration equipment, completeness of the calibration record, and reliability of the calibration results.

[0115] Taking the correlation with the target mechanical sensing calibration device as an example, an optional similarity metric can be calculated between the initial deviation correction state vector and the calibration data knowledge vector of the target device (obtained through the method mentioned above). For example, using cosine similarity as the metric, for an initial deviation correction state vector (V_{1i}) (where (i) represents the (i)th initial deviation correction state vector) and the calibration data knowledge vector (K) of the target device, the formula for calculating the cosine similarity (s) is:

[0116] (s=

[0117] frac{V_{1i}cdotK}{leftvertV_{1i}rightvert×leftvertKrightvert});

[0118] Where: (cdot) represents the vector dot product; (leftvertV_{1i}rightvert) and (leftvertKrightvert) represent the magnitudes of vectors (V_{1i}) and (K), respectively.

[0119] If the cosine similarity (s) is greater than a certain set threshold (s_{th}) (e.g. (s_{th}=0.8)), then the calibration deviation correction record corresponding to this initial deviation correction state vector is considered to have a high correlation and can be used as a calibration deviation correction record to be processed. Its initial deviation correction state vector becomes the deviation correction state vector after further processing (such as removing some redundant information or normalizing).

[0120] ② Example of processing procedure

[0121] For example, calculations show that the cosine similarity between several initial deviation correction state vectors and the calibration data knowledge vector of the target device is greater than (0.8). For one of the initial deviation correction state vectors (V_{1j}=[a_1, a_2, cdots, a_n]) ((j) represents the (j)th initial deviation correction state vector that meets the correlation requirement), if it is found that some elements are not very relevant to the main objective of the current deviation correction state vector (for example, the correction time element in the initial deviation correction state vector may not be very important when only the correlation between the correction result and the correction method is considered), these elements can be removed or given a lower weight.

[0122] For example, the weight of the first element in (V_{1j}) (e.g., a time-related element) is set to 0, and then the remaining elements are normalized. Let the vector composed of the remaining elements be (V_{1j}'=[a_2, cdots, a_n]), and the element (b_k) ((k=2, cdots, n)) of the normalized deviation correction state vector (V_{j}) is calculated using the formula: (b_k=frac{a_k}{sqrt{sum_{l=2}^{n}a_l^2}}). This yields the final deviation correction state vector (V_{j}), which is used for subsequent operations such as determining the association weight between the vector and the target mechanical sensing calibration device.

[0123] This approach offers several advantages. First, by utilizing the deviation correction state vector mining branch within StackedLSTM to perform knowledge vector mining on the deviation correction records, the powerful feature learning capabilities of StackedLSTM can be fully leveraged to extract valuable information from complex deviation correction records. For instance, in-depth mining of information such as correction methods, correction parameters, and correction results allows for a more comprehensive understanding of the essential characteristics of each correction operation. Second, by setting reasonable screening criteria to determine the final deviation correction state vector from the initial deviation correction state vector, the relevance to the target device can be improved, unnecessary information can be removed, and the resulting deviation correction state vector can be made more refined and effective. This helps to accurately determine the correlation weights between the target mechanical sensing calibration device and each calibration deviation correction record to be processed, thereby improving the accuracy and efficiency of the entire calibration process and enhancing the performance of the mechanical sensing calibration device.

[0124] In practical applications, the tuning method of the calibrated StackedLSTM network may include the following: obtaining a prior training example set, which includes: mechanical sensor calibration data examples, calibration deviation correction record authentication examples, and calibration deviation correction record perturbation examples corresponding to mechanical sensor calibration equipment examples; performing knowledge vector mining on the mechanical sensor calibration data examples through the calibration data knowledge vector mining branch in the StackedLSTM network to be tuned to obtain calibration data knowledge vector examples; performing knowledge vector mining on the calibration deviation correction record authentication examples through the first deviation correction state vector mining branch in the StackedLSTM network to be tuned to obtain deviation correction state vector authentication examples; and performing knowledge vector mining on the second deviation correction state vector mining branch in the StackedLSTM network to be tuned to obtain deviation correction state vector authentication examples. The knowledge vector mining branch performs knowledge vector mining on the calibration deviation correction record perturbation examples to obtain deviation correction state vector perturbation examples; it optimizes the attention semantics of the sensor performance expectation features corresponding to the calibration deviation correction record authentication examples to obtain first optimized training information; it optimizes the attention semantics of the sensor performance expectation features corresponding to the calibration deviation correction record perturbation examples to obtain second optimized training information; based on the calibration data knowledge vector examples, the deviation correction state vector authentication examples, the deviation correction state vector perturbation examples, the first optimized training information, and the second optimized training information, it determines the network tuning quality evaluation function; based on the network tuning quality evaluation function, it optimizes the neural network weights of the Stacked Long Short-Term Memory (StackedLSTM) network to be tuned until it meets the neural network tuning requirements, and then terminates to obtain the tuned StackedLSTM network.

[0125] In detail, determining the network calibration quality evaluation function based on the calibration data knowledge vector sample, the deviation correction state vector authentication sample, the deviation correction state vector perturbation sample, the first optimized training information, and the second optimized training information includes: determining a first correlation weight between the mechanical sensor calibration device sample and the calibration deviation correction record authentication sample based on the calibration data knowledge vector sample and the deviation correction state vector authentication sample; determining a second correlation weight between the mechanical sensor calibration device sample and the calibration deviation correction record perturbation sample based on the calibration data knowledge vector sample and the deviation correction state vector perturbation sample; obtaining a first iterative training error based on the first correlation weight, the first optimized training information, and the set iterative processing decision features; obtaining a second iterative training error based on the second correlation weight, the second optimized training information, and the set iterative processing decision features; and determining the network calibration quality evaluation function based on the first iterative training error and the second iterative training error.

[0126] In the following steps, the attention semantics of the sensor performance expectation features corresponding to the calibration deviation correction record authentication examples are optimized to obtain first optimized training information; the attention semantics of the sensor performance expectation features corresponding to the calibration deviation correction record perturbation examples are optimized to obtain second optimized training information, including: weighting the sensor performance expectation features corresponding to the calibration deviation correction record authentication examples and setting reinforcement features to obtain first sensor performance expectation features; processing the attention semantics of the first sensor performance expectation features through a channel attention mechanism to obtain first optimized training information; weighting the sensor performance expectation features corresponding to the calibration deviation correction record perturbation examples and setting reinforcement features to obtain second sensor performance expectation features; and processing the attention semantics of the second sensor performance expectation features through a channel attention mechanism to obtain second optimized training information.

[0127] The tuning method for StackedLSTM networks can be achieved through the following five steps.

[0128] I. Obtaining the Prior Training Example Set

[0129] ① Components of the prior training example set

[0130] The prior training example set contains a variety of sample data that are crucial for tuning the Stacked Long Short-Term Memory (LSTM) network. Among them, the mechanical sensing calibration data samples corresponding to the mechanical sensing calibration device examples are the foundational part. These data samples are similar to the data generated by actual mechanical sensing calibration devices during the calibration process, including different mechanical input quantities (such as force, pressure, torque, etc.) and the corresponding output signals of the sensors (voltage, current, etc.), and may also include data related to the calibration environment (such as temperature, humidity, etc.).

[0131] Certified calibration deviation correction records are certified examples of accurate and reliable calibration deviation correction records. These records detail various information during the deviation correction process for mechanical sensing devices, such as the type of correction (zero-point correction, gain correction, etc.), specific parameters during correction (such as correction coefficients), and post-correction result evaluation (measured values ​​of sensor performance parameters after correction).

[0132] Calibration deviation correction record disturbance samples are correction record samples that differ from certification samples. They may be correction records generated under certain conditions (such as the presence of certain noise interference, environmental fluctuations, or slight changes in correction parameters) and are used to test the network's adaptability and accuracy under different conditions.

[0133] ②Acquisition methods and data accuracy requirements

[0134] Examples of mechanical sensor calibration data can be obtained through simulated calibration experiments or by selecting representative portions from a large pool of existing calibration data. For simulated experiments, data acquisition must be performed strictly according to the actual calibration process and accuracy requirements. For example, in the calibration of simulated force sensors, the simulation accuracy of the mechanical input should be close to the actual measurement level (e.g., force simulation accuracy of ±0.01N), the acquisition accuracy of the sensor output signal should also meet the requirements (e.g., voltage acquisition accuracy of ±0.001V), and the simulation accuracy of environmental parameters should also be guaranteed (temperature simulation accuracy of ±0.1°C, humidity simulation accuracy of ±1%).

[0135] Calibration deviation correction record certification samples and disturbance samples can be obtained from the actual calibration operation record database. Certification samples are records that have undergone rigorous review and verification, while disturbance samples can be obtained by making controlled modifications to certification samples (such as changing calibration parameters within a certain range, introducing small-amplitude environmental fluctuations, etc.).

[0136] II. Knowledge Vector Mining Using the StackedLSTM to be Tuned

[0137] ① Operations for defining knowledge vector mining branches in data

[0138] For the knowledge vector mining branch of the calibration data in the StackedLSTM to be calibrated, mechanical sensing calibration data samples are used as input. For example, a mechanical sensing calibration data sample is a data sequence containing 1000 time steps, each time step containing force, voltage, temperature, and humidity information. Before being input into the StackedLSTM, this data needs to be preprocessed, such as normalizing the force, voltage, temperature, and humidity to the interval [-1, 1].

[0139] Then, the data enters each layer of the StackedLSTM LSTM. Taking the first layer of the LSTM as an example, for the input vector (x_t) at a single time step (t) (containing normalized force, voltage, temperature, and humidity information), the forget gate (f_t) is calculated as: (f_t = sigma(W_fcdot[h_{t-1}, x_t] + b_f)), where (W_f) is the weight matrix of the forget gate, (h_{t-1}) is the hidden state of the previous time step, (b_f) is the bias vector of the forget gate, and (sigma) is the sigmoid function. The updates of the input gate (i_t), output gate (o_t), and cell state (C_t), as well as the calculation of the hidden state (h_t), are all performed according to the standard LSTM calculation rules. After processing by multiple layers of LSTM, the output of the last LSTM layer is converted into a labeled data knowledge vector sample through a fully connected layer. For example, if the weight matrix of the fully connected layer is (W_{out}) and the bias vector is (b_{out}), then the formula for calculating the labeled data knowledge vector sample (V) is: (V=W_{out}cdoth_m+b_{out}), where (h_m) is the hidden state at the last time step.

[0140] ② Operation of the first deviation correction state vector mining branch

[0141] The first deviation correction state vector mining branch in the StackedLSTM to be calibrated performs knowledge vector mining on the calibration deviation correction record authentication examples. For example, the calibration deviation correction record authentication example is represented in vector form as ([t, p, m, i, r]), where (t) represents the correction time (in optional time encoding form), (p) represents the performance parameter targeted by the correction (encoded representation), (m) represents the correction method (encoded representation), (i) represents the correction input parameter, and (r) represents the evaluation of the result after correction.

[0142] This vector is input into the StackedLSTM of the first bias-corrected state vector mining branch, where the internal LSTM units process it according to the corresponding calculation rules. For example, the calculation of the forget gate, input gate, output gate, and cell state is similar to the previous ones. After processing through multiple LSTM layers, the output of the last LSTM layer is converted into a bias-corrected state vector authentication example through a fully connected layer. Let the weight matrix of the fully connected layer be (W_{out1}) and the bias vector be (b_{out1}), then the calculation formula for the bias-corrected state vector authentication example (V_{auth}) is: (V_{auth}=W_{out1}cdoth_{last1}+b_{out1}), where (h_{last1}) is the hidden state at the last time step of this branch.

[0143] ③ Operation of the second deviation correction state vector mining branch

[0144] The second deviation correction state vector mining branch performs knowledge vector mining on the perturbation samples of the calibration deviation correction records to be calibrated. The process is similar to the first deviation correction state vector mining branch, but because the characteristics of the input calibration deviation correction record perturbation samples may differ (e.g., data changes caused by perturbation factors), their internal parameters such as weight matrices and bias vectors may also differ. For example, the vector representation of a calibration deviation correction record perturbation sample is ([t', p', m', i', r']) (where the apostrophe-marked elements represent values ​​that may differ from the certified samples). After processing by StackedLSTM, the deviation correction state vector perturbation sample (V_{pert}) is obtained. The calculation process is similar to the first deviation correction state vector mining branch, except that a different weight matrix (W_{out2}) and bias vector (b_{out2}) are used. The deviation correction state vector perturbation sample is as follows:

[0145] (V_{pert}=W_{out2}cdoth_{last2}+b_{out2}), where (h_{last2}) is the hidden state of the last time step of this branch.

[0146] III. Attention semantic optimization yields optimized training information.

[0147] ① Processing of calibration deviation correction record certification samples

[0148] For the sensor performance expectation feature corresponding to the calibration deviation correction record certification sample, it is first weighted with the set enhancement feature to obtain the first sensor performance expectation feature. For example, if the sensor performance expectation feature is a vector (E=[e_1, e_2, cdots, e_n]) and the set enhancement feature is a vector (S=[s_1, s_2, cdots, s_n]), then the element of the first sensor performance expectation feature (E_1) is calculated as (e_{1i}=e_i×s_i)((i=1, 2, cdots, n)).

[0149] Then, the attention semantics of the first sensor performance expectation feature are processed through a channel attention mechanism to obtain the first optimized training information. The channel attention mechanism can adjust the weight of each channel by calculating the relationship between the channels of the feature map. For example, for the first sensor performance expectation feature (E_1), its channel attention weight vector (A=[a_1, a_2, cdots, a_n]) can be obtained by global average pooling and fully connected layer operations. Then, the elements of the first optimized training information (I_{auth}) are calculated as (i_{authi}=e_{1i}×a_i)((i=1,2,cdots,n)).

[0150] ② Processing of disturbance samples in calibration deviation correction records

[0151] The second sensor performance expectation feature is obtained by weighting the sensor performance expectation feature corresponding to the calibration deviation correction record disturbance sample and the set enhancement feature. For example, if the sensor performance expectation feature corresponding to the calibration deviation correction record disturbance sample is a vector (E'=[e_1', e_2', cdots, e_n']), then the element of the second sensor performance expectation feature (E_2) is calculated as (e_{2i}=e_i'×s_i)((i=1,2,cdots,n)).

[0152] The second optimized training information is obtained by processing the attention semantics of the expected performance features of the second sensor through the channel attention mechanism. Similarly, its channel attention weight vector (A'=[a_1', a_2', cdots, a_n']) is calculated, and then the elements of the second optimized training information (I_{pert}) are calculated as (i_{perti}=e_{2i}×a_i') ((i=1,2,cdots,n)).

[0153] IV. Determining the Network Calibration Quality Evaluation Function

[0154] ① Calculate the first association weight and the second association weight

[0155] The first association weight between the calibration data knowledge vector sample and the deviation correction state vector authentication sample is determined based on the calibration data knowledge vector sample and the deviation correction record authentication sample. An optional distance metric algorithm, such as the Euclidean distance algorithm, can be used. Let the calibration data knowledge vector sample be (V=[v_1, v_2, cdots, v_m]) and the deviation correction state vector authentication sample be (V_{auth}=[v_{auth1}, v_{auth2}, cdots, v_{authn}]) (for example, (m) and (n) can be appropriately processed to make them usable for calculation, such as padding or truncation), then the formula for calculating the first association weight (W_1) is: (W_1=frac{1}{1+sqrt{sum_{i=1}^{min(m,n)}(v_i-v_{authi})^2}}).

[0156] Similarly, the second association weight (W_2) between the mechanical sensing calibration equipment sample and the calibration deviation correction record disturbance sample is determined based on the calibration data knowledge vector sample and the deviation correction state vector disturbance sample. Let the deviation correction state vector disturbance sample be (V_{pert}=[v_{pert1},v_{pert2},cdots,v_{pertn}]), then:

[0157] (W_2=frac{1}{1+sqrt{sum_{i=1}^{min(m,n)}(v_i-v_{perti})^2}}).

[0158] ② Calculate the training error of the first iteration and the training error of the second iteration.

[0159] The first iterative training error is obtained based on the first association weight, the first optimized training information, and the set iterative processing decision features. For example, the set iterative processing decision features are a vector (D=[d_1, d_2, cdots, d_n]), and the first optimized training information is:

[0160] (I_{auth}=[i_{auth1}, i_{auth2}, cdots, i_{authn}]), then the formula for calculating the training error (E_1) of the first iteration is:

[0161] (E_1=sum_{i=1}^{n}(W_1×i_{authi}-d_i)^2).

[0162] The second iterative training error is obtained based on the second association weight, the second optimized training information, and the set iterative processing decision features. Let the second optimized training information be:

[0163] (I_{pert}=[i_{pert1}, i_{pert2}, cdots, i_{pertn}]), then the formula for calculating the training error (E_2) of the second iteration is:

[0164] (E_2=sum_{i=1}^{n}(W_2×i_{perti}-d_i)^2).

[0165] ③ Determine the network calibration quality evaluation function

[0166] The network tuning quality evaluation function is determined based on the training errors of the first and second iterations. For example, the network tuning quality evaluation function (Q) can be defined as: (Q=E_1+E_2). This function comprehensively considers the training errors obtained from the certified and perturbation examples, and can comprehensively evaluate the tuning quality of the network.

[0167] V. Neural Network Weight Optimization and Obtaining the Adjusted StackedLSTM

[0168] ① Neural network weight optimization process

[0169] The neural network weights of the StackedLSTM to be tuned are optimized based on the network tuning quality evaluation function (Q). During the optimization process, optimization algorithms such as gradient descent can be used. For example, for a certain weight (w) in the StackedLSTM, its update formula is: (w = w - alpha × frac{partialQ}{partialw}), where (alpha) is the learning rate (e.g., (alpha = 0.001)). This process updates all weights in the StackedLSTM (including weights in the calibration data knowledge vector mining branch, the first bias correction state vector mining branch, and the second bias correction state vector mining branch).

[0170] After each weight update, the network tuning quality evaluation function (Q) is recalculated to determine whether it meets the standards for neural network tuning. The standards can be that the value of (Q) is less than a certain set threshold (e.g., (Q_{th}=0.01)) or that optimization stops after a certain number of iterations (e.g., 1000 iterations).

[0171] ② Obtain the calibrated StackedLSTM

[0172] The optimization process terminates when the neural network meets the calibration requirements, and the resulting StackedLSTM is the calibrated StackedLSTM. This calibrated network can be used for subsequent operations such as mining calibration data knowledge vectors from mechanical sensor calibration data and mining deviation correction state vectors from deviation correction records, thereby providing accurate data support for the entire calibration deviation correction process of the mechanical sensor calibration equipment.

[0173] By applying the above embodiments, and utilizing a prior training example set containing various examples (mechanical sensor calibration data examples, calibration deviation correction record authentication examples, and perturbation examples) for network tuning, StackedLSTM can adapt to data under different conditions, improving the network's generalization ability. For example, the introduction of calibration deviation correction record perturbation examples allows the network to accurately extract information even when faced with data that may contain interference or changes.

[0174] A method based on association weights, optimized training information, and iterative processing of decision features to calculate the network tuning quality evaluation function comprehensively considers multiple factors, enabling a more accurate assessment of network tuning quality. Furthermore, optimizing the sensor performance expectation features through attention semantics to obtain optimized training information helps focus on key factors affecting performance, improving the network's ability to process sensor performance-related information.

[0175] The entire calibration process ensures that the StackedLSTM can more accurately mine vectors when used for subsequent operations, thereby improving the accuracy and efficiency of the entire mechanical sensing calibration equipment calibration deviation correction process.

[0176] In some preferred design approaches, determining the association weight between the target mechanical sensing calibration device and each calibration deviation correction record to be processed, based on the calibration data knowledge vector and the deviation correction state vector corresponding to each calibration deviation correction record to be processed, includes: for each calibration deviation correction record to be processed, determining the feature operation results of the calibration data knowledge vector and the deviation correction state vector corresponding to the target calibration deviation correction record to be processed; and determining the association weight between the target mechanical sensing calibration device and the target calibration deviation correction record to be processed based on the feature operation results.

[0177] Based on this, the optimization of the attention semantics of the sensor performance expectation features corresponding to each calibration deviation correction record to obtain the performance optimization output features of each sensor performance expectation feature includes: for each calibration deviation correction record to be processed, weighting the sensor performance expectation features corresponding to the target calibration deviation correction record and the set enhancement features to obtain the enhanced sensor performance expectation features; and processing the attention semantics of the enhanced sensor performance expectation features through a channel attention mechanism to obtain the performance optimization output features of the sensor performance expectation features corresponding to the target calibration deviation correction record.

[0178] Furthermore, the step of integrating the correlation weights between each calibration deviation correction record to be processed and the target mechanical sensing calibration device with the performance optimization output features of the corresponding sensor performance expectation features to obtain the screening priority indications corresponding to each calibration deviation correction record to be processed includes: for each calibration deviation correction record to be processed, the correlation weights between the target mechanical sensing calibration device and the target calibration deviation correction record to be processed, and the performance optimization output features of the sensor performance expectation features corresponding to the target calibration deviation correction record to be processed are globally processed to obtain the screening priority indication of the target calibration deviation correction record to be processed.

[0179] In practical applications, the following is a detailed explanation of the above design concept.

[0180] I. Determine the correlation weight between the target mechanical sensing calibration equipment and each calibration deviation correction record to be processed.

[0181] ① Determine the result of feature operation

[0182] For each calibration deviation correction record to be processed, it is necessary to determine the feature operation results of the calibration data knowledge vector and the deviation correction state vector corresponding to the calibration deviation correction record. For example, the calibration data knowledge vector is (V=[v_1, v_2, cdots, v_n]), and the deviation correction state vector is (U=[u_1, u_2, cdots, u_m]) (where (n) and (m) are the dimensions of the vector).

[0183] An exemplary feature operation could be the dot product of vectors. For example, calculating the dot product of the calibrated data knowledge vector (V) and the bias correction state vector (U) as the feature operation result (R) would be (R = VcdotU = sum_{i=1}^{min(n,m)}v_i × u_i). Another possible feature operation could be cosine similarity calculation, where cosine similarity (S) is calculated as (S = frac{VcdotU}{leftvertVrightvert × leftvertUrightvert}), where:

[0184] (leftvertVrightvert=sqrt{sum_{i=1}^{n}v_i^2}),

[0185] (leftvertUrightvert=sqrt{sum_{i=1}^{m}u_i^2}).

[0186] The results of these feature operations can, to some extent, reflect the close relationship between the calibration data knowledge vector and the deviation correction state vector.

[0187] ② Determine the association weight

[0188] The association weight between the target mechanical sensing calibration device and the calibration deviation correction record to be processed is determined based on the feature operation results. For example, when the dot product is used as the feature operation result (R), the dot product result can be converted into association weight (W) through a mapping function. An exemplary mapping function can be (W=frac{1}{1+e^{-R}}), where (e) is the natural constant. This function maps the dot product result (R) to the interval ((0,1)), making the value of the association weight (W) within this reasonable range. If cosine similarity (S) is used as the feature operation result, the association weight (W) can be directly defined as (W=S), because the value of cosine similarity itself is in the interval ([-1,1]) and can well represent the similarity between two vectors. Here, it can be regarded as the association weight, and only the value of ([0,1]) part is taken to represent the positive correlation.

[0189] II. Optimizing Sensor Performance Desired Features Through Attention Semantics to Obtain Performance-Optimized Output Features

[0190] ① Weighted enhancement of sensor performance expected characteristics

[0191] For each calibration deviation correction record to be processed, the expected sensor performance feature and the set enhancement feature corresponding to the target calibration deviation correction record are weighted. For example, the expected sensor performance feature is (E=[e_1, e_2, cdots, e_p]), and the set enhancement feature is (F=[f_1, f_2, cdots, f_p]).

[0192] The elements of the enhanced sensor performance expected feature (G) are calculated as (g_i = e_i × f_i) ((i = 1, 2, cdots, p)). For example, if an element (e_1) in the sensor performance expected feature (E) represents the sensor's sensitivity deviation at a certain force value, with a value of (0.01), and the corresponding element in the enhanced feature (F) is set to (f_1 = 2), then the corresponding element in the enhanced sensor performance expected feature (G) is (g_1 = 0.01 × 2 = 0.02).

[0193] ② Performance-optimized output features are obtained through channel attention mechanism processing.

[0194] The enhanced sensor performance expectation features (G) are processed using a channel attention mechanism to obtain performance-optimized output features. The channel attention mechanism adjusts the weights of each channel by calculating the relationships between channels in the feature map.

[0195] For example, the channel attention mechanism employs a method based on global average pooling and fully connected layers. First, global average pooling is performed on the enhanced sensor performance expectation feature (G) to obtain a (1×p) vector (H=[h_1, h_2, cdots, h_p]), where (h_i=frac{1}{q}sum_{j=1}^{q}g_{ij}) (for example, (G) has (q) samples or time steps). Then (H) is passed through a fully connected layer with a weight matrix of (W_{fc}=[w_{11}, w_{12}, cdots, w_{1p}; w_{21}, w_{22}, cdots, w_{2p}; cdots; w_{r1}, w_{r2}, cdots, w_{rp}]) (where (r) is the number of neurons in the fully connected layer), and a bias vector of (b=[b_1, b_2, cdots, b_p]).

[0196] The output vector after the fully connected layer is (I=[i_1, i_2, cdots, i_p]), where (i_k=sum_{l=1}^{p}w_{kl}h_l+b_k)((k=1, 2, cdots, p)). Finally, (I) is passed through an activation function (such as the sigmoid function (sigma)) to obtain the channel attention weight vector (A=[a_1, a_2, cdots, a_p]=sigma(I)). The elements of the performance-optimized output feature (O) are calculated as (o_i=g_i×a_i)((i=1, 2, cdots, p)).

[0197] III. Feature integration yields filtering priority indicators

[0198] ① Global processing yields filtering priority indicators

[0199] For each calibration deviation correction record to be processed, the correlation weight (W) between the target mechanical sensing calibration device and the target calibration deviation correction record to be processed, and the performance optimization output feature (O) of the sensor performance expectation feature corresponding to the target calibration deviation correction record to be processed are processed globally to obtain the screening priority indicator.

[0200] One approach to global processing is to perform element-wise multiplication of the association weights (W) and the performance-optimized output features (O), and then normalize the resulting vector. For example, given the association weights (W=0.8) and the performance-optimized output features (O=[o_1, o_2, cdots, o_p]), first calculate (P=[w×o_1, w×o_2, cdots, w×o_p]). Then normalize (P) and calculate the vector magnitude of (P):

[0201] (leftvertPrightvert=sqrt{sum_{i=1}^{p}(w×o_i)^2}), filter the elements of the priority indicator (Q) (q_i=frac{w×o_i}{leftvertPrightvert})((i=1,2,cdots,p)). This priority indicator (Q) combines information from correlation weights and performance optimization output characteristics, and can be used to determine the priority order of calibration deviation correction records to be processed.

[0202] This design has several advantages. First, when determining the association weights, the relationship between the calibration data knowledge vector and the deviation correction state vector is measured through specific feature calculations, accurately reflecting the degree of association between the two. Using methods such as dot product or cosine similarity calculations makes the determination of association weights more scientific. Second, when optimizing the expected features of sensor performance, the combination of weighted operations and channel attention mechanisms can highlight key information based on set enhancement features. The complex calculations of the channel attention mechanism accurately adjust the weights of each channel, improving the accuracy of sensor performance evaluation. Finally, the feature integration method for obtaining the screening priority indicator comprehensively considers the association weights and performance optimization output features. Through global processing, a comprehensive index is obtained, which helps to rationally determine the priority order of calibration deviation correction records to be processed, thereby improving the accuracy and efficiency of the entire calibration deviation correction process.

[0203] In one innovative design of the present invention, the tuned StackedLSTM network includes multiple LSTM layers, which are stacked vertically on top of each other. The multiple LSTM layers achieve gating information transmission and gradient vanishing mitigation through a gating mechanism.

[0204] Based on this, obtaining the prior training example set includes: obtaining the calibration deviation correction record queue example corresponding to the mechanical sensing calibration device example; filtering the noise calibration deviation correction records in the calibration deviation correction record queue example to obtain the deviation correction record denoising queue; and determining the prior training example set based on the deviation correction record denoising queue and the set augmentation strategy.

[0205] Furthermore, determining the prior training example set based on the deviation correction record denoising queue and the set augmentation strategy includes: for each deviation correction record example in the deviation correction record denoising queue, determining the initial performance expectation features based on the sensor performance expectation features corresponding to the target deviation correction record example and the set augmentation coefficient; and determining the prior training example set based on the target deviation correction record example and the initial performance expectation features.

[0206] Furthermore, determining the prior training example set based on the deviation correction record denoising queue and the set augmentation strategy includes: analyzing the deviation correction pattern corresponding to each deviation correction record sample in the deviation correction record denoising queue; determining a target deviation correction pattern whose number of deviation correction record samples is less than a set number based on the deviation correction pattern analysis results; performing sample augmentation on the deviation correction record denoising queue based on the target deviation correction pattern; and determining the prior training example set based on the sample augmentation results.

[0207] In this invention, an exemplary description of the above-described innovative design is provided below.

[0208] I. Structural characteristics of the calibrated StackedLSTM network

[0209] ① Vertical stacking of multiple LSTM layers

[0210] StackedLSTM consists of multiple LSTM layers stacked vertically on top of each other. LSTM (Long Short-Term Memory) is a special type of recurrent neural network (RNN) designed to solve the vanishing and exploding gradient problems encountered by traditional RNNs when processing long sequences of data. Each LSTM layer has its own internal structure and function.

[0211] In a vertically stacked LSTM structure, the output of the previous LSTM layer serves as the input to the next LSTM layer. For example, the first LSTM layer receives the raw input data (such as mechanical sensor calibration data samples and related data), performs internal calculations, and outputs an intermediate result. This intermediate result contains the feature information extracted by the first LSTM layer from the input data. This intermediate result is then used as the input to the second LSTM layer, which further extracts higher-level and more abstract features. This process continues, and stacking multiple LSTM layers can progressively delve deeper into the complex patterns within the data.

[0212] ② The gating mechanism enables gating information transmission and mitigates gradient vanishing.

[0213] LSTM uses gating mechanisms to control the transmission of information, including input gates, forget gates, and output gates. Taking an exemplary LSTM unit as an example, the forget gate (f_t) determines how much information from the cell state (C_{t-1}) at the previous time step is retained at the current time step (t). Its calculation formula is (f_t=sigma(W_fcdot[h_{t-1},x_t]+b_f)), where (W_f) is the weight matrix of the forget gate, (h_{t-1}) is the hidden state at the previous time step, (x_t) is the input at the current time step, (b_f) is the bias vector of the forget gate, and (sigma) is the sigmoid function.

[0214] The input gate (i_t) determines how much new information can be updated to the cell state at the current time step (i_t = sigma(W_icdot[h_{t-1}, x_t] + b_i)). The update formula for the cell state (C_t) is (C_t = f_t × C_{t-1} + i_t × tanh(W_Ccdot[h_{t-1}, x_t] + b_C)), where (W_C) is the weight matrix for cell state updates, and (b_C) is the corresponding bias vector. The output gate (o_t) determines how much information from the cell state (C_t) can be output to the hidden state (h_t) at the current time step (o_t = sigma(W_ocdot[h_{t-1}, x_t] + b_o)), and (h_t = o_t × tanh(C_t)).

[0215] In StackedLSTM, this gating mechanism plays a role in each LSTM layer, not only enabling the transfer of gating information between layers but also effectively mitigating the vanishing gradient problem. When calculating the gradient during backpropagation, the gating mechanism controls the flow of gradients, preventing them from vanishing or exploding too quickly, thus enabling the network to effectively learn long-term dependencies in long sequences of data.

[0216] II. The process of obtaining the prior training example set

[0217] ① Obtain a sample of the calibration deviation correction record queue and filter out noisy records.

[0218] First, obtain the calibration deviation correction record queue sample corresponding to the mechanical sensor calibration device sample. This queue sample contains a series of records of deviation correction performed on this mechanical sensor calibration device. These records may contain various information about the correction, such as the correction time, the sensor performance parameters targeted for correction, the correction method, and the results after correction.

[0219] However, these records may contain noise calibration deviation correction records, such as abnormal correction records caused by temporary instrument malfunctions, environmental interference, or other factors. These noise calibration deviation correction records need to be filtered to obtain a denoising queue of deviation correction records. An exemplary filtering method could be based on statistical analysis. For example, for a key indicator (such as the corrected sensor performance parameter value) in each deviation correction record, its mean (mu) and standard deviation (sigma) are calculated. If the value of this indicator in a record exceeds the range of ([mu-ksigma, mu+ksigma]) (where (k) is a set threshold, for example (k=3)), then the record is considered a noise record and is filtered out.

[0220] ② Determine the prior training example set based on the bias correction record denoising queue and the set augmentation strategy.

[0221] Determine the initial performance expectation features: For each deviation correction record sample in the deviation correction record denoising queue, determine the initial performance expectation features based on the sensor performance expectation features corresponding to the target deviation correction record sample and the set enhancement coefficient. For example, if the sensor performance expectation features corresponding to a deviation correction record sample are (E=[e_1, e_2, cdots, e_n]), and the enhancement coefficient is set to (s) (e.g., (s=2)), then the initial performance expectation features are (E'=[e_1×s, e_2×s, cdots, e_n×s]). This step can enhance the sensor performance expectation features according to actual needs, so as to highlight the influence of certain features in subsequent training.

[0222] The prior training example set is determined based on the deviation correction record samples and the initial performance expectation features: The prior training example set is determined according to the target deviation correction record samples and the initial performance expectation features. For example, information from the deviation correction record samples (such as correction time, correction method, etc.) is combined with the initial performance expectation features to form a sample in the prior training example set. Specifically, if the deviation correction record sample is (D=[t, m, r]) ((t) represents the correction time, (m) represents the correction method, and (r) represents the correction result), and the initial performance expectation features are (E'=[e_1', e_2', cdots, e_n']), then the combined sample can be ([t, m, r, e_1', e_2', cdots, e_n']), and these combined samples together constitute the prior training example set.

[0223] Sample augmentation: Analysis is performed on the deviation correction patterns corresponding to each deviation correction record sample in the deviation correction record denoising queue. Deviation correction patterns can be categorized based on factors such as the type of correction and the target sensor performance parameters. For example, correction patterns can be divided into zero-point correction patterns for linearity, gain correction patterns for sensitivity, etc.

[0224] Based on the analysis results of the deviation correction mode, the target deviation correction mode is determined to have fewer than the set number of deviation correction record samples. For example, if the set number is (N=10), and the analysis reveals that there are only 5 deviation correction record samples in a certain special correction mode for repeatability, which is less than the set number, then this mode is the target deviation correction mode.

[0225] Based on the target bias correction mode, sample augmentation is performed on the denoising queue of bias correction records. An exemplary sample augmentation method can be based on interpolation. For example, if the bias correction record samples under the target bias correction mode are ([x_1, y_1, z_1]) and ([x_2, y_2, z_2]) ((x) represents the input parameters during correction, (y) represents the parameters related to the correction method, and (z) represents the sensor performance parameters after correction), a new sample ([x_3, y_3, z_3]) can be obtained through linear interpolation, where:

[0226] (x_3=frac{x_1+x_2}{2})

[0227] (y_3=frac{y_1+y_2}{2}),

[0228] (z_3=frac{z_1+z_2}{2}).

[0229] This method augments the samples under the target bias correction mode, making the number of samples under different bias correction modes more balanced, thereby improving the quality of the prior training example set.

[0230] The prior training example set is determined based on the sample augmentation results. The augmented samples and the samples in the original bias correction record denoising queue (after operations such as combining with the initial performance expectation features) together constitute the prior training example set. This prior training example set will be used in the subsequent StackedLSTM tuning process.

[0231] This design offers several advantages. First, the StackedLSTM structure, with its multiple vertically stacked LSTM layers, utilizes a gating mechanism to facilitate gating information transmission and mitigate gradient vanishing. This allows the network to effectively process long sequences of data and uncover complex patterns, enhancing its ability to handle mechanical sensor calibration data and related information. Second, when acquiring the prior training example set, filtering outnoise calibration bias correction records removes interference from abnormal data, improving data quality. The initial performance expectation features are determined based on the sensor's desired characteristics and the set enhancement coefficients, allowing for the highlighting of key features according to actual needs. The sample augmentation operation balances the number of samples under different bias correction modes, making the prior training example set more representative, thereby improving the calibration effect of the StackedLSTM and ultimately enhancing the accuracy and efficiency of the entire mechanical sensor calibration equipment's processing flow.

[0232] In some extended embodiments of the present invention, after prioritizing each calibration deviation correction record to be processed according to the screening priority indication to obtain a calibration deviation correction record queue, the method further includes: determining a target calibration deviation correction record according to the calibration deviation correction record queue; generating a current calibration correction scheme for the target mechanical sensing calibration device using the target calibration deviation correction record; and performing calibration correction processing on the target mechanical sensing calibration device using the current calibration correction scheme.

[0233] In some further extended embodiments of the present invention, the target calibration deviation correction record is used to generate a current calibration correction scheme for the target mechanical sensing calibration device, including: obtaining the error source discrimination viewpoint and the offset capture discrimination viewpoint corresponding to the target mechanical sensing calibration device; pairing the error source discrimination viewpoint and the offset capture discrimination viewpoint in the target mechanical sensing calibration device based on the discrimination viewpoint logic characteristics between them to obtain a discrimination viewpoint set; and identifying unpaired offset capture discrimination viewpoints as outliers. The system first captures and discriminates viewpoints. Based on the semantic commonality value between the offset capture discriminative viewpoints in the discriminative viewpoint set and the outlier offset capture discriminative viewpoints, it determines the correction strategy features that match the outlier offset capture discriminative viewpoints. It then pairs the correction strategy features that match the outlier offset capture discriminative viewpoints with the outlier offset capture discriminative viewpoints to obtain a linkage calibration correction vector. Based on the linkage calibration correction vector, the discriminative viewpoint set, and the calibration deviation correction causal features corresponding to the target calibration deviation correction record, it generates the current calibration correction scheme corresponding to the target mechanical sensing calibration device.

[0234] The following is an exemplary description of the relevant technical solutions in relation to the above extended embodiments.

[0235] 1. Determine the target calibration deviation correction record based on the calibration deviation correction record queue and perform calibration correction processing.

[0236] ① Determine the target calibration deviation correction record

[0237] After prioritizing the calibration deviation correction records according to the screening priority indicators to obtain a calibration deviation correction record queue, the target calibration deviation correction record is determined based on this queue. Records in the calibration deviation correction record queue are arranged according to priority, with higher-priority records being more likely to be selected as target calibration deviation correction records. For example, in the queue, the priority of a record can be represented by a numerical value, with higher values ​​indicating higher priority. For instance, if there are 5 records in the calibration deviation correction record queue with priority values ​​of [0.8, 0.6, 0.5, 0.3, 0.2], the record with a priority value of 0.8 will be preferentially selected as the target calibration deviation correction record.

[0238] This selection process can be based on an exemplary selection rule, namely, selecting the record with the highest priority. In practical applications, more complex selection strategies can also be set according to specific needs, such as randomly selecting one record from those that meet a certain priority threshold (e.g., 0.7) to increase the flexibility of the selection.

[0239] ② Generate the current calibration and correction scheme and perform calibration and correction processing.

[0240] A current calibration scheme for the target mechanical sensor calibration equipment is generated using the selected target calibration deviation correction record. The target mechanical sensor calibration equipment is then calibrated using this current calibration scheme. For example, if the target calibration deviation correction record indicates that previous calibration operations adjusted the sensor gain with a gain adjustment factor of 1.05, the current calibration scheme may include a strategy to further adjust the gain, such as adjusting it to 1.03 or 1.07 based on the current measurement conditions of the target equipment. During the calibration process, the relevant parameters of the target mechanical sensor calibration equipment are adjusted according to the strategy in this scheme to improve its calibration accuracy.

[0241] II. Detailed process of generating the current calibration correction scheme using target calibration deviation correction records.

[0242] ① Obtain the error source discrimination viewpoint and offset capture discrimination viewpoint

[0243] This involves acquiring the error source identification and offset capture identification perspectives for the target mechanical sensing calibration equipment. The error source identification perspective may include various factors, such as the impact of environmental factors (e.g., temperature, humidity) on sensor measurement error, and the error caused by the equipment's own hardware characteristics (e.g., sensor linearity, sensitivity). For example, the error source identification perspective might be expressed as: due to temperature changes within the range of 5°C-10°C, the sensor's measurement error may increase by 1%-2%, which is based on extensive experimental data and theoretical analysis.

[0244] The offset capture and discrimination approach focuses on judging the offset of sensor measurements. For example, after a long period of operation, if the output voltage of a sensor deviates by 0.05V from its initial value, this offset may be caused by zero drift or gain changes. The offset capture and discrimination approach is to judge this offset phenomenon and its possible causes.

[0245] ② Based on the logical features of the discriminative viewpoints, a set of discriminative viewpoints is obtained by pairing them.

[0246] Based on the logical characteristics of the error source discrimination viewpoint and the offset capture discrimination viewpoint in the target mechanics sensing calibration equipment, the two are paired to obtain a set of discrimination viewpoints. The logical characteristics of the discrimination viewpoints can be based on factors such as causality and correlation. For example, if the error source discrimination viewpoint indicates that temperature changes cause zero-point drift of the sensor (the error source is temperature, and the result is zero-point drift), and the offset capture discrimination viewpoint also detects the zero-point drift phenomenon, then these two viewpoints can be paired according to this logical relationship.

[0247] For example, there are three error source discrimination viewpoints: (E_1), (E_2), and (E_3), and four offset capture discrimination viewpoints: (O_1), (O_2), (O_3), and (O_4). By analyzing the logical relationships between them, it is found that (E_1) and (O_1), and (E_2) and (O_3) are logically related. Therefore, the discrimination viewpoint set contains two pairs of discrimination viewpoints: ((E_1, O_1)) and ((E_2, O_3)).

[0248] ③ Determine the outlier bias capture and discrimination viewpoint and match the correction strategy features.

[0249] Unpaired offset capture discriminant views are identified as outlier offset capture discriminant views. In the example above, (O_2) and (O_4) are outlier offset capture discriminant views. Then, based on the semantic commonality values ​​between offset capture discriminant views in the discriminant view set and outlier offset capture discriminant views, correction strategy features that match the outlier offset capture discriminant views are determined.

[0250] The semantic commonality value of viewpoints can be calculated using methods such as vector space models. For example, the offset capture discriminant viewpoint can be represented as a vector, such as (O_1=[0.1, 0.2, 0.3]) (where the values ​​can represent different semantic feature weights), (O_2=[0.15, 0.1, 0.25]), and the semantic commonality value of viewpoints can be determined by calculating the cosine similarity of the vectors. Let the cosine similarity between (O_1) and (O_2) be:

[0251] (C=frac{O_1cdotO_2}{leftvertO_1rightvert×leftvertO_2rightvert}), where (cdot) represents the vector dot product, and (leftvertO_1rightvert) and (leftvertO_2rightvert) are the moduli of vectors (O_1) and (O_2), respectively.

[0252] Based on this semantic commonality value, if (C) is greater than a certain threshold (e.g., 0.6), they are considered to have high semantic commonality. The correction strategy feature matching (O_2) can be determined based on the correction strategy feature corresponding to the error source discrimination viewpoint (E_1) paired with (O_1) in the discrimination viewpoint set. For example, if the correction strategy feature corresponding to (E_1) is to adjust the sensor zero point with an adjustment coefficient of 0.9, then a similar zero-point adjustment strategy can be used for (O_2), but the adjustment coefficient may be fine-tuned based on the specific semantic commonality value, such as adjusting it to 0.92.

[0253] ④ Obtain the linkage calibration correction vector and generate the current calibration correction scheme.

[0254] The correction strategy features that match the outlier offset capture discrimination viewpoint are paired with the outlier offset capture discrimination viewpoint to obtain the linkage calibration correction vector. For example, for the outlier offset capture discrimination viewpoint (O_2) and its matching correction strategy feature (zero-point adjustment coefficient of 0.92), the linkage calibration correction vector can be expressed as ([O_2, 0.92]).

[0255] Based on the linkage calibration correction vector, the discriminant viewpoint set, and the causal characteristics of the calibration deviation correction corresponding to the target calibration deviation correction record, the current calibration correction scheme for the target mechanical sensing calibration device is generated. The causal characteristics of the calibration deviation correction can represent the causal relationship between the correction operation and the sensor performance change in the target calibration deviation correction record. For example, if the target calibration deviation correction record indicates that the previous correction operation was a gain adjustment, and this led to a change in the sensor's linearity, this is a causal characteristic of the calibration deviation correction.

[0256] Taking these factors into account, current calibration and correction schemes may include adjustment strategies for multiple parameters such as zero point and gain. For example, based on the discrimination point set, zero point adjustment may be performed for errors caused by temperature, zero point adjustment may be performed for specific offsets based on the linkage calibration and correction vector, and gain may be fine-tuned based on the causal characteristics of calibration deviation correction corresponding to the target calibration deviation correction record, thus forming a comprehensive current calibration and correction scheme.

[0257] By applying the above embodiments, the target calibration deviation correction record is determined through a calibration deviation correction record queue. This allows for the rational selection of the most suitable correction record based on priority, improving the scientific rigor and accuracy of target correction record determination. When generating the current calibration correction scheme, multiple factors are comprehensively considered, including error source discrimination and offset capture discrimination perspectives. Through complex pairing, outlier identification, and matching correction strategies, the sensor's state can be analyzed more comprehensively, leading to a more precise correction scheme. This scheme generation method, based on multiple discrimination perspectives and logical relationships, helps improve the calibration correction effect of mechanical sensing calibration equipment, thereby enhancing the equipment's measurement accuracy and reliability.

[0258] In summary, this invention, by determining the correlation weights between the target mechanical sensing calibration device and each calibration deviation correction record to be processed, can accurately measure the degree of connection between the two, making subsequent processing more targeted. Optimizing the attention semantics of the sensor performance expectation features yields performance optimization output features, helping to focus on factors more critical to sensor performance. Feature integration provides a screening priority indicator that comprehensively and reasonably reflects the importance of each calibration deviation correction record to be processed, taking into account both correlation weights and performance optimization output features. Prioritizing the records according to the screening priority indicator yields a queue of calibration deviation correction records, ultimately determining the target calibration deviation correction record to assist in generating the current calibration correction scheme, thus improving the efficiency and accuracy of calibration correction.

[0259] Furthermore, Figure 2 This is a schematic diagram of the structure of a calibration data analysis system 200 provided by the present invention. (See diagram below.) Figure 2 The calibration data analysis system 200 shown includes a processor 210, which can call and run computer programs from memory to implement the method of the present invention.

[0260] Optionally, such as Figure 2 As shown, the calibration data analysis system 200 may further include a memory 230. The processor 210 can retrieve and run computer programs from the memory 230 to implement the method of this invention.

[0261] The memory 230 can be a separate device independent of the processor 210, or it can be integrated into the processor 210.

[0262] Optionally, such as Figure 2 As shown, the calibration data analysis system 200 may also include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices. Specifically, it may send information or data to other devices or receive information or data sent by other devices.

[0263] Optionally, the calibration data analysis system 200 can implement the corresponding processes of the storage engine or components (such as processing modules) in the storage engine or the device on which the storage engine is deployed in the various methods of the present invention. For the sake of brevity, these will not be described in detail here.

[0264] It should be understood that the processor of the present invention may be an integrated circuit chip with signal processing capabilities.

[0265] It is understood that the memory in this invention may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. It should be noted that the memory in the systems and methods described herein is intended to include, but is not limited to, suitable types of memory.

[0266] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0267] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of the present invention is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0268] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0269] It should be understood that the formula parameters recorded in the above technical solution are in LaTeX format (an existing formula editing format), and those skilled in the art can undoubtedly reconstruct the corresponding original formula expression based on the formula parameters recorded in the above technical solution.

[0270] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention without departing from the spirit and scope of the present invention, and all of these modifications are within the scope of protection of the present invention.

Claims

1. A calibration data analysis method for mechanical sensors, characterized in that: The method comprises: Obtain a calibration data knowledge vector corresponding to the target mechanical sensor calibration device, and determine the association weight between the target mechanical sensor calibration device and each calibration deviation correction record to be processed according to the calibration data knowledge vector and the deviation correction state vector corresponding to each calibration deviation correction record to be processed; wherein the calibration data knowledge vector is a comprehensive characterization vector of the calibration data of the target mechanical sensor calibration device, and the calibration data knowledge vector covers various aspects of information in the calibration process, including input physical quantities, output signals, calibration environmental conditions, and properties of the calibration device itself; the deviation correction state vector describes the state information in the previous calibration deviation correction process, and the state information includes: the type of correction, the degree of correction, and the performance index of the sensor after correction; the type of correction includes zero point correction, gain correction, and nonlinear correction; the degree of correction includes the value of the correction coefficient; the performance index of the sensor after correction includes the linearity and sensitivity after correction; Optimizing the attention semantics of the expected sensor performance features corresponding to each calibration deviation correction record to be processed, and obtaining the performance optimization output features of each expected sensor performance feature; the expected sensor performance feature reflects the quantized relationship vector corresponding to the sensor performance parameter; The association weights between each calibration deviation correction record to be processed and the target mechanical sensor calibration device are respectively integrated with the performance optimization output features of the corresponding sensor performance expected features to obtain the screening priority indications corresponding to each calibration deviation correction record to be processed; According to the screening priority indication, each calibration deviation correction record to be processed is prioritized to obtain a calibration deviation correction record queue, and the calibration deviation correction record queue is used to determine a target calibration deviation correction record, and the target calibration deviation correction record is used to assist in generating a current calibration correction plan for the target mechanical sensing calibration device.

2. The method according to claim 1, characterized in that The step of obtaining the calibration data knowledge vector corresponding to the target mechanical sensor calibration device includes: Obtaining mechanical sensor calibration data corresponding to the target mechanical sensor calibration device; By using the calibration data knowledge vector mining branch in the calibrated stacked long short-term memory network StackedLSTM, knowledge vector mining is performed on the mechanical sensor calibration data to obtain the calibration data knowledge vector corresponding to the target mechanical sensor calibration device.

3. The method according to claim 1, characterized in that The method further comprises: Through the first bias correction state vector mining branch or the second bias correction state vector mining branch in the calibrated stacked long short-term memory network StackedLSTM, knowledge vector mining is performed on each calibrated bias correction record in the bias correction record set to obtain the initial bias correction state vector corresponding to each calibrated bias correction record; Determine the deviation correction state vectors corresponding to each calibration deviation correction record to be processed from each initial deviation correction state vector; The method for adjusting the adjusted stacked long short-term memory network StackedLSTM includes: Acquire a priori training example set, wherein the priori training example set includes: a mechanical sensor calibration data sample corresponding to a mechanical sensor calibration equipment sample, a calibration deviation correction record authentication sample, and a calibration deviation correction record disturbance sample; By using the calibration data knowledge vector mining branch in the stacked long short-term memory network StackedLSTM to be adjusted, the knowledge vector mining is performed on the mechanical sensor calibration data sample to obtain the calibration data knowledge vector sample; Through the first deviation correction state vector mining branch in the stacked long short-term memory network StackedLSTM to be adjusted, the calibration deviation correction record authentication sample is subjected to knowledge vector mining to obtain the deviation correction state vector authentication sample; By using the second bias correction state vector mining branch in the stacked long short-term memory network StackedLSTM to be calibrated, the knowledge vector mining is performed on the calibration bias correction record disturbance sample to obtain the bias correction state vector disturbance sample; The attention semantics of the expected characteristics of sensor performance corresponding to the calibration deviation correction record authentication sample is optimized to obtain the first optimized training information; the attention semantics of the expected characteristics of sensor performance corresponding to the calibration deviation correction record disturbance sample is optimized to obtain the second optimized training information; Determine a network calibration quality evaluation function according to the calibration data knowledge vector sample, the deviation correction state vector authentication sample, the deviation correction state vector disturbance sample, the first optimization training information and the second optimization training information; Optimizing the neural network weights of the stacked long short-term memory network Stacked LSTM to be calibrated according to the network calibration quality evaluation function until the neural network calibration meets the standard requirements, thereby obtaining a calibrated stacked long short-term memory network Stacked LSTM; The determining of the network calibration quality evaluation function according to the calibration data knowledge vector sample, the deviation correction state vector authentication sample, the deviation correction state vector disturbance sample, the first optimization training information and the second optimization training information includes: Determining a first association weight between a mechanical sensing calibration device sample and a calibration deviation correction record authentication sample based on the calibration data knowledge vector sample and the deviation correction state vector authentication sample; Determining a second association weight between a mechanical sensing calibration device sample and a calibration deviation correction record disturbance sample based on the calibration data knowledge vector sample and the deviation correction state vector disturbance sample; Obtaining a first iterative training error according to the first association weight, the first optimization training information and a set iterative processing decision feature; Obtaining a second iterative training error according to the second association weight, the second optimization training information and the set iterative processing decision feature; A network calibration quality evaluation function is determined according to the first iterative training error and the second iterative training error.

4. The method according to claim 1, characterized in that The determining of the association weight between the target mechanical sensor calibration device and each calibration deviation correction record to be processed based on the calibration data knowledge vector and the deviation correction state vectors respectively corresponding to each calibration deviation correction record to be processed comprises: For each calibration deviation correction record to be processed, determine the characteristic operation result of the calibration data knowledge vector and the deviation correction state vector corresponding to the calibration deviation correction record to be processed; determine the association weight between the target mechanical sensor calibration device and the calibration deviation correction record to be processed based on the characteristic operation result.

5. The method according to claim 1, characterized in that The optimization of the attention semantics of the expected sensor performance features corresponding to each calibration deviation correction record to be processed, and obtaining the performance optimization output features of each expected sensor performance feature, include: For each calibration deviation correction record to be processed, weighting the expected sensor performance characteristics and the set enhanced characteristics corresponding to the calibration deviation correction record to be processed to obtain the enhanced expected sensor performance characteristics; The attention semantics of the enhanced expected sensor performance characteristics are processed through a channel attention mechanism to obtain a performance optimization output feature of the expected sensor performance characteristics corresponding to the calibration deviation correction record to be processed.

6. The method according to any one of claims 1 to 5, characterized in that: The feature integration of the association weights between each calibration deviation correction record to be processed and the target mechanical sensor calibration device and the performance optimization output features of the corresponding sensor performance expected features to obtain the screening priority indications corresponding to each calibration deviation correction record to be processed, includes: For each calibration deviation correction record to be processed, the association weight between the target mechanical sensor calibration device and the calibration deviation correction record to be processed is globally processed with the performance optimization output characteristics of the expected characteristics of the sensor performance corresponding to the calibration deviation correction record to be processed, so as to obtain a screening priority indication of the calibration deviation correction record to be processed.

7. The method according to claim 3, characterized in that The calibrated stacked long short-term memory network StackedLSTM includes multiple LSTM layers, which are vertically stacked on each other. The multiple LSTM layers achieve gated information transmission and gradient vanishing mitigation through a gating mechanism.

8. The method according to claim 7, characterized in that The obtaining of a priori training example set comprises: Obtain a calibration deviation correction record queue sample corresponding to a force sensing calibration device sample; The noise calibration deviation correction records in the calibration deviation correction record queue samples are filtered to obtain the deviation correction record denoising queue; the prior training example set is determined according to the deviation correction record denoising queue and the set augmentation strategy; The step of recording a denoising queue based on the deviation correction and setting an augmentation strategy to determine a priori training example set includes: For each bias correction record sample in the bias correction record denoising queue, determine the initial performance expected characteristics according to the sensor performance expected characteristics corresponding to the bias correction record sample and the set reinforcement coefficient; determine the prior training example set according to the bias correction record sample and the initial performance expected characteristics; or, Analyze the deviation correction patterns corresponding to each deviation correction record sample in the deviation correction record denoising queue; determine a target deviation correction pattern in which the number of deviation correction record samples included is less than a set number based on the deviation correction pattern analysis results; perform sample augmentation on the deviation correction record denoising queue based on the target deviation correction pattern; and determine a priori training example set based on the sample augmentation results.

9. The method according to claim 7, characterized in that The optimization of the attention semantics of the expected characteristics of the sensor performance corresponding to the calibration deviation correction record authentication sample is performed to obtain the first optimized training information; The attention semantics of the expected features of the sensor performance corresponding to the calibration deviation correction record disturbance sample are optimized to obtain the second optimized training information, including: Weighting the expected sensor performance characteristics and the set enhanced characteristics corresponding to the calibration deviation correction record authentication sample to obtain the first expected sensor performance characteristics; Processing the attention semantics of the first sensor performance expected feature through a channel attention mechanism to obtain first optimized training information; The expected sensor performance characteristics and the set enhanced characteristics corresponding to the calibration deviation correction record disturbance sample are weighted to obtain the second expected sensor performance characteristics; the attention semantics of the second expected sensor performance characteristics are processed through the channel attention mechanism to obtain the second optimized training information.

10. A calibration data analysis system, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images