Digital employee semantic interaction enhancement method based on large language model and medium

Abstract

The invention relates to the technical field of semantic processing, in particular to a digital employee semantic interaction enhancement method based on a large language model and a medium. According to valve opening data at different moments in a neighborhood range of each moment and flow measurement data distribution, flow measurement compensation data after delay compensation at each moment are obtained, and whether an actuator and a sensor are communicated or not is judged; if yes, according to the valve opening degree data at different moments in the local range of each moment and the flow measurement compensation data distribution after delay compensation, obtaining an operator adjustment feature label at each moment; if not, setting the operator adjustment feature tag at each moment as a fault mode; and on the basis of the operator adjustment feature tag and the actuator-sensor connection state, constructing a constraint prompt word, and inputting the constraint prompt word into the large language model to generate interaction feedback information. The physical consistency of semantic interaction is improved by accurately converting physical facts into semantic constraints.

Description

Technical Field

[0001] This invention relates to the field of semantic processing technology, specifically to a method and medium for enhancing semantic interaction among digital employees based on a large language model. Background Technology

[0002] In industrial fluid control scenarios, operators need to adjust actuators such as valves and pumps in real time according to process requirements. With the development of Large Language Modeling (LLM) technology, LLM-based digital employees are being introduced into industrial settings to assist operators in understanding operating conditions and providing decision-making suggestions.

[0003] In existing technologies, general-purpose large language models perform inference based on text probability when processing industrial time-series data, but lack prior knowledge of the dynamic characteristics of physical systems, especially the transmission delay of fluid pipelines and the response lag caused by fluid inertia. This leads to auxiliary system outputs of alarms or shutdown suggestions that do not match the actual situation on site, interfering with normal production scheduling and resulting in poorer effectiveness of semantic interaction. Summary of the Invention

[0004] To address the technical problem that the lack of prior knowledge about the dynamic latency characteristics of physical systems leads to poorer semantic interaction effectiveness, this invention aims to provide a method and medium for enhancing semantic interaction among digital employees based on a large language model. The specific technical solution adopted is as follows: This invention proposes a method for enhancing the semantic interaction of digital employees based on a large language model, the method comprising: Acquire valve opening data and flow measurement data at every moment in the fluid control scenario; Based on the valve opening data and flow measurement data distribution within the neighborhood of each moment, the flow measurement compensation data after delay compensation is obtained at each moment, and it is determined whether the actuator-sensor is connected; if connected, the regulation-response isotropic product at each moment is obtained based on the valve opening data and flow measurement compensation data distribution within the local range of each moment. Based on the integral distribution of regulation-response homogeneity at different times within the neighborhood of each time moment, the integral value of the trend-following regulation intensity and the counter-trend regulation reversal count at each time moment are obtained, and the trend-following intensity threshold and the counter-trend reversal threshold are obtained; based on the trend-following intensity threshold and the counter-trend reversal threshold, as well as the integral value of the trend-following regulation intensity and the counter-trend regulation reversal count at each time moment, the operator regulation feature label at each time moment is obtained. If the connection is not established, the operator adjustment feature label at each time step is set to the fault mode; constraint prompts are constructed based on the operator adjustment feature labels and the actuator-sensor connectivity status, and the constraint prompts are input into the large language model to generate interactive feedback information.

[0005] Furthermore, the method for obtaining the flow measurement compensation data includes: Based on the valve opening data at different times within the neighborhood of each time, determine whether it is in a quiescent steady state; If in a quiescent steady state, obtain the transmission delay time of the previous moment; if not in a quiescent steady state, obtain the cumulative value of the product between the valve opening data at different times in the neighborhood of each moment and the flow measurement data after any preset delay time, as the degree of cross-correlation; select the preset delay time with the highest degree of cross-correlation among all preset delay times and take the corresponding preset delay time as the transmission delay time. In the sequence of traffic measurement data at different times within the neighborhood of each time moment, the traffic measurement data after the transmission delay time at each time moment is obtained as the traffic measurement compensation data after delay compensation at each time moment.

[0006] Furthermore, the method for determining whether a state is in a silent steady state includes: Obtain the fluctuation level of valve opening data for all times within the neighborhood of each time. If the fluctuation level is greater than or equal to the preset quiescent threshold, it is determined that the valve is not in a quiescent steady state. If the fluctuation level is less than the preset silent threshold, it is determined that the system is in a silent steady state.

[0007] Furthermore, the method for determining whether the actuator and sensor are connected includes: If the system is in a static steady state, determine that the actuator and sensor are connected. If not in a quiescent steady state, obtain the correlation coefficient of the sequence formed by the valve opening data of all times in the neighborhood of each time and the corresponding flow measurement compensation data after delay compensation. If the correlation coefficient is greater than or equal to the preset connectivity threshold, determine that the actuator-sensor is connected; otherwise, determine that the actuator-sensor is not connected.

[0008] Furthermore, the method for obtaining the regulation-response simultaneity product includes: For each moment, the valve opening data or the flow measurement compensation data after delay compensation within a local range is used as the analysis data. First-order differential fitting is performed on the analysis data corresponding to different moments to obtain the data change rate. The product of the rate of change of the valve opening data and the flow measurement compensation data is obtained as the regulation-response unidirectional product.

[0009] Furthermore, the method for obtaining the integral value of the trend-following intensity and the reverse-trending reversal count includes: If the regulation-response product at any given moment is greater than a preset regulation-response product threshold, the corresponding moment is taken as the trend moment; the regulation-response product of each trend moment is obtained and multiplied by the time difference between each trend moment and the previous adjacent moment to obtain the first product; the cumulative value of the first product corresponding to all trend moments in the neighborhood of each moment is calculated as the trend regulation intensity integral value. If the regulation-response product is less than the preset regulation-response product threshold at a certain moment, and the rate of change of valve opening data between the corresponding moment and the previous moment is not equal, the condition value at the corresponding moment is set to a positive integer 1; otherwise, the condition value is set to 0. The condition values ​​of all moments in the neighborhood of each moment are accumulated and used as the counter-trend regulation reversal count.

[0010] Furthermore, the method for obtaining the trend strength threshold and the counter-trend reversal threshold includes: The mean intensity and standard deviation of the adaptive intensity integral value at all times within the neighborhood of each time are obtained. The product of the preset first confidence coefficient and the standard deviation of intensity is obtained. The sum of the product and the mean intensity is calculated as the adaptive intensity threshold. Obtain the reversal mean and reversal standard deviation of the counter-trend adjustment reversal count for all times in the neighborhood of each time step. Obtain the product of the preset second confidence coefficient and the reversal standard deviation. Calculate the sum of the product result and the reversal mean as the counter-trend reversal threshold.

[0011] Furthermore, the method for obtaining the operator-adjusted feature labels includes: If the integral value of the trend adjustment intensity at each moment is greater than the trend adjustment intensity threshold, and the counter-trend adjustment reversal count is less than or equal to the counter-trend reversal threshold, the operator adjustment feature label at the corresponding moment is set to expert aggressive mode. If the counter-trend adjustment reversal count at any given moment exceeds the counter-trend reversal threshold, the operator adjustment feature label for the corresponding moment will be set to Novice Panic Mode. If the integral value of the trend adjustment intensity at each moment is less than or equal to the trend adjustment intensity threshold, and the counter-trend adjustment reversal count is less than or equal to the counter-trend reversal threshold, the operator adjustment feature label at the corresponding moment is set to standard mode.

[0012] Furthermore, the correlation coefficient is obtained using the Pearson correlation coefficient.

[0013] The present invention also proposes a digital employee semantic interaction enhancement medium based on a large language model, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the steps of the digital employee semantic interaction enhancement method based on a large language model.

[0014] The present invention has the following beneficial effects: This invention obtains the delayed-compensated flow measurement data for each moment based on the valve opening data and flow measurement data distribution within the neighborhood of each moment. This eliminates timing misalignment caused by transmission delay and determines whether the actuator-sensor is connected, avoiding misjudging normal delayed responses as equipment disconnection or actuator failure due to data asynchrony. If connected, the invention obtains the regulation-response unidirectional product for each moment based on the valve opening data and the delayed-compensated flow measurement data distribution within the local area of ​​each moment. This helps identify forward regulation and reverse regulation, avoiding false alarms caused by parameter overshoot. Based on the regulation-response unidirectional product distribution at different moments, the invention obtains the operator regulation feature label for each moment, achieving a physical distinction between expert aggressive operation and novice panic oscillation. If disconnected, the operator regulation feature label for each moment is set to a fault mode. Constraint prompts are constructed based on the operator regulation feature labels and the actuator-sensor connectivity status. These constraint prompts are input into a large language model to generate interactive feedback information, enhancing the physical interpretability of the large model's interactive content. This invention improves the physical consistency of semantic interaction by accurately transforming physical facts into semantic constraints. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for enhancing the semantic interaction of digital employees based on a large language model, as provided in one embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for acquiring flow measurement compensation data according to an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a digital employee semantic interaction enhancement method and medium based on a large language model proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for a digital employee semantic interaction enhancement method and medium based on a large language model provided by this invention.

[0020] Please see Figure 1 The diagram illustrates a flowchart of a method for enhancing the semantic interaction of digital employees based on a large language model, according to an embodiment of the present invention. The method specifically includes: Step S1: Obtain valve opening data and flow measurement data for the fluid control scenario at each moment.

[0021] In the embodiments of the present invention, considering that when skilled operators make large-scale aggressive adjustments in order to quickly switch working conditions, or novice operators make frequent oscillating corrections due to panic, the model often misjudges the non-steady-state fluctuations caused by human intent as equipment failures based only on the instantaneous deviation of the values, thus interfering with normal production scheduling; firstly, valve controllers and flow meters are installed in the industrial fluid control scenario to obtain valve opening data and flow measurement data at each moment.

[0022] It should be noted that, considering the different physical dimensions and significant differences in numerical magnitude of the data, direct calculation would lead to an imbalance in subsequent feature weights; therefore, it is necessary to perform maximum and minimum value normalization on the data. Within the dimensionless range, it accurately reflects the proportion of the equipment's action amplitude relative to its physical limits, eliminating the influence between dimensions. That is, the valve opening range boundary and flow measurement range boundary preset in the system are selected and normalized. The specific methods are well known to those skilled in the art and will not be described in detail here.

[0023] It should be noted that, in one embodiment of the present invention, the time interval is determined according to the system's preset sampling frequency, such as 1 second corresponding to one time moment for data acquisition; in other embodiments of the present invention, the time interval can be set according to specific circumstances, and will not be limited or elaborated here.

[0024] Step S2: Based on the valve opening data and flow measurement data distribution within the neighborhood of each moment, obtain the flow measurement compensation data after delay compensation at each moment, and determine whether the actuator-sensor is connected; if connected, obtain the regulation-response unidirectional product at each moment based on the valve opening data and flow measurement compensation data distribution within the local range of each moment.

[0025] The valve is active, while the flow rate is passive. The response is retrieved based on the time the control command occurs. To ensure both perform mathematical calculations at the same point in time, the time difference must be eliminated. Based on the valve opening data and flow rate measurement data distribution within the neighborhood of each moment, the delayed flow rate measurement compensation data for each moment is obtained.

[0026] Preferably, in one embodiment of the present invention, the method for obtaining flow measurement compensation data is described in [reference needed]. Figure 2 It illustrates a flowchart of a method for acquiring flow measurement compensation data, including: Step S201: Determine whether the valve is in a quiescent steady state based on the valve opening data at different times within the neighborhood of each time.

[0027] Preferably, in one embodiment of the present invention, the method for determining whether a state is in a quiescent steady state includes: Obtain the fluctuation level of valve opening data for all times within the neighborhood of each time. If the fluctuation level is greater than or equal to the preset quiescent threshold, it is determined that the valve is not in a quiescent steady state. If the fluctuation level is less than the preset silent threshold, it is determined that the system is in a silent steady state.

[0028] It should be noted that, in one embodiment of the present invention, the degree of fluctuation is represented by calculating the standard deviation of the valve opening data of all times in the neighborhood of each time. The larger the standard deviation, the greater the degree of fluctuation, and the smaller the standard deviation, the smaller the degree of fluctuation. In other embodiments of the present invention, variance can also be used to represent the degree of fluctuation. The specific means are well known to those skilled in the art and will not be described in detail here.

[0029] It should be noted that, in one embodiment of the present invention, the size of the neighborhood range is a range consisting of a number of historical moments based on each moment, wherein the number ranges from 10 to 20; the preset silence threshold can be set to any value in the range of 0.01 to 0.05 based on relevant historical experience; in other embodiments of the present invention, the size of the neighborhood range and the preset silence threshold can be set according to specific circumstances, and will not be limited or elaborated here.

[0030] Step S202: If in a quiescent steady state, obtain the transmission delay time of the previous moment; if not in a quiescent steady state, obtain the cumulative value of the product between the valve opening data at different times in the neighborhood of each moment and the flow measurement data after any preset delay time, as the degree of cross-correlation; select the preset delay time with the largest cross-correlation among all preset delay times and take the corresponding preset delay time as the transmission delay time.

[0031] It should be noted that the system is assigned an initial transmission delay time of 0 at startup. Subsequently, the valve opening data in the neighborhood range of each time moment is analyzed sequentially. If it is in a quiescent steady state, the original transmission delay time is maintained; if it is not in a quiescent steady state, the cross-correlation degree is analyzed to obtain a new transmission delay time. The subsequent time moments are analyzed continuously to obtain the transmission delay time of the current moment.

[0032] It should be noted that, in the embodiments of the present invention, the allowed delay time range is preset based on relevant historical experience. ,in, This indicates the maximum delay time set by the system, such as 5 times. Each of these values ​​is taken as the preset delay time for analysis.

[0033] Step S203: In the sequence of all traffic measurement data corresponding to each moment in the neighborhood, obtain the traffic measurement data after the transmission delay time after each moment, and use it as the traffic measurement compensation data after delay compensation at each moment.

[0034] It should be noted that, taking an example, the preset neighborhood range includes 10 time points. For the traffic measurement data within the neighborhood range at each time point, 0.2, 0.4, 0.3, 0.5, 0.6, 0.45, 0.35, 0.25, 0.1, 0.7, if the transmission delay time is 3, then the first time point corresponds to 0.5 after delay compensation, the second time point corresponds to 0.6 after delay compensation, the third time point corresponds to 0.45 after delay compensation, and so on, until the end of the sequence. Only the acquired data is analyzed; that is, the sequence of traffic measurement compensation data within the neighborhood range at each time point after delay compensation is 0.5, 0.6, 0.45, 0.35, 0.25, 0.1, 0.7.

[0035] In actual working conditions, there may be hard faults such as sensor damage, actuator linkage breakage, or loss of air supply. Under fault conditions, valve commands cannot trigger the expected flow changes. By analyzing the changes in valve opening data and flow measurement compensation data at different times, it can be determined whether the actuator and sensor are connected.

[0036] Preferably, in one embodiment of the present invention, the method for determining whether the actuator and sensor are connected includes: If the system is in a static steady state, determine that the actuator and sensor are connected. If not in a quiescent steady state, obtain the correlation coefficient of the sequence formed by the valve opening data of all times in the neighborhood of each time and the corresponding flow measurement compensation data after delay compensation. If the correlation coefficient is greater than or equal to the preset connectivity threshold, determine that the actuator-sensor is connected; otherwise, determine that the actuator-sensor is not connected.

[0037] Based on this, the larger the correlation coefficient, the greater the consistency between the changes in valve opening data and flow measurement compensation data. The more effectively the actuator and sensor are correlated, the more normal the system loop is, and the more interconnected they are.

[0038] It should be noted that, in the embodiments of the present invention, the correlation coefficient is obtained by the Pearson correlation coefficient. The larger the Pearson correlation coefficient, the more similar the changes of the two sets of sequences are, and the larger the correlation coefficient is. The specific means are well known to those skilled in the art and will not be described in detail here.

[0039] It should be noted that, in one embodiment of the present invention, the preset connectivity threshold is 0.6; in other embodiments of the present invention, the preset connectivity threshold can be set according to specific circumstances, and will not be limited or elaborated here.

[0040] Skilled operators tend to apply energy in the same direction to quickly switch operating conditions, in line with system inertia. Novice operators, however, tend to apply reverse braking and make repeated corrections when system inertia is too high. Analyzing valve opening data at different times and the distribution of flow measurement compensation data after delay compensation can reflect data changes. Analyzing the valve adjustment direction and flow change direction reveals that if both are positive, it indicates an increase in valve opening and flow, meaning the operator is inputting control energy into the system in line with the current inertial direction. If the flow change direction is opposite to the valve adjustment direction, it indicates the operator is consuming operational energy to curb the system's current inertial trend. Quantifying the regulation-response unidirectional product helps to understand the trends in valve adjustment and flow change directions. Based on the distribution of valve opening data and flow measurement compensation data at different times within a local area at each moment, the regulation-response unidirectional product at each moment can be obtained.

[0041] Preferably, in one embodiment of the present invention, the method for obtaining the regulation-response simultaneity product includes: For each moment, the valve opening data or the flow measurement compensation data after delay compensation within a local range is used as the analysis data. First-order differential fitting is performed on the analysis data corresponding to different moments to obtain the data change rate. The product of the rate of change of the valve opening data and the flow measurement compensation data is obtained as the regulation-response unidirectional product.

[0042] It should be noted that the Savitzky-Golay digital filter is used to perform convolution calculations on the data to achieve first-order differential fitting, which reduces the influence of noise and helps to more accurately reflect the real operation trend. The specific methods are well known to those skilled in the art and will not be elaborated here.

[0043] It should be noted that, in one embodiment of the present invention, the size of the local range is a range consisting of a number of historical moments based on each moment, wherein the number ranges from 5 to 7; in other embodiments of the present invention, the size of the local range can be specifically set according to the specific circumstances, which will not be elaborated here.

[0044] Step S3: Based on the integral distribution of regulation-response homogeneity at different times within the neighborhood of each time moment, obtain the integral value of the trend-following regulation intensity and the counter-trend regulation reversal count at each time moment, and obtain the trend-following intensity threshold and the counter-trend reversal threshold; based on the trend-following intensity threshold and the counter-trend reversal threshold, as well as the integral value of the trend-following regulation intensity and the counter-trend regulation reversal count at each time moment, obtain the operator regulation feature label at each time moment.

[0045] The regulation-response unidirectional product reflects the integral distribution of whether the valve regulation direction and the flow response direction are consistent. The more positive the regulation-response unidirectional product, the greater the energy invested in accelerating the change of the driving fluid, and the greater the integral value of the unidirectional regulation intensity. The more negative the regulation-response unidirectional product, the smaller the value, indicating hesitant, repetitive, or oscillating operation behavior when counteracting fluid inertia, and the greater the reverse count of the reverse regulation. Based on the integral distribution of regulation-response unidirectional products in the neighborhood of each moment, the integral value of the unidirectional regulation intensity and the reverse regulation count of each moment are obtained, and the unidirectional intensity threshold and the reverse reversal threshold are obtained.

[0046] Preferably, in one embodiment of the present invention, the method for obtaining the integral value of the trend-following intensity and the reverse count of the counter-trend-following reversal includes: If the regulation-response product at any given moment is greater than a preset regulation-response product threshold, the corresponding moment is taken as the trend moment; the regulation-response product of each trend moment is obtained and multiplied by the time difference between each trend moment and the previous adjacent moment to obtain the first product; the cumulative value of the first product corresponding to all trend moments in the neighborhood of each moment is calculated as the trend regulation intensity integral value. If the regulation-response product is less than the preset regulation-response product threshold at a certain moment, and the rate of change of valve opening data between the corresponding moment and the previous moment is not equal, the condition value at the corresponding moment is set to a positive integer 1; otherwise, the condition value is set to 0. The condition values ​​of all moments in the neighborhood of each moment are accumulated and used as the counter-trend regulation reversal count.

[0047] It should be noted that, in the embodiments of the present invention, the time difference between each trend moment and the previous adjacent moment represents the value of each trend moment minus the value of the previous adjacent moment; when the trend direction is the same, the adjustment-response unidirectional product is a positive number; when the trend direction is opposite, the adjustment-response unidirectional product is a negative number. In order to distinguish between trend and trend for analysis, the preset unidirectional product threshold is set to 0.

[0048] Preferably, in one embodiment of the present invention, the method for obtaining the trend strength threshold and the counter-trend reversal threshold includes: The mean intensity and standard deviation of the adaptive intensity integral value at all times within the neighborhood of each time are obtained. The product of the preset first confidence coefficient and the standard deviation of intensity is obtained. The sum of the product and the mean intensity is calculated as the adaptive intensity threshold. Obtain the reversal mean and reversal standard deviation of the counter-trend adjustment reversal count for all times in the neighborhood of each time step. Obtain the product of the preset second confidence coefficient and the reversal standard deviation. Calculate the sum of the product result and the reversal mean as the counter-trend reversal threshold.

[0049] It should be noted that, considering the parameter setting based on the statistical principle of three standard deviations in the normal distribution, the threshold can dynamically adapt to the process characteristics of different industrial sites, avoiding misjudgments caused by a fixed threshold; in the embodiments of the present invention, considering that high-energy behavior that significantly deviates from the conventional statistical law under the condition of following the trend will be considered to be an aggressive state, a higher threshold is required for differentiation, and the range of the first confidence coefficient is preset based on relevant historical experience. Any value in the range; reversals are rare in normal operation, so even a slight abnormal reversal requires further analysis, and the range of the second confidence coefficient should be preset based on relevant historical experience. Any value in the range; the specific means are technical means well known to those skilled in the art, and will not be elaborated here.

[0050] The integral value of trend-following adjustment strength is a typical characteristic of expert aggressive trading, while the reversal count of counter-trend adjustment is a typical characteristic of novice panic trading. By analyzing the changes in the integral value of trend-following adjustment strength and the reversal count of counter-trend adjustment relative to the corresponding thresholds at each time point, we can better understand the significance of these characteristics and help to accurately determine the operator's adjustment characteristic label. Based on the trend-following strength threshold and the reversal threshold, as well as the integral value of trend-following adjustment strength and the reversal count of counter-trend adjustment at each time point, we can obtain the operator's adjustment characteristic label at each time point.

[0051] Preferably, in one embodiment of the present invention, the operator adjustment feature tag acquisition method includes: If the integral value of the trend adjustment intensity at each moment is greater than the trend adjustment intensity threshold, and the counter-trend adjustment reversal count is less than or equal to the counter-trend reversal threshold, the operator adjustment feature label at the corresponding moment is set to expert aggressive mode. If the counter-trend adjustment reversal count at any given moment exceeds the counter-trend reversal threshold, the operator adjustment feature label for the corresponding moment will be set to Novice Panic Mode. If the integral value of the trend adjustment intensity at each moment is less than or equal to the trend adjustment intensity threshold, and the counter-trend adjustment reversal count is less than or equal to the counter-trend reversal threshold, the operator adjustment feature label at the corresponding moment is set to standard mode.

[0052] Based on this, the expert aggressive mode reflects a high-energy-driven, low-frequency correction operation mode, in which the operator makes large adjustments using system inertia and the operation is stable during the braking phase; the novice panic mode reflects a frequent change of direction operation mode, in which high-frequency reversals exceeding the statistical normal range occur during the braking phase; the standard mode reflects normal stable operation, fine-tuning, or the system being in a steady state, with operation data within the normal fluctuation range.

[0053] Step S4: If the connection is not established, set the operator adjustment feature label at each time step to the fault mode; construct constraint prompt words based on the operator adjustment feature label and the actuator-sensor connectivity status, input the constraint prompt words into the large language model, and generate interactive feedback information.

[0054] Leveraging the powerful natural language generation capabilities of large language models, the physical data calculated by the preceding modules can be transformed into natural language feedback, which helps to solve the semantic offset problem in industrial scenarios.

[0055] It should be noted that the constraint template corresponding to the operator adjustment feature label is called from the preset structured template library, and the integral value of the follow-up adjustment intensity or the reverse count of the counter-current adjustment is filled into the preset data slot of the constraint template to generate scene constraint prompt words containing physical evidence data. The system acquires the user's query text, appends scenario constraint prompts to it to form a reasoning context, and inputs this context into a pre-trained large language model. This allows the large language model to generate feedback text containing physical evidence data under the logical boundary constraints of the scenario constraint prompts. For example, in expert aggressive mode, if the system detects the current moment's integral value of the adaptive adjustment intensity, confirming it matches a normal transient response under high-energy initiation, it provides positive feedback, affirming the operation's effectiveness and confirming it's not a device malfunction. In oscillation mode, if the system detects the number of reverse adjustments performed in a short period, indicating the current flow fluctuation is operational oscillation rather than a device malfunction, it suggests pausing operations for 5 seconds to allow fluid inertia to stabilize before fine-tuning.

[0056] Based on this, the present invention realizes the lossless transfer of physical calculation results to the semantic space, ensuring that the responses of digital employees are strictly anchored to physical facts, and achieving enhanced physical consistency of semantic interaction under industrial non-steady-state conditions.

[0057] In summary, this invention obtains the delayed-compensated flow measurement data for each moment based on the valve opening data and flow measurement data distribution within the neighborhood of each moment, and determines whether the actuator-sensor is connected. If connected, it obtains the operator adjustment feature label for each moment based on the valve opening data and the delayed-compensated flow measurement data distribution within the local area of ​​each moment. If disconnected, it sets the operator adjustment feature label for each moment to a fault mode. Constraint prompts are constructed based on the operator adjustment feature labels and the actuator-sensor connectivity status. These constraint prompts are then input into a large language model to generate interactive feedback information. This invention improves the physical consistency of semantic interaction by accurately transforming physical facts into semantic constraints.

[0058] This invention also proposes a digital employee semantic interaction enhancement medium based on a large language model, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any one of the steps of a digital employee semantic interaction enhancement method based on a large language model.

[0059] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0060] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for enhancing the semantic interaction of digital employees based on a large language model, characterized in that, The method includes: Acquire valve opening data and flow measurement data at every moment in the fluid control scenario; Based on the valve opening data and flow measurement data distribution within the neighborhood of each moment, the flow measurement compensation data after delay compensation is obtained at each moment, and it is determined whether the actuator-sensor is connected; if connected, the regulation-response isotropic product at each moment is obtained based on the valve opening data and flow measurement compensation data distribution within the local range of each moment. Based on the integral distribution of regulation-response homogeneity at different times within the neighborhood of each time moment, the integral value of the trend-following regulation intensity and the counter-trend regulation reversal count at each time moment are obtained, and the trend-following intensity threshold and the counter-trend reversal threshold are obtained; based on the trend-following intensity threshold and the counter-trend reversal threshold, as well as the integral value of the trend-following regulation intensity and the counter-trend regulation reversal count at each time moment, the operator regulation feature label at each time moment is obtained. If the connection is not established, the operator adjustment feature label at each time step is set to the fault mode; constraint prompts are constructed based on the operator adjustment feature labels and the actuator-sensor connectivity status, and the constraint prompts are input into the large language model to generate interactive feedback information.

2. The method for enhancing the semantic interaction of digital employees based on a large language model according to claim 1, characterized in that, The method for obtaining the flow measurement compensation data includes: Based on the valve opening data at different times within the neighborhood of each time, determine whether it is in a quiescent steady state; If in a quiescent steady state, obtain the transmission delay time of the previous moment; if not in a quiescent steady state, obtain the cumulative value of the product between the valve opening data at different times in the neighborhood of each moment and the flow measurement data after any preset delay time, as the degree of cross-correlation; select the preset delay time with the highest degree of cross-correlation among all preset delay times and take the corresponding preset delay time as the transmission delay time. Within the neighborhood of each time step, the traffic measurement data after the transmission delay time is obtained from the sequence of all traffic measurement data, and used as the traffic measurement compensation data after delay compensation at each time step.

3. The method for enhancing the semantic interaction of digital employees based on a large language model according to claim 2, characterized in that, The method for determining whether a state is in a silent steady state includes: Obtain the fluctuation level of valve opening data for all times within the neighborhood of each time. If the fluctuation level is greater than or equal to the preset quiescent threshold, it is determined that the valve is not in a quiescent steady state. If the fluctuation level is less than the preset silent threshold, it is determined that the system is in a silent steady state.

4. The method for enhancing the semantic interaction of digital employees based on a large language model according to claim 3, characterized in that, The method for determining whether the actuator and sensor are connected includes: If the system is in a static steady state, determine that the actuator and sensor are connected. If not in a quiescent steady state, obtain the correlation coefficient of the sequence formed by the valve opening data of all times in the neighborhood of each time and the corresponding flow measurement compensation data after delay compensation. If the correlation coefficient is greater than or equal to the preset connectivity threshold, determine that the actuator-sensor is connected; otherwise, determine that the actuator-sensor is not connected.

5. The method for enhancing the semantic interaction of digital employees based on a large language model according to claim 1, characterized in that, The method for obtaining the regulation-response simultaneous product includes: For each moment, the valve opening data or the flow measurement compensation data after delay compensation within a local range is used as the analysis data. First-order differential fitting is performed on the analysis data corresponding to different moments to obtain the data change rate. The product of the rate of change of the valve opening data and the flow measurement compensation data is obtained as the regulation-response unidirectional product.

6. The method for enhancing the semantic interaction of digital employees based on a large language model according to claim 5, characterized in that, The methods for obtaining the integral value of the trend-following intensity and the reverse-trending reversal count include: If the regulation-response product at any given moment is greater than a preset regulation-response product threshold, the corresponding moment is taken as the trend moment; the regulation-response product of each trend moment is obtained and multiplied by the time difference between each trend moment and the previous adjacent moment to obtain the first product; the cumulative value of the first product corresponding to all trend moments in the neighborhood of each moment is calculated as the trend regulation intensity integral value. If the regulation-response product is less than the preset regulation-response product threshold at a certain moment, and the rate of change of valve opening data between the corresponding moment and the previous moment is not equal, the condition value at the corresponding moment is set to a positive integer 1; otherwise, the condition value is set to 0. The condition values ​​of all moments in the neighborhood of each moment are accumulated and used as the counter-trend regulation reversal count.

7. The method for enhancing the semantic interaction of digital employees based on a large language model according to claim 1, characterized in that, The methods for obtaining the trend strength threshold and the counter-trend reversal threshold include: The mean intensity and standard deviation of the adaptive intensity integral value at all times within the neighborhood of each time are obtained. The product of the preset first confidence coefficient and the standard deviation of intensity is obtained. The sum of the product and the mean intensity is calculated as the adaptive intensity threshold. Obtain the reversal mean and reversal standard deviation of the counter-trend adjustment reversal count for all times in the neighborhood of each time step. Obtain the product of the preset second confidence coefficient and the reversal standard deviation. Calculate the sum of the product result and the reversal mean as the counter-trend reversal threshold.

8. The method for enhancing the semantic interaction of digital employees based on a large language model according to claim 1, characterized in that, The method for obtaining the operator-adjusted feature labels includes: If the integral value of the trend adjustment intensity at each moment is greater than the trend adjustment intensity threshold, and the counter-trend adjustment reversal count is less than or equal to the counter-trend reversal threshold, the operator adjustment feature label at the corresponding moment is set to expert aggressive mode. If the counter-trend adjustment reversal count at any given moment exceeds the counter-trend reversal threshold, the operator adjustment feature label for the corresponding moment will be set to Novice Panic Mode. If the integral value of the trend adjustment intensity at each moment is less than or equal to the trend adjustment intensity threshold, and the counter-trend adjustment reversal count is less than or equal to the counter-trend reversal threshold, the operator adjustment feature label at the corresponding moment is set to standard mode.

9. The method for enhancing the semantic interaction of digital employees based on a large language model according to claim 4, characterized in that, The correlation coefficient is obtained using the Pearson correlation coefficient.

10. A digital employee semantic interaction enhancement medium based on a large language model, the medium comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the digital employee semantic interaction enhancement method based on any one of claims 1 to 9.

Images