Intelligent office table parameter personalization setting method and system based on knowledge graph

Abstract

The application discloses a knowledge graph-based intelligent office table parameter individualization setting method and system, relates to the field of ergonomics intelligent control technology, and comprises the following steps: collecting operation signals generated by user operation of a man-machine interaction device and extracting multi-mode operation features; obtaining a user current working state with state confidence based on multi-mode operation feature fusion classification; combining the working state with a pre-constructed knowledge graph containing a field general knowledge layer and an individual individualization layer to carry out parameter fusion decision, obtaining individualized desktop control parameters of the intelligent office table; and controlling the intelligent office table to complete desktop height adjustment according to the individualized desktop control parameters. The application solves the technical problems that the existing intelligent lifting office table is difficult to realize parameter self-adaptive configuration according to real-time working behavior of a user, the user adaptability is poor, new users lack effective individualized parameters, and the control strategy is prone to misadjustment.

Description

Technical Field

[0001] This invention relates to the field of ergonomic intelligent control technology, specifically to a method and system for personalized parameter setting of intelligent office desks based on knowledge graphs. Background Technology

[0002] Height-adjustable smart desks enable users to switch between sitting and standing postures, effectively alleviating physical strain caused by prolonged fixed postures while working, and are increasingly widely used in modern office settings. Currently, most smart desks on the market are controlled by fixed preset parameters or general ergonomic strategies, requiring users to manually set sitting and standing heights and the sitting-to-standing switching cycle, or the device triggers height adjustment according to fixed time thresholds.

[0003] However, existing control models have significant limitations in actual office use. Users' work behaviors and task types are dynamically changing. Fixed-time triggers or methods relying entirely on manual settings struggle to match users' real-time work rhythms, easily leading to inappropriate adjustment timing. This can interrupt users' focused work or fail to respond promptly to their needs when they are fatigued. Furthermore, different users have significantly different physical conditions and usage habits; standardized parameters only fit the average user and cannot meet the personalized needs of individuals.

[0004] In addition, existing products have many practical problems: new users have no historical usage data when using the device for the first time, and can only directly apply general parameters, which presents a problem of cold start for personalized parameters; the device mostly only records the user's manual adjustment operations, making it difficult to find the true usage intention behind the operation; when there is a deviation in the judgment of the user's work scenario, unreasonable misadjustment is likely to occur, and there is a lack of complete safety constraints and smooth adjustment mechanisms, so the overall user experience needs to be further improved. Summary of the Invention

[0005] This invention provides a knowledge graph-based method and system for personalized parameter setting of smart office desks, addressing the technical problems of existing smart height-adjustable office desks, such as difficulty in adapting parameters to real-time user work behavior, poor user adaptability, lack of effective personalized parameters for new users, and the tendency for control strategies to be mis-adjusted.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] In a first aspect, the present invention provides a method for personalized parameter setting of a smart office desk based on a knowledge graph, including:

[0008] Collect operation signals generated by users operating human-computer interaction devices and extract multi-mode operation features;

[0009] Based on multimodal operational features, a fusion classification is performed to determine the user's current work status, where the work status is associated with a state confidence level.

[0010] By combining work status with a pre-built knowledge graph, parameter fusion decision-making is performed to obtain personalized desktop control parameters for smart desks. The knowledge graph includes a domain-general knowledge layer and an individual personalized layer.

[0011] Based on personalized desktop control parameters, the smart office desk can adjust its desktop height.

[0012] Secondly, this invention provides a knowledge graph-based intelligent office desk parameter personalization setting system, including:

[0013] The operation feature extraction module is used to collect operation signals generated by users operating human-computer interaction devices and extract multi-mode operation features;

[0014] The work status determination module is used to perform fusion classification based on multi-modal operation features to determine the user's current work status, wherein the work status is associated with a state confidence level;

[0015] The parameter fusion decision module is used to combine the working status with the pre-built knowledge graph to perform parameter fusion decision-making and obtain personalized desktop control parameters for the smart office desk. The knowledge graph includes a domain-general knowledge layer and an individual personalized layer.

[0016] The desktop height control module is used to control the smart office desk to adjust its desktop height according to personalized desktop control parameters.

[0017] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0018] By sensing users' real-time work behavior through multi-modal operational features, and introducing a knowledge graph with a two-layer structure to complete parameter fusion decision-making, it takes into account both general ergonomic knowledge and individual user preferences. Through state confidence level-based adjustment of parameter range, it reduces the risk of misadjustment caused by recognition uncertainty, while enabling personalized parameter cold start in new user scenarios and continuously iterating individual preference knowledge based on user manual intervention. It breaks away from the limitations of traditional desks with fixed parameters or complete reliance on manual settings, achieving dynamic adaptive output of desktop height parameters, improving adaptability to different users; and by setting ergonomic safety constraints and smooth transition adjustments, it ensures the safety and stability of the desk adjustment process, improving the actual user experience of the smart desk. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process for personalizing the parameters of a smart office desk based on a knowledge graph, provided in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of three-layer parameter fusion provided in an embodiment of the present invention;

[0021] Figure 3 This is a schematic diagram of the structure of a knowledge graph-based intelligent office desk parameter personalization setting system provided in an embodiment of the present invention.

[0022] In the attached diagram, the components represented by each number are as follows:

[0023] Operation feature extraction module 11, working status determination module 12, parameter fusion decision module 13, desktop height control module 14. Detailed Implementation

[0024] This invention provides a method and system for personalized parameter setting of intelligent office desks based on knowledge graphs, which specifically solves the technical problems of existing intelligent height-adjustable office desks, such as difficulty in achieving adaptive parameter configuration based on real-time user work behavior, poor user adaptability, lack of effective personalized parameters for new users, and easy misadjustment of control strategies.

[0025] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0026] Example 1, as Figure 1 As shown, this invention provides a method for personalized parameter setting of a smart office desk based on a knowledge graph, including:

[0027] S1: Collect the operation signals generated by the user's operation of the human-computer interaction device and extract multi-mode operation features;

[0028] Step S1 in the method provided in this embodiment of the invention includes:

[0029] S1.1: Collect human-computer interaction device signals generated by the user through the keyboard, and extract keyboard operation features, wherein the keyboard operation features include at least two of the following: keystroke frequency, keystroke duration, keystroke interval, and input continuity ratio.

[0030] In this embodiment, the user's keyboard operation signals are obtained through the human-machine interface device event interface of the computer terminal. Each keyboard operation signal includes at least a key value identifier, a key press timestamp, and a key release timestamp. For the same valid key operation, the key release timestamp is subtracted from the key press timestamp to obtain the key duration of that key operation.

[0031] Furthermore, a preset keyboard behavior statistics window is set up. The number of valid keystrokes completed by the user is counted within this window, and the number of valid keystrokes is divided by the duration of the preset keyboard behavior statistics window to obtain the keystroke frequency for that specific window. That is:

[0032] Keystroke frequency = Number of valid keystrokes in the statistics window ÷ Duration of the statistics window.

[0033] The duration of the preset keyboard behavior statistics window is determined based on the rate of change in the user's work status, the frequency of keyboard events, and the real-time requirements for work status recognition. In one specific embodiment of this example, the duration of the statistics window can be 60 seconds.

[0034] For two consecutive valid keystrokes, the keystroke interval between the two keystrokes is obtained by subtracting the keystroke release time of the previous valid keystroke from the keystroke press time of the subsequent valid keystroke.

[0035] Furthermore, within the current keyboard behavior statistics window, all adjacent keystroke intervals are arranged in ascending order, and the keystroke interval in the middle position is taken as the representative keystroke interval of the current statistics window; when the number of keystroke intervals is even, the average of the two middle keystroke intervals is taken as the representative keystroke interval.

[0036] Furthermore, to determine whether the user is in a continuous input state, this embodiment determines a continuous input interval threshold based on the user's own historical keyboard operation data.

[0037] Specifically, during the initial use of the smart desk, valid keyboard operation data from a preset number of keyboard behavior statistics windows are continuously collected. Representative keystroke intervals for each statistics window are obtained, and the average and standard deviation of these representative keystroke intervals are calculated. The average of these representative keystroke intervals is added to the standard deviation by a preset multiple to obtain the user's corresponding continuous input interval threshold. That is:

[0038] Continuous input interval threshold = average value of representative keystroke intervals in user history + preset multiple × standard deviation of representative keystroke intervals in user history.

[0039] The preset multiplier is determined based on the dispersion of historical keyboard behavior data and the requirement to distinguish continuous input behavior. In this embodiment, the preset multiplier is preferably 2.

[0040] Within the current keyboard behavior statistics window, when the interval between adjacent keystrokes is less than or equal to the continuous input interval threshold, the corresponding adjacent keystroke pair is identified as a continuous input keystroke pair; when the interval between adjacent keystrokes is greater than the continuous input interval threshold, the corresponding adjacent keystroke pair is identified as a non-continuous input keystroke pair.

[0041] Dividing the number of consecutive keystroke pairs by the total number of all adjacent keystroke pairs yields the input continuity ratio, i.e.:

[0042] Input continuity ratio = Number of consecutive keystroke pairs ÷ Number of all adjacent keystroke pairs.

[0043] For example, if 240 valid keystrokes are detected within a specific statistical window, forming 239 pairs of adjacent keystrokes, of which 198 pairs meet the continuous input condition, then the keystroke frequency is calculated based on the duration of the statistical window, and the input continuity ratio = 198 ÷ 239 ≈ 82.85%.

[0044] S1.2: Collect human-computer interaction device signals generated by the user through the mouse, and extract mouse operation features, wherein the mouse operation features include at least two of the following: cursor movement speed, click frequency, cursor still time ratio, and scrolling frequency;

[0045] In this embodiment, mouse movement events, mouse button events, and scroll wheel events are obtained through the human-machine interface device event interface of the computer terminal. For two adjacent valid mouse movement events, the cursor positions and timestamps corresponding to the two events are obtained, and the cursor movement speed is determined based on the distance between the two cursor positions and the time interval between the two mouse movement events. Wherein:

[0046] Cursor movement distance = ;

[0047] Cursor movement speed = cursor movement distance ÷ time interval between adjacent mouse movement events.

[0048] Within the current mouse behavior statistics window, all valid mouse movement events are arranged in ascending order of cursor movement speed, and the cursor movement speed in the middle position is taken as the representative cursor movement speed of the current statistics window.

[0049] Furthermore, the number of valid mouse clicks is counted in the mouse behavior statistics window, and the number of valid mouse clicks is divided by the duration of the statistics window to obtain the click frequency, i.e.:

[0050] Click frequency = Number of valid mouse clicks in the statistics window ÷ Duration of the statistics window.

[0051] Furthermore, in order to identify cursor stationary behavior during user mouse operation, this embodiment determines the cursor stationary period based on the cursor position change and time interval between adjacent mouse movement events.

[0052] When the cursor movement distance between adjacent mouse movement events is less than a preset stationary distance threshold, and the time interval between two mouse movement events is greater than a preset stationary time threshold, the corresponding time period is determined as the cursor stationary period.

[0053] The stationary distance threshold and stationary time threshold are determined based on the mouse event acquisition accuracy, cursor movement characteristics, and user's historical mouse operation data. In one specific embodiment of this example, the stationary distance threshold can be 2 screen pixels, and the stationary time threshold can be 500 ms.

[0054] The duration of all cursor still periods within the current mouse behavior statistics window is summed up, and the total cursor still duration is divided by the duration of the statistics window to obtain the cursor still duration ratio, i.e.:

[0055] Cursor stillness duration ratio = Sum of all cursor stillness durations within the statistics window ÷ Statistics window duration.

[0056] For example, if five cursor stillness periods are detected in a statistical window, with durations of 1.2s, 0.8s, 2.1s, 1.5s and 0.9s respectively, then the total cursor stillness duration is 6.5s; if the duration of the statistical window is 60s, then the cursor stillness duration ratio = 6.5 ÷ 60 ≈ 10.83%.

[0057] Furthermore, the number of valid scroll wheel operations within the current mouse behavior statistics window is counted, and the number of valid scroll wheel operations is divided by the duration of the statistics window to obtain the scrolling frequency, i.e.:

[0058] Scrolling frequency = Number of valid scroll wheel operations within the statistics window ÷ Duration of the statistics window.

[0059] S1.3: Collect the application window context signal corresponding to the currently active application in the foreground, and extract the application context features, wherein the application context features include at least the application category and the application switching frequency;

[0060] In this embodiment, the application identifier, window identifier, window focus time, and window focus loss time that are currently in the foreground state are obtained through the foreground window recognition interface provided by the operating system, and the current foreground application is mapped to a preset application category according to the application identifier.

[0061] In this embodiment, the preset application categories include office editing, information browsing, program development, communication interaction, meeting collaboration, and other application categories.

[0062] Within the current application context statistics window, when the application category of the current foreground application is different from the application category of the previous valid foreground application, this change is identified as an application category switching event.

[0063] Divide the number of application category switching events by the duration of the application context statistics window to obtain the application switching frequency, i.e.:

[0064] Application switching frequency = Number of times application categories are switched within the statistics window ÷ Duration of the statistics window.

[0065] For example, within a specific statistics window, a user is sequentially in an application environment of office editing, information browsing, office editing, spreadsheet, and information browsing, resulting in a total of 4 application category switches. If the duration of the statistics window is 60 seconds, then the application switching frequency = 4 ÷ 60 ≈ 0.067 times / second.

[0066] Furthermore, to exclude non-user-initiated task switching behaviors such as system notification windows and short-term pop-ups, when the duration of an application window gaining foreground focus is less than the preset minimum effective foreground dwell time, this window change will not be counted in the effective application category switching count.

[0067] The minimum effective foreground dwell time is determined based on the user's historical application window dwell data. Specifically, a preset number of foreground application dwell times are collected during the initial user usage phase. The dwell times are arranged in ascending order, and the dwell time corresponding to a preset percentile position is selected as the initial dwell time reference value. When the initial dwell time reference value is less than a preset lower limit, the preset lower limit is used as the minimum effective foreground dwell time. In this embodiment, the preset lower limit can be 500ms to reduce the impact of instantaneous pop-ups on the application switching frequency calculation.

[0068] S1.4: Merge the keyboard operation features, mouse operation features, and application context features into multi-modal operation features.

[0069] In this embodiment, after obtaining keyboard operation features, mouse operation features, and application context features respectively, the three types of features are associated according to their corresponding acquisition time and processed in a unified data format to obtain multi-mode operation features.

[0070] Specifically, keyboard operation features, mouse operation features, and application context features acquired within the same time frame are combined accordingly. Keyboard operation features are used to characterize the user's keyboard input behavior, mouse operation features are used to characterize the user's mouse interaction behavior, and application context features are used to characterize the application environment in which the user is currently located.

[0071] For continuous numerical features with different dimensions, normalization is performed based on the historical statistical range of the corresponding features to ensure that features from different sources are on a uniform data scale. For discrete features such as application categories, a pre-defined category code is used for representation.

[0072] For example, within the same feature statistics time range, the keystroke frequency, key duration and input continuity ratio in keyboard operation features are obtained, the cursor movement speed, click frequency and cursor stillness ratio in mouse operation features are obtained, and the application category and application switching frequency in application context features are obtained. Then, the above features are processed uniformly and combined in a preset order to obtain the corresponding multi-mode operation features.

[0073] Therefore, keyboard operation features, mouse operation features, and application context features are unified into multimodal operation features, providing input data for subsequent fusion classification based on multimodal operation features.

[0074] In summary, this step collects keyboard and mouse human-computer interaction signals and context signals of active application windows through the computer terminal interface, extracting keyboard operation features, mouse operation features, and application context features respectively; performs statistical calculations, threshold discrimination, and filtering on various features to remove invalid interference data such as pop-ups, and then merges them after time correlation, format normalization, and encoding to output multi-modal operation features, providing standardized input data for subsequent multi-modal fusion classification and recognition of user working status.

[0075] S2: Based on multi-modal operation features, perform fusion classification to determine the user's current work status, where the work status is associated with state confidence.

[0076] Step S2 in the method provided in this embodiment of the invention includes:

[0077] S2.1: Based on the preset time window length and sliding step size, perform sliding window truncation on the multi-mode operation features to generate the feature vector of the current time window;

[0078] After obtaining continuous multimodal operation features, the multimodal operation features are truncated by a sliding window according to the preset time window length and sliding step size.

[0079] The time window length is used to determine the continuous operation time range referenced for current working state identification, and the sliding step size is used to determine the update interval between adjacent time windows. The time window length and sliding step size are determined based on the duration of the user's working state, the rate of change of operational characteristics, and the real-time requirements of working state identification. In this embodiment, the time window length is 300 seconds, and the sliding step size is 60 seconds.

[0080] Specifically, at the current time point, multi-mode operation features within a continuous range of 300 seconds are acquired and arranged in chronological order to form a feature sequence corresponding to the current time window; then, keyboard operation features, mouse operation features, and application context features in the feature sequence are combined in a preset feature order to generate a feature vector for the current time window.

[0081] After a sliding step, the time window is moved forward by 60 seconds. The multi-modal operation features corresponding to the earliest 60 seconds in the original time window are deleted, and the multi-modal operation features corresponding to the latest 60 seconds are added, thereby generating an updated time window feature vector.

[0082] For example, when the time window length is 300s and the sliding step size is 60s, the first time window covers the operational features of the 300s before the current moment; after 60s, the second time window moves forward 60s, and there is a 240s data overlap between the two adjacent time windows, so that the operational behavior within a continuous time range can be used to identify the working status.

[0083] S2.2: Using keyboard operation features and mouse operation features in the feature vector as the first modal input, and application context features in the feature vector as the second modal input, a Bayesian fusion classification model is used to jointly infer the first modal input and the second modal input to calculate the posterior probability of each preset candidate working state.

[0084] In this embodiment, multiple candidate working states are preset, and probabilistic correlations are established between the first modal input, the second modal input, and each candidate working state.

[0085] The first modal input consists of keyboard operation features and mouse operation features in the current time window feature vector, which are used to characterize the user's current human-computer interaction behavior; the second modal input consists of application context features in the current time window feature vector, which are used to characterize the application task environment in which the user is currently located.

[0086] In this embodiment, candidate working states can be preset according to the actual application scenario of the smart desk. For example, they can be set to focused work state, regular work state, communication and collaboration state, and non-work state. The specific number and categories of candidate working states can be set according to the actual application scenario.

[0087] In the Bayesian fusion classification model building stage, the correspondence between the second modality input and each candidate working state is first statistically analyzed based on historical labeled working state samples to determine the prior probability of inferring each candidate working state from the second modality input.

[0088] Specifically, for any candidate working state, the number of samples in the historical samples that belong to the candidate working state and have the corresponding second modality application context features is counted, and this number is normalized with the total number of historical samples to obtain the corresponding working state prior probability.

[0089] In this embodiment, historical user operation samples with work status annotations are used for statistical analysis. Historical samples can be derived from manual status annotation results during actual user use or pre-organized work status annotation data.

[0090] Furthermore, to avoid a candidate working state appearing infrequently in historical samples, resulting in an excessively low prior probability, this embodiment employs a smoothing process. Specifically, a preset smoothing count is added to the sample count of each candidate working state, and the sample count of each candidate working state after adding the smoothing count is normalized.

[0091] The value of the smoothing count is determined based on the size of the historical sample and the number of candidate working states. In this embodiment, when there are 4 categories of candidate working states and the size of the historical sample is large, 1 can be used as the smoothing count for each candidate working state.

[0092] Furthermore, the likelihood probability of the first modal input under each candidate working state condition is determined based on historical labeled working state samples.

[0093] Specifically, for any candidate working state, the probability distribution of the feature value corresponding to the first modal input in the historical samples belonging to the candidate working state is statistically analyzed. Based on the probability of the first modal input feature value appearing in the probability distribution in the current time window, the likelihood probability of the first modal input under the candidate working state condition is determined.

[0094] For continuous first-mode features, their mean and standard deviation can be statistically analyzed based on historical samples, and the likelihood probability of the current first-mode input under the corresponding candidate working state conditions can be calculated using the corresponding probability density function.

[0095] For example, for the keystroke frequency in a certain candidate working state, if the historical sample statistics show that the average keystroke frequency for that working state is 4.2 times / s and the standard deviation is 0.8 times / s, then the corresponding likelihood probability is determined based on the degree of matching between the keystroke frequency in the current time window and the historical distribution.

[0096] For mouse operation features, the likelihood probability under each candidate working state condition can be determined in the same way.

[0097] Furthermore, the keyboard operation features and mouse operation features in the first modality are jointly processed to obtain the comprehensive likelihood probability of the first modality input under each candidate working state condition.

[0098] In one specific implementation, when the features in the first modality are considered to be conditionally independent under a given operating state, the overall likelihood probability of the first modality input under a candidate operating state is the product of the likelihood probabilities corresponding to each first modality feature, that is:

[0099] The overall likelihood probability of the first mode = the likelihood probability of keyboard operation features × the likelihood probability of mouse operation features.

[0100] When the first modality contains multiple keyboard operation features and multiple mouse operation features, the conditional likelihood probabilities corresponding to each feature are multiplied.

[0101] Based on this, the posterior probability of each candidate working state is calculated according to the prior probability of the working state determined by the second modal input and the likelihood probability of the first modal input under each working state condition. Specifically:

[0102] The posterior probability of a candidate working state = the prior probability of the working state determined by the second modal input × the likelihood probability of the first modal input under the condition of the candidate working state ÷ the sum of the above products corresponding to all candidate working states.

[0103] For example, the preset candidate work states include focused work state, routine work state, collaborative communication state, and non-work state. For the current time window, the prior probabilities corresponding to the four candidate work states are determined to be 0.45, 0.30, 0.15, and 0.10 based on application context features; based on the matching degree between the current keyboard operation features and mouse operation features and the historical feature distributions of each candidate work state, the corresponding first modality comprehensive likelihood probabilities are obtained to be 0.60, 0.50, 0.20, and 0.05, respectively. Therefore, the unnormalized probabilities corresponding to the four candidate work states are: focused work state: 0.45 × 0.60 = 0.270, routine work state: 0.30 × 0.50 = 0.150, collaborative communication state: 0.15 × 0.20 = 0.030, and non-work state: 0.10 × 0.05 = 0.005. The sum of the products corresponding to all candidate work states is 0.270 + 0.150 + 0.030 + 0.005 = 0.455. Therefore, the posterior probability of the focused work state is 0.270 ÷ 0.455 ≈ 0.593, the posterior probability of the regular work state is 0.150 ÷ ​​0.455 ≈ 0.330, the posterior probability of the communication and collaboration state is 0.030 ÷ 0.455 ≈ 0.066, and the posterior probability of the non-work state is 0.005 ÷ 0.455 ≈ 0.011.

[0104] S2.3: Determine the candidate working state with the highest posterior probability as the user's current working state, and use the corresponding posterior probability as the state confidence.

[0105] Specifically, when the state confidence level is not lower than a preset first confidence threshold, the working state is transmitted to participate in parameter fusion decision-making; when the state confidence level is not lower than a preset second confidence threshold but lower than the first confidence threshold, the working state is transmitted to participate in parameter fusion decision-making, and the parameter adjustment range in the parameter fusion decision-making is reduced to a preset reduction ratio; when the state confidence level is lower than the second confidence threshold, the current desktop control parameters are kept unchanged; wherein, the first confidence threshold is greater than the second confidence threshold.

[0106] After obtaining the posterior probability corresponding to each candidate work state, the posterior probabilities of all candidate work states are compared. The candidate work state with the highest posterior probability is determined as the user's current work state, and the posterior probability corresponding to this candidate work state is determined as the state confidence.

[0107] For example, based on the above calculation results, the posterior probability corresponding to the focused work state is approximately 0.593, which is the highest among all candidate work states. Therefore, the user's current work state is determined to be the focused work state, and approximately 0.593 is used as the corresponding state confidence level.

[0108] Furthermore, to avoid low-confidence work status recognition results directly causing significant changes in the control parameters of the smart office desk, this embodiment sets a first confidence threshold and a second confidence threshold based on the state confidence level, and determines the transmission method of the work status based on the interval in which the state confidence level is located.

[0109] The first confidence threshold and the second confidence threshold are determined based on the consistency between the historical work status identification results and the actual work status labeling results.

[0110] Specifically, the classification accuracy corresponding to different state confidence intervals is obtained using historical labeled samples, and confidence thresholds that meet preset accuracy requirements are determined for each interval. The higher confidence threshold is defined as the first confidence threshold, and the lower confidence threshold is defined as the second confidence threshold, ensuring that the first confidence threshold is greater than the second confidence threshold. In this embodiment, the first confidence threshold can be set to 0.65, and the second confidence threshold can be set to 0.40.

[0111] When the state confidence is not lower than the first confidence threshold, it indicates that the current working state has high recognition reliability. Therefore, the working state is passed on to participate in the subsequent parameter fusion decision and the parameter fusion decision is executed according to the normal parameter adjustment range.

[0112] When the state confidence is not lower than the second confidence threshold but lower than the first confidence threshold, it indicates that the current working state has a certain degree of recognition reliability, but there is still a high degree of uncertainty. Therefore, the working state is also passed on to participate in the parameter fusion decision, while the parameter adjustment range in the parameter fusion decision is reduced.

[0113] Specifically, a reduction ratio is preset, and the reduction ratio is determined based on the historical parameter adjustment effect under different state recognition confidence levels.

[0114] The magnitude of parameter adjustment = the magnitude of original parameter adjustment determined by parameter fusion decision × the reduction ratio.

[0115] In this embodiment, the reduction ratio can be set to 0.50, that is, the original parameter adjustment range is reduced to 50% of the original range. For example, if the parameter fusion decision determines to adjust the desktop height by 20mm, and the current state confidence level is between the second confidence threshold and the first confidence threshold, then the actual adjustment range = 20 × 0.50 = 10mm.

[0116] When the state confidence is lower than the second confidence threshold, it is considered that the current working state identification result is insufficient to support the adjustment of the desktop control parameters. Therefore, the current working state is not passed on to participate in the parameter fusion decision, and the current desktop control parameters are kept unchanged.

[0117] In summary, this step involves using a sliding window to extract features from continuous multimodal operations by setting the time window length and sliding step size, assembling the feature vectors. It distinguishes between two modal inputs: human-computer interaction behavior and application task environment, and employs a Bayesian fusion classification model to perform joint inference. The posterior probability of each candidate working state is calculated, and the highest probability item is selected as the current working state, with the state confidence score output. Based on a two-level confidence threshold, tiered processing is applied: high confidence scores directly participate in decision-making; medium confidence scores participate in decision-making while reducing the adjustment range; and low confidence scores maintain the original parameters, reducing the risk of unreasonable adjustments due to recognition uncertainty.

[0118] S3: Combine the working status with the pre-built knowledge graph to make parameter fusion decisions and obtain personalized desktop control parameters for smart desks. The knowledge graph includes a domain-general knowledge layer and an individual personalized layer.

[0119] Step S3 in the method provided in this embodiment of the invention includes:

[0120] S3.1: Based on the domain-wide general knowledge layer, match and obtain the domain-default desktop parameters corresponding to the working status;

[0121] First, based on the user's current working status determined in step S2, retrieve the domain default desktop parameters corresponding to the working status from the domain general knowledge layer.

[0122] The domain-specific general knowledge layer stores the general relationships between different working states and desktop parameters. The domain's default desktop parameters include at least recommended sitting height, recommended standing height, recommended switching cycle, and recommended standing percentage.

[0123] For example, the following mapping between working states and domain default desktop parameters is established in advance:

[0124] Work status Recommended sitting height Recommended standing height Recommended switching cycle Recommended standing ratio Focus on work 720mm 1080mm 45min 40% Normal working status 720mm 1070mm 50min 35% Communication and Collaboration Status 730mm 1080mm 40min 45%

[0125] The above parameters were determined based on ergonomic experimental data, historical user data, and actual operating data of the smart office desk. Specifically, the recommended sitting and standing heights can be calibrated based on actual user data, while the recommended switching cycle and recommended standing percentage can be determined based on historical sitting and standing behavior data for corresponding work states.

[0126] For example, when step S2 determines that the user is currently focused on work, the corresponding recommended sitting posture height, recommended standing posture height, recommended switching cycle and recommended standing percentage are obtained from the domain general knowledge layer, and the matched domain default desktop parameters are used as the baseline parameters for subsequent personalized parameter fusion.

[0127] S3.2: Based on the individual personalization layer, match and obtain the personalized correction knowledge of the current user. The personalized correction knowledge is represented by the individual preference node in the individual personalization layer. The individual preference node includes at least the user's standing tolerance time, preferred sitting height, preferred standing height, preference switching cycle, preferred standing percentage, and preference data reliability.

[0128] After obtaining the domain's default desktop parameters, the user's personalized correction knowledge is retrieved from the individual personalization layer based on the user identifier corresponding to the current user.

[0129] Among them, the personalized correction knowledge is represented by the individual preference node in the individual personalization layer. The individual preference node includes at least the user's standing tolerance time, preferred sitting height, preferred standing height, preference switching cycle, preferred standing percentage, and preference data reliability.

[0130] Among them, the reliability of preference data is used to characterize the stability of the corresponding individual's preference knowledge, and is determined based on the user's historical effective usage data.

[0131] Specifically, for any user preference parameter, the number of valid samples corresponding to that preference parameter in the historical valid usage records, the stability of the samples, and the proportion of recent usage data are counted, and the reliability of the preference data is determined based on the above indicators.

[0132] In one specific implementation, the sufficiency of effective samples, the stability of preference parameters, and the consistency of recent data are each normalized to a value between 0 and 1, and then weighted and summed according to a preset ratio. The specific calculation is as follows:

[0133] Reliability of preference data = Sufficient quantity of effective samples × 0.4 + Stability of preference parameters × 0.3 + Consistency of recent data × 0.3.

[0134] The weights 0.4, 0.3, and 0.3 are determined based on the degree of influence of the three evaluation factors on the stability of individual preferences. For example, by comparing the correlation between different factors and the degree of subsequent preference retention using historical validation data, it is determined that the sufficiency of the effective sample has the greatest impact, and therefore it is assigned a weight of 0.4, while the other two factors are each assigned a weight of 0.3.

[0135] Furthermore, the deviation between user preference parameters and domain default parameters is calculated, and the deviation is adjusted using the reliability of preference data, thereby avoiding over-correction of domain default parameters when individual preference data is insufficient or has low stability.

[0136] For the height parameter, specifically:

[0137] Personalized sitting posture height correction amount = (user preferred sitting posture height - recommended sitting posture height) × preference data reliability.

[0138] Personalized standing posture height correction amount = (user preferred standing posture height - recommended standing posture height) × preference data reliability.

[0139] The same reliability adjustment method is used for switching cycles and standing percentage:

[0140] Personalized switching cycle adjustment amount = (user preference switching cycle - recommendation switching cycle) × preference data reliability.

[0141] Personalized standing percentage correction = (user preference standing percentage - recommended standing percentage) × preference data reliability.

[0142] For example, if the current user's preferred sitting posture height is 710mm, the domain's default recommended sitting posture height is 720mm, and the preference data reliability is 0.8, then the personalized sitting posture height correction amount = (710-720)×0.8 = -8mm. Therefore, in the user's current working state, the personalized corrected sitting posture height = 720-8 = 712mm.

[0143] S3.3: Based on the current time period, obtain the corresponding context correction parameters;

[0144] Furthermore, the user's current time period is determined based on the current system time, and the context correction parameters corresponding to the current time period are obtained from the context knowledge.

[0145] In this embodiment, the working hours can be divided into multiple time periods in advance, such as the morning working period, the afternoon working period, and the evening working period, and corresponding desktop parameter correction rules can be established for different working periods.

[0146] For example, for the afternoon work period, based on historical user data, it is determined that users are more likely to have increased continuous sitting time during this period. Therefore, a context correction parameter of increasing the standing rate by 5 percentage points can be set. For the evening work period, a smaller desktop adjustment range can be set based on actual usage data.

[0147] The specific value of the context correction parameter can be determined based on the statistical relationship between historical desktop usage data and user work status data within the corresponding time period.

[0148] For example, the average deviation between the actual standing percentage of users and the baseline standing percentage in each time period is calculated, and the average deviation is used as the initial context correction amount for the corresponding time period, and adjusted according to the historical verification results.

[0149] S3.4: Perform three-layer parameter fusion on the domain default desktop parameters, personalized correction knowledge and context correction parameters to generate personalized desktop control parameters. The personalized desktop control parameters include at least the desktop target height value and the height adjustment cycle.

[0150] like Figure 2 As shown, after obtaining the domain default desktop parameters, personalized correction knowledge, and context correction parameters, the three-layer parameter fusion is performed in the manner of using the domain default parameters as the baseline, overlaying personalized correction knowledge, and further overlaying context correction parameters to generate personalized desktop control parameters.

[0151] Specifically, for the target desktop height value, the recommended sitting or standing height from the domain's default desktop parameters is first used as the baseline parameter. Then, a personalized height correction adjusted for the reliability of preference data is added, and finally, a contextual height correction corresponding to the current time period is added. That is:

[0152] Personalized desktop target height value = Domain default desktop height + Personalized height correction amount + Context height correction amount.

[0153] For the height adjustment cycle:

[0154] Personalized height adjustment cycle = Domain default switching cycle + Personalized switching cycle correction amount + Context switching cycle correction amount.

[0155] Regarding the percentage of standing:

[0156] Personalized standing percentage = Domain default standing percentage + Personalized standing percentage correction amount + Context standing percentage correction amount.

[0157] For example, the current user is focused on work. The default parameters for the domain are: recommended sitting height 720mm, recommended standing height 1080mm, recommended switching cycle 45min, and recommended standing percentage 40%. The user's individual preferences are: preferred sitting height 710mm, preferred standing height 1090mm, preferred switching cycle 40min, and preferred standing percentage 50%, with a corresponding preference data reliability of 0.8. Therefore, the personalized sitting height correction is (710-720)×0.8=-8mm, the personalized standing height correction is (1090-1080)×0.8=8mm, the personalized switching cycle correction is (40-45)×0.8=-4min, and the personalized standing percentage correction is (50%-40%)×0.8=8 percentage points. Assuming the current work period is afternoon, the contextual correction parameters based on historical time period data are: height correction 2mm, switching cycle correction -2min, and standing percentage correction 3 percentage points. The final results are: personalized sitting height = 720 - 8 + 2 = 714 mm, personalized standing height = 1080 + 8 + 2 = 1090 mm, personalized height adjustment cycle = 45 - 4 - 2 = 39 min, and personalized standing percentage = 40% + 8% + 3% = 51%.

[0158] Furthermore, based on the user's working state determined in the current step, the corresponding personalized desktop target height is selected. For example, when the user's current working state corresponds to a seated working state, the personalized seated height is used as the desktop target height value; when the user's current working state corresponds to a standing working state, the personalized standing height is used as the desktop target height value.

[0159] Step S3 in the method provided in this embodiment of the invention further includes:

[0160] Define node types and attribute sets for each node type. Node types should include at least working status nodes, ergonomic rule nodes, domain default desktop parameter nodes, manual control log nodes, and individual preference nodes.

[0161] Instantiate each preset work status category as a work status node, instantiate the preset ergonomic rules as ergonomic rule nodes, and instantiate the default desktop parameters corresponding to each work status category as domain default desktop parameter nodes.

[0162] Define relationship types, which should include at least recommendation relationships, constraint relationships, personalized overlay relationships, and preference update relationships;

[0163] Establish recommendation relationships between work status nodes and domain default desktop parameter nodes, and constraint relationships between ergonomic rule nodes and domain default desktop parameter nodes;

[0164] Based on cross-user behavior similarity matching and preference propagation, an initial individual personalization layer for new users is constructed.

[0165] First, define the node types of knowledge objects involved in the personalized settings of the smart office desk parameters, and configure corresponding attribute sets for different node types.

[0166] In this embodiment, the node types include at least working status nodes, ergonomic rule nodes, domain default desktop parameter nodes, manual control log nodes, and individual preference nodes.

[0167] The system includes the following nodes: Work Status Node: Represents a preset user work status category; its attributes include at least a work status identifier and a work status name. Ergonomic Rule Node: Represents the ergonomic constraints corresponding to the smart desk parameters; its attributes include at least a rule identifier, constraint parameter type, and constraint parameter range. Domain Default Desktop Parameter Node: Represents the default desktop parameters corresponding to the work status; its attributes include at least recommended sitting height, recommended standing height, recommended switching cycle, and recommended standing percentage. Manual Control Log Node: Represents the operation records formed by the user's manual control of the smart desk. Individual Preference Node: Represents the individual preferences formed by the current user; its attributes include at least the user's standing tolerance time, preferred sitting height, preferred standing height, preference switching cycle, preferred standing percentage, and preference data reliability.

[0168] Furthermore, the preset knowledge is instantiated according to the corresponding node type.

[0169] Specifically, the preset work status categories are instantiated as work status nodes, the preset ergonomic rules are instantiated as ergonomic rule nodes, and the default desktop parameters corresponding to each work status category are instantiated as domain default desktop parameter nodes.

[0170] For example, for a preset working state, the working state is instantiated as a corresponding working state node, and the recommended sitting posture height, recommended standing posture height, recommended switching cycle and recommended standing percentage corresponding to the working state are stored in the corresponding domain default desktop parameter node, so that the working state node can be associated with and retrieve the corresponding domain default desktop parameters.

[0171] Furthermore, the types of relationships between different nodes are defined. In this embodiment, the relationship types include at least recommendation relationships, constraint relationships, personalized coverage relationships, and preference update relationships.

[0172] Among them, the recommendation relationship is used to characterize the correspondence between the working status node and the domain default desktop parameter node; the constraint relationship is used to characterize the constraint relationship between the ergonomic rule node and the domain default desktop parameter node; the personalized coverage relationship is used to characterize the personalized modification association between the individual preference node and the domain default desktop parameter node; and the preference update relationship is used to characterize the preference update association between the manual control log node and the individual preference node.

[0173] Establish a recommendation relationship between work status nodes and domain default desktop parameter nodes, so that after obtaining the user's current work status, the corresponding domain default desktop parameters can be obtained by matching the corresponding work status nodes.

[0174] Furthermore, a constraint relationship is established between the ergonomic rule node and the domain default desktop parameter node, so that the domain default desktop parameter is associated with the corresponding ergonomic rule, thereby obtaining the corresponding ergonomic constraint information when using the domain default desktop parameter in the future.

[0175] At the same time, a personalized overlay relationship is established between individual preference nodes and domain default desktop parameter nodes, as well as a preference update relationship between manually controlled log nodes and individual preference nodes, so that individual preferences can be associated with domain default desktop parameters, and user-manually controlled operation records can be associated with the corresponding individual preferences.

[0176] Based on this, a knowledge graph is formed based on the above nodes and relationships. Among them, the work status node, ergonomic rule node, and domain default desktop parameter node are used to constitute general domain knowledge, the individual preference node is used to constitute individual personalized knowledge, and the manual control log node is used to associate the user's historical manual control records.

[0177] Step S3 in the method provided in this embodiment of the invention further includes:

[0178] For each user, a user behavior feature vector is constructed. The user behavior feature vector includes at least the average keystroke frequency, working state distribution entropy, concentration of main working periods, average input continuity ratio, manual operation frequency, and standing early termination rate.

[0179] Calculate the cosine similarity of behavioral feature vectors between new users and existing users, where the individual personalization layer of existing users has been established and the reliability of preference data has reached a preset confidence threshold.

[0180] Select the first preset number of existing users with the highest cosine similarity as similar users;

[0181] The product of the cosine similarity of each similar user and the reliability of the preference data is used as the propagation weight. The personalized correction knowledge of each similar user is weighted, summed, and normalized to obtain the propagation preference vector.

[0182] Based on the cold start hybrid weight coefficient, the propagation preference vector and the zero offset vector are weighted and mixed to obtain the initial personalized correction knowledge of new users and construct the initial individual personalization layer;

[0183] The cold start hybrid weight coefficient is determined based on the ratio of the number of manual control logs accumulated by the new user to the preset minimum effective data volume, and is positively correlated with the ratio.

[0184] First, construct a corresponding user behavior feature vector for each user. The user behavior feature vector includes at least: average keystroke frequency, work state distribution entropy, concentration of main work periods, average input continuity ratio, manual operation frequency, and standing early termination rate.

[0185] Among them, the average keystroke frequency is used to characterize the average input intensity of a user during effective keyboard operation, specifically:

[0186] Average keystroke frequency = Total number of valid keystrokes within the statistical period ÷ Valid keyboard operation time.

[0187] For example, if the effective keyboard operation time of a user is 100 hours within a preset statistical period and the cumulative effective keystrokes are 360,000, then the average keystroke frequency is 360,000 ÷ 360,000 s, which is 1 keystroke / s.

[0188] To determine the work state distribution entropy, we first calculate the percentage of time a user spends in each preset work state within the statistical period. Then, we calculate the work state distribution entropy based on the percentage of time spent in each work state. Specifically:

[0189] The entropy of the working state distribution is the negative of the sum of the products of the time proportion of each working state and the logarithm of the corresponding time proportion.

[0190] For example, if the percentage of time a user spends in a focused work state, a routine work state, a communication and collaboration state, and a non-work state during the statistical period is 0.50%, 0.30%, 0.15%, and 0.05%, respectively, then the corresponding work state distribution entropy can be calculated based on the above time percentages.

[0191] To determine the concentration of primary work periods, the day is divided into multiple preset time periods. The effective working hours of users within each time period are calculated, and the time period with the highest effective working hours is identified as the primary work period. Specifically, the concentration of primary work periods is as follows:

[0192] Concentration of main working periods = Effective working hours of main working periods ÷ Effective working hours of all periods.

[0193] For example, if a user's total effective working time is 8 hours, with 5 hours of effective working time in the morning (which is the highest among all time periods), then the concentration of the main working time period is 5 ÷ 8 = 0.625.

[0194] The average input continuity ratio is calculated based on the input continuity ratio of users across multiple valid statistical time ranges, specifically as follows:

[0195] Average input continuity ratio = Sum of input continuity ratios for each valid statistical time range ÷ Number of valid statistical time ranges.

[0196] Regarding the frequency of manual operations, the relationship between the number of times users actively adjusted desktop parameters and the effective usage time was analyzed, specifically as follows:

[0197] Manual operation frequency = Number of times desktop parameters are manually adjusted within the statistical period ÷ Effective usage time.

[0198] To determine the early termination rate of standing, the number of standing events where the actual standing time is less than the preset target standing time is counted, specifically:

[0199] Early termination rate of standing events = Number of early termination events ÷ Total number of valid standing events.

[0200] After obtaining the above six types of user behavior features, the statistical range of each feature is determined based on existing user historical behavior data, and each user behavior feature is normalized to eliminate the impact of differences in the scale of different features on subsequent similarity calculations. Then, the corresponding user behavior feature vectors are formed by combining them in a preset order.

[0201] Furthermore, for new users, the accumulated human-computer interaction data and manual control logs of the new users are obtained, and user behavior feature vectors corresponding to the new users are constructed in the manner described above.

[0202] At the same time, users whose individual personalization layers have been established and whose preference data reliability reaches a preset confidence threshold are selected from existing users as candidate existing users.

[0203] The preset confidence threshold is used to filter existing users with sufficient reliable knowledge of individual preferences. Its specific value is determined based on the consistency between historical user preference data and subsequent actual usage behavior.

[0204] For example, based on historical verification data, if the preset reliability requirement can be met when the reliability of the preference data reaches 0.70, then 0.70 can be used as the preset reliability threshold in this embodiment.

[0205] Furthermore, the cosine similarity of the behavioral feature vectors between the new user and each candidate existing user is calculated. Specifically:

[0206] Cosine similarity = Sum of the products of corresponding elements of the new user behavior feature vector and the existing user behavior feature vector ÷ (Square root of the sum of squares of the elements of the new user behavior feature vector × Square root of the sum of squares of the elements of the existing user behavior feature vector).

[0207] For example, the normalized behavioral feature vector of a new user is (0.60, 0.70, 0.80, 0.65, 0.30, 0.20), and the normalized behavioral feature vector of an existing user is (0.65, 0.68, 0.75, 0.70, 0.32, 0.18). The cosine similarity between the two is approximately 0.997, calculated in the manner described above.

[0208] The preset number is determined based on the preference propagation error corresponding to different numbers of similar users in historical verification data. For example, if different numbers of similar users are verified, and a lower overall propagation error is obtained when 5 similar users are selected, then 5 can be used as the preset number in this embodiment.

[0209] Furthermore, personalized correction knowledge is obtained from the individualized layers of similar users, and the product of the cosine similarity between each similar user and the new user and the reliability of the corresponding preference data is used as the propagation weight. Specifically:

[0210] Propagation weight = cosine similarity × reliability of preference data.

[0211] For example, if the cosine similarity between a similar user and a new user is 0.90, and the reliability of the similar user's preference data is 0.80, then the propagation weight = 0.90 × 0.80 = 0.72.

[0212] After obtaining the propagation weights for each similar user in the same manner, the personalized modified knowledge of each similar user is weighted, summed, and normalized to obtain the propagation preference vector. Specifically:

[0213] The propagation preference vector = the sum of the products of the personalized modified knowledge of each similar user and the corresponding propagation weight ÷ the sum of the propagation weights of each similar user.

[0214] For example, for the preferred sitting posture height correction, the propagation weights of the three similar users are 0.72, 0.68 and 0.60, respectively, and the corresponding sitting posture height corrections are -8mm, -5mm and -10mm, respectively. Then the propagation preferred sitting posture height correction = [0.72×(-8)+0.68×(-5)+0.60×(-10)]÷(0.72+0.68+0.60)≈-7.54mm.

[0215] In the same way, other personalized corrections in the propagation preference vector can be obtained.

[0216] Furthermore, based on the cold start hybrid weight coefficient, the propagation preference vector and the zero offset vector are weighted and mixed to obtain the initial personalized correction knowledge of new users.

[0217] The zero offset vector represents a state where all personalization adjustments are zero, meaning that the domain default desktop parameters are not personalized based on cross-user preference propagation.

[0218] The cold start hybrid weighting coefficient is determined based on the ratio of the number of manually controlled logs accumulated by the new user to the preset minimum effective data volume, and is positively correlated with this ratio. Specifically:

[0219] Cold start hybrid weight coefficient = number of manually controlled logs ÷ the smaller of the preset minimum effective data volume and 1.

[0220] Therefore, when the number of manual control logs accumulated by a new user is lower than the preset minimum effective data volume, the cold start hybrid weight coefficient gradually increases as the number of manual control logs increases; when the number of manual control logs reaches or exceeds the preset minimum effective data volume, the cold start hybrid weight coefficient reaches 1.

[0221] For example, if the minimum effective data volume is preset to 20 records, and a new user has already accumulated 4 manual control logs, then the cold start mixed weight coefficient = 4 ÷ 20 = 0.20. If the new user has accumulated 25 manual control logs, then the cold start mixed weight coefficient = the smaller value between 25 ÷ 20 and 1 = 1.

[0222] Furthermore, the propagation preference vector and the zero offset vector are weighted and mixed, specifically as follows:

[0223] Initial personalized correction knowledge = cold start mixed weight coefficient × propagation preference vector + (1 - cold start mixed weight coefficient) × zero offset vector.

[0224] For example, the propagation preference for sitting posture height correction is -7.54mm. The new user has already accumulated 4 manual control logs. The preset minimum effective data volume is 20 logs. Then the cold start hybrid weight coefficient is 0.20 and the zero offset is 0. Thus, the initial personalized sitting posture height correction is -7.54×0.20+0×0.80=-1.508mm.

[0225] Other personalized correction values ​​are obtained in the same way, and the obtained initial personalized correction knowledge is used as the individual preference information of the new user to construct the initial individual personalization layer of the new user.

[0226] Step S3 in the method provided in this embodiment of the invention further includes:

[0227] Acquire records of manual intervention operations performed by users on the smart desk and the operation context information when the manual intervention operations occurred;

[0228] For each preset operation intent category, the attribution confidence is calculated based on the corresponding feature conditions in the operation context information;

[0229] The operation intention category with the highest attribution confidence was used as the final operation intention category for the manual intervention operation record;

[0230] Based on the correction information corresponding to the final operational intent category, the individual personalized layer of the knowledge graph is updated using an exponential moving average method in conjunction with the preset learning rate, and the reliability of the preference data is updated simultaneously.

[0231] Specifically, it acquires records of manual intervention operations performed by users on the smart desk, as well as the corresponding operation context information when the manual intervention operations occur.

[0232] The operation context information may include the desktop height during the operation, the duration of continuous work before the operation, the input continuity ratio before the operation, the application switching frequency before and after the operation, and the application category of the currently active application.

[0233] For each manual intervention record, the attribution confidence is calculated based on the feature conditions in the operation context information corresponding to the corresponding operation intent category, according to the preset operation intent category.

[0234] In this embodiment, the preset operation intent categories include at least standing intolerance, flow protection, emergency situations, and severe discomfort. Different operation intent categories are configured with corresponding feature conditions, allowing the same operation context information to be used for attribution judgments for different operation intent categories.

[0235] For any operation intent category, the feature condition matching score is determined based on the degree of matching of the corresponding feature conditions in the operation context information, and then weighted and summed according to the corresponding weights to obtain the attribution confidence of that operation intent category. Specifically:

[0236] Attribution confidence = the sum of the products of the conditional matching scores of each feature and their corresponding weights.

[0237] Among them, the feature condition matching score is used to characterize the degree of matching between the current operation context information and the corresponding operation intention category, and the value range is 0 to 1; the weight corresponding to each feature condition is determined according to the ability of the corresponding feature condition to distinguish different operation intention categories in the historical manual intervention operation data, and is normalized so that the sum of the weights corresponding to the same operation intention category is 1.

[0238] For example, for the category of high discomfort, the desktop height during operation, the continuous working time before operation, and the input continuity ratio before operation correspond to the corresponding feature conditions, with matching scores of 0.90, 0.70, and 0.80, and corresponding weights of 0.50, 0.30, and 0.20, respectively. Therefore, the attribution confidence of the high discomfort category is 0.90×0.50+0.70×0.30+0.80×0.20=0.83.

[0239] In the same manner, the attribution confidence scores for the standing intolerance category, flow protection category, emergency situation category, and high discomfort category were calculated respectively.

[0240] After obtaining the attribution confidence of each operational intent category, the operational intent category with the highest attribution confidence is determined, and this operational intent category is taken as the final operational intent category of the corresponding manual intervention operation record.

[0241] Furthermore, the highest attribution confidence score is compared with a preset attribution confidence score threshold.

[0242] The attribution confidence threshold is determined based on the attribution results of historical manual intervention operations. Specifically, the accuracy of operation intent recognition within different attribution confidence intervals can be statistically analyzed, and the attribution confidence threshold can be determined according to preset recognition accuracy requirements. For example, if historical verification results show that an attribution confidence of 0.75 meets the preset recognition accuracy requirements, then 0.75 can be used as the attribution confidence threshold in this embodiment.

[0243] When the highest attribution confidence level is not lower than the preset attribution confidence level threshold, a definite attribution label is assigned to the manual intervention operation record; when the highest attribution confidence level is lower than the attribution confidence level threshold, the manual intervention operation record is marked as having an undetermined attribution.

[0244] For example, if the attribution confidence scores for the standing intolerance category, flow protection category, emergency situation category, and high discomfort category corresponding to a certain manual intervention are 0.82, 0.63, 0.35, and 0.76, respectively, then the standing intolerance category is determined as the final intention category. Since the highest attribution confidence score of 0.82 is not lower than 0.75, a specific attribution label is assigned to this manual intervention.

[0245] For manual intervention operations with definite attribution labels, the individual personalized layer in the knowledge graph is updated using an exponential moving average method based on the correction information corresponding to the final operation intention category and the preset learning rate.

[0246] The final operation intent category is associated with the corresponding correction information, which is used to characterize the individual desktop parameter correction results corresponding to the operation intent.

[0247] The individual preference parameters to be updated are as follows:

[0248] Updated individual preference parameters = learning rate × correction information corresponding to this operation + (1 - learning rate) × individual preference parameters before update.

[0249] The learning rate is used to control the degree of influence of this manual intervention on the existing individual preferences, and its value is determined according to the preset learning rules.

[0250] For example, if the user's current preferred standing posture height is 1085mm, and the correction information corresponding to this manual intervention is 1070mm, with a preset learning rate of 0.2, then the updated preferred standing posture height = 0.2 × 1070 + (1 - 0.2) × 1085 = 1082mm.

[0251] While completing the individual personalization layer update, the reliability of the preference data is updated simultaneously.

[0252] Specifically, the reliability update amount is determined based on the attribution confidence level of this manual intervention, and the update is performed based on the reliability of the preference data before the update. The reliability update amount is jointly determined by the attribution confidence level and the preset reliability update coefficient.

[0253] Reliability update amount = Attribution confidence of this manual intervention operation × Preset reliability update coefficient.

[0254] The reliability of the updated preference data = the reliability of the preference data before the update + the smaller value between the reliability update amount and 1.

[0255] For example, if the attribution confidence of a manual intervention with a defined attribution label is 0.82, the preset reliability update coefficient is 0.1, and the reliability of the preference data before the update is 0.70, then the reliability update amount = 0.1 × 0.82 = 0.082. The reliability of the updated preference data = 0.70 + 0.082 = 0.782.

[0256] When the highest attribution confidence is lower than the preset attribution confidence threshold, the manual intervention is marked as an undetermined attribution. Therefore, the individual's personalization layer is not updated based on the manual intervention, nor is the reliability of the preference data updated based on the intervention. This avoids erroneous impacts on existing individual preference knowledge when the intention of the intervention cannot be reliably determined.

[0257] In summary, this step utilizes a pre-constructed two-layer knowledge graph to perform parameter fusion decision-making. First, it matches the domain's default desktop parameters corresponding to the current working state from the domain-general knowledge layer as a baseline. Then, it reads user-personalized correction knowledge adjusted for preference data reliability from the individual personalization layer. Simultaneously, it obtains the context-corrected parameters for the current time period, completing the three-layer parameter fusion to obtain personalized desktop control parameters. This step also defines and instantiates knowledge graph nodes and relationships, utilizes cross-user behavior similarity to initialize new user cold-start preferences, and performs intent attribution for user manual interventions. Based on the attribution results, iteratively updates the individual personalization layer and preference data reliability.

[0258] S4: Control the smart office desk to perform desktop height adjustment based on personalized desktop control parameters.

[0259] Step S4 in the method provided in this embodiment of the invention includes:

[0260] Perform constraint boundary checks on the personalized desktop control parameters and tighten the parameter values ​​that exceed the preset ergonomic safety range.

[0261] Based on the preset approach rate and combined with the current desktop control parameters, a weighted fusion-based smooth transition process is performed on the personalized desktop control parameters after constraint boundary checks to obtain the actual execution parameter sequence.

[0262] Based on the actual execution parameter sequence, the smart office desk is driven to adjust its desktop height.

[0263] Specifically, after obtaining the personalized desktop control parameters, a constraint boundary check is first performed on the personalized desktop control parameters, and parameter values ​​that exceed the preset ergonomic safety range are safely limited.

[0264] The preset ergonomic safety range is used to limit the allowable value range of personalized desktop control parameters, setting at least a minimum and maximum allowable value for the target desktop height. The ergonomic safety range is determined according to preset ergonomic rules and corresponds to the aforementioned ergonomic rule nodes.

[0265] Regarding the desktop target height value, when the desktop target height in the personalized desktop control parameters is lower than the minimum allowable value of the preset ergonomic safety range, the desktop target height will be reduced to the minimum allowable value; when the desktop target height is higher than the maximum allowable value of the preset ergonomic safety range, the desktop target height will be reduced to the maximum allowable value; when the desktop target height is within the ergonomic safety range, the original personalized desktop target height will remain unchanged.

[0266] For example, the desktop height corresponding to the ergonomic safety range is 650mm to 1250mm. When the target height of the generated personalized desktop is 1280mm, since 1280mm exceeds the maximum allowable value of 1250mm, it is safely limited to 1250mm. When the target height of the personalized desktop is 700mm, since 700mm is within the safety range, it remains unchanged at 700mm.

[0267] The specific values ​​of the ergonomic safety range are determined according to ergonomic rules. For example, the effective usage range corresponding to different ergonomic parameters in historical user data can be statistically analyzed, and the minimum and maximum allowable values ​​can be determined in combination with preset safety margins.

[0268] After completing the constraint boundary check and safety narrowing, according to the preset approach rate and combined with the current desktop control parameters, a weighted fusion-based smooth transition process is performed on the personalized desktop control parameters after the constraint boundary check to obtain the actual execution parameter sequence.

[0269] Among them, the current desktop control parameters are used to characterize the desktop control parameters that the smart office desk is currently actually executing, and the personalized desktop control parameters after constraint boundary checks are used as the target control parameters for this adjustment; the approach rate is used to control the speed at which the current desktop control parameters change towards the target control parameters.

[0270] In one specific implementation, the fusion weight used for each adjustment is determined based on the parameter difference between the current desktop control parameters and the personalized desktop control parameters after constraint boundary checks, and the fusion weight corresponding to the target control parameters is gradually increased according to a preset approach rate, so that the actual execution parameters gradually approach the target control parameters.

[0271] Specifically, the parameters of the current execution can be obtained by weighted fusion of the current desktop control parameters and the target control parameters:

[0272] The actual execution parameters are calculated as follows: Current desktop control parameters × Current parameter weights + Personalized desktop control parameters after constraint boundary checks × Target parameter weights.

[0273] Where, the weight of the current parameter plus the weight of the target parameter equals 1.

[0274] The target parameter weights are determined based on a preset approach rate. The larger the approach rate, the faster the target parameter weights increase, and the faster the actual executed parameters approach the target control parameters; the smaller the approach rate, the slower the target parameter weights increase, thus making the desktop height adjustment process smoother.

[0275] The approach rate is determined based on the actual height adjustment response characteristics of the smart desk and the smoothness of height adjustment during user operation.

[0276] For example, by testing the height change range during desktop height adjustment at different approach rates, when a certain approach rate can meet the adjustment response requirements while avoiding excessive continuous adjustment range, it is taken as the preset approach rate.

[0277] For example, if the current desktop height is 720mm, and the personalized desktop target height after constraint boundary checks is 1080mm, and the weight of the first target parameter corresponding to the preset approach rate is 0.25, then the actual execution parameter obtained from the first smooth transition is 720 × 0.75 + 1080 × 0.25 = 810mm. In the next parameter update, the target parameter weight is increased according to the preset approach rate, reaching 0.50, so the actual execution parameter is 810 × 0.50 + 1080 × 0.50 = 945mm. Weighted fusion continues according to the preset approach rate, gradually bringing the actual execution parameter from the current desktop height closer to the personalized desktop target height after safety constraint, thus forming the actual execution parameter sequence.

[0278] After obtaining the actual execution parameter sequence, the smart office desk is driven to adjust the desktop height according to the actual execution parameter sequence.

[0279] Specifically, the actual execution parameter sequence is sent sequentially to the height adjustment actuator of the smart desk in the order of generation, causing the height adjustment actuator to gradually adjust the desktop height according to the actual execution parameter sequence. As the actual execution parameters gradually approach the personalized desktop control parameters after constraint boundary checks, the actual desktop height of the smart desk synchronously approaches the target desktop height.

[0280] For example, the target height of the personalized desktop after constraint boundary check is 1080mm, and the current desktop height is 720mm. After weighted fusion, the actual execution parameters are obtained in sequence, such as 810mm, 945mm, 1012.5mm and 1046.25mm. The smart office desk is then driven to adjust its height in the above order, so that the desktop height gradually approaches 1080mm from 720mm.

[0281] Step S4 in the method provided in this embodiment of the invention further includes:

[0282] Fatigue-related signals in multi-mode operation features are screened, and the cumulative fatigue amount is continuously calculated by weighted summation method;

[0283] When the accumulated fatigue reaches the preset fine-tuning trigger threshold, the height fine-tuning increment is calculated based on the accumulated fatigue.

[0284] When the height adjustment increment is not lower than the preset minimum perceptible height change threshold, a progressive height adjustment is performed based on the current desktop height until the desktop height increases by the height adjustment increment, and the fatigue accumulation is partially reset.

[0285] Specifically, fatigue-related signals are screened from multi-mode operation features, and the cumulative fatigue amount is continuously calculated by weighted summation.

[0286] Among them, fatigue-related signals are operational features related to user fatigue status within the multi-mode operational features. Based on historical user operation data, the degree of correlation between each multi-mode operational feature and fatigue-related behaviors can be statistically analyzed, and operational features with a correlation degree reaching a preset requirement are identified as fatigue-related signals.

[0287] For example, historical data shows that a decrease in keystroke frequency, an increase in keystroke interval, an increase in cursor stillness time, and a decrease in input continuity ratio are significantly associated with user fatigue. Therefore, these characteristics can be used as fatigue-related signals.

[0288] Furthermore, the cumulative fatigue amount is calculated based on the fatigue contribution and weight corresponding to each fatigue-related signal. Specifically:

[0289] Current fatigue accumulation = previous fatigue accumulation + sum of the products of fatigue contribution of each fatigue-related signal and its corresponding weight.

[0290] The weights of each fatigue-related signal are determined based on the degree of correlation between the corresponding fatigue-related signal and the fatigue state in the historical user operation data, and are normalized so that the sum of each weight is 1.

[0291] For example, three fatigue-related signals are obtained through filtering: keystroke frequency change, cursor stillness duration change, and input continuity ratio change, with corresponding weights of 0.40, 0.25, and 0.35, respectively. The current fatigue contribution amounts for each fatigue-related signal are 0.60, 0.40, and 0.50, respectively. Therefore, the increase in the current fatigue accumulation is 0.60 × 0.40 + 0.40 × 0.25 + 0.50 × 0.35 = 0.505. If the previous fatigue accumulation was 4.5, then the current fatigue accumulation is 4.5 + 0.505 = 5.005.

[0292] Furthermore, when the accumulated fatigue reaches the preset fine-tuning trigger threshold, the height fine-tuning increment is calculated based on the accumulated fatigue.

[0293] The fine-tuning trigger threshold is used to determine whether the accumulated fatigue level has reached the point where the desktop height needs to be adjusted. Its value is calibrated based on historical user operation data. Specifically, the correlation between different accumulated fatigue levels and users actively adjusting the desktop height is statistically analyzed, and the accumulated fatigue level that can achieve the preset trigger accuracy rate is determined as the fine-tuning trigger threshold.

[0294] For example, if historical data shows that when the cumulative fatigue level reaches 5, the probability of the user actively adjusting the desktop height reaches a preset requirement, then 5 can be used as the fine-tuning trigger threshold.

[0295] The fine-tuning increment is determined based on the degree to which the current accumulated fatigue exceeds the fine-tuning trigger threshold, specifically:

[0296] The increment for fine-tuning is calculated as follows: (Current accumulated fatigue - Fine-tuning trigger threshold) × Preset fine-tuning ratio.

[0297] The preset fine-tuning ratio is determined based on the correspondence between the degree of fatigue accumulation and the amount of change in the user's acceptable height in historical user data.

[0298] For example, if the fine-tuning trigger threshold is 5, the current fatigue accumulation is 7, and the preset fine-tuning ratio is 10mm / unit, then the height fine-tuning increment = (7-5)×10 = 20mm.

[0299] Furthermore, the height fine-tuning increment is compared with a preset minimum perceptible height change threshold.

[0300] The minimum perceptible height change threshold is determined based on the height adjustment resolution of the smart desk and the user's actual perception of different height changes. For example, if user testing determines that a 10mm height change can achieve a preset perceptible proportion, then 10mm is used as the minimum perceptible height change threshold in this embodiment.

[0301] If the height adjustment increment is lower than the minimum perceptible height change threshold, the current height adjustment will not be performed; if the height adjustment increment is not lower than the minimum perceptible height change threshold, a progressive height adjustment will be performed based on the current desktop height, increasing the desktop height by the height adjustment increment.

[0302] For example, if the current desktop height is 720mm and the height adjustment increment is 20mm, then the target height for the adjustment is determined to be 740mm. The desktop height is then progressively adjusted according to the approach rate of the aforementioned desktop control parameters, gradually increasing from 720mm to 740mm. The height adjustment ends when the actual desktop height reaches the target height or the preset height adjustment termination tolerance.

[0303] After completing this height fine-tuning, the fatigue accumulation is partially reset so that the fatigue accumulation level that was used to trigger this height fine-tuning will no longer participate in the subsequent fine-tuning triggering judgment according to the original accumulation level.

[0304] Specifically, the reset amount is determined based on the degree of fatigue accumulation corresponding to this height fine-tuning, and the reset amount is deducted from the current fatigue accumulation amount to obtain the partially reset fatigue accumulation amount.

[0305] For example, if the current fatigue accumulation is 7, and this fine-tuning has been triggered and completed, then a portion of the fatigue accumulation will be reset according to the preset reset rules; if the reset amount is 5, then the fatigue accumulation after partial reset = 7 - 5 = 2.

[0306] The fatigue accumulation is then updated based on the fatigue correlation signal in the multi-mode operation features. When the fatigue accumulation reaches the fine-tuning trigger threshold again and the calculated height fine-tuning increment reaches the minimum perceptible height change threshold, the next progressive height fine-tuning is executed.

[0307] In summary, this step performs constraint boundary checks on personalized desktop control parameters, implementing safety limits for parameters exceeding ergonomic safety ranges. Then, combined with the current control parameters of the desk, a weighted fusion smooth transition is performed according to a preset approach rate to generate the actual execution parameter sequence, driving the device to complete desktop height adjustment step by step, avoiding abrupt height changes. Simultaneously, this step also extracts fatigue-related signals from multi-mode operation features, weighted and accumulating fatigue accumulation. When trigger conditions are met and the height fine-tuning increment reaches a perceptible threshold, progressive fatigue fine-tuning is performed, and the fatigue accumulation is partially reset, achieving assisted height adjustment based on the user's fatigue state.

[0308] Example 2, as Figure 3 As shown, this invention provides a knowledge graph-based intelligent office desk parameter personalization setting system, including:

[0309] The operation feature extraction module 11 is used to collect operation signals generated by the user's operation of the human-computer interaction device and extract multi-mode operation features;

[0310] The work status determination module 12 is used to perform fusion classification based on multi-modal operation features to determine the user's current work status, wherein the work status is associated with a state confidence level.

[0311] The parameter fusion decision module 13 is used to combine the working status with the pre-built knowledge graph to perform parameter fusion decision-making and obtain personalized desktop control parameters for the smart office desk. The knowledge graph includes a domain general knowledge layer and an individual personalized layer.

[0312] The desktop height control module 14 is used to control the smart office desk to perform desktop height adjustment according to personalized desktop control parameters.

[0313] In this embodiment of the invention, the operation feature extraction module 11 is further configured to:

[0314] Collect human-computer interaction device signals generated by the user through the keyboard, and extract keyboard operation features, wherein the keyboard operation features include at least two of the following: keystroke frequency, keystroke duration, keystroke interval, and input continuity ratio;

[0315] Collect signals from human-computer interaction devices generated by the user through the mouse, and extract mouse operation features, including at least two of the following: cursor movement speed, click frequency, cursor still time ratio, and scrolling frequency.

[0316] Collect the application window context signal corresponding to the currently active foreground application and extract application context features, which include at least the application category and application switching frequency;

[0317] The combined output of keyboard operation features, mouse operation features, and application context features is a multi-modal operation feature.

[0318] In this embodiment of the invention, the working status determination module 12 is further configured to:

[0319] Based on the preset time window length and sliding step size, the multi-modal operation features are truncated by sliding window to generate the feature vector of the current time window;

[0320] The keyboard operation features and mouse operation features in the feature vector are used as the first modal input, and the application context features in the feature vector are used as the second modal input. A Bayesian fusion classification model is used to jointly infer the first modal input and the second modal input to calculate the posterior probability of each preset candidate working state.

[0321] The candidate working state with the highest posterior probability is determined as the user's current working state, and the corresponding posterior probability is used as the state confidence.

[0322] Specifically, when the state confidence level is not lower than a preset first confidence threshold, the working state is transmitted to participate in parameter fusion decision-making; when the state confidence level is not lower than a preset second confidence threshold but lower than the first confidence threshold, the working state is transmitted to participate in parameter fusion decision-making, and the parameter adjustment range in the parameter fusion decision-making is reduced to a preset reduction ratio; when the state confidence level is lower than the second confidence threshold, the current desktop control parameters are kept unchanged; wherein, the first confidence threshold is greater than the second confidence threshold.

[0323] In this embodiment of the invention, the parameter fusion decision module 13 is further configured to:

[0324] Based on the domain-wide general knowledge layer, the domain default desktop parameters corresponding to the working status are matched and obtained.

[0325] Based on the individual personalization layer, the personalized correction knowledge of the current user is obtained through matching. The personalized correction knowledge is represented by the individual preference nodes in the individual personalization layer. The individual preference nodes include at least the user's standing tolerance time, preferred sitting height, preferred standing height, preference switching cycle, preferred standing percentage, and preference data reliability.

[0326] Based on the current time period, obtain the corresponding context correction parameters;

[0327] The default desktop parameters, personalized correction knowledge, and context correction parameters are fused into three layers to generate personalized desktop control parameters. The personalized desktop control parameters include at least the target desktop height value and the height adjustment cycle.

[0328] The construction of knowledge graphs includes:

[0329] Define node types and attribute sets for each node type. Node types should include at least working status nodes, ergonomic rule nodes, domain default desktop parameter nodes, manual control log nodes, and individual preference nodes.

[0330] Instantiate each preset work status category as a work status node, instantiate the preset ergonomic rules as ergonomic rule nodes, and instantiate the default desktop parameters corresponding to each work status category as domain default desktop parameter nodes.

[0331] Define relationship types, which should include at least recommendation relationships, constraint relationships, personalized overlay relationships, and preference update relationships;

[0332] Establish a recommendation relationship between work status nodes and the domain default desktop parameter nodes, and a constraint relationship between ergonomic rule nodes and the domain default desktop parameter nodes;

[0333] Based on cross-user behavior similarity matching and preference propagation, an initial individual personalization layer for new users is constructed.

[0334] Among them, based on cross-user behavior similarity matching and preference propagation, an initial individual personalization layer for new users is constructed, including:

[0335] For each user, a user behavior feature vector is constructed. The user behavior feature vector includes at least the average keystroke frequency, working state distribution entropy, concentration of main working periods, average input continuity ratio, manual operation frequency, and standing early termination rate.

[0336] Calculate the cosine similarity of behavioral feature vectors between new users and existing users, where the individual personalization layer of existing users has been established and the reliability of preference data has reached a preset confidence threshold.

[0337] Select the first preset number of existing users with the highest cosine similarity as similar users;

[0338] The product of the cosine similarity of each similar user and the reliability of the preference data is used as the propagation weight. The personalized correction knowledge of each similar user is weighted, summed, and normalized to obtain the propagation preference vector.

[0339] Based on the cold start hybrid weight coefficient, the propagation preference vector and the zero offset vector are weighted and mixed to obtain the initial personalized correction knowledge of new users and construct the initial individual personalization layer;

[0340] The cold start hybrid weight coefficient is determined based on the ratio of the number of manual control logs accumulated by the new user to the preset minimum effective data volume, and is positively correlated with the ratio.

[0341] Acquire records of manual intervention operations performed by users on the smart desk and the operation context information when the manual intervention operations occurred;

[0342] For each preset operation intent category, the attribution confidence is calculated based on the corresponding feature conditions in the operation context information;

[0343] The operation intention category with the highest attribution confidence was used as the final operation intention category for the manual intervention operation record;

[0344] Based on the correction information corresponding to the final operational intent category, the individual personalized layer of the knowledge graph is updated using an exponential moving average method in conjunction with the preset learning rate, and the reliability of the preference data is updated simultaneously.

[0345] In this embodiment of the invention, the desktop height control module 14 is further configured to:

[0346] Perform constraint boundary checks on the personalized desktop control parameters and tighten the parameter values ​​that exceed the preset ergonomic safety range.

[0347] Based on the preset approach rate and combined with the current desktop control parameters, a weighted fusion-based smooth transition process is performed on the personalized desktop control parameters after constraint boundary checks to obtain the actual execution parameter sequence.

[0348] Based on the actual execution parameter sequence, the smart office desk is driven to adjust its desktop height.

[0349] Fatigue-related signals in multi-mode operation features are screened, and the cumulative fatigue amount is continuously calculated by weighted summation method;

[0350] When the accumulated fatigue reaches the preset fine-tuning trigger threshold, the height fine-tuning increment is calculated based on the accumulated fatigue.

[0351] When the height adjustment increment is not lower than the preset minimum perceptible height change threshold, a progressive height adjustment is performed based on the current desktop height until the desktop height increases by the height adjustment increment, and the fatigue accumulation is partially reset.

[0352] In summary, each module of this invention performs the steps of the aforementioned method. The operation feature extraction module completes the acquisition of multi-source human-computer interaction signals and the extraction of multi-mode operation features; the working state determination module obtains the user's working state with confidence through sliding window and Bayesian fusion classification, and processes the decision participation mode in a graded manner according to two-level confidence thresholds; the parameter fusion decision module completes knowledge matching and three-layer parameter fusion based on a two-layer knowledge graph, while simultaneously realizing knowledge graph construction, cold start of new user preferences, and iterative update of preferences based on manual intervention intent attribution; the desktop height control module completes parameter security verification, smooth transition processing, device driving, and auxiliary fine-tuning based on fatigue state. This system relies on the coordinated cooperation of each module to achieve automated and personalized adaptive adjustment of the desktop parameters of the smart office desk.

[0353] It should be noted that the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0354] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0355] This specification and accompanying drawings are merely illustrative examples of the invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its scope. Therefore, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is intended to include these modifications and modifications.

Claims

1. A method for personalized parameter setting of intelligent office desks based on knowledge graphs, characterized in that, include: S1: Collect the operation signals generated by the user's operation of the human-computer interaction device and extract multi-mode operation features; S2: Based on the multi-modal operation features, perform fusion classification to determine the user's current working state, wherein the working state is associated with a state confidence level; S3: Combining the working state with the pre-built knowledge graph, perform parameter fusion decision-making to obtain personalized desktop control parameters for the smart office desk, wherein the knowledge graph includes a domain-general knowledge layer and an individual personalized layer; S4: Based on the personalized desktop control parameters, control the smart office desk to perform desktop height adjustment.

2. The method for personalized parameter setting of a smart office desk based on knowledge graphs as described in claim 1, characterized in that, Collecting operation signals generated by the user's operation of the human-computer interaction device and extracting multi-modal operation features, step S1 includes: S1.1: Collect human-computer interaction device signals generated by the user through the keyboard, and extract keyboard operation features, wherein the keyboard operation features include at least two of the following: keystroke frequency, keystroke duration, keystroke interval, and input continuity ratio; S1.2: Collect human-computer interaction device signals generated by the user through the mouse, and extract mouse operation features, wherein the mouse operation features include at least two of the following: cursor movement speed, click frequency, cursor still time ratio, and scrolling frequency; S1.3: Collect the application window context signal corresponding to the currently active foreground application and extract the application context features, wherein the application context features include at least the application category and the application switching frequency; S1.4: Merge and output the keyboard operation features, the mouse operation features, and the application context features as the multi-mode operation features.

3. The method for personalized parameter setting of a smart office desk based on knowledge graphs as described in claim 1, characterized in that, Based on the multimodal operation features, a fusion classification is performed to determine the user's current work status, wherein the work status is associated with a state confidence level. Step S2 includes: S2.1: Based on the preset time window length and sliding step size, the multi-mode operation features are truncated by a sliding window to generate the feature vector of the current time window; S2.2: Using the keyboard operation features and mouse operation features in the feature vector as the first modal input, and the application context features in the feature vector as the second modal input, a Bayesian fusion classification model is used to perform joint inference on the first modal input and the second modal input to calculate the posterior probability of each preset candidate working state; S2.3: Determine the candidate working state with the highest posterior probability as the user's current working state, and use the corresponding posterior probability as the state confidence. Specifically, when the state confidence level is not lower than a preset first confidence threshold, the working state is transmitted to participate in parameter fusion decision-making; when the state confidence level is not lower than a preset second confidence threshold but lower than the first confidence threshold, the working state is transmitted to participate in parameter fusion decision-making, and the parameter adjustment range in the parameter fusion decision is reduced to a preset reduction ratio; when the state confidence level is lower than the second confidence threshold, the current desktop control parameters are kept unchanged; wherein, the first confidence threshold is greater than the second confidence threshold.

4. The method for personalized parameter setting of a smart office desk based on knowledge graphs as described in claim 1, characterized in that, Combining the described working state with a pre-built knowledge graph, parameter fusion decision-making is performed to obtain personalized desktop control parameters for the smart office desk. The knowledge graph includes a domain-wide general knowledge layer and an individual personalized layer. Step S3 includes: S3.1: Based on the general knowledge layer of the domain, match and obtain the domain default desktop parameters corresponding to the working state; S3.2: Based on the individual personalization layer, match and obtain the personalized correction knowledge of the current user, wherein the personalized correction knowledge is represented by the individual preference node in the individual personalization layer, and the individual preference node includes at least the user's standing tolerance time, preferred sitting height, preferred standing height, preference switching cycle, preferred standing percentage and preference data reliability. S3.3: Based on the current time period, obtain the corresponding context correction parameters; S3.4: Perform three-layer parameter fusion on the domain default desktop parameters, the personalized correction knowledge, and the context correction parameters to generate the personalized desktop control parameters, wherein the personalized desktop control parameters include at least the desktop target height value and the height adjustment cycle.

5. The method for personalized parameter setting of a smart office desk based on knowledge graphs as described in claim 1, characterized in that, The construction of the knowledge graph includes: Define node types and attribute sets for each node type. The node types include at least working status nodes, ergonomic rule nodes, domain default desktop parameter nodes, manual control log nodes, and individual preference nodes. Each preset work status category is instantiated as a work status node, each preset ergonomic rule is instantiated as an ergonomic rule node, and each default desktop parameter corresponding to each work status category is instantiated as a domain default desktop parameter node. Define relationship types, which include at least recommendation relationships, constraint relationships, personalized coverage relationships, and preference update relationships; Establish a recommendation relationship between the working status node and the domain default desktop parameter node, and a constraint relationship between the ergonomic rule node and the domain default desktop parameter node; Based on cross-user behavior similarity matching and preference propagation, an initial individual personalization layer for new users is constructed.

6. The method for personalized parameter setting of a smart office desk based on knowledge graphs as described in claim 5, characterized in that, Based on cross-user behavior similarity matching and preference propagation, an initial individual personalization layer for new users is constructed, including: For each user, a user behavior feature vector is constructed. The user behavior feature vector includes at least the average keystroke frequency, working state distribution entropy, concentration of main working periods, average input continuity ratio, manual operation frequency, and standing early termination rate. Calculate the cosine similarity of behavioral feature vectors between new users and existing users, wherein the individual personalization layer of the existing users has been established and the reliability of the preference data has reached a preset confidence threshold; Select the first preset number of existing users with the highest cosine similarity as similar users; The product of the cosine similarity of each similar user and the reliability of the preference data is used as the propagation weight. The personalized correction knowledge of each similar user is weighted, summed, and normalized to obtain the propagation preference vector. Based on the cold start hybrid weight coefficient, the propagation preference vector and the zero offset vector are weighted and mixed to obtain the initial personalized correction knowledge of the new user, and the initial individual personalization layer is constructed. The cold start hybrid weight coefficient is determined based on the ratio of the number of manual control logs accumulated by the new user to the preset minimum effective data volume, and is positively correlated with the ratio.

7. The method for personalized parameter setting of a smart office desk based on knowledge graphs as described in claim 1, characterized in that, Also includes: The fatigue-related signals in the multi-mode operation features are filtered out, and the fatigue accumulation amount is continuously calculated by weighted summation method. When the accumulated fatigue reaches the preset fine-tuning trigger threshold, the height fine-tuning increment is calculated based on the accumulated fatigue. When the height adjustment increment is not lower than the preset minimum perceptible height change threshold, a progressive height adjustment is performed based on the current desktop height until the desktop height increases by the height adjustment increment, and the fatigue accumulation is partially reset.

8. The method for personalized parameter setting of a smart office desk based on knowledge graphs as described in claim 1, characterized in that, Based on the personalized desktop control parameters, controlling the smart office desk to perform desktop height adjustment also includes: Perform constraint boundary checks on the personalized desktop control parameters and safely limit parameter values ​​that exceed the preset ergonomic safety range; According to the preset approach rate, combined with the current desktop control parameters, the personalized desktop control parameters after constraint boundary checks are subjected to a weighted fusion-based smooth transition process to obtain the actual execution parameter sequence. Based on the actual execution parameter sequence, the smart office desk is driven to adjust its desktop height.

9. The method for personalized parameter setting of a smart office desk based on knowledge graphs as described in claim 1, characterized in that, Also includes: Acquire records of manual intervention operations performed by the user on the smart office desk and the operation context information when the manual intervention operations occurred; For each preset operation intent category, the attribution confidence is calculated based on the corresponding feature conditions in the operation context information; The operation intention category with the highest attribution confidence was used as the final operation intention category for the manual intervention operation record; Based on the correction information corresponding to the final operation intention category, the individual personalization layer of the knowledge graph is updated using an exponential moving average method in conjunction with the preset learning rate, and the reliability of the preference data is updated simultaneously.

10. A knowledge graph-based intelligent office desk parameter personalization setting system, characterized in that, A method for personalizing the parameters of a smart office desk based on a knowledge graph, as described in any one of claims 1 to 9, includes: The operation feature extraction module is used to collect operation signals generated by users operating human-computer interaction devices and extract multi-mode operation features; The work status determination module is used to perform fusion classification based on the multi-modal operation features to determine the user's current work status, wherein the work status is associated with a state confidence level; The parameter fusion decision module is used to combine the working state with the pre-built knowledge graph to perform parameter fusion decision-making and obtain personalized desktop control parameters for the smart office desk. The knowledge graph includes a domain-general knowledge layer and an individual personalized layer. The desktop height control module is used to control the smart office desk to adjust its desktop height according to the personalized desktop control parameters.

Images