An instruction semantic guided elevator heterogeneous data target fusion method

Abstract

The present application relates to the elevator technical field, especially a kind of instruction semantic guidance's elevator heterogeneous data target fusion method.Elevator control log is extracted in current discrete control instruction time stamp as instruction trigger time;With the instruction response time window based on instruction trigger time, intercept sensor monitoring signal as instruction response signal;Current discrete control instruction generates one-dimensional semantic vector through semantic coding;Instruction response signal generates two-dimensional time sequence feature matrix through time sequence convolution feature extraction and down-sampling;One-dimensional semantic vector is used as query vector, to generate key matrix and value matrix with two-dimensional time sequence feature matrix, calculate attention weight distribution and target interaction feature;With the attention weight distribution to determine adaptive gating coefficient, fuse one-dimensional semantic vector and target interaction feature, after processing obtain the target fusion feature fingerprint for elevator running state evaluation.The present application can realize the efficient, stable fusion of elevator control log and sensor monitoring signal.

Description

Technical Field

[0001] This invention relates to the field of elevator technology, and in particular to a method for targeted fusion of heterogeneous elevator data guided by instruction semantics. Background Technology

[0002] As elevator equipment evolves towards higher reliability, intelligence, and lower maintenance, the demand for monitoring the operational status of key components such as elevator traction machines and brakes is constantly increasing. Elevator IoT monitoring systems typically require simultaneous acquisition of control logs output from the elevator main control board and sensor monitoring signals mounted on mechanical components. Control logs are mostly semi-structured data, containing discrete control commands, equipment status, and timestamps; sensor monitoring signals are mostly continuous time-series data, such as traction machine vibration signals, acceleration signals, and current signals. These two types of data differ significantly in data format, sampling frequency, time granularity, and physical meaning, making direct synchronous fusion challenging.

[0003] Existing methods typically use a sliding window of fixed length to capture sensor monitoring signals on a global time axis, and then correlate these captured signal segments with control logs for analysis. While simple to implement, this approach is unsuitable for the actual operating characteristics of elevators. Elevators experience numerous periods of standby, constant speed operation, or inconspicuous status changes during operation. Sensor signals during these periods usually lack diagnostic value for specific component actions. Continuously processing all signals using a global window approach would introduce a large amount of irrelevant data, increase the processing burden on edge computing devices, and reduce the specificity of subsequent status analysis.

[0004] Furthermore, the physical response times for different control commands vary. For example, actions such as brake opening and contactor engagement have short durations, while processes like door opening, closing, starting, braking, and leveling take longer. If sensor signals are extracted based on the actual response process of each command, the resulting signal segments often have inconsistent lengths. For fusion methods that require matrix operations, feature splicing, or attention calculations, inconsistent input lengths can easily lead to inconsistent feature dimensions, affecting the normal progress of subsequent fusion calculations.

[0005] Meanwhile, the elevator environment is complex, and sensor monitoring signals are easily affected by electromagnetic interference, mechanical shock, loose wiring, abnormal acquisition modules, or sensor disconnection. Under these circumstances, sensor data may exhibit problems such as strong noise, waveform distortion, and missing valid signals. Existing fusion methods typically rely on the combined input of control logs and sensor monitoring signals. When the quality of sensor-side data deteriorates, it can easily affect the stability of the fused features and the reliability of the status judgment results.

[0006] Therefore, it is necessary to propose a method for fusion of heterogeneous elevator data that can effectively intercept control commands, uniformly represent response signals of different lengths, and maintain basic fusion stability under sensor malfunctions. Summary of the Invention

[0007] The purpose of this invention is to propose a command semantic-guided targeted fusion method for heterogeneous elevator data. Without relying on a global sliding time window to continuously scan sensor monitoring signals, the method extracts the corresponding command response signal based on the triggering time of the elevator control command. Through multi-scale feature extraction, adaptive length alignment, and an adaptive gating mechanism based on attention information entropy, the method achieves stable fusion of control logs and sensor monitoring signals, providing targeted and robust fused feature fingerprint data for elevator operation status assessment, anomaly detection, and fault diagnosis.

[0008] To achieve this objective, the present invention adopts the following technical solution:

[0009] A method for targeted fusion of heterogeneous elevator data guided by instruction semantics includes the following steps:

[0010] S1: Obtain elevator control logs and sensor monitoring signals, and extract the timestamp of the current discrete control command in the control log as the command trigger time;

[0011] S2: Based on the command triggering time, construct a command response time window according to a preset pre-set time and a preset post-set time, and extract the sensor monitoring signal within the command response time window as the command response signal;

[0012] S3: Semantically encode the current discrete control command to generate a one-dimensional semantic vector representing the control intent;

[0013] S4: Perform temporal convolution feature extraction and downsampling on the instruction response signal to generate a two-dimensional temporal feature matrix;

[0014] S5: Use the one-dimensional semantic vector as the query vector, perform linear projection on the two-dimensional temporal feature matrix to generate a key matrix and a value matrix, and obtain the attention weight distribution and targeted interaction features through attention calculation;

[0015] S6: Calculate the normalized information entropy based on the attention weight distribution, and determine the adaptive gating coefficient through the normalized information entropy;

[0016] S7: Based on the adaptive gating coefficient, the one-dimensional semantic vector and the targeted interaction feature are fused together. After residual splicing, normalization processing and low-dimensional mapping, a targeted fusion feature fingerprint for elevator operation status evaluation is output.

[0017] Further, step S2 includes:

[0018] S21: Triggered by the aforementioned instruction Based on the time reference, according to the preset pre-set duration and preset post-processing duration Calculate the start time respectively and termination time According to the start time and termination time Constructing instruction response time window ;

[0019] S22: The start timestamp of the data stream monitored by the sensor The zero-point time of the discrete sampling sequence will be used as the instruction response time window. The start time and termination time Convert them into starting sampling point indices respectively. and the index of the terminating sampling point ;

[0020] S23: Based on the starting sampling point index and the index of the terminating sampling point The sensor monitors the signal. Perform interval truncation to obtain a time series length of... Command response signal .

[0021] Further, step S3 includes:

[0022] S31: After deduplicating the historical discrete control commands in the elevator control log, obtain a set of commands containing C different discrete control command string codes. Then, assign a unique integer feature index to each discrete control command string code according to a preset sorting rule; establish a command encoding dictionary with the discrete control command string code as the query key and the corresponding integer feature index as the mapping value.

[0023] The target feature index is obtained by inputting the current discrete control instruction string code into the instruction encoding dictionary. Based on the one-hot encoding rule, through the target feature index... Build length is One-hot sparse eigenvectors ;

[0024] S32: Constructing a cross-modal embedding matrix It is used to store the primary continuous semantic representations corresponding to each standard discrete control command, and the number of columns is equal to the preset primary semantic vector length. The uniquely heated sparse feature vector With the cross-modal embedding matrix Perform matrix multiplication to obtain primary continuous semantic vectors. ;

[0025] S33: For the primary continuous semantic vector Linear projection and nonlinear activation are performed to generate a one-dimensional semantic vector representing the control intent. .

[0026] Further, step S4 includes:

[0027] S41: Constructing a system containing A one-dimensional temporally dilated convolutional network with parallel branches, and for the first... Each parallel branch is configured with a corresponding hole rate. The instruction response signal with a time series length of T is input into the one-dimensional temporal dilated convolutional network, and each parallel branch is processed according to the corresponding dilation rate. Temporal feature extraction is performed, and zero-padding is used during the feature extraction process to obtain N sets of intermediate temporal feature sequences of length T.

[0028] S42: In the feature channel dimension, the aforementioned The intermediate time-series feature sequences of the group are concatenated and spliced ​​to generate a dimension of size. Multi-scale fusion feature matrix By using a linear projection that includes a channel alignment projection matrix and a bias vector, the feature channel dimension of the multi-scale fused feature matrix is ​​mapped from N to a target latent space dimension D consistent with the feature dimension of the one-dimensional semantic vector, generating a channel alignment feature matrix of dimension T×D. ;

[0029] S43: Align the channels with the feature matrix Input a one-dimensional adaptive max pooling layer and perform downsampling along the time step dimension to reduce the length of the time series from... Unified transformation to a preset target sequence length And maintain the target latent space dimension The generated dimension size remains unchanged. Two-dimensional time series feature matrix .

[0030] Furthermore, in step S41, the parallel branches are arranged according to the corresponding void ratio. The method for extracting temporal features is as follows:

[0031] The first in the one-dimensional temporally dilated convolutional network Several parallel branches, indexed at discrete time steps. Discrete feature elements output at the location The following is calculated using one-dimensional dilated convolution:

[0032] ,

[0033] in, This represents the length of a one-dimensional dilated convolution kernel, i.e., the number of kernel weights involved in each convolution operation; This represents the sampling point number inside the convolution kernel, satisfying... ; Indicates the first Hollowness rate of parallel branches; Indicates from the first Starting at each time step, according to the void ratio The interval selected The signal values ​​at each sampling point; Indicates the first In the convolution kernel of the nth parallel branch Each weighting coefficient; Indicates the first The bias terms corresponding to each parallel branch; This indicates a modified linear unit activation function, which performs non-linear activation processing on the convolution calculation results; Represents the discrete time step index, and ;

[0034] Traversing the first Index of all discrete time steps in each parallel branch Discrete feature elements output at each time step Arranged in chronological order, the time series has a length of [length missing]. The Intermediate time series feature sequences ;right Each parallel branch performs the above processing, resulting in... Intermediate time series feature sequences of the group: ;

[0035] In step S42, the channel alignment feature matrix Located in the time step Target Channel Discrete feature elements at the location The calculation formula is:

[0036] ,

[0037] in, In the multi-scale fusion feature matrix, the first... The time step, the first Feature values ​​at each feature channel; Indicates the first channel alignment projection matrix. Line 1 The weighted element of the column; Represents the first in the bias vector One biased element; Represents the dimension of the target's latent space. Represents the target channel index, satisfying ;

[0038] Traverse all time steps and all target channel indexes The calculated discrete feature elements Arrange and combine them in the time sequence dimension and the target channel dimension to form a complete channel alignment feature matrix;

[0039] In step S43, the two-dimensional time-series feature matrix The Middle The target time step and the first Discrete target elements at each target channel Calculated using the following pooling sampling formula:

[0040] ,

[0041] in, This indicates that the one-dimensional adaptive max pooling layer has a time series length of [missing information]. The above is the generation of the first The pooling interval, divided according to the target time steps, is expressed as follows:

[0042] ,

[0043] in, Represents the channel alignment feature matrix Located in the middle The time step, the first Feature values ​​at each target channel; Represents the channel alignment feature matrix The length of the time series; Indicates the preset target time series length; Represents the latent space dimension of the target; This represents the target time step index after downsampling, and ; Indicates the target channel index, and ; Indicates the pooling interval Take the maximum value within the range; This represents the floor function; This represents the function for rounding up.

[0044] Further, step S5 includes:

[0045] S51: Define the one-dimensional semantic direction as the query vector Q in the cross-modal attention mechanism; and project it using a pre-defined key matrix. Sum projection matrix Linear matrix projection is performed on the two-dimensional time-series feature matrix to generate the key matrix. Sum matrix ;

[0046] S52: Combine the query vector Q with the transpose key matrix The similarity score sequence is obtained by performing matrix dot product operation, and after being scaled by a scaling factor, the attention weight distribution is generated by the Softmax normalization function.

[0047] S53: Apply the attention weight distribution to the value matrix Weighted aggregation is performed along the time step dimension to generate targeted interactive features associated with the semantic features of the current discrete control command.

[0048] Furthermore, in step S52, a scaling factor is used. The similarity score sequence is scaled, and an attention weight distribution is generated using the Softmax normalization function. The calculation formula is:

[0049] ,

[0050] in, This represents the distribution of attention weights; Represents the latent space dimension of the target; This is a scaling factor used to normalize the scale of the dot product similarity results; This represents the exponential normalization function, which is used to convert the similarity scores at each time step into non-negative attention weights, and makes the sum of the attention weights at all time steps equal to 1.

[0051] Further, step S6 includes:

[0052] S61: Extract the first value from the attention weight distribution. Attention weights corresponding to each time step Based on the information entropy calculation method, the attention weight distribution is... Attention weights in Entropy calculation is performed to obtain the original attention information entropy, which characterizes the degree of dispersion of the attention distribution. ;

[0053] S62: The original attention information entropy... Dimension elimination and scaling are performed to calculate the normalized information entropy. ;

[0054] S63: Preset a high-entropy failure detection threshold in the system operating parameters. and gate slope coefficient The normalized information entropy Input to based The nonlinear gated mapping function is constructed to generate adaptive gate coefficients. .

[0055] Furthermore, in step S63, the formula for calculating the adaptive gating coefficient is:

[0056] ,

[0057] in, Denotes the adaptive gating coefficient, and ; This represents the normalized information entropy, used to characterize the degree of dispersion in the attention weight distribution; This represents the threshold for determining high-entropy failure. Indicates the gating slope coefficient; This represents the natural exponential function.

[0058] Further, step S7 includes:

[0059] S71: Perform gated attenuation processing on the targeted interaction features according to the adaptive gating coefficient to generate controlled interaction features;

[0060] S72: The one-dimensional semantic vector generated in step S3 is used as the semantic residual benchmark and concatenated with the controlled interaction feature in the feature channel dimension to generate a cross-modal concatenated feature vector.

[0061] S73: Concatenate the cross-modal feature vector Perform layer normalization to generate a voltage-stabilized feature tensor;

[0062] S74: Construct the output projection matrix and output bias vector, perform low-dimensional linear mapping on the voltage regulation feature tensor, and generate a targeted fusion feature fingerprint for elevator operation status assessment.

[0063] The technical solution provided by this invention may include the following beneficial effects:

[0064] 1. Construct an instruction response time window based on the elevator control instruction trigger time, and only capture sensor monitoring signals related to the current control instruction. Avoid blindly sliding the window to process global continuous signals, thereby reducing invalid background data and redundant calculations, and improving the targeting and processing efficiency of heterogeneous data fusion.

[0065] 2. By using multi-scale dilated convolution and adaptive pooling to extract features and unify the length of the command response signal, the variable-length response signals corresponding to different control commands are converted into two-dimensional temporal feature matrices of consistent size, thereby solving the problem of inconsistent feature dimensions caused by inconsistent response times of different commands and ensuring that subsequent fusion calculations can be performed stably.

[0066] 3. The discrete control command is encoded into a one-dimensional semantic vector representing the control intention, and the attention calculation of the sensor temporal features is guided by the semantic vector. This enables the fusion process to focus on the key physical response segments related to the current control command, reduce the interference of irrelevant temporal information on the fusion result, and improve the relevance and interpretability of the fusion features.

[0067] 4. An adaptive gating mechanism is constructed based on the information entropy of attention weight distribution. When the sensor signal quality is good and the key response is clear, the temporal interaction features are fully utilized. When the sensor signal is affected by noise or the effective response is not obvious, uncertain features are suppressed. Combined with semantic residuals, normalization and low-dimensional mapping, a stable targeted fusion feature fingerprint is output, thereby improving the robustness of the fusion results under abnormal conditions.

[0068] Through the above-mentioned technical means, the present invention can achieve efficient and stable fusion of control logs and sensor monitoring signals during elevator operation, and output targeted fusion feature fingerprints with clear command correlation and state representation capabilities, providing a reliable data foundation for elevator operation status assessment, anomaly detection and fault diagnosis, and has good engineering practical value and promotion prospects. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating the elevator heterogeneous data targeted fusion method guided by instruction semantics of the present invention.

[0070] Figure 2 This is a schematic diagram of obtaining elevator control logs and sensor monitoring signals in step S1;

[0071] Figure 3 This is a flowchart illustrating the process of obtaining the instruction trigger time, constructing the instruction response time window, and intercepting the instruction response signal in step S2.

[0072] Figure 4 This is a flowchart illustrating the process of generating a one-dimensional semantic vector representing the control intent in step S3.

[0073] Figure 5 This is a flowchart illustrating step S4, which involves extracting temporal convolutional features from the instruction response signal and downsampling to generate a two-dimensional temporal feature matrix.

[0074] Figure 6This is a flowchart illustrating the process of obtaining attention weight distribution and targeted interaction features in step S5.

[0075] Figure 7 This is a flowchart illustrating the process of determining the adaptive gating coefficient in step S6.

[0076] Figure 8 This is a flowchart illustrating the process of obtaining the targeted fusion feature fingerprint in step S7. Detailed Implementation

[0077] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the present invention.

[0078] Reference Figure 1-8 An embodiment of the present invention provides a method for targeted fusion of heterogeneous elevator data guided by instruction semantics, comprising the following steps:

[0079] S1: Obtain elevator control logs and sensor monitoring signals, and extract the timestamp of the current discrete control command in the control log as the command trigger time;

[0080] S2: Based on the command triggering time, construct a command response time window according to a preset pre-set time and a preset post-set time, and extract the sensor monitoring signal within the command response time window as the command response signal;

[0081] S3: Semantically encode the current discrete control command to generate a one-dimensional semantic vector representing the control intent;

[0082] S4: Perform temporal convolution feature extraction and downsampling on the instruction response signal to generate a two-dimensional temporal feature matrix;

[0083] S5: Use the one-dimensional semantic vector as the query vector, perform linear projection on the two-dimensional temporal feature matrix to generate a key matrix and a value matrix, and obtain the attention weight distribution and targeted interaction features through attention calculation;

[0084] S6: Calculate the normalized information entropy based on the attention weight distribution, and determine the adaptive gating coefficient through the normalized information entropy;

[0085] S7: Based on the adaptive gating coefficient, the one-dimensional semantic vector and the targeted interaction feature are fused together. After residual splicing, normalization processing and low-dimensional mapping, a targeted fusion feature fingerprint for elevator operation status evaluation is output.

[0086] This invention introduces a heterogeneous data targeted fusion mechanism guided by instruction semantics. Instead of blindly sliding a window across sensor monitoring signals at the signal processing front end, it uses the trigger time of discrete control commands in the elevator control log as the time base, constructing an instruction response time window by combining preset pre- and post-time durations. Only sensor monitoring signals related to the discrete control commands within this window are extracted as instruction response signals, thereby reducing interference from irrelevant background data such as standby and constant-speed operation on subsequent fusion calculations. Furthermore, this invention encodes discrete control commands into one-dimensional semantic vectors representing control intentions, and performs multi-scale temporal feature extraction and adaptive length alignment on the instruction response signals, enabling different control commands to achieve targeted fusion. Corresponding command response signals of inconsistent lengths can be uniformly converted into two-dimensional temporal feature matrices of consistent size. Based on this, the two-dimensional temporal feature matrices are guided by the one-dimensional semantic vector to perform cross-modal attention calculations, obtaining targeted interaction features associated with the current discrete control command. Adaptive gating coefficients are determined based on the information entropy of the attention weight distribution. When the key physical response is clear, the sensor temporal features are fully utilized. When the sensor signal is affected by noise, the effective response is not obvious, or abnormal fluctuations occur, uncertain interaction information is suppressed. This results in the output of targeted fusion feature fingerprints with clear command correlation and high stability, providing a reliable data foundation for elevator operation status assessment, anomaly detection, and fault diagnosis.

[0087] In this embodiment of the invention, the sensor monitoring signal is illustrated using vibration signals collected on the surface of the elevator traction machine as an example. Specifically, a high-frequency vibration sensor is installed on the surface of the elevator traction machine (such as the traction machine base or a rigid housing near the traction sheave), and the vibration signals of the traction machine are continuously collected as sensor monitoring signals during elevator operation. (Refer to...) Figure 2 Step S1 involves synchronously acquiring the control log output from the elevator control cabinet's main control board and the traction machine vibration monitoring signal collected by the vibration sensor via a data acquisition device. The control log includes discrete control commands such as start, brake, leveling, brake open, brake closed, traction machine start, and traction machine stop, along with their corresponding timestamps. During elevator operation, the system continuously receives the control log and traction machine vibration monitoring signal. When a discrete control command is detected, the timestamp corresponding to the current discrete control command is extracted as the command trigger time. The timestamp represents the time point when the discrete control command occurs, using the format "hour:minute:second.millisecond," such as "14:30:15.100" representing 14:30:15:100 on the current day, used for subsequent time window construction and sampling index conversion. As an example, the data acquisition device acquires the current discrete control command string code CMD_BRAKE_OPEN, representing brake open, and accurately extracts the command trigger time corresponding to this code. =14:30:15.100.

[0088] Step S2 of this invention is used to extract the corresponding command response signal from the continuous sensor monitoring signal according to the command triggering time, so as to characterize the physical response changes of key elevator components before and after the discrete control command is triggered, reduce interference from irrelevant background data, and provide targeted input for subsequent time-series feature extraction, cross-modal fusion, and operation status evaluation. (Refer to...) Figure 3 In one embodiment of the present invention, step S2 includes:

[0089] S21: Triggered by the aforementioned instruction Based on the time reference, according to the preset pre-set duration and preset post-processing duration Calculate the start time respectively and termination time According to the start time and termination time Constructing instruction response time window ;

[0090] S22: The start timestamp of the data stream monitored by the sensor The zero-point time of the discrete sampling sequence will be used as the instruction response time window. The start time and termination time Convert them into starting sampling point indices respectively. and the index of the terminating sampling point The calculation formula is as follows:

[0091] ,

[0092] ,

[0093] in, This represents the floor function; This represents the floor function; The sampling frequency of the sensor monitoring signal; This represents the total number of sampling points for the sensor-monitored signals. Used to index the starting sampling point Limit the range to no less than 0 to prevent the starting sampling point index from going out of bounds; Used to index the termination sampling point Limited to no more than Within the specified range, prevent the termination sampling point index from going out of bounds;

[0094] S23: Based on the starting sampling point index and the index of the terminating sampling point The sensor monitors the signal. Perform interval truncation to obtain a time series length of... Command response signal The calculation formula is as follows:

[0095] ,

[0096] in, Indicates the sensor monitoring signal at the first The signal value at each sampling point Indicates the sampling point index; This indicates that the index of the truncated sampling point is greater than or equal to Less than or equal to Sensor monitoring signals; time series length .

[0097] In this embodiment of the invention, when performing step S21, the system first reads the trigger time of the current discrete control command from the elevator control log. The discrete control commands can be control commands that cause changes in the physical state of the elevator, such as starting, braking, leveling, brake opening, brake closing, traction machine starting, and traction machine stopping. To simultaneously cover the state changes before and after the discrete control commands are issued, the system presets a pre-load time. and preset post-processing duration The preset pre-set time is used to retain a short-term baseline state before the instruction is triggered, and the preset post-set time is used to cover the dynamic response process after the instruction is triggered. The preset pre-set time... and preset post-processing duration The start time is predetermined based on the discrete control command type, the response duration of the monitored component, and the range of changes before and after command triggering in historical sensor monitoring signals. and termination time The calculation formula is as follows:

[0098] ,

[0099] .

[0100] As an example, when the trigger time of the brake opening command in the elevator control log is... The preset pre-roll time is The preset post-processing time is At this point, the starting time can be calculated as The termination time is The constructed instruction response time window is as follows: .

[0101] If the data stream start timestamp of the vibration sensor monitoring signal is The sampling frequency is The total number of sampling points is ;at this time, , The starting sampling point index is:

[0102] ;

[0103] The index of the termination sampling point is:

[0104] .

[0105] Based on this, vibration data between the 15050th and 15550th sampling points were extracted from the vibration sensor monitoring signals to obtain a command response signal with a length of 501 sampling points. This command response signal includes the reference vibration state of the traction machine within 0.05 seconds before the door closing command is triggered, and the vibration change process caused by the traction machine brake action, traction machine load change, or traction system response within 0.45 seconds after the door closing command is triggered. This provides a data foundation for subsequently establishing the correspondence between discrete control commands and the physical response of the traction machine.

[0106] Step S3 of this invention converts the current discrete control command from its string form in the log into a one-dimensional semantic vector that can participate in computation. Through command encoding, cross-modal embedding, and semantic projection processing, different types of control commands are converted into continuously numerical control intent representations, thereby solving the problem that discrete control commands are difficult to directly fuse with sensor monitoring signals. Therefore, subsequent steps can use the one-dimensional semantic vector as semantic guidance information to perform targeted feature extraction and fusion of command response signals, improving the targeting and interpretability of heterogeneous data fusion. (Refer to...) Figure 4 In one embodiment of the present invention, step S3 includes:

[0107] S31: After deduplicating the historical discrete control instructions in the elevator control log, the instruction set contains C different discrete control instruction string codes. Then, a unique integer feature index is assigned to each discrete control instruction string code according to the preset sorting rules. An instruction encoding dictionary is established with the discrete control instruction string code as the query key and the corresponding integer feature index as the mapping value.

[0108] The target feature index is obtained by inputting the current discrete control instruction string code into the instruction encoding dictionary. Based on the one-hot encoding rule, through the target feature index... Build length is One-hot sparse eigenvectors ;

[0109] S32: Constructing a cross-modal embedding matrix It is used to store the primary continuous semantic representations corresponding to each standard discrete control command, and the number of columns is equal to the preset primary semantic vector length. The uniquely heated sparse feature vector With the cross-modal embedding matrix Perform matrix multiplication to obtain primary continuous semantic vectors. Among them, the cross-modal embedding matrix The initial element values ​​are determined according to preset initialization rules, preferably within a preset value range. The internal parameters are generated randomly using a uniform distribution to ensure that the initial embedding parameters have small amplitudes and initial differences in both positive and negative directions;

[0110] S33: For the primary continuous semantic vector Linear projection and nonlinear activation are performed to generate a one-dimensional semantic vector representing the control intent. .

[0111] In this embodiment of the invention, discrete control instruction string codes that appeared during historical operation are first extracted from the elevator control log, which may include door opening instructions. Door closing command Traction machine start command Traction machine stop command Brake opening command Brake closing command Leveling start command and leveling end command Standard instruction codes that can characterize elevator control actions. The following instruction code dictionary can be established:

[0112]

[0113] If the current discrete control command is At that time, the system inputs the instruction string code into the instruction encoding dictionary for querying, and obtains the corresponding target feature index as 4. Subsequently, a length of [missing information] is constructed based on the one-hot encoding rule. One-hot sparse eigenvectors .when At that time, the uniquely hot sparse eigenvector is:

[0114] ,

[0115] In this configuration, the element at index 4 has a value of 1, while the other elements have a value of 0, indicating that the current discrete control command belongs to... Categories. Through this process, discrete control commands of different categories are converted into mutually distinguishable numerical codes.

[0116] In step S32, the cross-modal embedding matrix The number of lines is equal to the total number of standard discrete control instruction types in the instruction prior dictionary. The number of columns is equal to the preset length of the primary semantic vector. As an example, take , Then the cross-modal embedding matrix for A two-dimensional matrix, where each row corresponds to a primary continuous semantic representation of a standard discrete control instruction. The one-hot sparse eigenvectors... With the cross-modal embedding matrix Perform matrix multiplication to obtain primary continuous semantic vectors. :

[0117] ,

[0118] ,

[0119] The matrix multiplication operation is equivalent to the operation based on the target feature index. Extract the cross-modal embedding matrix The corresponding row vectors in the code transform discrete, isolated text symbols into primary continuous semantic vectors with continuous numerical representations; for the current control command... The target feature index is 4, therefore the primary continuous semantic vector is equivalent to extracting the cross-modal embedding matrix. The 16-dimensional row vector in the 4th row of the diagram.

[0120] Step S33 is used to enable the semantic features of the control instructions to match the subsequent temporal feature matrix in terms of feature dimensions. The system further processes the primary continuous semantic vector. Linear projection and nonlinear activation are performed to generate a one-dimensional semantic vector representing the control intent. :

[0121] ,

[0122] ,

[0123] in, The semantic projection weight matrix is ​​predefined, and the initial element values ​​are determined according to predefined initialization rules, preferably within a predefined range. The internal parameters are generated randomly using a uniform distribution to ensure that the initial embedding parameters have small amplitudes and initial differences in both positive and negative directions; The initial element values ​​are determined according to the preset semantic projection bias vector. Each element can be initialized to 0, a preset constant value, or a historical sample calibration value. It is a Gaussian error linear unit function, belonging to nonlinear activation functions; The feature dimension of the one-dimensional semantic vector; and Used to set the feature dimension as Primary continuous semantic vectors Mapped to feature dimension One-dimensional semantic vector Furthermore, the parameters remain fixed during the semantic projection process.

[0124] Step S4 of this invention is used to process the temporal features of the intercepted command response signal. Since the response processes corresponding to different control commands may differ in duration and fluctuation patterns, directly using the original signal is not convenient for subsequent fusion calculations. Therefore, features such as local impact, vibration changes, and attenuation response in the command response signal are extracted through temporal convolution, and then converted into a uniform-sized two-dimensional temporal feature matrix through downsampling. This preserves the main dynamic features of the elevator's physical response after command triggering and provides a consistent and structurally stable temporal input for subsequent fusion calculations with the command semantic vector. (Refer to...) Figure 5 In one embodiment of the present invention, step S4 includes:

[0125] S41: Constructing a system containing A one-dimensional temporally dilated convolutional network with parallel branches, and for the first... Each parallel branch is configured with a corresponding hole rate. The instruction response signal with a time series length of T is input into the one-dimensional temporal dilated convolutional network, and each parallel branch is processed according to the corresponding dilation rate. Temporal feature extraction is performed, and zero-padding is used during the feature extraction process to obtain N sets of intermediate temporal feature sequences of length T.

[0126] S42: In the feature channel dimension, the aforementioned The intermediate time-series feature sequences of the group are concatenated and spliced ​​to generate a dimension of size. Multi-scale fusion feature matrix ; through a channel-aligned projection matrix and bias vector The linear projection maps the feature channel dimension of the multi-scale fused feature matrix from N to the target latent space dimension D, which is consistent with the feature dimension of the one-dimensional semantic vector, generating a channel-aligned feature matrix of dimension T×D. ;

[0127] S43: Align the channels with the feature matrix Input a one-dimensional adaptive max pooling layer and perform downsampling along the time step dimension to reduce the length of the time series from... Unified transformation to a preset target sequence length And maintain the target latent space dimension The generated dimension size remains unchanged. Two-dimensional time series feature matrix .

[0128] Specifically, in step S41, the parallel branches are arranged according to the corresponding hole ratio. The method for extracting temporal features is as follows:

[0129] The first in the one-dimensional temporally dilated convolutional network Several parallel branches, indexed at discrete time steps. Discrete feature elements output at the location The following is calculated using one-dimensional dilated convolution:

[0130] ,

[0131] in, This represents the length of a one-dimensional dilated convolution kernel, i.e., the number of kernel weights involved in each convolution operation; This represents the sampling point number inside the convolution kernel, satisfying... ; Indicates the first Hollowness rate of parallel branches; Indicates from the first Starting at each time step, according to the void ratio The interval selected The signal values ​​at each sampling point; Indicates the first In the convolution kernel of the nth parallel branch Each weighting coefficient, preferably with the initial range set to [value]. For each weight position m, a real number is randomly generated from the range according to a uniform distribution as the corresponding weight coefficient; Indicates the first The bias term corresponding to each parallel branch is preferably initialized to 0, and the next best setting is... Real numbers randomly generated within a uniform distribution within the range are used to perform translation correction on the weighted summation result of the convolution of the parallel branch; This indicates a modified linear unit activation function, which performs non-linear activation processing on the convolution calculation results; Represents the discrete time step index, and ;

[0132] Traversing the first Index of all discrete time steps in each parallel branch Discrete feature elements output at each time step Arranged in chronological order, the time series has a length of [length missing]. The Intermediate time series feature sequences ;right Each parallel branch performs the above processing, resulting in... Intermediate time series feature sequences of the group: ;

[0133] In step S42, the channel alignment feature matrix Located in the time step Target Channel Discrete feature elements at the location The calculation formula is:

[0134] ,

[0135] in, In the multi-scale fusion feature matrix, the first... The time step, the first Feature values ​​at each feature channel; Indicates the first channel alignment projection matrix. Line 1 For the weighted elements of a column, it is preferable to set the initial range to 1. For each weight element, a real number is randomly generated from the range according to a uniform distribution to represent the weighted contribution when the nth multi-scale convolutional channel feature is mapped to the dth target latent space dimension. Represents the first in the bias vector Each bias element is preferably initialized to 0, and the next best setting is... Real numbers randomly generated in a uniform distribution within the range are used to perform translation correction on the linear projection result mapped to the d-th target latent space dimension, so as to improve the adaptability of the channel alignment feature matrix to time series features of different amplitude levels. Represents the dimension of the target's latent space. Represents the target channel index, satisfying ;

[0136] Traverse all time steps and all target channel indexes The calculated discrete feature elements Arranged and combined along the temporal and target channel dimensions, a complete channel alignment feature matrix is ​​formed, wherein the target latent space dimension... Set as the one-dimensional semantic vector in step S3 The feature dimensions are consistent;

[0137] In step S43, the two-dimensional time-series feature matrix The Middle The target time step and the first Discrete target elements at each target channel Calculated using the following pooling sampling formula:

[0138] ,

[0139] in, This indicates that the one-dimensional adaptive max pooling layer has a time series length of [missing information]. The above is the generation of the first The pooling interval, divided according to the target time steps, is expressed as follows:

[0140] ,

[0141] in, Represents the channel alignment feature matrix Located in the middle The time step, the first Feature values ​​at each target channel; Represents the channel alignment feature matrix The length of the time series; Indicates the preset target time series length; Represents the latent space dimension of the target; This represents the target time step index after downsampling, and ; Indicates the target channel index, and ; Indicates the pooling interval Take the maximum value within the range; This represents the floor function; This represents the function for rounding up.

[0142] As an example, if the length obtained in step S3 is... Command response signal for each sampling point The system construction includes A one-dimensional temporally dilated convolutional network with parallel branches, where the length of the one-dimensional dilated convolutional kernel is set to... Configure different void ratios for these three parallel branches. , , These are used to extract local impact features, mesoscale vibration and wave features, and attenuation or resonance features over a longer time range, respectively. Zero-padding is applied to the boundaries during feature extraction in each branch, resulting in three sets of intermediate time-series feature sequences with a constant length of 501. , , .

[0143] Subsequently, these three sets of intermediate time-series feature sequences were... , , Cascaded concatenation is performed along the feature channel dimension to generate a dimension of size [missing information]. Multi-scale fusion feature matrix Each row corresponds to a time step, and each column corresponds to a scale feature extracted by a dilated convolution branch. Then, the target latent space dimension is set. and introduce a size of Channel alignment projection matrix and a bias vector of length 32 The multi-scale fusion feature matrix is ​​mapped from 3 feature channels to 32 target latent space dimensions, generating a dimension of [missing information]. Channel alignment feature matrix This is so that it can be fused and calculated with the one-dimensional semantic vector, which also has a feature dimension of 32, in the subsequent process.

[0144] Finally, the dimension is... Channel alignment feature matrix Input a one-dimensional adaptive max pooling layer, and set the preset target sequence length to . The one-dimensional adaptive max-pooling layer aligns the feature matrix along the time step dimension to the channel. Downsampling is performed to uniformly transform the time series length from 501 to 64, while keeping the target latent space dimension unchanged at 32, resulting in a generated dimension of [missing value]. Two-dimensional time series feature matrix .in, Each row corresponds to a target time segment after downsampling, and each column corresponds to a target latent space feature dimension, so that instruction response signals of different lengths can be converted into time-series inputs with a unified structure.

[0145] Step S5 of this invention primarily addresses the difficulty in accurately aligning and effectively fusing the semantic features of discrete control commands with the features of continuous temporal operation. By using a one-dimensional semantic vector as a query vector and projecting a two-dimensional temporal feature matrix into a key matrix and a value matrix, an attention mechanism is employed to calculate the correlation between the current control command and the operational features of each time step. This adaptively highlights the key temporal segments most relevant to the current command while suppressing irrelevant or weakly correlated temporal information. The effect is to generate targeted interactive features oriented towards the current control semantics, improving the targeting and expressive power of cross-modal feature fusion, and providing a more accurate and effective feature foundation for subsequent state recognition, fault diagnosis, or control logic analysis. (Refer to...) Figure 6 In one embodiment of the present invention, step S5 includes:

[0146] S51: Define the one-dimensional semantic direction as the query vector Q in the cross-modal attention mechanism; and project it using a pre-defined key matrix. Sum projection matrix Linear matrix projection is performed on the two-dimensional time-series feature matrix to generate the key matrix. Sum matrix The calculation formula is:

[0147] ,

[0148] ,

[0149] ,

[0150] in, This represents the one-dimensional semantic vector. The resulting query vector is used to characterize the semantic features of the current discrete control command; The key matrix is ​​used in attention calculations in conjunction with the query vector. Similarity matching is performed to determine the degree of correlation between the temporal features of each time step and the semantic features of the current discrete control command; The value matrix is ​​used to perform weighted aggregation according to the attention weights corresponding to each time step in the attention calculation to generate temporal fusion features associated with the semantic features of the current discrete control command; the element values ​​in the key projection matrix and the value projection matrix are determined according to the preset initialization rules, preferably by uniform distribution random generation within the preset value range [-0.125, 0.125], so that the initial projection parameters of the key feature space and the value feature space maintain a small amplitude and have initial differences in positive and negative directions;

[0151] S52: Combine the query vector Q with the transpose key matrix A similarity score sequence is obtained by performing matrix dot product operations, which is used to characterize the degree of matching between the semantic features of the current discrete control command and the temporal features of each time step; after scaling by a scaling factor, the attention weight distribution is generated by the Softmax normalization function.

[0152] S53: Utilizing the aforementioned attention weight distribution For the value matrix Weighted aggregation is performed along the time step dimension to generate targeted interaction features associated with the semantic features of the current discrete control command. The calculation formula is as follows:

[0153] .

[0154] Specifically, in step S52, a scaling factor is used. The similarity score sequence is scaled, and an attention weight distribution is generated using the Softmax normalization function. The calculation formula is:

[0155] ,

[0156] in, This represents the distribution of attention weights; Represents the latent space dimension of the target; This is a scaling factor used to normalize the scale of the dot product similarity results; This represents the exponential normalization function, which is used to convert the similarity scores at each time step into non-negative attention weights, and makes the sum of the attention weights at all time steps equal to 1.

[0157] As an example: the dimensional dimensions generated in step S3 A one-dimensional semantic vector Determined as the query vector in the cross-modal attention mechanism Therefore, the query vector The dimension is This is used to characterize the semantic features of the current discrete control command; at the same time, the size is set to... The key projection matrix Sum projection matrix Subsequently, the dimension size is... Two-dimensional time series feature matrix Projection matrix with bond respectively Sum projection matrix Perform linear matrix projection to generate the key matrix. Sum matrix .because The dimension is , and All dimensions are Therefore, the key matrix obtained after projection Sum matrix All dimensions are Among them, the key matrix The feature base, value matrix, is used to characterize the similarity matching between the temporal features at each time step and the semantic features of the current control command. Used to provide temporal feature content corresponding to each time step in the subsequent weighted aggregation process.

[0158] Then, the dimension size is key matrix Perform transpose to obtain the dimension with transpose bond matrix and set the dimension size to query vector With the transpose key matrix Performing a matrix dot product operation yields a matrix with dimensions of... The similarity score sequence is generated. The 64 score values ​​in this sequence correspond to 64 time steps in the two-dimensional temporal feature matrix, representing the degree of matching between the semantic features of the current discrete control command and the operational features of each time step. Subsequently, a scaling factor is applied. The similarity score sequences are scale-normalized, and the dimensionality is generated using the Softmax normalization function. Attention weight distribution This ensures that the attention weights corresponding to the 64 time steps are all non-negative, and that the sum of all attention weights is 1.

[0159] Finally, using the dimension size as Attention weight distribution For dimensions of Value matrix Weighted aggregation is performed along the time step dimension, generating a dimension of size [missing information]. Targeted interaction features .in, Each element is obtained by weighted summation of corresponding value features at 64 time steps according to attention weights. Thus, the system can adaptively highlight the key time segments most relevant to the current discrete control command based on its semantic features, suppressing the influence of background noise, stable operation segments, or weakly correlated response segments. This generates targeted interactive features oriented towards the current control semantics, providing a structurally unified and semantically more relevant fusion feature foundation for subsequent adaptive gating fusion, state recognition, fault diagnosis, or control logic analysis.

[0160] Step S6 of this invention is used to evaluate the reliability of the attention fusion result and determine the gating adjustment strength in subsequent feature fusion accordingly. Specifically, by calculating the normalized information entropy of the attention weight distribution, it is determined whether the attention weights are concentrated in a few key time steps. When the attention weights are relatively concentrated, it indicates that the physical response corresponding to the current discrete control command is relatively clear, and the extracted temporal features have high reference value and can be retained in subsequent fusion. When the attention weight distribution is relatively dispersed, it indicates that the current command response signal lacks clear key response segments, or the sensor monitoring signal is affected by environmental noise, abnormal disturbances, etc. In this case, by increasing the adaptive gating coefficient, the influence of weakly correlated or low-reliability temporal features on the fusion result is reduced. Thus, step S6 can adaptively evaluate and adjust the effectiveness of temporal features before the fusion of semantic features and physical temporal features, improving the stability, anti-interference ability, and command correlation of the fused features. (Refer to...) Figure 7 In one embodiment of the present invention, step S6 includes:

[0161] S61: Extract the first value from the attention weight distribution. Attention weights corresponding to each time step ,and satisfy Based on the information entropy calculation method, the attention weight distribution is... Attention weights in Entropy calculation is performed to obtain the original attention information entropy, which characterizes the degree of dispersion of the attention distribution. The calculation formula is:

[0162] ,

[0163] in, This indicates the length of the time series in the attention weight distribution, which is consistent with the preset target time series length in step S4; This indicates the time step index, which is consistent with the target time step index in step S4, and ; Represents the natural logarithm function; Represents the smallest positive number used to avoid zero-valued logarithmic operations, preferably It is used to maintain the numerical stability of logarithmic operations when the attention weight is zero or close to zero, and its value is much smaller than the effective numerical range of the attention weight in a single time step, so it will not have a substantial impact on the calculation result of the normalized information entropy.

[0164] S62: The original attention information entropy... Dimension elimination and scaling are performed to calculate the normalized information entropy. Its algebraic normalized mapping formula is:

[0165] ,

[0166] Wherein, the normalized information entropy The value is strictly constrained and converges to Within the closed interval, in a fundamental physical sense, it serves as a virtual probe to determine whether the sensor monitoring signal acquired in real-time is completely contaminated and distorted by environmental noise.

[0167] S63: Preset a high-entropy failure detection threshold in the system operating parameters. and gate slope coefficient Among them, the high-entropy failure determination threshold The gating slope coefficient is used to determine whether the attention weight distribution is excessively dispersed. Used to control the response sensitivity of the gating coefficient as a function of the normalized information entropy; the normalized information entropy Input to based The nonlinear gated mapping function is constructed to generate adaptive gate coefficients. .

[0168] Specifically, the formula for calculating the adaptive gating coefficient is as follows:

[0169] ,

[0170] in, Denotes the adaptive gating coefficient, and ; This represents the normalized information entropy, used to characterize the degree of dispersion in the attention weight distribution; The threshold for determining high entropy failure is preferably 0.75, which is used to determine attention-distracted states with high normalized information entropy as high entropy failure states. The gating slope coefficient is preferably 10, which is used to make the adaptive gating coefficient produce a significant but not excessively abrupt nonlinear response when the normalized information entropy is close to the high entropy failure judgment threshold, thereby timely improving the gating suppression strength when the attention distribution changes from concentrated to dispersed. This represents the natural exponential function.

[0171] When the normalized information entropy Less than the high-entropy failure determination threshold When the attention weight distribution is relatively concentrated, the key physical responses in the current instruction response signal are relatively clear, and the adaptive gating coefficient is... The value is relatively small; when the normalized information entropy Greater than or close to the high-entropy failure determination threshold When the attention weight distribution tends to be dispersed, the effective physical response in the current instruction response signal is not obvious or there is strong background noise interference, the adaptive gating coefficient Increase the gating coefficient to suppress proportional filtering interference; thus, the adaptive gating coefficient... It can dynamically adjust the retention or suppression ratio in the subsequent feature fusion process according to the dispersion of attention weight distribution, so as to reduce the influence of temporal features that are weakly correlated with the current discrete control command on the fusion result.

[0172] As an example: Step S5 generates a dimension with... Attention weight distribution The attention weight distribution include The attention weights for each time step are: ,and Each attention weight is a non-negative number, and the sum of all attention weights is 1; the attention weight distribution is based on the information entropy calculation method. Entropy calculation is performed to obtain the original attention information entropy. Specifically, the original attention information entropy The calculation formula is:

[0173] ,

[0174] in, Represents the smallest positive number used to avoid zero-valued logarithmic operations, preferably It is used to maintain the numerical stability of logarithmic operations when the attention weight is zero or close to zero, and its value is much smaller than the effective numerical range of the attention weight in a single time step, so it will not have a substantial impact on the calculation result of the normalized information entropy.

[0175] Subsequently, to eliminate the influence of time series length on the information entropy value, the system adjusts the original attention information entropy. Normalization is performed to generate normalized information entropy. Due to the attention weight distribution in this example include There are 1 time step, therefore the maximum information entropy is 1. Normalized information entropy The calculation formula is:

[0176] ,

[0177] Among them, normalized information entropy The values ​​are strictly constrained and converge to Within the closed interval. Therefore, the normalized information entropy... No longer affected by variations in the target time series length, it is instead used to uniformly characterize the dispersion of attention weight distribution. For example, when attention weights are concentrated on a few key time steps, A value of 0.50 indicates that the system can relatively clearly locate the key response segment corresponding to the current control command; when the attention weights are close to a uniform distribution, A value of 0.85 indicates that the attention mechanism has difficulty identifying a clear key response region from the current time-series characteristics, and the current sensor monitoring signal may have strong noise interference or weak correlation with the current discrete control command.

[0178] Finally, the system uses the normalized information entropy... Calculate the adaptive gating coefficient Preset high-entropy failure detection threshold. Preset gate slope system The formula for calculating the nonlinear gated mapping function is:

[0179] ,

[0180] when Substituting into the above formula, we can obtain... This indicates that the current attention weight distribution is relatively concentrated, the key physical responses in the instruction response signal are relatively clear, and the subsequent fusion process can retain more temporal physical features; when Substituting into the above formula, we can obtain... This indicates that the current attention weight distribution tends to be dispersed, reducing the reliability of the temporal physical response. Subsequent fusion processes need to increase the gating suppression ratio to reduce the impact of weakly correlated or noise-contaminated temporal features on the fusion result. Therefore, step S6 can transform the concentration of the attention distribution into a calculable gating adjustment, providing a reliable basis for subsequent adaptive gating fusion, thereby improving the noise resistance, stability, and instruction targeting of cross-modal fusion features.

[0181] Step S7 of this invention is used to robustly fuse instruction semantic features and targeted interaction features. This step adjusts the participation level of the targeted interaction features based on an adaptive gating coefficient, reducing the impact of uncertain temporal response information on the fusion result when attention distribution is scattered and interaction relationships are unclear. Simultaneously, a one-dimensional semantic vector is used as the residual benchmark and concatenated with the controlled interaction features, ensuring that the fusion result retains both the semantic information of the current control instruction and the physical response features related to that instruction. Subsequently, normalization is used to reduce the scale differences between feature channels, and a structurally unified and compactly expressed targeted fusion feature fingerprint is generated through low-dimensional mapping. These processes improve the stability and discriminative power of the fused features, providing more reliable feature input for subsequent elevator operation status assessment. (Refer to...) Figure 8 In one embodiment of the present invention, step S7 includes:

[0182] S71: Based on the adaptive gating coefficient, the targeted interaction features... Perform gating attenuation processing to generate controlled interaction features. The calculation formula is as follows:

[0183] ,

[0184] in, This represents the feature retention coefficient, used to adjust the retention ratio of the targeted interaction features based on the dispersion of the attention weight distribution; when the attention weight distribution is relatively concentrated, The controlled interaction features are smaller, and retain more targeted interaction information; when the attention weight distribution is more dispersed, As the controlled interaction features are increased, the uncertain interaction information in the controlled interaction features is further suppressed;

[0185] S72: The one-dimensional semantic vector generated in step S3... As a semantic residual benchmark, and the controlled interaction features Concatenated concatenation is performed along the feature channel dimension to generate cross-modal concatenated feature vectors. The calculation formula is as follows:

[0186] ,

[0187] in, This indicates splicing along the feature channel dimension; This represents a one-dimensional semantic vector corresponding to the current discrete control command, used to preserve the semantic prior information of the control command; through the above residual concatenation method, while preserving the semantic features of the original discrete control command, time-series response features related to the command semantics are introduced to provide input for subsequent fusion feature voltage regulation processing;

[0188] S73: Concatenate the cross-modal feature vector Layer normalization is performed to reduce the numerical scale differences between different feature channels, generating a voltage-stabilized feature tensor. The calculation formula is as follows:

[0189] ,

[0190] in, This represents the cross-modal splicing feature vector. Mean value along the feature channel dimension; This represents the cross-modal splicing feature vector. Variance in the feature channel dimension; This represents the smallest positive number used to avoid division by zero, and its preferred value is... It can also be based on the calculation precision. to Select from the range; This represents the preset channel scaling factor vector, with dimension 1. This is used to adjust the amplitude of each channel after normalization. All elements in the set value are preferably set to 1; This represents the preset channel translation coefficient vector, with dimension 1. This is used to perform channel-by-channel offset adjustment on the normalized channel features. All elements in the array are preferably set to 0; This represents element-wise multiplication.

[0191] S74: Construct the output projection matrix and output bias vector For the voltage regulation characteristic tensor Perform low-dimensional linear mapping to generate targeted fusion feature fingerprints for elevator operation status assessment. The calculation formula is:

[0192] ,

[0193] in, This represents the feature dimension of the low-dimensional representation space, and Output projection matrix The projection parameter matrix is ​​preset, and the element values ​​in the matrix are determined according to preset initialization rules, preferably within a preset value range. The internal structure is generated randomly with a uniform distribution to avoid abnormal amplification of fused features during low-dimensional mapping, while maintaining a certain degree of distinguishability between different output channels. The initial element values ​​of the preset projection bias vector are determined according to preset initialization rules, preferably initialized to 0, and second-most preferably set to 0. Real numbers randomly generated within a uniform distribution within the range are used to perform translation correction on low-dimensional feature results.

[0194] As an example: when the target's latent space dimension At that time, a one-dimensional semantic vector Targeted interaction features If the adaptive gating coefficients calculated in step S6 are used... Then the feature retention coefficient For the targeted interaction features Gating attenuation processing is performed to obtain controlled interaction features. This indicates that the attention distribution is relatively concentrated, thus the retention rate of targeted interaction features is relatively high.

[0195] Subsequently, the one-dimensional semantic vector With the controlled interaction features Concatenated along the feature channel dimension yields a cross-modal concatenated feature vector. The first 32 channels are used to retain the semantic prior information of the current discrete control command, while the last 32 channels are used to characterize the physical timing response information associated with the control command.

[0196] Next, calculate the cross-modal concatenation feature vector. The mean and variance of the 64 channel features are used to generate a voltage regulation feature tensor through layer normalization. The calculation formula is as follows:

[0197] ,

[0198] Wherein, the channel scaling factor vector All elements in the vector are set to 1, and the channel translation coefficient vector is... All elements in the normalization process are set to 0. Therefore, this normalization process is mainly used to reduce the numerical scale difference between the semantic feature channel and the temporal response feature channel, so that the subsequent low-dimensional mapping input remains stable.

[0199] Finally, let the low-dimensional representation space have a dimension. Construct the output projection matrix and output bias vector ; where the output projection matrix Each element in the preset value range The internal bias vector is generated randomly using a uniform distribution, and the output bias vector is... All elements in the formula are set to 0. A low-dimensional linear mapping is performed on the voltage regulation feature tensor, and the calculation formula is as follows:

[0200] ,

[0201] Finally, the targeted fusion feature fingerprint was obtained. This targeted fusion feature fingerprint, while retaining the semantic information of the current discrete control command, introduces physical response features related to the command, and through gating attenuation, residual splicing, normalization and low-dimensional mapping processing, forms a fusion feature representation that is dimensionally unified, numerically stable and easy to call in subsequent elevator operation status assessment.

[0202] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0203] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0204] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0205] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a multimedia terminal device (which may be a mobile phone, computer, television receiver, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0206] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for targeted fusion of heterogeneous elevator data guided by instruction semantics, characterized in that, Includes the following steps: S1: Obtain elevator control logs and sensor monitoring signals, and extract the timestamp of the current discrete control command in the control log as the command trigger time; S2: Based on the command triggering time, construct a command response time window according to a preset pre-set time and a preset post-set time, and extract the sensor monitoring signal within the command response time window as the command response signal; S3: Semantically encode the current discrete control command to generate a one-dimensional semantic vector representing the control intent; S4: Perform temporal convolution feature extraction and downsampling on the instruction response signal to generate a two-dimensional temporal feature matrix; S5: Use the one-dimensional semantic vector as the query vector, perform linear projection on the two-dimensional temporal feature matrix to generate a key matrix and a value matrix, and obtain the attention weight distribution and targeted interaction features through attention calculation; S6: Calculate the normalized information entropy based on the attention weight distribution, and determine the adaptive gating coefficient through the normalized information entropy; S7: Based on the adaptive gating coefficient, the one-dimensional semantic vector and the targeted interaction feature are fused together. After residual splicing, normalization processing and low-dimensional mapping, a targeted fusion feature fingerprint for elevator operation status evaluation is output. Step S3 includes: S31: After deduplicating the historical discrete control commands in the elevator control log, obtain a set of commands containing C different discrete control command string codes. Then, assign a unique integer feature index to each discrete control command string code according to a preset sorting rule; establish a command encoding dictionary with the discrete control command string code as the query key and the corresponding integer feature index as the mapping value. The target feature index is obtained by inputting the current discrete control instruction string code into the instruction encoding dictionary. Based on the one-hot encoding rule, through the target feature index... Build length is One-hot sparse eigenvectors ; S32: Constructing a cross-modal embedding matrix It is used to store the primary continuous semantic representations corresponding to each standard discrete control command, and the number of columns is equal to the preset primary semantic vector length. The uniquely heated sparse feature vector With the cross-modal embedding matrix Perform matrix multiplication to obtain primary continuous semantic vectors. ; S33: For the primary continuous semantic vector Linear projection and nonlinear activation are performed to generate a one-dimensional semantic vector representing the control intent. .

2. The method according to claim 1, characterized in that, Step S2 includes: S21: Triggered by the aforementioned instruction Based on the time reference, according to the preset pre-set duration and preset post-processing duration Calculate the start time respectively and termination time According to the start time and termination time Constructing instruction response time window ; S22: The start timestamp of the data stream monitored by the sensor The zero-point time of the discrete sampling sequence will be used as the instruction response time window. The start time and termination time Convert them into starting sampling point indices respectively. and the index of the terminating sampling point ; S23: Based on the starting sampling point index and the index of the terminating sampling point The sensor monitors the signal. Perform interval truncation to obtain a time series length of... Command response signal .

3. The method according to claim 1, characterized in that, Step S4 includes: S41: Constructing a system containing A one-dimensional temporally dilated convolutional network with parallel branches, and for the first... Each parallel branch is configured with a corresponding hole rate. The instruction response signal with a time series length of T is input into the one-dimensional temporal dilated convolutional network, and each parallel branch is processed according to the corresponding dilation rate. Temporal feature extraction is performed, and zero-padding is used during the feature extraction process to obtain N sets of intermediate temporal feature sequences of length T. S42: In the feature channel dimension, the aforementioned The intermediate time-series feature sequences of the group are concatenated and spliced ​​to generate a dimension of size. Multi-scale fusion feature matrix By using a linear projection that includes a channel alignment projection matrix and a bias vector, the feature channel dimension of the multi-scale fused feature matrix is ​​mapped from N to a target latent space dimension D consistent with the feature dimension of the one-dimensional semantic vector, generating a channel alignment feature matrix of dimension T×D. ; S43: Align the channels with the feature matrix Input a one-dimensional adaptive max pooling layer and perform downsampling along the time step dimension to reduce the length of the time series from... Unified transformation to a preset target sequence length And maintain the target latent space dimension The generated dimension size remains unchanged. Two-dimensional time series feature matrix .

4. The method according to claim 3, characterized in that, In step S41, the parallel branches are arranged according to the corresponding void ratio. The method for extracting temporal features is as follows: The first in the one-dimensional temporally dilated convolutional network Several parallel branches, indexed at discrete time steps. Discrete feature elements output at the location The following is calculated using one-dimensional dilated convolution: ， in, This represents the length of a one-dimensional dilated convolution kernel, i.e., the number of kernel weights involved in each convolution operation; This represents the sampling point number inside the convolution kernel, satisfying... ; Indicates the first Hollowness rate of parallel branches; Indicates from the first Starting at each time step, according to the void ratio The interval selected The signal values ​​at each sampling point; Indicates the first In the convolution kernel of the nth parallel branch Each weighting coefficient; Indicates the first The bias terms corresponding to each parallel branch; This indicates a modified linear unit activation function, which performs non-linear activation processing on the convolution calculation results; Represents the discrete time step index, and ; Traversing the first Index of all discrete time steps in each parallel branch Discrete feature elements output at each time step Arranged in chronological order, the time series has a length of [length missing]. The Intermediate time series feature sequences ;right Each parallel branch performs the above processing, resulting in... Intermediate time series feature sequences of the group: ; In step S42, the channel alignment feature matrix Located in the time step Target Channel Discrete feature elements at the location The calculation formula is: ， in, In the multi-scale fusion feature matrix, the first... The time step, the first Feature values ​​at each feature channel; Indicates the first channel alignment projection matrix. Line 1 The weighted element of the column; Represents the first in the bias vector One biased element; Represents the dimension of the target's latent space. Represents the target channel index, satisfying ; Traverse all time steps and all target channel indexes The calculated discrete feature elements Arrange and combine them in the time sequence dimension and the target channel dimension to form a complete channel alignment feature matrix; In step S43, the two-dimensional time-series feature matrix The Middle The target time step and the first Discrete target elements at each target channel Calculated using the following pooling sampling formula: ， in, This indicates that the one-dimensional adaptive max pooling layer has a time series length of [missing information]. The above is the generation of the first The pooling interval, divided according to the target time steps, is expressed as follows: ， in, Represents the channel alignment feature matrix Located in the middle The time step, the first Feature values ​​at each target channel; Represents the channel alignment feature matrix The length of the time series; Indicates the preset target time series length; Represents the latent space dimension of the target; This represents the target time step index after downsampling, and ; Indicates the target channel index, and ; Indicates the pooling interval Take the maximum value within the range; This represents the floor function; This represents the function for rounding up.

5. The method according to claim 1, characterized in that, Step S5 includes: S51: Determine the one-dimensional semantic vector as the query vector Q in the cross-modal attention mechanism; through a preset key projection matrix. Sum projection matrix Linear matrix projection is performed on the two-dimensional time-series feature matrix to generate the key matrix. Sum matrix ; S52: Combine the query vector Q with the transpose key matrix The similarity score sequence is obtained by performing matrix dot product operation, and after being scaled by a scaling factor, the attention weight distribution is generated by the Softmax normalization function. S53: Apply the attention weight distribution to the value matrix Weighted aggregation is performed along the time step dimension to generate targeted interactive features associated with the semantic features of the current discrete control command.

6. The method according to claim 5, characterized in that, In step S52, a scaling factor is used. The similarity score sequence is scaled, and an attention weight distribution is generated using the Softmax normalization function. The calculation formula is: ， in, This represents the distribution of attention weights; Represents the latent space dimension of the target; This is a scaling factor used to normalize the scale of the dot product similarity results; This represents the exponential normalization function, which is used to convert the similarity scores at each time step into non-negative attention weights, and makes the sum of the attention weights at all time steps equal to 1.

7. The method according to claim 1, characterized in that, Step S6 includes: S61: Extract the first value from the attention weight distribution. Attention weights corresponding to each time step Based on the information entropy calculation method, the attention weight distribution is... Attention weights in Entropy calculation is performed to obtain the original attention information entropy, which characterizes the degree of dispersion of the attention distribution. ; S62: The original attention information entropy... Dimension elimination and scaling are performed to calculate the normalized information entropy. ; S63: Preset a high-entropy failure detection threshold in the system operating parameters. and gate slope coefficient The normalized information entropy Input to based The nonlinear gated mapping function is constructed to generate adaptive gate coefficients. .

8. The method according to claim 7, characterized in that, In step S63, the formula for calculating the adaptive gating coefficient is: ， in, Denotes the adaptive gating coefficient, and ; This represents the normalized information entropy, used to characterize the degree of dispersion in the attention weight distribution; This represents the threshold for determining high-entropy failure. Indicates the gating slope coefficient; This represents the natural exponential function.

9. The method according to claim 1, characterized in that, Step S7 includes: S71: Perform gated attenuation processing on the targeted interaction features according to the adaptive gating coefficient to generate controlled interaction features; S72: The one-dimensional semantic vector generated in step S3 is used as the semantic residual benchmark and concatenated with the controlled interaction feature in the feature channel dimension to generate a cross-modal concatenated feature vector. S73: Concatenate the cross-modal feature vector Perform layer normalization to generate a voltage-stabilized feature tensor; S74: Construct the output projection matrix and output bias vector, perform low-dimensional linear mapping on the voltage regulation feature tensor, and generate a targeted fusion feature fingerprint for elevator operation status assessment.

Images