English equipment name segmentation and filling method and system for intelligent building

Abstract

The invention provides an English equipment name segmentation and filling method and system for a smart building. The method comprises the following steps: forming an initial equipment name mode by adopting a text clustering algorithm based on a character similarity distance; aiming at different modes, training a neural language model to calculate a character transition probability; generating a pseudo label through the distribution of the character transition probability; a neural network integration algorithm is adopted to train a segmentation model based on a pseudo tag, and the equipment name is segmented; and filling the segmentation fragments according to dictionary rules. According to the method, the English equipment names can be segmented into meaningful fragments without manual intervention, the corresponding fragments are filled according to the dictionary rules, the required massive equipment names are generated, the limitation of domain knowledge is solved, the labor cost is reduced, the working efficiency is improved, and wide application of the intelligent building technology is further promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of smart buildings, and in particular relates to a method and system for segmenting and filling English device names for smart buildings. Background Art

[0002] Device name segmentation aims to divide the string of device names into several semantic segments, which is an important task for implementing smart building technologies because these technologies fundamentally rely on understanding the context of device data. Device names are usually an alphanumeric character sequence containing multiple segments, each of which encodes key context of the device, such as building name, location, device type, etc.

[0003] Currently, device name segmentation requires domain knowledge and tedious manual work due to its building specificity. Device names are created by building vendors, and they often adopt different structures and vocabularies in different buildings. Usually, to build a device name segmentation model, a technician with domain expertise is required to understand these device names and then design rules to segment and annotate these names. There are no universal predefined parsing rules such as regular expressions for device names. Therefore, from the perspective of cost and efficiency, it remains a major obstacle to the widespread application of smart building technologies.

[0004] Despite recent progress in applying active learning and transfer learning to device name interpretation, these methods still require human annotation and are not fully automated. Therefore, there is a need to design automated segmentation and filling solutions for English device names. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method and system for segmenting and filling English device names for smart buildings. Given the English names of all devices in a building, the English device names can be segmented into meaningful segments without human intervention, and the corresponding segments can be filled according to dictionary rules to generate the required massive device names, thereby solving the limitations of domain knowledge, reducing labor costs, improving work efficiency, and further promoting the widespread application of smart building technology.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention proposes a method for segmenting and filling English device names for smart buildings, characterized in that it comprises the following steps:

[0008] S1: Using a text clustering algorithm based on character similarity distance to form an initial device name pattern;

[0009] S2: For different modes, train the neural language model to calculate the character transition probability;

[0010] S3: Generate pseudo labels through the distribution of character transition probabilities;

[0011] S4: A segmentation model based on pseudo labels is trained using a neural network ensemble algorithm to segment device names.

[0012] S5: Fill the segmented segments according to the dictionary rules.

[0013] Optionally, step S1 specifically includes the following sub-steps:

[0014] S11: Set the similarity measure between strings;

[0015] S12: Consider the character string of each device name as one category, and calculate the similarity distance between any two character strings;

[0016] S13: merge the two classes with the smallest distance into a new class and recalculate the similarity distance between the strings;

[0017] S14: Repeat S13 until all the classes are finally merged into a preset number of classes, each class corresponding to a different device name pattern.

[0018] Optionally, in step S2, the neural language model is Char-RNN; the process of calculating the character transition probability is as follows:

[0019] Given a device name character sequence X = {x1, x2, ..., xN} of length N, Char-RNN learns the probability p(x i+1 |x1,x2,...,x i ), and observe the transition from x1 to x i After the characters are obtained, a hidden state vector h is obtained. i , using a softmax layer applied to h i To predict the character transition probability

[0020]

[0021] In the formula, w c and w c′ are the linear transformations of characters c and c′ respectively, and x i+1 The cross entropy between the one-hot encodings of is used as the loss function for the characters.

[0022] Optionally, step S3 specifically includes the following sub-steps:

[0023] S31: Generate character transition probability Histogram of ;

[0024] S32: setting character transfer labels to "connect" and "split", indicating that the characters are connectable and split, respectively; calculating a connection accuracy curve and a segmentation accuracy curve, wherein the connection accuracy curve refers to the proportion of "connection" transfers among all character transfers exceeding a certain threshold, and the segmentation accuracy curve refers to the proportion of "segmentation" transfers among all character transfers below a certain threshold, and the threshold is a value of a confidence interval of a probability density function;

[0025] S33: Determine the turning point t1 of the connection accuracy curve and the turning point t2 of the segmentation accuracy curve through the distribution of character transition probability and the connection accuracy curve and the segmentation accuracy curve, and generate three types of pseudo labels "segmentation", "unknown", and "connection", among which the pseudo label of the characters in [0, t1] is "segmentation", the pseudo label of the characters in [t1, t2] is "unknown", and the pseudo label of the characters in [t2, 1] is "connection".

[0026] Optionally, step S4 specifically includes the following sub-steps:

[0027] S41: Subset sampling is performed on the generated “connection” and “segmentation” pseudo labels to generate a training set;

[0028] S42: Using intelligent optimization neural network algorithm to train the segmentation model;

[0029] S43: An ensemble learning approach is used to divide unknown labels into “connection” and “segmentation” labels.

[0030] Optionally, step S42 specifically includes the following sub-steps:

[0031] S421: Initialization of binary classification neural network model parameters;

[0032] S422: Output the predicted label and calculate the error loss between the predicted label and the pseudo label;

[0033] S423: Optimize model parameters using Levy's flying giant armadillo optimization algorithm;

[0034] S424: Calculate the prediction label error, and save the segmentation model when the error meets the conditions.

[0035] Optionally, in step S423, the Levy flying giant armadillo optimization algorithm is specifically:

[0036] S4231: Initialize the giant armadillo population parameters, specifically:

[0037] The population is modeled according to the following formula:

[0038]

[0039] Where X is the giant armadillo population matrix, X i is the i-th giant armadillo member, which is a candidate solution, x i，d For X i The d-th dimension decision variable in the search space, N is the number of giant armadillos, and m is the number of decision variables;

[0040] The initial position of the giant armadillo in the problem-solving space is randomly initialized using the following formula:

[0041] x i,d =lb d +r·(ub d -lb d );

[0042] Where r is a random number in the interval [0, 1], lb d ,ub d are the lower and upper bounds of the d-th dimension decision variable respectively;

[0043] The set of evaluation values ​​of the objective function is expressed as follows:

[0044]

[0045] Where F is the vector of the evaluation objective function, F i is the evaluation objective function based on the i-th giant armadillo member;

[0046] S4232: Based on the exploration phase of simulating a giant armadillo moving towards a termite mound, specifically:

[0047] The set of candidate termite mounds for each member of the colony is as follows:

[0048] TM i =(X k : F k <F i , k≠i}, i=1,2,...,N, k∈{1,2,...,N};

[0049] Where TM i is the set of candidate termite mound locations of the i-th giant armadillo, X k is the population member whose objective function value is better than the i-th giant armadillo, F k is the value of its objective function;

[0050] Based on modeling the movement of giant armadillos toward termite mounds, the mathematical representation of the new position of each member of the population is as follows:

[0051]

[0052] In the formula, STM iis the termite mound selected for the i-th giant armadillo, STM i，j is its j-th dimension, is the new position calculated by the i-th giant armadillo during the attack phase, is its jth dimension, r i，j is a random number in the interval [0, 1], I i，j is a number randomly chosen to be 1 or 2;

[0053] A new position replaces the corresponding previous position if it improves the value of the objective function according to the following equation:

[0054]

[0055] In the formula, is the new position calculated by the i-th giant armadillo during the attack phase, is the value of its objective function;

[0056] S4233: Use the Levy flight strategy to simulate the development phase of giant armadillos digging termite mounds for predation, specifically:

[0057] The process of updating the positions of giant armadillo population members introduced with Levy Flight is as follows:

[0058]

[0059] In the formula, is the new position calculated by the i-th giant armadillo during the digging phase, is its jth dimension, r i，j is a random number in the interval [0, 1], t is the number of iterations, is the value of its objective function;

[0060] L(λ) represents the random search path of Levy flight, as follows:

[0061]

[0062] In the formula, λ is the power, μ and ν follow a random normal distribution, expressed as:

[0063]

[0064] Optionally, in step S43, the ensemble learning method is a method of performing weighted averaging on multiple binary classification neural network models.

[0065] Optionally, step S5 specifically includes the following sub-steps:

[0066] S51: Selecting a segment to be filled from the segmented segments;

[0067] S52: Setting filling rules, including English characters to be filled, numerical increment rules, and the number of device names to be generated;

[0068] S53: Automatically generate the required number of device names according to the filling rules.

[0069] In a second aspect, the present invention provides a system for segmenting and filling English device names for smart buildings, which is characterized by comprising:

[0070] A text clustering module, for implementing a text clustering algorithm based on character similarity distance to form an initial device name pattern;

[0071] The neural language module is used to train the neural language model to calculate the character transition probability;

[0072] A label generation module is used to generate pseudo labels through the distribution of character transition probabilities;

[0073] The name segmentation module is used to train a pseudo-label-based segmentation model using a neural network ensemble algorithm to segment device names;

[0074] Name filling module, used to fill the segmentation fragments according to the dictionary rules.

[0075] The beneficial effects of the present invention are as follows: in the present invention, a new string similarity distance formula is proposed for text clustering, and the device name patterns are initially classified to provide initial labels for downstream segmentation tasks and increase model accuracy; a new intelligent optimization neural network integration algorithm is proposed to train the segmentation model, which, combined with text clustering, can segment English device names into meaningful segments without human intervention and maintain high accuracy; a method and system for segmenting and filling English device names for smart buildings are proposed, which fills the corresponding segments according to dictionary rules to generate the required massive device names, solves the limitations of domain knowledge, reduces labor costs, improves work efficiency, and further promotes the widespread application of smart building technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 The invention discloses a flow chart of a method for segmenting and filling English device names for smart buildings.

[0077] Figure 2 It is a flowchart of the intelligent optimization neural network integration algorithm for training the segmentation model.

[0078] Figure 3 is a flowchart of the Levy Flying Giant Armadillo optimization algorithm during training.

[0079] Figure 4 It is a structural diagram of the English device name segmentation and filling system used in smart buildings. DETAILED DESCRIPTION

[0080] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0081] Example 1

[0082] This embodiment provides a method for segmenting and filling English device names for smart buildings. Figure 1 Flowchart of the method for segmenting and filling English device names for smart buildings in this embodiment, Figure 1 As shown, the method includes:

[0083] Step S1: using a text clustering algorithm based on character similarity distance to form an initial device name pattern.

[0084] Step S2: For different modes, train the neural language model to calculate the character transition probability.

[0085] Step S3: Generate pseudo labels through the distribution of character transition probabilities.

[0086] Step S4: Use an intelligent optimization neural network integration algorithm to train a pseudo-label-based segmentation model to segment the device names.

[0087] Step S5: Fill the segmented segments according to the dictionary rules.

[0088] Specifically, in step S1, the text clustering algorithm process includes:

[0089] S11: String similarity measurement. The Dice distance and Jaro-Winker distance formula are used to calculate the similarity d between any two device name strings. dice and d jw .

[0090] S12: Consider each device name string as a class and calculate the similarity distance d between each string char The formula is:

[0091]

[0092] S13: Merge the two classes with the smallest distance into a new class and recalculate d char .

[0093] S14: until all the classes are finally merged into a preset number of classes, that is, different device name patterns.

[0094] Specifically, in step S2, the neural language model can adopt any character-level neural language model. In this embodiment, the neural language model takes Char-RNN as an example. The character transition probability calculation process is:

[0095] Given a device name character sequence of length N, X = {x1, x2, ..., x N}. Char-RNN learns the probability of observing the next character given all previous characters, i.e.: p(x i+1 |x1,x2,...,x i ). Each character x i Get an embedding vector X i (The input data of the Char-RNN algorithm needs to be in vector form, so the character sequence must be converted into a vector, that is, an embedding vector) and observe the i After the characters are obtained, a hidden state vector h is obtained. i . A softmax layer is applied to h i To predict the distribution of the entire character set The formula is as follows:

[0096]

[0097] In the formula, w c and w c′ are the linear transformations of characters c and c′ respectively. and x i+1 The cross entropy between the one-hot encodings of is used as the loss function for that character.

[0098] Specifically, in step S3, the pseudo label generation process includes:

[0099] S31: Generate character transition probability Histogram.

[0100] S32: Assuming that the character transition labels are "connection" and "segmentation", calculate the connection accuracy curve and the segmentation accuracy curve respectively. The connection accuracy curve refers to the proportion of "connection" transitions among all character transitions exceeding a certain threshold, and the segmentation accuracy curve refers to the proportion of "segmentation" transitions among all character transitions below a certain threshold.

[0101] For example, a string of characters: Ilikeit means "I like it". Connection means that the two characters before and after can be connected together, such as "like": 1 and i will be detected as connected, i and k, k and e are the same. Segmentation means that the two characters before and after are separated, such as I and 1 will be detected as segmented, e and i are the same.

[0102] The connection accuracy curve refers to the proportion of "connected" transfers in all character transfers that exceed a certain threshold, and the segmentation accuracy curve refers to the proportion of "segmented" transfers in all character transfers below a certain threshold. The threshold is a value of the confidence interval of the probability density function. It is a statistical concept that describes the uncertainty of probability estimation. For example: assuming that the connection ratio is 0.9, the corresponding threshold is 0.77. Therefore, when it exceeds 0.77, the proportion of "connected" transfers in all character transfers is 0.9.

[0103] S33: Since the "turning point" on the segmentation accuracy curve is highly correlated with the peak value in the histogram, the thresholds t1 and t2 can be determined by the distribution of the character transfer probability and the accuracy curve to generate three types of pseudo labels: "segmentation", "unknown", and "connection". For example: point t2 is the turning point of the segmentation accuracy curve, one of the high peak values ​​of the probability density, and the connection accuracy is greater than the segmentation accuracy; point t1 is the turning point of the connection accuracy curve, one of the low peak values ​​of the probability density, and the segmentation accuracy is greater than the connection accuracy. Then the character pseudo label in [0, t1] is "segmentation", the character pseudo label in [t1, t2] is "unknown", and the character pseudo label in [t2, 1] is "connection".

[0104] Figure 2 Flowchart of the intelligent optimization neural network integration algorithm training segmentation model in this embodiment. Figure 2 As shown, step S4 includes:

[0105] S41: Subset sampling is performed on the generated connection and segmentation pseudo labels to generate a training set. In order to balance the data, the ratio of connection and segmentation pseudo label sampling is set to 10:1; here, the subset is the division of the data set. For example, there are 100 pseudo-label data, and 70 of them are taken as training data. These 70 are a subset of the 100.

[0106] S42: Use intelligent optimization neural network algorithm for training.

[0107] S43: An ensemble learning method is used, that is, a weighted average method is used to divide unknown labels into connection and segmentation labels.

[0108] Specifically, in step S42, the intelligent optimization neural network algorithm training process includes:

[0109] S421: Initialize the parameters of the binary classification neural network model.

[0110] S422: Output the predicted label and calculate the error loss between the predicted label and the pseudo label.

[0111] S423: Optimize model parameters using the Levy flying giant armadillo optimization algorithm.

[0112] S424: Calculate the prediction label error, and save the segmentation model when the error meets the conditions.

[0113] Figure 3 FIG. 4 is a flow chart of the optimization algorithm of Levy's flying giant armadillo during the training process according to this embodiment, as shown in FIG. Figure 3 As shown, step S423 includes:

[0114] S4231: Initialization of giant armadillo population parameters.

[0115] S4232: Based on the exploratory phase of a simulated giant armadillo moving toward a termite mound.

[0116] S4233: Development phases of predation based on simulations of giant armadillos digging into termite mounds using a Levy flight strategy.

[0117] Step S4231 is specifically as follows: The giant armadillo optimization algorithm is based on the search ability of its members in the problem-solving space, and can provide a suitable solution to the optimization problem in an iterative process. Each giant armadillo member determines the value of the decision variable of the problem according to its position in the problem-solving space. Therefore, as a member of the population, each giant armadillo is a candidate solution to the problem modeled mathematically by a vector. The giant armadillos together constitute the population of the algorithm, and the population can be modeled mathematically by a matrix according to the following formula.

[0118]

[0119] Where X is the giant armadillo population matrix, X i is the i-th giant armadillo member, i.e., the candidate solution, x i , d is the d-th dimension decision variable in the search space, N is the number of giant armadillos, and m is the number of decision variables.

[0120] At the beginning of the algorithm execution, the initial position of the giant armadillo in the problem-solving space needs to be randomly initialized using the following formula. The mathematical expression is as follows:

[0121] x i，d =lb d +r·(ub d -lb d );

[0122] Where r is a random number in the interval [0, 1], lb d ,ub d are the lower and upper bounds of the d-th dimension decision variable, respectively.

[0123] Since the position of each giant armadillo in the problem-solving space represents a candidate solution to the problem, the value of the objective function can be evaluated corresponding to each giant armadillo. Accordingly, the set of evaluation values ​​of the objective function can be expressed using the following formula.

[0124]

[0125] Where F is the vector of the evaluation objective function, F i is the evaluation objective function based on the i-th giant armadillo member. The best value obtained by the objective function corresponds to the best member, i.e., the best candidate solution. Since the position of the giant armadillo in the problem space is updated in each iteration, the best member also needs to be updated based on the comparison of the updated value of the objective function. At the end of the algorithm implementation, the position of the best member obtained during the algorithm iteration is taken as the solution to the problem.

[0126] Step S4232 is specifically as follows: based on the change in the position of the giant armadillo when it moves toward the termite mound, the position of the population member in the problem-solving space is updated. Modeling this attack process causes the position of the giant armadillo to change widely, thereby increasing the algorithm's exploration ability in global search management. For each population member representing a giant armadillo, the position of other population members with better objective function values ​​is considered to be a termite mound. The mathematical formula for the set of candidate termite mounds for each member of the population is as follows:

[0127] TM i =(X k :F k <F i ,k≠i},i=1,2,...,N,k∈{1,2,...,N};

[0128] Where TM i is the set of candidate termite mound locations of the i-th giant armadillo, X k is the population member whose objective function value is better than the i-th giant armadillo, F k is the value of its objective function.

[0129] The giant armadillo randomly selects a candidate termite mound and attacks it. Based on modeling the giant armadillo's movement toward the termite mound, the mathematical expression for the new position of each member of the population is as follows:

[0130]

[0131] In the formula, STM i is the termite mound selected for the i-th giant armadillo, STM i，j is its j-th dimension, is the new position calculated by the i-th giant armadillo during the attack phase, is its jth dimension, r i，j is a random number in the interval [0, 1], I i，j is a number randomly chosen to be 1 or 2.

[0132] If the new position improves the value of the objective function according to the following formula, then the new position replaces the corresponding previous position.

[0133]

[0134] In the formula, is the new position calculated by the i-th giant armadillo during the attack phase, is the value of its objective function.

[0135] Step S4233 is specifically as follows: based on the simulation of the giant armadillo digging termite mounds to prey on termites, the positions of the population members in the problem-solving space are updated. With the goal of preying on termites, the digging process of the giant armadillo is modeled, so that the position of the giant armadillo changes slightly, thereby improving the local search capability of the algorithm.

[0136] Lévy flight is a random walk with a probability distribution of step length. Introducing Lévy flight into the process of updating the positions of population members of giant armadillos and disturbing the current optimal solution can enhance the local escape ability and detect the farther solution in the algorithm space. The process of updating the positions of population members of giant armadillos with Lévy flight is as follows:

[0137]

[0138] In the formula, is the new position calculated by the i-th giant armadillo during the digging phase, is its jth dimension, r i，j is a random number in the interval [0,1], t is the number of iterations, is the value of its objective function.

[0139] L(λ) represents the random search path of Levy flight, as shown in the formula:

[0140]

[0141] In the formula, λ is the power, which can be understood as a constant, generally taking the value of 1.5, and μ and ν obey the random normal distribution, which can be expressed as:

[0142]

[0143] Specifically, in step S5, the dictionary rule filling process includes:

[0144] S51: Select a segment to be filled from the segmented segments.

[0145] S52: Set filling rules. Including but not limited to the English characters to be filled, the number increment rule, and the number of device names to be generated. For example: the English characters to be filled in segment 1 are "ABC", the English characters to be filled in segment 2 are "XYZ", the number increment rule is "increase by 1", and the required number is "5".

[0146] S53: Automatically generate the required number of device names according to the filling rules. For example, according to the example rules of S52, the generated device names are "ABC-XYZ1", "ABC-XYZ2", "ABC-XYZ3", "ABC-XYZ4", "ABC-XYZ5".

[0147] Example 2

[0148] This embodiment provides a system for segmenting and filling English device names for smart buildings, which is used to execute the method for segmenting and filling English device names for smart buildings in Example 1. Figure 4 : is a structural diagram of the English device name segmentation and filling system for smart buildings in this embodiment, such as Figure 4 As shown, the system 100 includes the following modules:

[0149] The text clustering module 101 is used to implement a text clustering algorithm based on character similarity distance to form an initial device name pattern.

[0150] The neural language module 102 is used for calculating character transition probabilities using a neural language model.

[0151] The label generation module 103 is used to generate pseudo labels according to the distribution of character transition probabilities.

[0152] The name segmentation module 104 is used to intelligently optimize the neural network integration algorithm to train a pseudo-label-based segmentation model to segment the device names.

[0153] The name filling module 105 is used to fill the segmented segments according to the dictionary rules.

[0154] In this embodiment, the specific functions and execution processes of each module are the same as those of the method in Embodiment 1, and thus will not be described in detail here.

[0155] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A method for segmenting and filling English device names for smart buildings, characterized in that: The following steps are involved: S1: Using a text clustering algorithm based on character similarity distance to form an initial device name pattern; S2: For different modes, train the neural language model to calculate the character transition probability; S3: Generate pseudo labels through the distribution of character transition probabilities; S4: A segmentation model based on pseudo labels is trained using a neural network ensemble algorithm to segment device names. S5: Fill the segmented segments according to the dictionary rules.

2. The method for segmenting and filling English device names for smart buildings as claimed in claim 1, characterized in that: Step S1 specifically includes the following sub-steps: S11: Set the similarity measure between strings; S12: Consider the character string of each device name as one category, and calculate the similarity distance between any two character strings; S13: merge the two classes with the smallest distance into a new class and recalculate the similarity distance between the strings; S14: Repeat S13 until all the classes are finally merged into a preset number of classes, each class corresponding to a different device name pattern.

3. The method for segmenting and filling English device names for smart buildings as claimed in claim 1, characterized in that: In step S2, the neural language model is Char-RNN; the process of calculating the character transition probability is as follows: Given a device name character sequence X = {x1, x2, ..., xN} of length N, Char-RNN learns the probability p(x i+1 |x1,x2,...,x i ), and observe the transition from x1 to x i After the characters are obtained, a hidden state vector h is obtained. i , using a softmax layer applied to h i To predict the character transition probability In the formula, w c and w c′ are the linear transformations of characters c and c′ respectively, and x i+1 The cross entropy between the one-hot encodings of is used as the loss function for the characters.

4. The method for segmenting and filling English device names for smart buildings as claimed in claim 1, characterized in that: Step S3 specifically includes the following sub-steps: S31: Generate character transition probability Histogram of ; S32: setting character transfer labels to "connect" and "split", indicating that the characters are connectable and split, respectively; calculating a connection accuracy curve and a segmentation accuracy curve, wherein the connection accuracy curve refers to the proportion of "connection" transfers among all character transfers exceeding a certain threshold, and the segmentation accuracy curve refers to the proportion of "segmentation" transfers among all character transfers below a certain threshold, and the threshold is a value of a confidence interval of a probability density function; S33: Determine the turning point t1 of the connection accuracy curve and the turning point t2 of the segmentation accuracy curve through the distribution of character transition probability and the connection accuracy curve and the segmentation accuracy curve, and generate three types of pseudo labels "segmentation", "unknown", and "connection", among which the pseudo label of the characters in [0, t1] is "segmentation", the pseudo label of the characters in [t1, t2] is "unknown", and the pseudo label of the characters in [t2, 1] is "connection".

5. The method for segmenting and filling English device names for smart buildings as claimed in claim 4, characterized in that: Step S4 specifically includes the following sub-steps: S41: Subset sampling is performed on the generated "connection" and "segmentation" pseudo labels to generate a training set; S42: Using intelligent optimization neural network algorithm to train the segmentation model; S43: An ensemble learning method is used to divide unknown labels into "connection" and "segmentation" labels.

6. The method for segmenting and filling English device names for smart buildings as claimed in claim 5, characterized in that: Step S42 specifically includes the following sub-steps: S421: Initialization of binary classification neural network model parameters; S422: Output the predicted label and calculate the error loss between the predicted label and the pseudo label; S423: Optimize model parameters using Levy's flying giant armadillo optimization algorithm; S424: Calculate the prediction label error, and save the segmentation model when the error meets the conditions.

7. A method for segmenting and filling English device names for smart buildings as claimed in claim 6, characterized in that: In step S423, the Levy flying giant armadillo optimization algorithm is specifically: S4231: Initialize the giant armadillo population parameters, specifically: The population is modeled according to the following equation: Where X is the giant armadillo population matrix, X i is the i-th giant armadillo member, which is a candidate solution, X i , d is X i The d-th dimension decision variable in the search space, N is the number of giant armadillos, and m is the number of decision variables; The initial position of the giant armadillo in the problem-solving space is randomly initialized using the following formula: x i ,d=lb d +r·(ub d -lb d ); Where r is a random number in the interval [0, 1], lb d ,ub d are the lower and upper bounds of the d-th dimension decision variable respectively; The set of evaluation values ​​of the objective function is expressed as follows: Where F is the vector of the evaluation objective function, F i is the evaluation objective function based on the i-th giant armadillo member; S4232: Based on the exploration phase of simulating a giant armadillo moving towards a termite mound, specifically: The set of candidate termite mounds for each member of the colony is as follows: TM i =(X k :F k <F i ,k≠i},i=1,2,...,N,k∈{1,2,...,N}; Where TM i is the set of candidate termite mound locations of the i-th giant armadillo, Xk is the population member with a better objective function value than the i-th giant armadillo, and Fk is its objective function value; Based on modeling the movement of giant armadillos toward termite mounds, the mathematical representation of the new position of each member of the population is as follows: In the formula, STM i is the termite mound selected for the i-th giant armadillo, STM i，j is its j-th dimension, is the new position calculated by the i-th giant armadillo during the attack phase, is its jth dimension, r i，j is a random number in the interval [0, 1], I i，j is a number randomly chosen to be 1 or 2; A new position replaces the corresponding previous position if it improves the value of the objective function according to the following equation: In the formula, is the new position calculated by the i-th giant armadillo during the attack phase, is the value of its objective function; S4233: Use the Levy flight strategy to simulate the development phase of giant armadillos digging termite mounds for predation, specifically: The process of updating the positions of giant armadillo population members introduced with Levy Flight is as follows: In the formula, is the new position calculated by the i-th giant armadillo during the digging phase, is its jth dimension, r i，j is a random number in the interval [0, 1], t is the number of iterations, is the value of its objective function; L(λ) represents the random search path of Levy flight, as follows: In the formula, λ is the power, μ and ν follow a random normal distribution, expressed as:

8. The method for segmenting and filling English device names for smart buildings as claimed in claim 5, characterized in that: In step S43, the ensemble learning method is a method of performing weighted averaging on multiple binary classification neural network models.

9. The method for segmenting and filling English device names for smart buildings as claimed in claim 1, characterized in that: Step S5 specifically includes the following sub-steps: S51: Selecting a segment to be filled from the segmented segments; S52: Setting filling rules, including English characters to be filled, numerical increment rules, and the number of device names to be generated; S53: Automatically generate the required number of device names according to the filling rules.

10. A system for segmenting and filling English equipment names for smart buildings, characterized in that: include: A text clustering module, for implementing a text clustering algorithm based on character similarity distance to form an initial device name pattern; The neural language module is used to train the neural language model to calculate the character transition probability; A label generation module is used to generate pseudo labels through the distribution of character transition probabilities; The name segmentation module is used to train a pseudo-label-based segmentation model using a neural network ensemble algorithm to segment device names; Name filling module, used to fill the segmentation fragments according to the dictionary rules.

Images