Park early warning system based on knowledge graph

Abstract

The invention discloses a park early warning system based on a knowledge graph, which relates to the technical field of park early warning and comprises a sound acquisition unit, a signal decomposition unit, a representation vector determination unit, a classifier combination unit, a parameter optimization unit and an early warning level output unit. According to the method, a plurality of binary classifiers containing different belonging category judgments are arranged through the classifier combination unit, the classifiers are utilized to perform collaborative analysis on the belonging categories of the samples, so that the recognition result can be more suitable for the actual situation, unknown parameters of the binary classifiers are optimized by using the parameter optimization unit, and the recognition accuracy is improved. The optimal value of each unknown parameter is screened out, the accuracy of the classifier in the operation process is further improved, the early warning level output unit derives the generation reason and the early warning level of the current sound signal step by step according to the relation path in the knowledge graph, and it is ensured that managers can take measures in time for intervention.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of park early warning, in particular to a park early warning system based on a knowledge graph. BACKGROUND

[0002] The park early warning system based on the knowledge graph integrates multiple sound signal data, determines the belonging category and source location of different sound signals, timely locates abnormal sound data, realizes real-time risk perception, visual early warning and multi-terminal collaborative response, and improves the park safety prevention and control capability. In the application number 202311003775.X invention patent, an "event early warning efficient management method and system based on park management" is disclosed. The invention belongs to the field of park safety management and relates to data analysis technology. It is used to solve the problem that the existing event early warning efficient management method cannot analyze the suitability of the processing personnel according to the event attributes and potential harm degree. It includes an efficient management platform, which is communicatively connected with an event analysis module, a processing recommendation module and a storage module. The event analysis module is used to analyze the park early warning event. After the event analysis module receives the park early warning event, it obtains the event attributes and involved data of the park early warning event. The invention can analyze the park early warning event, filter the marked events from the historical early warning events according to the event attributes and involved data of the park early warning event, and then analyze and calculate the harm performance value according to the harm parameters of the marked events.

[0003] The above-mentioned prior art solves the problem of being unable to analyze the suitability of the processing personnel according to the event attributes and potential harm degree, but the system cannot analyze the sound signal separately when running, which makes it difficult to locate the sound source in a short time when glass breaking sound, explosion sound and gunshot sound occur inside the park. At the same time, the footsteps sound in different time periods cannot be set to different early warning levels, which makes it difficult for the management personnel to take timely measures to intervene. SUMMARY

[0004] The purpose of the present application is to provide a park early warning system based on a knowledge graph to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solution: a park early warning system based on a knowledge graph, comprising a classifier combination unit, a parameter optimization unit and an early warning level output unit.

[0006] The sound acquisition unit sets the total number of layers of the intrinsic mode function of the signal, reads all processed sound signals, combines the sound signals with randomly generated different normal noises to obtain multiple combined sound signals, uses the difference method to determine the time series data of the maximum point and the time series data of the minimum point corresponding to the combined sound, calculates the polynomial corresponding to each time interval in the time series data step by step, and uses the polynomial to construct the maximum value connection curve and the minimum value connection curve.

[0007] The signal decomposition unit calculates the average value at different times based on the maximum value connection curve and the minimum value connection curve, subtracts the average value from the combined sound signal to obtain a new sound signal, determines the signal intrinsic mode function and the combined sound signal for the next cycle based on the new sound signal, and repeats the operation until the signal intrinsic mode function of all layers of the current sound signal is calculated.

[0008] The representation vector determination unit acquires the intrinsic mode functions of the signals at each layer in all processed audio signals, calculates the signal representation coefficients of each frame in the mode functions, determines the short-time energy, zero-crossing rate, and time interval from the initial value to the peak value of each frame in all audio signals, and combines the short-time energy corresponding to each frame signal in the intrinsic mode functions of each layer to obtain the attribute representation vector of the audio signal.

[0009] Preferably, the sound acquisition unit includes a sound processing module, a noise addition module, a signal analysis module, and a curve generation module. The sound processing module acquires different types of raw sound signals from the AudioSet website, including shouts, glass breaking sounds, explosions, gunshots, footsteps, and other noises. It samples, pre-emphasizes, frames, windows, and denoises each raw sound signal to obtain the processed sound signal. Before sampling, the highest frequency of the sampled raw signal is counted, ensuring that each sampling frequency is greater than or equal to twice the highest frequency of the signal. During framing, the raw signal is divided into 20ms frames with a 10ms step size. The noise addition module is set to a maximum loop count n. max The total number of layers ρ of the intrinsic mode functions of the signal max Then, read all processed audio signals. For the current number of combinations An initialization operation is performed, where h0 represents the number of sound signals, and n1 represents the number of combinations corresponding to the first sound signal ξ1(t). This represents the h0th sound signal. The corresponding number of combinations will determine the k-th sound signal ξ. k (t) and randomly generated normal noise combining, the kth combined sound signal ξ" is obtained k (t), wherein k represents a parameter, and n represents the number of combinations at the time k automatically adding one, the signal analysis module obtains the kth combined sound signal ξ' k (t), and the maximum point time sequence data corresponding to ξ' k (t) is determined using a difference method and minimum point time sequence data wherein h1 represents the number of maximum points, and α, β represent parameters, and the maximum value corresponding to the first time is determined according to the first time and the maximum value corresponding to the second time is determined according to the second time corresponding polynomials are constructed After that, the polynomials corresponding to each time interval are calculated step by step all coefficients in each polynomial are determined according to a preset condition, and the maximum value connection curve is constructed by the curve generation module using the polynomials corresponding to different time intervals. Similarly, the polynomials of each time interval are calculated through the minimum point time sequence data, and the minimum value connection curve is constructed according to the polynomials, wherein the preset condition is specifically:

[0010] the first derivative value in is equal to the first derivative value in , and the first derivative value in is equal to the first derivative value in ;

[0011] the second derivative value in is equal to the second derivative value in ;

[0012] Preferably, the signal decomposition unit comprises an average value calculation module, a point-by-point subtraction module, a point number statistics module, a loop execution module and a function generation module, the average value calculation module determines the maximum value and the minimum value at different times according to the maximum value connection curve and the minimum value connection curve, and calculates the corresponding average value using the maximum value and the minimum value at the same time, the point-by-point subtraction module subtracts the kth combined sound signal ξ" k (t) from the average value at each time to obtain a new sound signal ξ" k (t), wherein ​The point counting module sets a first threshold ε1 and a second threshold ε2, counts the total number of points N contained in the time series data of the maximum and minimum points, and determines the combined sound signal ξ′. k The number of zero-value signal points N0 contained in (t) is if And if |N-N0|≤ε2, then the new sound signal ξ″ k (t) is used as the eigenmode function of the signal, and ξ′ k (t) and ξ″ k Subtracting (t) from each other, the resulting signal is used as the combined sound signal for the next cycle, and the current layer number ρ i Automatically increment by one; otherwise, add a new sound signal ξ″. k (t) serves as the combined sound signal for the next round of the loop, and the loop execution module repeats the operation until the current layer number ρ. i Equal to the total number of floors ρ max Then, the sound signal ξ k (t) and randomly generated normal noise Combine them, and then combine them the number of times n was combined at that time. k Automatically increment by one, and calculate the signal intrinsic mode function values ​​corresponding to different layer numbers, repeating the same operation until the number of combinations n. k Greater than or equal to the maximum number of loops n max The loop ends, and the function generation module reads the sound signal ξ. k (t) is combined with the intrinsic mode functions of the signal at each layer after being combined with randomly generated normal noise. The average function is then calculated based on the intrinsic mode functions of the signal at each layer under different combination numbers, and this average function is used as the sound signal ξ. k (t) The final signal eigenmode functions of each layer.

[0013] Preferably, the representation vector determination unit includes a frame calculation module, a coefficient output module, a ratio analysis module, and a vector combination module. The frame calculation module acquires the intrinsic mode functions (EMFs) of each layer within all processed audio signals, performs frame division, and divides the EMFs into 20ms frames with a step size of 10ms. It calculates the power spectrum of each frame within the EMF. The coefficient output module analyzes the power spectrum using a triangular filter to obtain the log-Mel energy. The log-Mel energy is then subjected to a discrete cosine transform to determine the signal representation. The ratio analysis module determines the short-time energy, zero-crossing rate, and time interval from the initial value to the peak value for each frame in all sound signals, and counts the short-time energy corresponding to each frame in the intrinsic mode function of each layer of signals. The sound signal is matched with each frame in the intrinsic mode function of the signal, and the ratio of their energies in the same time period is calculated. The vector combination module combines the short-time energy, zero-crossing rate, time interval from the initial value to the peak value, signal characterization coefficients contained in the intrinsic mode function of each layer of signals, and energy ratio of each frame in each sound signal to obtain the attribute characterization vector of the sound signal.

[0014] Preferably, the classifier combination unit includes a sample generation module, a classifier construction module, a label setting module, and a category output module. The sample generation module acquires the attribute representation vectors of multiple sound signals, labels the category of each sound signal, where the categories include shouting, glass breaking, explosion, gunshot, footsteps, and other noise. It combines the attribute representation vectors and categories of individual sound signals and stores them as samples in a 8:2 ratio in a first set and a second set, with the number of samples in each category being equal within each set. The classifier construction module constructs multiple binary classifiers a1,…,a1. j ,…,a 15 Train a binary classifier using any two samples from different categories in the first set. Repeat this process until all binary classifiers are trained. Then, transfer all samples from the second set to a1,...,a j ,...,a 15 In the middle, determine a1,...,a j ,...,a 15 The accuracy of a1,…,a j ,…,a 15 Sort the data according to accuracy and select the binary classifier a with the highest accuracy. j As the first binary classifier, the label setting module is in binary classifier a j Includes category b x and b y The determination is made after the sample has undergone a jBefore the determination, a corresponding estimated category set is set, and the initial estimated category set contains the numbers of all belonging categories. The sample is then transferred to a. j In the middle, the category analysis algorithm is used to determine a j If the output label value is positive, then delete b from the estimated class set of the current sample. y Number them and list all those that do not contain 'b'. y For a binary classifier that determines the class, if the label is negative, then remove b from the estimated class set of the current sample. x Number them and list all those that do not contain 'b'. x The binary classifier is determined by sorting all listed binary classifiers according to their accuracy. The binary classifier with the highest accuracy is selected for further determination. This process is repeated until only one number remains in the estimated class set. The class of the current sample is then output according to the number. The specific category analysis algorithm is as follows:

[0015] f(H) = sign(∑(ω·exp(-g||H)) E -H F || 2 )-b))

[0016] Where g represents the radial coefficient, ω represents the adjustment parameter, b represents the bias term, and H E H represents the attribute representation vector of the E-th training sample. F Let E and F represent the attribute representation vector of the F-th sample to be predicted, where E and F represent parameters.

[0017] Preferably, the parameter optimization unit includes a candidate node construction module and an initial score setting module. After determining all the unknown parameters in the binary classifier, the candidate node construction module sets a corresponding limit range for each unknown parameter. Within the limit range, the current unknown parameter is taken multiple times, and the values ​​of different types of unknown parameters are selected one by one and stored in the candidate nodes. The initial score setting module obtains multiple candidate nodes, and each candidate node has one and only one value for each unknown parameter. There is no case where the values ​​of the candidate nodes are completely the same. The value of each unknown parameter in the candidate node is used as the actual parameter value of the binary classifier, the accuracy under this set of parameter values ​​is calculated, and it is used as the initial score of the candidate node. The change coefficients of all candidate nodes are set to zero.

[0018] Preferably, the parameter optimization unit further includes a weight initialization module and an optimal parameter determination module. The weight initialization module initializes the weight coefficients, correlation factors, and maximum number of iterations. Based on the initial score of each candidate node, it determines the local optimal score of a single node and the overall optimal score of all nodes. The optimal parameter determination module uses a parameter update algorithm to adjust all parameter values ​​and change coefficients within each node. It uses the value of each unknown parameter in the current candidate node as the actual parameter value of the binary classifier, calculates the new score of each candidate node, updates the local optimal score of a single node and the overall optimal score of all nodes based on the current score, calculates the new weight coefficients using a weight analysis algorithm, and repeats the operation until the number of iterations reaches the maximum number of iterations. Finally, it outputs the parameter value corresponding to the overall optimal score and uses it as the actual parameter value of the binary classifier.

[0019] Preferably, the warning level output unit includes a knowledge graph generation module, a cause determination module, and a warning execution module. The knowledge graph generation module acquires historical abnormal sound records of a designated area, wherein the abnormal sound records include the area number, device number, sound category, collection time, cause, and warning level. A knowledge graph is constructed based on the historical abnormal sound records. The cause determination module acquires the current sound signal, determines its attribute representation vector and category, analyzes the corresponding area number using the device number that acquired the sound signal, and uses the relationship path in the knowledge graph to reason about the area number, device number, sound category, and collection time to derive the cause and warning level of the current sound signal. The warning execution module outputs the cause and warning level of the sound signal through a visual interface and takes corresponding measures according to different warning levels.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] 1. This invention collects different types of raw sound signals through a sound acquisition unit, adds random normal noise to the raw sound signals, so that the individual mode functions of the subsequent sound signals contain only one frequency scale attribute representation information. It statistically analyzes all the maxima and minima in the sound signals after adding noise, thereby constructing the maxima connection curve and the minima connection curve. This design can predict the maxima and minima at each time point, which is convenient for solving the next round of new sound signals. The signal decomposition unit determines the signal intrinsic mode functions of each layer of the final sound signal, and calculates the mean value using different combinations of sound signals, thereby ensuring that the obtained signal intrinsic mode functions are more reasonable and reliable. The representation vector determination unit combines the different information of a single sound signal, and the obtained attribute representation vector is beneficial to the model's analysis of its category.

[0022] 2. This invention sets up multiple binary classifiers with different category determinations through a classifier combination unit. These classifiers are used to perform collaborative analysis on the category to which the sample belongs, so that the recognition results can better match the actual situation. The parameter optimization unit optimizes the unknown parameters of the binary classifiers and selects the best value for each unknown parameter, further improving the accuracy of the classifier during operation. The warning level output unit derives the cause of the current sound signal and the warning level step by step according to the relationship path in the knowledge graph, ensuring that managers can take timely measures to intervene. Attached Figure Description

[0023] Figure 1 A schematic diagram of the overall system flow is provided for embodiments of the present invention;

[0024] Figure 2 This is an internal module block diagram of the sound acquisition unit provided in an embodiment of the present invention;

[0025] Figure 3 This is an internal module block diagram of the signal decomposition unit provided in an embodiment of the present invention;

[0026] Figure 4 This is an internal module block diagram of the characterization vector determination unit provided in an embodiment of the present invention;

[0027] Figure 5 This is an internal module block diagram of the parameter optimization unit provided in an embodiment of the present invention.

[0028] In the diagram: 1. Sound Acquisition Unit; 101. Sound Processing Module; 102. Noise Addition Module; 103. Signal Analysis Module; 104. Curve Generation Module; 2. Signal Decomposition Unit; 201. Average Value Calculation Module; 202. Point-by-Point Subtraction Module; 203. Point Count Module; 204. Loop Execution Module; 205. Function Generation Module; 3. Representation Vector Determination Unit; 301. Frame Calculation Module; 302. Coefficient Output Module; 303. Ratio Analysis Module; 304. Vector Combination Module; 4. Classifier Combination Unit; 401. Sample Generation Module; 402. Classifier Construction Module; 403. Label Setting Module; 404. Category Output Module; 5. Parameter Optimization Unit; 501. Candidate Node Construction Module; 502. Initial Score Setting Module; 503. Weight Initialization Module; 504. Optimal Parameter Determination Module; 6. Warning Level Output Unit; 601. Atlas Generation Module; 602. Cause Determination Module; 603. Warning Execution Module. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Please see Figures 1-5 The present invention provides a technical solution: a park early warning system based on knowledge graph, including a classifier combination unit 4, a parameter optimization unit 5 and an early warning level output unit 6;

[0031] Sound acquisition unit 1, after setting the total number of layers of the intrinsic mode function of the signal, reads all processed sound signals, combines the sound signals with randomly generated different normal noises to obtain multiple combined sound signals, uses the difference method to determine the time series data of the maximum point and the time series data of the minimum point corresponding to the combined sound, calculates the polynomial corresponding to each time interval in the time series data step by step, and uses the polynomial to construct the maximum value connection curve and the minimum value connection curve;

[0032] Signal decomposition unit 2 calculates the average value at different times based on the maximum value connection curve and the minimum value connection curve. It subtracts the average value from the combined sound signal to obtain a new sound signal. Based on the new sound signal, it determines the signal intrinsic mode function and the combined sound signal for the next cycle. The operation is repeated until the signal intrinsic mode function of all layers of the current sound signal is calculated.

[0033] The characterization vector determination unit 3 acquires the intrinsic mode functions of the signals of each layer in all processed sound signals, calculates the signal characterization coefficients of each frame in the mode functions, determines the short-time energy, zero-crossing rate, and time interval from the initial value to the peak value of each frame in all sound signals, and combines the short-time energy corresponding to each frame signal in the intrinsic mode functions of each layer to obtain the attribute characterization vector of the sound signal.

[0034] The sound acquisition unit 1 includes a sound processing module 101, a noise addition module 102, a signal analysis module 103, and a curve generation module 104. The sound processing module 101 acquires different types of raw sound signals from the AudioSet website, including shouts, glass-breaking sounds, explosions, gunshots, footsteps, and other noises. It samples, pre-emphasizes, frames, windows, and denoises each raw sound signal to obtain the processed sound signal. Before sampling, the highest frequency of the sampled raw signal is counted, ensuring that each sampling frequency is greater than or equal to twice the highest frequency of the signal. During framing, the raw signal is divided into 20ms frames with a 10ms step size. The noise addition module 102 sets a maximum loop count n. max The total number of layers ρ of the intrinsic mode functions of the signal max Then, read all processed audio signals. For the current number of combinations An initialization operation is performed, where h0 represents the number of sound signals, and n1 represents the number of combinations corresponding to the first sound signal ξ1(t). This represents the h0th sound signal. The corresponding number of combinations will determine the k-th sound signal ξ. k (t) and randomly generated normal noise By combining them, the k-th combined sound signal ξ′ is obtained. k (t), where k represents the parameter, and n represents the number of combinations at that time. k Automatic increment; signal analysis module 103 acquires the k-th combined sound signal ξ′. k (t), ξ′ is determined using the finite difference method. k The time series data corresponding to the maximum point of (t) and time series data of minimum points in h1 represents the number of maxima, α and β represent parameters, and the value is determined based on the first time step. Corresponding maximum value and the second moment Corresponding maximum value Construct the corresponding polynomial Then, the polynomial corresponding to each time interval is calculated step by step. Based on preset conditions, all coefficients within each polynomial are determined. The curve generation module 104 constructs a maximum value connection curve using polynomials corresponding to different time intervals. Similarly, the polynomials for each time interval are calculated from the time series data of the minimum points, and a minimum value connection curve is constructed based on the polynomials. The specific preset conditions are as follows:

[0035] exist The first derivative value is equal to The first derivative value in, and exist The first derivative value is equal to The first derivative value in;

[0036] exist The value of the second derivative in is equal to exist The value of the second derivative in;

[0037] Signal decomposition unit 2 includes an average value calculation module 201, a point-by-point subtraction module 202, a point counting module 203, a loop execution module 204, and a function generation module 205. The average value calculation module 201 determines the maximum and minimum values ​​at different times based on the maximum and minimum value connection curves, and calculates the corresponding average value using the maximum and minimum values ​​at the same time. The point-by-point subtraction module 202 subtracts the k-th combined sound signal ξ′. k (t) and the average value at each time step Perform the subtraction operation to obtain the new sound signal ξ″ k (t), where The point counting module 203 sets a first threshold ε1 and a second threshold ε2, counts the total number of points N contained in the time series data of the maximum and minimum points, and determines the combined sound signal ξ′. k The number of zero-value signal points N0 contained in (t) is if And if |N-N0|≤ε2, then the new sound signal ξ″ k (t) is used as the eigenmode function of the signal, and ξ′ k (t) and ξ″ k Subtracting (t) from each other, the resulting signal is used as the combined sound signal for the next cycle, and the current layer number ρ i Automatically increment by one; otherwise, add a new sound signal ξ″. k (t) serves as the combined sound signal for the next round of the loop. The loop execution module 204 repeats the operation until the current layer number ρ. i Equal to the total number of floors ρ max Then, the sound signal ξ k (t) and randomly generated normal noise Combine them, and then combine them the number of times n was combined at that time. k Automatically increment by one, and calculate the signal intrinsic mode function values ​​corresponding to different layer numbers, repeating the same operation until the number of combinations n. k Greater than or equal to the maximum number of loops n max The loop ends, and function generation module 205 reads the sound signal ξ. k(t) is combined with the intrinsic mode functions of the signal at each layer after being combined with randomly generated normal noise. The average function is then calculated based on the intrinsic mode functions of the signal at each layer under different combination numbers, and this average function is used as the sound signal ξ. k (t) The final signal eigenmode functions of each layer;

[0038] The representation vector determination unit 3 includes a frame calculation module 301, a coefficient output module 302, a ratio analysis module 303, and a vector combination module 304. The frame calculation module 301 acquires the intrinsic mode functions (EMFs) of each layer within all processed audio signals and performs frame division. During frame division, the EMFs are divided into 20ms frames with a step size of 10ms. The power spectrum of each frame within the EMF is calculated. The coefficient output module 302 analyzes the power spectrum using a triangular filter to obtain the log-Mel energy. The log-Mel energy is then subjected to a discrete cosine transform to determine the... The signal characterization coefficient and ratio analysis module 303 determines the short-time energy, zero-crossing rate, and time interval from the initial value to the peak value of each frame in all sound signals, and counts the short-time energy corresponding to each frame of the signal in the intrinsic mode function of each layer of the signal. The sound signal is matched with each frame in the intrinsic mode function of the signal, and the ratio of the two energies in the same time period is calculated. The vector combination module 304 combines the short-time energy, zero-crossing rate, time interval from the initial value to the peak value, signal characterization coefficients contained in the intrinsic mode function of each layer of the sound signal, and energy ratio of each frame in each sound signal to obtain the attribute characterization vector of the sound signal.

[0039] The classifier combination unit 4 includes a sample generation module 401, a classifier construction module 402, a label setting module 403, and a category output module 404. The sample generation module 401 acquires the attribute representation vectors of multiple sound signals and labels the category of each sound signal, including shouting, glass breaking, explosion, gunshot, footsteps, and other noises. It combines the attribute representation vectors and categories of individual sound signals and stores them as samples in an 8:2 ratio in the first and second sets, with an equal number of samples for each category within each set. The classifier construction module 402 constructs multiple binary classifiers a1,…,a1. j ,...,a 15 Train a binary classifier using any two samples from different categories in the first set. Repeat this process until all binary classifiers are trained. Then, transfer all samples from the second set to a1,...,a j ,...,a 15 In the middle, determine a1,...,a j ,...,a 15 The accuracy of a1,...,a j,...,a 15 Sort the data according to accuracy and select the binary classifier a with the highest accuracy. j As the first binary classifier, the label setting module 403 is in binary classifier a j Includes category b x and b y The determination is made after the sample has undergone a j Before the determination, a corresponding estimated category set is set, and the initial estimated category set contains the numbers of all belonging categories. The sample is then transferred to a. j In the middle, the category analysis algorithm is used to determine a j If the output label value is positive, then delete b from the estimated class set of the current sample. y Number them and list all those that do not contain 'b'. y For a binary classifier that determines the class, if the label is negative, then remove b from the estimated class set of the current sample. x Number them and list all those that do not contain 'b'. x The binary classifier is determined, and the category output module 404 sorts all the listed binary classifiers according to their accuracy. The binary classifier with the highest accuracy is selected for further determination. This process is repeated until only one number remains in the estimated category set. The category of the current sample is then output according to the number. The specific category analysis algorithm is as follows:

[0040] f(H) = sign(∑(ω·exp(-g||H)) E -H F || 2 )-b))

[0041] Where g represents the radial coefficient, ω represents the adjustment parameter, b represents the bias term, and H E H represents the attribute representation vector of the E-th training sample. F Let E and F represent the attribute representation vector of the F-th sample to be predicted, and let E and F represent the parameters.

[0042] The parameter optimization unit 5 includes a candidate node construction module 501 and an initial score setting module 502. After the candidate node construction module 501 determines all the unknown parameters in the binary classifier, it sets a corresponding limit range for each unknown parameter. Within the limit range, the current unknown parameter is taken multiple times, and the values ​​of different types of unknown parameters are selected one by one and stored in the candidate nodes. The initial score setting module 502 obtains multiple candidate nodes, and each unknown parameter in each candidate node has one and only one value. There is no case where the values ​​between candidate nodes are completely the same. The value of each unknown parameter in the candidate node is used as the actual parameter value of the binary classifier. The accuracy under this set of parameter values ​​is calculated and used as the initial score of the candidate node. The change coefficient of all candidate nodes is set to zero.

[0043] The parameter optimization unit 5 also includes a weight initialization module 503 and an optimal parameter determination module 504. The weight initialization module 503 initializes the weight coefficients, correlation factors, and maximum number of iterations. Based on the initial score of each candidate node, it determines the local optimal score of a single node and the overall optimal score of all nodes. The optimal parameter determination module 504 uses a parameter update algorithm to adjust all parameter values ​​and change coefficients within each node. It uses the value of each unknown parameter in the current candidate node as the actual parameter value of the binary classifier, calculates the new score for each candidate node, updates the local optimal score of a single node and the overall optimal score of all nodes based on the current score, calculates the new weight coefficients using a weight analysis algorithm, and repeats the operation until the maximum number of iterations is reached. Finally, it outputs the parameter value corresponding to the overall optimal score, using it as the actual parameter value of the binary classifier. The parameter update algorithm is as follows:

[0044]

[0045] Among them, D M-1 B represents the value of the candidate node parameter in the (M-1)th iteration. M-1 D represents the change coefficient of the candidate node in the (M-1)th cycle. M B represents the value of the candidate node parameter in the Mth iteration. M D represents the change coefficient of the candidate node in the Mth cycle. X D represents the local best score of a single node. Y G1 and G2 represent the overall best score for all nodes, G1 and G2 represent the correlation factors, and M represents the current iteration number.

[0046] The specific weighting analysis algorithm is as follows:

[0047]

[0048] Where κ represents the weighting coefficient, κ max κ represents the maximum value of the weight coefficient. min M represents the minimum value of the weighting coefficient. max M represents the maximum number of iterations, and M represents the current number of iterations.

[0049] The warning level output unit 6 includes a knowledge graph generation module 601, a cause determination module 602, and a warning execution module 603. The knowledge graph generation module 601 acquires historical abnormal sound records of a designated area. The abnormal sound records include the area number, device number, sound category, collection time, cause, and warning level. A knowledge graph is constructed based on the historical abnormal sound records. After acquiring the current sound signal, the cause determination module 602 determines its attribute representation vector and category. It analyzes the corresponding area number using the device number that acquired the sound signal and uses the relationship path in the knowledge graph to reason about the area number, device number, sound category, and collection time to derive the cause and warning level of the current sound signal. The warning execution module 603 outputs the cause and warning level of the sound signal through a visual interface and takes corresponding measures according to different warning levels.

[0050] Working Principle: This invention processes the original sound signal through the sound processing module 101 in the sound acquisition unit 1, combines the sound signal with randomly generated normal noise using the noise addition module 102 to obtain a combined sound signal, determines the time series data of the maximum and minimum points of the combined sound signal through the signal analysis module 103, analyzes the maximum and minimum connection curves using the curve generation module 104, calculates the corresponding average value using the maximum and minimum values ​​at the same time through the average value calculation module 201 in the signal decomposition unit 2, and uses the point-by-point subtraction module 20... 2. A new sound signal is obtained. The point counting module 203 determines the sound signal combination for the next round of loop. The loop execution module 204 calculates the signal intrinsic mode function values ​​corresponding to different layers in each loop. The function generation module 205 outputs the final signal intrinsic mode functions for each layer of the sound signal. The frame calculation module 301 in the representation vector determination unit 3 obtains the power spectrum of the signal in each frame within the mode function. The coefficient output module 302 determines the signal representation coefficients. The ratio analysis module 303 obtains the energy ratio contained in the signal intrinsic mode functions of each layer. The vector combination module 304 outputs the sound signal. The attribute representation vector of the sound signal is combined with its category by the sample generation module 401 in the classifier combination unit 4 to form a sample. The classifier construction module 402 sorts the binary classifiers according to their accuracy and selects the binary classifier with the highest accuracy as the first binary classifier. The label setting module 403 lists new binary classifiers according to the label values. The category output module 404 determines the category of the current sample. The candidate node construction module 501 in the parameter optimization unit 5 selects the values ​​of different types of unknown parameters one by one and stores them in the candidate nodes. The scoring setting module 502 initializes the scores and change coefficients of candidate nodes. The weight initialization module 503 sets the local best score and the overall best score. The optimal parameter determination module 504 outputs the parameter values ​​corresponding to the overall best score and uses them as the actual parameter values ​​of the binary classifier. The knowledge graph is constructed by the graph generation module 601 in the warning level output unit 6 based on historical abnormal sound records. The cause determination module 602 analyzes the cause of the current sound signal and the warning level. The warning execution module 603 takes corresponding measures according to different warning levels.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A knowledge graph-based early warning system for a park, comprising a classifier combination unit (4), a parameter optimization unit (5), and an early warning level output unit (6), characterized in that: The sound acquisition unit (1) sets the total number of layers of the intrinsic mode function of the signal, reads all processed sound signals, combines the sound signals with randomly generated different normal noises to obtain multiple combined sound signals, uses the difference method to determine the time series data of the maximum point and the time series data of the minimum point corresponding to the combined sound, calculates the polynomial corresponding to each time interval in the time series data step by step, and uses the polynomial to construct the maximum connection curve and the minimum connection curve; The signal decomposition unit (2) calculates the average value at different times based on the maximum value connection curve and the minimum value connection curve, subtracts the average value from the combined sound signal to obtain a new sound signal, determines the signal intrinsic mode function and the combined sound signal for the next cycle based on the new sound signal, and repeats the operation until the signal intrinsic mode function of all layers of the current sound signal is calculated. The characterization vector determination unit (3) acquires the intrinsic mode functions of the signals of each layer in all processed sound signals, calculates the signal characterization coefficients of each frame in the mode functions, determines the short-time energy, zero-crossing rate and time interval from the initial value to the peak value of each frame in all sound signals, and the short-time energy corresponding to each frame signal in the intrinsic mode functions of each layer, and combines them to obtain the attribute characterization vector of the sound signal.

2. The knowledge graph-based park early warning system according to claim 1, characterized in that: The sound acquisition unit (1) includes a sound processing module (101), a noise addition module (102), a signal analysis module (103), and a curve generation module (104). The sound processing module (101) acquires different types of raw sound signals from the AudioSet website, and performs sampling, pre-emphasis, framing, windowing, and noise reduction operations on each raw sound signal to obtain the processed sound signal. The noise addition module (102) sets a maximum number of loops n. max The total number of layers ρ of the intrinsic mode functions of the signal max Then, read all processed audio signals. For the current number of combinations An initialization operation is performed, where h0 represents the number of sound signals, and n1 represents the number of combinations corresponding to the first sound signal ξ1(t). This represents the h0th sound signal. The corresponding number of combinations will determine the k-th sound signal ξ. k (t) and randomly generated normal noise By combining them, the k-th combined sound signal ξ is obtained. k ′(t), where k represents the parameter, and n represents the number of combinations at that time. k Automatically incremented by one, the signal analysis module (103) acquires the k-th combined sound signal ξ. k ξ′(t) is determined using the finite difference method. k Time series data corresponding to the maximum point of ′(t) and time series data of minimum points in h1 represents the number of maxima, α and β represent parameters, and the value is determined based on the first time step. Corresponding maximum value and the second moment Corresponding maximum value Construct the corresponding polynomial f1 α After (t), the polynomial corresponding to each time interval is calculated step by step. All coefficients in each polynomial are determined according to the preset conditions. The curve generation module (104) uses the polynomials corresponding to different time intervals to construct the maximum value connection curve. Similarly, the polynomials of each time interval are deduced from the time series data of the minimum point, and the minimum value connection curve is constructed according to the polynomial.

3. The knowledge graph-based park early warning system according to claim 1, characterized in that: The signal decomposition unit (2) includes an average value calculation module (201), a point-by-point subtraction module (202), a point counting module (203), a loop execution module (204), and a function generation module (205). The average value calculation module (201) determines the maximum and minimum values ​​at different times based on the maximum and minimum value connection curves, and calculates the corresponding average value using the maximum and minimum values ​​at the same time. The point-by-point subtraction module (202) subtracts the k-th combined sound signal ξ. k ′(t) and the average value at each time step Perform the subtraction operation to obtain the new sound signal ξ. k "(t), where The point counting module (203) sets a first threshold ε1 and a second threshold ε2, counts the total number of points N contained in the time series data of the maximum and minimum points, and determines the combined sound signal ξ. k The number of zero-value signal points N0 contained in ′(t) is... And if |N-N0|≤ε2, then the new sound signal ξ k "(t) is used as the eigenmode function of the signal, and ξ is used as the eigenmode function of the signal. k ′(t) and ξ k Subtracting '(t) from the given signal, the resulting signal is used as the combined sound signal for the next cycle, and the current layer number ρ i Automatically increment by one, otherwise add a new sound signal ξ. k "(t) serves as the combined sound signal for the next round of the loop, and the loop execution module (204) repeats the operation until the current layer number ρ. i Equal to the total number of floors ρ max Then, the sound signal ξ k (t) and randomly generated normal noise Combine them, and then combine them the number of times n was combined at that time. k Automatically increment by one, and calculate the signal intrinsic mode function values ​​corresponding to different layer numbers, repeating the same operation until the number of combinations n. k Greater than or equal to the maximum number of loops n max The loop ends, and the function generation module (205) reads the sound signal ξ. k (t) is combined with the intrinsic mode functions of the signal at each layer after being combined with randomly generated normal noise. The average function is then calculated based on the intrinsic mode functions of the signal at each layer under different combination numbers, and this average function is used as the sound signal ξ. k (t) The final signal eigenmode functions of each layer.

4. The knowledge graph-based park early warning system according to claim 1, characterized in that: The representation vector determination unit (3) includes a frame calculation module (301), a coefficient output module (302), a ratio analysis module (303), and a vector combination module (304). The frame calculation module (301) obtains the intrinsic mode functions of the signal at each layer in all processed sound signals, performs frame division operations on them, and calculates the power spectrum of the signal in each frame within the mode function. The coefficient output module (302) analyzes the power spectrum using a triangular filter to obtain the log-Mel energy, and performs a discrete cosine transform on the log-Mel energy to determine the signal representation coefficients. The ratio analysis module... Block (303) determines the short-time energy, zero-crossing rate, and time interval from the initial value to the peak value of each frame in all sound signals, and counts the short-time energy corresponding to each frame signal in the intrinsic mode function of each layer signal. It matches the sound signal with each frame in the intrinsic mode function of the signal and calculates the ratio of their energies in the same time period. The vector combination module (304) combines the short-time energy, zero-crossing rate, time interval from the initial value to the peak value, signal characterization coefficients contained in the intrinsic mode function of each layer signal, and energy ratio of each frame in each sound signal to obtain the attribute characterization vector of the sound signal.

5. The knowledge graph-based park early warning system according to claim 1, characterized in that: The classifier combination unit (4) includes a sample generation module (401), a classifier construction module (402), a label setting module (403), and a category output module (404). The sample generation module (401) obtains the attribute representation vectors of multiple sound signals, labels the category of each sound signal, combines the attribute representation vector and category of a single sound signal, and stores them as samples in a 8:2 ratio in the first set and the second set respectively. The classifier construction module (402) constructs multiple binary classifiers a1,...,a j ,...,a 15 Train a binary classifier using any two samples from different categories in the first set. Repeat this process until all binary classifiers are trained. Then, transfer all samples from the second set to a1,...,a j ,…,a 15 In the middle, determine a1,…,a j ,…,a 15 The accuracy of a1,…,a j ,…,a 15 Sort the data according to accuracy and select the binary classifier a with the highest accuracy. j As the first binary classifier, the label setting module (403) is in binary classifier a j Includes category b x and b y The determination is made after the sample undergoes a j Before the determination, a corresponding estimated category set is set, and the initial estimated category set contains the numbers of all belonging categories. The sample is then transferred to a. j In the middle, the category analysis algorithm is used to determine a j If the output label value is positive, then delete b from the estimated class set of the current sample. y Number them and list all those that do not contain 'b'. y For a binary classifier, if the label is negative, then remove b from the estimated class set of the current sample. x Number them and list all those that do not contain 'b'. x The binary classifier is determined, and the category output module (404) sorts all the listed binary classifiers according to their accuracy, selects the binary classifier with the highest accuracy and continues to determine, repeating the operation until only one number remains in the estimated category set, and outputs the category to which the current sample belongs according to the number.

6. The knowledge graph-based park early warning system according to claim 1, characterized in that: The parameter optimization unit (5) includes a candidate node construction module (501) and an initial score setting module (502). After determining all the unknown parameters in the binary classifier, the candidate node construction module (501) sets a corresponding limit range for each unknown parameter. Within the limit range, the current unknown parameter is taken multiple times, and the values ​​of different types of unknown parameters are selected one by one and stored in the candidate nodes. The initial score setting module (502) obtains multiple candidate nodes, takes the value of each unknown parameter in the candidate node as the actual parameter value of the binary classifier, calculates the accuracy under this set of parameter values, takes it as the initial score of the candidate node, and sets the change coefficient of all candidate nodes to zero.

7. The knowledge graph-based park early warning system according to claim 6, characterized in that: The parameter optimization unit (5) further includes a weight initialization module (503) and an optimal parameter determination module (504). The weight initialization module (503) initializes the weight coefficients, correlation factors, and maximum number of iterations. Based on the initial score of each candidate node, it determines the local optimal score of a single node and the overall optimal score of all nodes. The optimal parameter determination module (504) uses a parameter update algorithm to adjust all parameter values ​​and change coefficients within each node. It takes the value of each unknown parameter in the current candidate node as the actual parameter value of the binary classifier, calculates the new score of each candidate node, updates the local optimal score of a single node and the overall optimal score of all nodes based on the current score, calculates the new weight coefficient using a weight analysis algorithm, repeats the operation until the number of iterations reaches the maximum number of iterations, outputs the parameter value corresponding to the overall optimal score, and uses it as the actual parameter value of the binary classifier.

8. The knowledge graph-based park early warning system according to claim 1, characterized in that: The warning level output unit (6) includes a graph generation module (601), a cause determination module (602), and a warning execution module (603). The graph generation module (601) acquires historical abnormal sound records of a designated park, wherein the abnormal sound records include the area number, device number, sound category, collection time, cause, and warning level. A knowledge graph is constructed based on the historical abnormal sound records. After acquiring the current sound signal, the cause determination module (602) determines its attribute representation vector and category. It analyzes the corresponding area number using the device number that acquired the sound signal and uses the relationship path in the knowledge graph to reason about the area number, device number, sound category, and collection time to derive the cause and warning level of the current sound signal. The warning execution module (603) outputs the cause and warning level of the sound signal through a visual interface and takes corresponding measures according to different warning levels.

Images