Alarm device based on artificial intelligence voice analysis

Abstract

The invention provides an alarm device based on artificial intelligence voice analysis, and relates to the technical field of voice analysis. An audio acquisition sensor is arranged in a key monitoring area through an audio sensing arrangement module; an environment fluctuation identification module obtains an area environment fluctuation coefficient of the key monitoring area; an alarm threshold configuration module determines a regional alarm threshold based on the regional environment fluctuation coefficient; the abnormal keyword library module traverses the historical voice data record sequence extracted by the historical voice data acquisition module to construct a regional abnormal keyword library; the voice anomaly analyzer module constructs a regional intelligent voice anomaly analyzer based on a regional anomaly keyword library and a regional alarm threshold value; and the anomaly recognition module performs anomaly recognition on the real-time voice data by using the regional intelligent voice anomaly analyzer to obtain alarm information. According to the method and the device, the technical problem of frequent false alarm and missing alarm caused by lack of accurate identification means for audio content in a complex environment in the prior art is solved, and the alarm accuracy and the response speed are improved.

Description

Technical Field

[0001] The present application relates to the field of speech analysis technology, and in particular to an alarm device based on artificial intelligence speech analysis. Background Art

[0002] In public places such as campuses, parks and other crowded areas, safety issues have always been an important part of management and monitoring. In order to detect abnormal situations in time and protect personnel safety, more and more alarm devices are widely used in these places. Audio-based alarm devices have become one of the common monitoring methods due to their ability to capture changes in environmental sounds. Such devices use audio sensors to capture changes in sound in the environment, and usually adopt an alarm method based on a fixed threshold, that is, when the environmental noise or sound exceeds the preset alarm threshold, the alarm is automatically triggered to achieve the effect of early warning. However, the fixed alarm threshold cannot adapt to changes in different environments and time. For example, in an environment with dense crowds and high background noise, unnecessary alarms may be triggered due to misjudgment of background sounds, resulting in false alarms; conversely, if the noise is low or the environment is quiet, the real abnormal sound may not be detected in time, resulting in missed alarms. Secondly, existing alarm devices usually rely only on simple volume or frequency changes for judgment, lack analysis of specific voice content, and cannot identify complex voice anomalies, such as disputes, emergency calls, etc. This makes the existing alarm devices perform poorly in the face of complex voice environments, unable to accurately identify potential dangers or emergencies, and restricts the safety of public places. Summary of the invention

[0003] The present application provides an alarm device based on artificial intelligence speech analysis, which solves the technical problem that traditional alarm devices lack the means to accurately identify audio content in complex environments, resulting in frequent false alarms and missed alarms, and achieves the technical effect of improving the accuracy and response speed of alarms in public places.

[0004] In view of the above problems, the present application provides an alarm device based on artificial intelligence voice analysis, the device comprising: an audio sensor deployment module, used to determine Q key monitoring areas of the target park, and to deploy audio acquisition sensors based on the Q regional scene information of the Q key monitoring areas to obtain Q audio acquisition sensor deployment arrays, wherein each audio acquisition sensor regional deployment array is for a key monitoring area, and Q is an integer greater than or equal to 1; an environmental fluctuation identification module, used to collect Q regional environmental feature information sequences of the Q key monitoring areas within a preset monitoring window, identify the degree of information fluctuation of the Q regional environmental feature information sequences, and obtain Q regional environmental fluctuation coefficients; an alarm threshold configuration module, used to determine Q regional alarm thresholds based on the size of the Q regional environmental fluctuation coefficients, wherein the area The domain environment fluctuation coefficient is proportional to the regional alarm threshold; a historical voice data acquisition module is used to extract the voice data collected by the Q audio acquisition sensor arrays within a preset historical window to obtain Q historical voice data record sequences; an abnormal keyword library module is used to traverse the Q historical voice data record sequences to construct an abnormal voice data keyword library to obtain Q regional abnormal keyword libraries; a voice abnormality analyzer module is used to construct Q regional intelligent voice abnormality analyzers based on Q regional abnormal keyword libraries and Q regional alarm thresholds; an abnormality recognition module is used to perform real-time voice data acquisition on the Q audio acquisition sensor arrays to obtain Q real-time voice data, use the Q regional intelligent voice abnormality analyzers to perform abnormality recognition on the Q real-time voice data, and obtain alarm information according to the Q recognition results.

[0005] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0006] The audio sensor deployment module determines the key monitoring area of ​​the target park and deploys audio collection sensors based on the regional scene information, providing a targeted layout for subsequent audio data collection to ensure that effective audio data in the key monitoring area can be obtained. The environmental fluctuation recognition module collects regional environmental feature information and identifies its fluctuation degree, which can perceive the changes in the environment in real time, provide a basis for the reasonable setting of the alarm threshold, and avoid false alarms or missed alarms caused by environmental changes. The alarm threshold configuration module determines the alarm threshold according to the environmental fluctuation coefficient, and can dynamically adjust the threshold size according to the environmental conditions of different regions to improve the alarm accuracy and environmental adaptability. The historical voice data acquisition module extracts historical voice data, which is the basic material for building an abnormal keyword library and provides a data basis for the subsequent accurate identification of abnormal voices. The abnormal keyword library module traverses the historical voice data to build an abnormal keyword library, which can establish an abnormal voice pattern related to the actual situation in the region and provide specific standards for subsequent abnormal analysis. The regional intelligent voice abnormality analyzer is constructed based on the abnormal keyword library and the alarm threshold through the voice abnormality analyzer module, which can perform targeted abnormal voice analysis according to the conditions of different regions. Real-time voice data is collected through the anomaly recognition module and anomalies are identified using the constructed regional intelligent voice anomaly analyzer. Finally, alarm information is obtained based on the recognition results, achieving a rapid response to abnormal events and ensuring that the alarm device responds within the first time.

[0007] In summary, this application, through audio collection, environmental fluctuation analysis, construction of an abnormal keyword library and intelligent voice anomaly analysis, customizes the deployment of audio sensors in different monitoring areas, and dynamically adjusts the alarm threshold in combination with the environmental fluctuation coefficient, it can adapt to different environmental conditions and significantly reduce false alarms and missed reports. In addition, by constructing an abnormal keyword library of historical voice data and an intelligent voice anomaly analyzer, it is possible to accurately identify abnormal signals in complex voices, thereby improving the accuracy and response speed of the alarm. On the whole, it not only enhances the adaptability to complex environments, but also enhances the accurate recognition of voice content, realizes more intelligent and timely abnormal warnings, and effectively improves the level of security in public places.

[0008] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 A schematic diagram of the structure of an alarm device based on artificial intelligence speech analysis provided in an embodiment of the present application.

[0010] Figure 2 A schematic diagram of a process for obtaining an abnormal keyword library of Q regions in an alarm device based on artificial intelligence speech analysis provided in an embodiment of the present application.

[0011] Figure 3 A schematic diagram of the process of constructing Q regional intelligent voice anomaly analyzers in the alarm device based on artificial intelligence voice analysis provided in the embodiment of the present application.

[0012] Explanation of the accompanying drawings: audio sensor deployment module 10, environmental fluctuation recognition module 20, alarm threshold configuration module 30, historical voice data collection module 40, abnormal keyword library module 50, voice abnormality analyzer module 60, abnormality recognition module 70. DETAILED DESCRIPTION

[0013] The embodiment of the present application solves the technical problem that traditional alarm devices lack the means to accurately identify audio content in complex environments, resulting in frequent false alarms and missed alarms, by providing an alarm device based on artificial intelligence voice analysis, thereby achieving the technical effect of improving the accuracy and response speed of alarms in public places.

[0014] like Figure 1 As shown, the embodiment of the present application provides an alarm device based on artificial intelligence speech analysis, the device comprising:

[0015] The audio sensor deployment module 10 is used to determine Q key monitoring areas of the target park, deploy audio acquisition sensors based on Q area scene information of the Q key monitoring areas, and obtain Q audio acquisition sensor deployment arrays, wherein each audio acquisition sensor area deployment array is for one key monitoring area, and Q is an integer greater than or equal to 1.

[0016] Specifically, the target park refers to a specific park that needs to be monitored for security, such as schools and industrial parks. The key monitoring area is the area in the target park where the security situation needs to be focused on, such as stairs and laboratories in schools, which may have higher security risks. Regional scene information is information about the environment, functions, etc. of the key monitoring area. For example, the classroom area scene information of a school includes the size of the space, the flow of people, the daily noise level, etc. This information helps to reasonably deploy audio collection sensors.

[0017] First, according to the overall planning and security needs analysis of the target park, determine the set of areas that need to be monitored in the target park. For the convenience of subsequent description, Q refers to the number of key monitoring areas, and Q is an integer greater than or equal to 1. Then, based on the scene information of these key monitoring areas (such as the size, shape, ambient noise level, etc. of the area), determine the optimal layout and number of audio acquisition sensors. For example, in a large conference room, sensors need to be deployed in four corners to ensure full coverage. In this way, each key monitoring area will have a dedicated audio acquisition sensor array to ensure that the sound in that area can be captured.

[0018] Through the audio sensor deployment module 10, audio collection sensors are deployed in key areas of the target park in a targeted manner to ensure that effective audio data in key monitoring areas is obtained, providing a basic data source for subsequent operations such as abnormality identification, thereby improving the effectiveness and targeting of audio data collection.

[0019] The environmental fluctuation identification module 20 is used to collect Q regional environmental characteristic information sequences of the Q key monitoring areas within a preset monitoring window, identify the degree of information fluctuation of the Q regional environmental characteristic information sequences, and obtain Q regional environmental fluctuation coefficients.

[0020] Specifically, the preset monitoring window is a pre-set time period during which the environmental characteristic information of the key monitoring area is collected and analyzed, which can be set to 1 day or 1 week, etc. The regional environmental characteristic information sequence is a series of information on the environmental characteristics of the key monitoring area collected within the preset monitoring window, and these characteristics include noise intensity, sound frequency distribution, etc. The degree of information fluctuation refers to the fluctuation of the environmental characteristic information in the time series.

[0021] In the preset monitoring window, use audio acquisition sensors and other equipment to collect regional environmental characteristic information sequences of Q key monitoring areas to obtain Q regional environmental characteristic information sequences. For example, in the playground area of ​​the school campus, in the preset monitoring window of morning exercises, the sounds collected include environmental characteristic information such as students' noise and broadcasting. Then, the video analysis software is used to identify the degree of fluctuation of the Q regional environmental characteristic information, and generate the corresponding Q regional environmental fluctuation coefficients. For example, compare the changes in noise intensity at different time points, calculate the fluctuation of the regional environmental characteristic information of each key monitoring area in the preset monitoring window, and generate the corresponding regional environmental fluctuation coefficient. The regional environmental fluctuation coefficient is a quantitative representation of the degree of environmental fluctuation, reflecting the strength of the sound fluctuation in the environment. A high fluctuation coefficient means that the environmental sound changes greatly.

[0022] The environmental fluctuation recognition module 20 can sense the environmental changes in the key monitoring area in real time, accurately obtain the regional environmental fluctuation coefficient, provide a basis for the reasonable setting of the alarm threshold, and avoid false alarms or missed alarms caused by environmental fluctuations.

[0023] The alarm threshold configuration module 30 is used to determine Q regional alarm thresholds based on the magnitudes of the Q regional environmental fluctuation coefficients, wherein the regional environmental fluctuation coefficient is proportional to the regional alarm threshold.

[0024] Specifically, the regional alarm threshold is a numerical standard set for each key monitoring area. When the audio-related data exceeds this standard, an alarm will be triggered.

[0025] The Q regional alarm thresholds are determined according to the size of the Q regional environmental fluctuation coefficients obtained by the environmental fluctuation identification module 20. The regional environmental fluctuation coefficient is proportional to the regional alarm threshold. The alarm threshold of the region with a large fluctuation coefficient will be increased accordingly. For example, at the entrance of a campus with a large flow of people, the regional environmental fluctuation coefficient is large, and the alarm threshold may be set higher than that of a quiet library. A simple mathematical calculation tool or configuration software can be used to determine the alarm threshold based on the proportional relationship.

[0026] By dynamically setting the alarm threshold through the alarm threshold configuration module 30, the alarm threshold can be dynamically adjusted according to the environmental conditions of different areas, adapting to the security monitoring needs in different environments, reducing false alarms caused by environmental changes, and improving the accuracy of alarms.

[0027] The historical voice data collection module 40 is used to extract the voice data collected by the Q audio collection sensor arrays within a preset historical window to obtain Q historical voice data record sequences.

[0028] Specifically, the preset historical window is similar to the preset monitoring window, which is a specific time period in the past, used to collect historical voice data, such as the past week or month. The voice data collected within the preset historical window is extracted from the array of Q audio acquisition sensors to obtain Q historical voice data record sequences. For example, in the library area of ​​the school campus, in the past month (preset historical window), voice data such as readers' conversations and book flipping are extracted from the voice data collected by the deployed sensors. The Q historical voice data record sequences provide a large amount of historical data for building an abnormal keyword library, so that the device can learn and identify abnormal sound patterns based on historical data, thereby improving the accuracy of abnormal voice recognition.

[0029] The abnormal keyword library module 50 is used to traverse the Q historical voice data record sequences to construct an abnormal voice data keyword library to obtain Q regional abnormal keyword libraries.

[0030] Specifically, the regional abnormal keyword library is a database containing keywords in historical voice data that are determined to be possibly related to abnormal situations. Traverse Q historical voice data record sequences, perform data mining and voice analysis, build an abnormal voice data keyword library for each key monitoring area, and obtain Q regional abnormal keyword libraries. For example, in an industrial park, if words related to production abnormalities such as "machine failure" appear multiple times in historical voice data, they will be included in the regional abnormal keyword library. By constructing a regional abnormal keyword library through the abnormal keyword library module 50, abnormal voices that may be in dangerous situations can be more accurately identified, providing a more accurate basis for abnormal voice recognition.

[0031] The voice anomaly analyzer module 60 is used to construct Q regional intelligent voice anomaly analyzers based on Q regional anomaly keyword libraries and Q regional alarm thresholds.

[0032] Specifically, the regional intelligent voice anomaly analyzer is an analysis tool built by combining the regional abnormal keyword library and the regional alarm threshold. It is used to perform anomaly analysis on real-time voice data to determine whether it contains abnormal keywords, and then determine whether there is an abnormal situation. Based on Q regional abnormal keyword libraries and Q regional alarm thresholds, Q regional intelligent voice anomaly analyzers are constructed using programming algorithms, machine learning models, etc. For example, you can use Python to write a program, input parameters such as keyword libraries and alarm thresholds, and build an analyzer. You can also use Q regional abnormal keyword libraries and Q regional alarm thresholds to train machine learning models (such as neural networks) to build an intelligent voice anomaly analyzer. The Q regional intelligent voice anomaly analyzers correspond to Q key monitoring areas one by one, and are used to identify abnormal voices in specific key monitoring areas, further improving the accuracy of the alarm.

[0033] By constructing a regional intelligent voice anomaly analyzer through the voice anomaly analyzer module 60, it is possible to perform targeted abnormal voice analysis according to the conditions of different regions, thereby improving the accuracy and effectiveness of the abnormal voice analysis.

[0034] The anomaly identification module 70 is used to collect real-time voice data from the array of the Q audio collection sensors to obtain Q real-time voice data, use the Q regional intelligent voice anomaly analyzers to perform anomaly identification on the Q real-time voice data, and obtain alarm information based on the Q identification results.

[0035] Specifically, real-time voice data is the voice data collected by the audio acquisition sensor at the current moment, which can reflect the actual sound conditions in the current key monitoring area. An array of Q audio acquisition sensors is arranged to collect real-time voice data to obtain Q real-time voice data. Then, the constructed Q regional intelligent voice anomaly analyzers are used to identify anomalies in these real-time voice data. For example, if keywords such as "fire" or "explosion" are detected in a laboratory on a campus and exceed the alarm threshold of the area, the module will generate alarm information, such as specific light alarms, sound alarms, etc.

[0036] The abnormality recognition module 70 can effectively recognize the real-time voice data and accurately determine whether there are abnormal conditions in the key monitoring area, thereby issuing alarm information in time to ensure the safety of the target park.

[0037] Further, such as Figure 2 As shown, the abnormal keyword library module 50 is also used to perform the following steps:

[0038] Step P51: extracting the first historical voice data record at the top from the Q historical voice data record sequences respectively, to obtain Q first historical voice data records.

[0039] Step P52: extracting keywords related to abnormal situations and sentences containing the keywords from the Q first historical voice data records, and obtaining Q first abnormal keyword sets and Q first abnormal key sentence sets.

[0040] Step P53: Add the Q first abnormal keyword sets and the Q first abnormal key sentence sets into the Q first memory banks respectively.

[0041] Step P54: Using the Q first memory banks as indexes, keyword association extraction is performed on the Q second historical voice data records located in the second position extracted from the Q historical voice data record sequences, and combined with the sentences containing the extracted keywords, Q second abnormal keyword sets and Q second abnormal key sentence sets are obtained.

[0042] Step P55: Add the Q second abnormal keyword sets and the Q second abnormal key sentence sets into the first memory bank set respectively to update the memory bank, and obtain Q second memory banks.

[0043] Step P56: Based on the Q second memory banks, keyword association recognition is performed on the Q historical voice data record sequences until the end of the sequence is reached, thereby obtaining Q target memory banks.

[0044] Step P57: Obtain the Q regional abnormal keyword libraries based on the Q target memory libraries.

[0045] Specifically, the memory library is a library used to store abnormal keyword sets and key sentence sets, and is an intermediate storage structure in the process of building an abnormal keyword library.

[0046] In the Q historical voice data record sequences, the first voice data record of each sequence is located and extracted, thereby obtaining Q first historical voice data records. For example, there are 3 historical voice data record sequences, and the first voice data record is extracted from the 3 sequences.

[0047] The Q first historical voice data records are analyzed in detail, and the keywords related to the abnormal situation and the sentences containing these keywords are extracted through natural language processing technology (such as text analysis, semantic analysis, etc.), forming Q first abnormal keyword sets and Q first abnormal key sentence sets respectively. For example, using technologies such as word vector models, keywords such as "danger" and "accident" that may be related to abnormal situations are identified, and sentences containing these keywords are found.

[0048] The Q first abnormal keyword sets and Q first abnormal key sentence sets obtained above are respectively stored in Q first memory banks, each of which contains a first abnormal keyword set and a first abnormal key sentence set extracted from a historical voice data record sequence corresponding to a key monitoring area.

[0049] Using the Q first memory banks as indexes, keyword association extraction is performed on the Q second historical voice data records. That is, according to the first abnormal keyword and the first abnormal key sentence in the first memory bank, the keywords related thereto are searched in the second historical voice data record, and combined with the sentences, Q second abnormal keyword sets and Q second abnormal key sentence sets are obtained. In the process of searching for abnormal keywords and abnormal key sentences, keyword matching algorithms, text similarity calculation algorithms (such as cosine similarity algorithms), etc. can be used to determine the degree of association between keywords, so as to perform association extraction.

[0050] The Q second abnormal keyword sets and Q second abnormal key sentence sets are added to the Q first memory banks respectively, and the first memory banks are updated according to the update rules such as merging and deduplication to obtain Q second memory banks. For example, if there are new keywords in the second abnormal keyword set, they are added to the structure corresponding to the first memory bank, and if there are repeated keywords, they are not added. Through this update operation, the memory bank can be continuously evolved and improved with the addition of new data.

[0051] Based on the Q second memory banks, the Q historical voice data record sequences are continuously identified by keyword association. According to the aforementioned association extraction method, the subsequent data in the historical voice data record sequence is gradually processed, and the memory bank is continuously updated. Each time, the next historical voice data record is processed based on the current memory bank, and this process is repeated until the entire sequence is processed, and finally Q target memory banks are obtained.

[0052] According to the information in the Q target memory banks, the keywords that best represent the abnormal situation are extracted from the target memory banks, and the final Q regional abnormal keyword banks are sorted and screened.

[0053] By gradually associating and updating the memory library to build a regional abnormal keyword library, keywords related to abnormal situations can be mined from historical voice data more comprehensively and accurately. Compared with the simple traversal construction method, this method can take into account the association between keywords and continuously optimize the extraction of keywords using the previous memory library information, thereby improving the quality of the abnormal keyword library and further improving the accuracy of speech anomaly analysis.

[0054] Further, step P54 includes:

[0055] Step P541: Using the Q first abnormal keyword sets stored in the Q first memory banks as indexes, keyword similarity matching is performed on the Q second historical voice data records.

[0056] Step P542: extract keywords whose matching results meet a preset similarity threshold, and obtain Q matching second abnormal keyword sets.

[0057] Step P543: Using the Q first abnormal key sentence sets stored in the Q first memory banks as indexes, semantic similarity matching is performed on the Q second historical voice data records.

[0058] Step P544: extract the sentences whose matching results meet the preset semantic similarity threshold, and obtain a set of Q matching abnormal sentences.

[0059] Step P545: extract keywords related to abnormal situations from the Q first abnormal key sentence sets respectively to obtain Q semantically matching second abnormal keyword sets.

[0060] Step P546: performing a union of the Q matching second abnormal keyword sets and the Q semantically matching second abnormal keyword sets to obtain the Q second abnormal keyword sets.

[0061] Specifically, the preset similarity threshold is a numerical standard set in advance when performing keyword similarity matching. Only when the similarity between keywords reaches this standard, the keywords are considered to be matched.

[0062] For each set of first abnormal keywords in the first memory bank, similarity is calculated between the first abnormal keyword set and the corresponding keywords in the second historical voice data record. For example, the first abnormal keyword (such as "fire") and the keywords in the second historical voice data record (such as "on fire" and "flame") are converted into vector form using a word vector model, and then the similarity between the two vectors is calculated using the cosine similarity formula.

[0063] For the calculated multiple keyword similarity results, the keywords that meet the preset similarity threshold are screened out. For example, if the preset similarity threshold is 0.8, then all keywords with similarity greater than or equal to 0.8 are extracted. These keywords that meet the preset similarity threshold are aggregated to form Q matching second abnormal keyword sets.

[0064] For each first abnormal key sentence set in the first memory bank, the semantic similarity is calculated between it and the corresponding sentence in the second historical voice data record. A semantic analysis model based on deep learning, such as a pre-trained BERT model, can be used to input the sentence into the model to obtain a semantic vector, and then calculate the similarity between the vectors to obtain multiple key sentence similarity results. From the multiple key sentence similarity results, sentences that meet the preset semantic similarity threshold are selected to obtain Q matching abnormal sentence sets, and the Q matching abnormal sentence sets are used as Q second abnormal key sentence sets.

[0065] For each set of first abnormal key sentences, keywords related to the abnormal situation are extracted through natural language processing technology (such as part-of-speech tagging, named entity recognition, etc.). For example, in the sentence "I saw the library on fire", "on fire" is identified as a keyword related to the abnormal situation through part-of-speech tagging, thereby obtaining Q semantically matching second abnormal keyword sets.

[0066] The keywords in the Q matching second abnormal keyword sets and the Q semantically matching second abnormal keyword sets are merged, and repeated keywords are removed, so as to finally obtain Q second abnormal keyword sets.

[0067] By calculating the keyword similarity and semantic similarity to construct the second abnormal keyword set, the keywords related to the abnormal situation can be more accurately extracted from the second historical voice data record. Keyword similarity matching can directly find words similar to existing abnormal keywords, and semantic similarity matching can mine potential related information from the sentence level. Combined with the keywords extracted from the first abnormal key sentence set, the second abnormal keyword set obtained can more comprehensively and accurately reflect the words related to the abnormal situation, laying the foundation for the subsequent construction of a more complete regional abnormal keyword library, thereby improving the accuracy of speech anomaly recognition.

[0068] Further, step P57 includes:

[0069] The abnormal keywords with the top m occurrence frequencies in the Q target memory libraries are respectively extracted to obtain Q regional abnormal keyword libraries, where m is an integer greater than or equal to 3.

[0070] Specifically, for each of the Q target memory banks, the frequency of occurrence of each abnormal keyword is counted. The data in the target memory bank can be traversed and a counter can be used to record the number of times each keyword appears. Then, all abnormal keywords are sorted from high to low according to the frequency of occurrence. Finally, the abnormal keywords ranked in the top m are extracted, which are considered to be the words that best represent abnormal situations. Here, m is a preset parameter that can be adjusted according to actual needs to balance the coverage and accuracy of the keyword library. In order to ensure the number of regional abnormal keyword libraries, m is an integer greater than or equal to 3. These high-frequency abnormal keywords are respectively classified into the corresponding Q regions to form Q regional abnormal keyword libraries. For example, in a target memory bank in the school campus, it is counted that "fire" appears 10 times, "emergency" appears 7 times, and "rupture" appears 5 times. When m=3, the abnormal keywords "fire", "emergency" and "rupture" are extracted to form the regional abnormal keyword library.

[0071] By extracting the regional abnormal keyword library based on the frequency of occurrence, the abnormal keywords that appear frequently in the historical voice data can be preferentially selected. These high-frequency keywords can better represent the abnormal situations that may occur in the target park, making the regional abnormal keyword library more representative and targeted. In this way, in the subsequent voice anomaly analysis, the abnormal voices related to these high-frequency abnormal keywords can be more accurately identified, which improves the efficiency and accuracy of voice anomaly recognition and alarm.

[0072] Further, such as Figure 3 As shown, the speech anomaly analyzer module 60 is also used to perform the following steps:

[0073] Step P61: According to a preset disturbance scale, interference data is randomly generated for the Q regional abnormal keyword libraries to obtain Q abnormal regional speech data sets.

[0074] Step P62: Perform abnormality identification on the Q abnormal area voice data sets in combination with the Q area alarm thresholds to obtain Q abnormality identification results.

[0075] Step P63: Use the Q abnormal area speech data sets and the Q abnormal recognition results to perform supervised training on the network layer constructed based on the feedforward neural network until the training converges, thereby obtaining the trained Q area intelligent speech anomaly analyzers.

[0076] Specifically, the preset disturbance scale is a pre-set standard for determining the degree or range of interference when generating interference data. For example, the preset disturbance scale is 0.5, which means that the fluctuation amplitude when generating interference data is within 0.5 times of a certain benchmark. According to the preset disturbance scale, the Q regional abnormal keyword libraries are randomly generated for interference data. This process can use data enhancement technology, such as adding background noise to the voice data, changing the tone or speed, etc., to simulate various situations that may occur in the real environment. Exemplary: For a campus area, if the abnormal keyword "emergency" appears in the voice data, some background conversation sounds, traffic sounds or other common noises are added through interference data generation to simulate the voice situation in different environments. The interference data generated for each regional abnormal keyword library are summarized separately to obtain Q abnormal regional voice data sets. By generating data with interference, the training data can be made more diverse, more complex actual situations can be simulated, and the adaptability of the regional intelligent voice anomaly analyzer to various situations can be improved.

[0077] Each data in the Q abnormal area speech data sets is compared with the corresponding Q regional alarm thresholds to determine whether the alarm threshold is reached, thereby obtaining Q abnormal recognition results. For example, if the abnormal degree corresponding to a data in the abnormal area speech data exceeds the regional alarm threshold, then the abnormal recognition result of the data is abnormal, otherwise it is normal. Combining the regional alarm threshold for abnormal recognition provides accurate supervision data for the training of the regional intelligent speech abnormality analyzer, so that the regional intelligent speech abnormality analyzer can accurately judge abnormal situations according to the set standards.

[0078] Using the deep learning framework, a feedforward neural network is constructed. For each network layer of the constructed feedforward neural network, Q abnormal area speech data sets are taken as input data, and Q abnormal recognition results are taken as supervision labels (i.e., correct answers) to perform supervised training on the feedforward neural network. During the training process, the weights of the network layer are adjusted according to the difference between the predicted results and the correct labels (such as using loss functions such as mean square error), and this process is iterated continuously until the training converges (i.e., the loss function no longer decreases significantly), and finally the trained Q regional intelligent speech anomaly analyzers are obtained. Through supervised training, the trained regional intelligent speech anomaly analyzers have higher accuracy and reliability, so that abnormal situations can be identified more accurately in actual speech anomaly analysis.

[0079] Furthermore, the audio sensor deployment module 10 of the embodiment of the present application is also used to perform the following steps:

[0080] Step P11: extracting features of the Q regional scene information according to preset scene features to obtain Q regional scene features.

[0081] Step P12: Using a sensor deployment plan identifier to identify the Q regional scene features, and obtaining Q regional sensor deployment plans.

[0082] Step P13: Perform regional sensor deployment according to the deployment positions and deployment quantities of the audio collection sensors in the Q regional sensor deployment schemes to obtain the Q audio collection sensor deployment arrays.

[0083] Specifically, the preset scene features are some pre-set feature types or patterns related to the regional scene, which are used to describe and distinguish the parameters of different monitoring area features, such as space size, personnel flow direction, whether it is a closed space, etc. The regional scene features are the scene features unique to each area extracted from the Q regional scene information according to the preset scene features. For example, a key monitoring area is a school classroom, and its regional scene features may include small space, personnel flow concentrated in class breaks, and semi-enclosed space. The sensor layout plan identifier is a tool or algorithm that can analyze and identify regional scene features to derive a sensor layout plan suitable for the area. It can work based on predefined rules, models or machine learning algorithms. The regional sensor layout plan is a plan for how to lay out audio acquisition sensors for each area, including information such as the sensor layout location (such as the four corners of the classroom, etc.) and the number of layouts (such as 3 sensors are required).

[0084] According to the preset scene features, the scene information of Q areas is extracted respectively. The scene features can be extracted using data analysis tools and feature engineering algorithms. For example, image recognition technology is used to determine the flow of people in the area, and sound analysis tools are used to determine the background noise level of the area. Through feature extraction, the scene features of each area can be obtained, providing a basis for the subsequent sensor deployment.

[0085] The Q regional scene features are input into the sensor deployment plan identifier. This sensor deployment plan identifier contains rule-based judgment logic or pre-trained machine learning models (such as decision tree models, etc.). Taking the school library area as an example, if the regional scene features are large space, less personnel flow and quiet, the sensor deployment plan identifier determines the plan of deploying a small number of sensors in several key locations of the library (such as entrances, bookshelf aisles, etc.) based on the internal rules or models, thereby obtaining Q regional sensor deployment plans.

[0086] According to the layout positions and layout quantities specified in the Q regional sensor layout plans, the actual layout of audio collection sensors is carried out in each corresponding area. For example, according to the regional sensor layout plan of a conference room, the specified number of audio collection sensors are installed in the corners and positions specified in the plan, and finally Q audio collection sensor layout arrays are obtained. It is ensured that each area can obtain appropriate audio collection coverage, thereby improving the collection quality and coverage of sound data, enabling the alarm device to capture abnormal sound events more accurately, and improving the responsiveness and accuracy of the alarm.

[0087] Furthermore, the environmental fluctuation identification module 20 is further configured to perform the following steps:

[0088] Step S21: according to a preset environmental feature extraction scale set, traverse the Q regional environmental feature information sequences to perform asynchronous information extraction, and obtain Q asynchronous extraction information subsequence sets.

[0089] Step S22: performing mean processing on the Q asynchronously extracted information subsequence sets respectively to obtain Q asynchronously extracted information mean value sets.

[0090] Step S23: Based on the Q asynchronously extracted information mean value sets, cross-time series information fluctuation identification is performed to obtain the Q regional environmental fluctuation coefficients.

[0091] Specifically, the preset environmental feature extraction scale set is a set of pre-set standards used to determine the scope or degree of extracting asynchronous information from the regional environmental feature information sequence. For example, this set can include different scale standards such as time scale (such as extracting information every 5 minutes) and spatial scale (such as environmental features per 10 square meters of area).

[0092] For each of the Q regional environmental feature information sequences, traverse according to the standards in the preset environmental feature extraction scale set. For example, if the preset environmental feature extraction scale set stipulates that asynchronous information extraction is performed at a time interval of every 10 minutes, then data points are extracted from the regional environmental feature information sequence at intervals of 10 minutes to form multiple asynchronous information subsequences, and finally Q asynchronous extraction information subsequence sets are obtained. Each subsequence contains information extracted from a specific regional environmental feature information sequence in a specific asynchronous manner (such as data within 10 minutes).

[0093] For each of the Q asynchronously extracted information subsequence sets, the average values ​​of all elements therein are calculated to obtain Q asynchronously extracted information mean value sets. Each asynchronously extracted information mean value set contains multiple means, each mean value corresponds to an asynchronously extracted information subsequence, and contains the means of all data in the subsequence.

[0094] The data in the Q asynchronously extracted information mean sets are used to identify cross-time series information fluctuations. For example, the fluctuation calculation method in time series analysis can be used, such as calculating the difference between adjacent means, or a more complex statistical model (such as an autoregressive moving average model) can be used to analyze the fluctuation of the data, and finally Q regional environmental fluctuation coefficients can be obtained.

[0095] Through the above steps, the environmental fluctuation identification module 20 can accurately identify the fluctuation of the environmental characteristic information of each monitoring area. By quantifying the environmental fluctuation, the alarm threshold can be set more reasonably, reducing false alarms and missed alarms, and improving the reliability and effectiveness of the alarm device.

[0096] Further, step S23 includes:

[0097] The fluctuation variances are calculated for the Q asynchronously extracted information mean sets respectively to obtain Q fluctuation variances.

[0098] The Q fluctuation variances are divided by the sum of the Q fluctuation variances to obtain Q regional environmental fluctuation coefficients.

[0099] Specifically, when calculating the regional environmental fluctuation coefficient, for each of the Q asynchronously extracted information mean value sets, first use the variance calculation formula to calculate and obtain Q fluctuation variances. Then, calculate the sum of the Q fluctuation variances. Divide each fluctuation variance by the sum of the fluctuation variances to obtain Q regional environmental fluctuation coefficients.

[0100] By calculating the fluctuation variance, the degree of fluctuation of each asynchronously extracted information mean set can be accurately quantified. Comparing the fluctuation variance with the sum to obtain the regional environmental fluctuation coefficient, the degree of fluctuation of each region can be normalized to a relative value, so that the environmental fluctuations between different regions can be directly compared, which is convenient for identifying areas with larger or smaller fluctuations in overall environmental monitoring and management.

[0101] In summary, the alarm device based on artificial intelligence speech analysis provided by the embodiment of the present application has the following technical effects:

[0102] Through the collaborative work of multiple modules, combined with artificial intelligence voice analysis and environmental fluctuation recognition, accurate and intelligent abnormal voice detection and alarm functions are realized. First, the audio sensor layout module 10 optimizes the sensor layout according to the regional scene information, ensuring that there is a suitable audio acquisition array in each key monitoring area, thereby improving the coverage and accuracy of the acquisition. Then, the environmental fluctuation recognition module 20 performs fluctuation analysis on the environmental feature information sequence to obtain the environmental fluctuation coefficient of each area, providing a basis for the adaptive adjustment of the subsequent alarm threshold, and reducing the possibility of false alarms and missed alarms. The alarm threshold configuration module 30 flexibly adjusts the alarm sensitivity of each area according to the change of the environmental fluctuation coefficient to ensure that the alarm device can work normally in a dynamic environment. The abnormal keyword library module 50 establishes an accurate abnormal voice keyword library by extracting keywords and matching similarity of historical voice data, so that the alarm device can identify complex voice anomalies such as disputes and calls for help. The voice anomaly analyzer combines the trained intelligent model with the anomaly recognition of real-time voice data to further improve the accuracy and response speed of anomaly detection.

[0103] Overall, the embodiment of the present application realizes intelligent anomaly recognition based on environmental fluctuations and voice content through the collaboration of multiple modules, significantly improves the intelligence and adaptability of the alarm device, can accurately trigger the alarm under different environmental conditions, and improves the alarm accuracy and response speed in public places.

[0104] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An alarm device based on artificial intelligence voice analysis, characterized in that: The device comprises: An audio sensor deployment module is used to determine Q key monitoring areas of the target park, deploy audio acquisition sensors based on Q regional scene information of the Q key monitoring areas, and obtain Q audio acquisition sensor deployment arrays, wherein each audio acquisition sensor regional deployment array is for one key monitoring area, and Q is an integer greater than or equal to 1; An environmental fluctuation identification module is used to collect Q regional environmental characteristic information sequences of the Q key monitoring areas within a preset monitoring window, identify the degree of information fluctuation of the Q regional environmental characteristic information sequences, and obtain Q regional environmental fluctuation coefficients; An alarm threshold configuration module, used to determine Q regional alarm thresholds based on the magnitudes of the Q regional environmental fluctuation coefficients, wherein the regional environmental fluctuation coefficient is proportional to the regional alarm threshold; A historical voice data collection module is used to extract the voice data collected by the Q audio collection sensor arrays within a preset historical window to obtain Q historical voice data record sequences; An abnormal keyword library module is used to traverse the Q historical voice data record sequences to construct an abnormal voice data keyword library to obtain Q regional abnormal keyword libraries; A voice anomaly analyzer module is used to build Q-region intelligent voice anomaly analyzers based on Q-region abnormal keyword libraries and Q-region alarm thresholds; The anomaly recognition module is used to collect real-time voice data from the array of the Q audio acquisition sensors to obtain Q real-time voice data, use the Q regional intelligent voice anomaly analyzers to perform anomaly recognition on the Q real-time voice data, and obtain alarm information based on the Q recognition results.

2. The alarm device based on artificial intelligence speech analysis as claimed in claim 1, characterized in that: The execution steps of the abnormal keyword library module include: Extracting the first historical voice data record at the top from the Q historical voice data record sequences respectively to obtain Q first historical voice data records; Extracting keywords related to abnormal situations and sentences containing the keywords from the Q first historical voice data records to obtain Q first abnormal keyword sets and Q first abnormal key sentence sets; Adding the Q first abnormal keyword sets and the Q first abnormal key sentence sets into the Q first memory banks respectively; Using the Q first memory banks as indexes, performing keyword association extraction on the Q second historical voice data records located at the second position extracted from the Q historical voice data record sequences, and combining the sentences containing the extracted keywords to obtain Q second abnormal keyword sets and Q second abnormal key sentence sets; Add the Q second abnormal keyword sets and the Q second abnormal key sentence sets to the first memory bank set respectively to update the memory bank, and obtain Q second memory banks; Based on the Q second memory banks, keyword association recognition is performed on the Q historical voice data record sequences until the end of the sequence is reached, thereby obtaining Q target memory banks; The Q regional abnormal keyword libraries are obtained based on the Q target memory libraries.

3. The alarm device based on artificial intelligence speech analysis as claimed in claim 2, characterized in that: The execution steps of the abnormal keyword library module include: Using the Q first abnormal keyword sets stored in the Q first memory banks as indexes, performing keyword similarity matching on the Q second historical voice data records; Extract keywords whose matching results meet a preset similarity threshold, and obtain a set of Q matching second abnormal keywords; Using the Q first abnormal key sentence sets stored in the Q first memory banks as indexes, performing semantic similarity matching on the Q second historical voice data records; Extract the sentences whose matching results meet the preset semantic similarity threshold, and obtain a set of Q matching abnormal sentences; Extracting keywords related to abnormal situations from the Q first abnormal key sentence sets respectively to obtain Q semantically matching second abnormal keyword sets; The Q matching second abnormal keyword sets and the Q semantic matching second abnormal keyword sets are unioned to obtain the Q second abnormal keyword sets.

4. The alarm device based on artificial intelligence speech analysis as claimed in claim 3, characterized in that: The execution steps of the abnormal keyword library module include: The abnormal keywords with the top m occurrence frequencies in the Q target memory libraries are respectively extracted to obtain Q regional abnormal keyword libraries, where m is an integer greater than or equal to 3.

5. The alarm device based on artificial intelligence speech analysis as claimed in claim 1, characterized in that: The execution steps of the speech anomaly analyzer module include: According to a preset disturbance scale, randomly generate interference data for the Q regional abnormal keyword libraries to obtain Q abnormal regional speech data sets; Performing abnormality recognition on the Q abnormal area voice data sets in combination with the Q area alarm thresholds to obtain Q abnormality recognition results; The network layer constructed based on the feedforward neural network is supervised and trained using the Q abnormal area speech data sets and the Q abnormal recognition results respectively until the training converges, thereby obtaining the trained Q area intelligent speech anomaly analyzers.

6. The alarm device based on artificial intelligence speech analysis as claimed in claim 1, characterized in that: The execution steps of the audio sensor deployment module include: Extracting features from the Q regional scene information according to preset scene features to obtain Q regional scene features; Using a sensor deployment scheme identifier to identify the Q regional scene features, and obtain Q regional sensor deployment schemes; Perform regional sensor deployment according to the deployment positions and deployment quantities of the audio collection sensors in the Q regional sensor deployment schemes to obtain the Q audio collection sensor deployment arrays.

7. The alarm device based on artificial intelligence speech analysis as claimed in claim 1, characterized in that: The execution steps of the environmental fluctuation identification module include: According to a preset environmental feature extraction scale set, traverse the Q regional environmental feature information sequences to perform asynchronous information extraction to obtain Q asynchronous extraction information subsequence sets; Performing mean processing on the Q asynchronously extracted information subsequence sets respectively to obtain Q asynchronously extracted information mean value sets; Based on the Q asynchronously extracted information mean value sets, cross-time series information fluctuation identification is performed to obtain the Q regional environmental fluctuation coefficients.

8. The alarm device based on artificial intelligence speech analysis as claimed in claim 7, characterized in that: The execution steps of the environmental fluctuation identification module include: Performing fluctuation variance calculations on the Q asynchronously extracted information mean value sets respectively to obtain Q fluctuation variances; The Q fluctuation variances are divided by the sum of the Q fluctuation variances to obtain Q regional environmental fluctuation coefficients.

Images